Prediction-Learning Algorithm for Efficient Energy ... - MDPI

energies Article

Prediction-Learning Algorithm for Efficient Energy Consumption in Smart Buildings Based on Particle Regeneration and Velocity Boost in Particle Swarm Optimization Neural Networks Sehrish Malik and DoHyeun Kim * Computer Engineering Department, Jeju National University, Jeju-si 63243, Korea; [email protected] * Correspondence: [email protected]; Tel.: +82-64-754-3658 Received: 20 April 2018; Accepted: 14 May 2018; Published: 17 May 2018

Abstract: Electricity, the most important form of energy and an indispensable resource, primarily for commercial and residential smart buildings, faces challenges requiring its hyper efficient consumption and production. Therefore, accurate energy consumption predictions are required in order to manage and optimize the energy consumption of smart buildings. Many studies have taken advantage of the power and robustness of neural networks (NN) when it comes to accurate predictions. A few studies have also used the particle swarm optimization (PSO) algorithm along with NNs to enhance and optimize the predictions. In this work, we study prediction learning using PSO-based neural networks (PSO-NN) and propose modifications in order to increase prediction accuracy. Our proposed modifications are re-generation based PSO-NN (R-PSO-NN) and velocity boost-based PSO-NN (VB-PSO-NN). The performance metrics used are: prediction accuracy, number of particles used, and number of epochs required. We compare the results of NN, PSO-NN, R-PSO-NN and VB-PSO-NN based on the performance metrics. Keywords: energy prediction; smart homes; neural networks (NN); particle swarm optimization (PSO)

1. Introduction With the growing global population and escalating industrialization, the thirst for energy in the world is reaching aberrant levels. According to the World Energy Council [1], per capita global energy demand will spike in 2030, and electricity demand is predicted to double by 2060; eventually necessitating larger investments in smart buildings, households, or smart infrastructures in order to ensure energy-efficient environments. Electricity, the preeminent source of energy for lighting, cooling and different appliances working in smart homes, is a rapidly developing source of energy. In Korea, the total consumption of electricity in residential and commercial sector is about 40.5% more than the total liquid natural gas (LNG) and city gas consumption in the residential and commercial sector [2]. Being the most important form of energy, electricity is facing a continuous challenge of increased demand worldwide. According to the U.S. Energy Information Administration (EIA), a country’s energy use, mainly electricity use, is linked to its economic growth. For growing economies in the world, the increasing population is becoming a vital reason for the rise in electricity demand. The International Energy agency (IEA) has also stated that there is a strong correlation between a country’s energy usage and its wealth. It has also predicted an increase of 28% in electricity usage over the period of 2015–2040, mainly in residential and commercial buildings [3]. Electricity is generated by many different energy sources as coal, natural gas, petroleum fuel, fossil fuel, nuclear, hydro, renewable fuels, geothermal and others. Typically, it takes 3 Energies 2018, 11, 1289; doi:10.3390/en11051289

www.mdpi.com/journal/energies

Energies 2018, 11, x FOR PEER REVIEW

2 of 21

generated Energies 2018,by 11, many 1289

different energy sources as coal, natural gas, petroleum fuel, fossil fuel, nuclear, 2 of 21 hydro, renewable fuels, geothermal and others. Typically, it takes 3 to 3.3 units of source energy to produce one unit of site energy of electricity [4]. Due to the expensive sources and generation to 3.3 units of source energy produce one unit ofto site energygas. of electricity [4].need Due to expensive overheads, electricity costs to more in comparison natural Hence, we to the optimize our sources and generation overheads, electricity costs more in comparison to natural gas. Hence, we need electricity energy consumption through timely and accurate energy predictions. to optimize our electricity consumption and accurate energy predictions. In recent years, ourenergy homes have beenthrough gettingtimely smarter with connected and remotely In recent years, our homes have been getting smarter with connected and remotely communicating communicating devices. In order to save household energy, the Internet of Things (IoT) also called devices. orderdisruptor” to save household energy, Internet digital of Things (IoT) also the “next great the “nextIngreat is emerging as a the significant revolution [5].called In South Korea, the disruptor” is emerging as a significant digital revolution [5]. In South Korea, the smart home industry smart home industry has come a long way since 2003. It was the time when the apartment owners has comeona long since 2003. It was time when the apartment ownersplanned focusedtoonbuild homea focused homeway networking systems at the a larger scale and the government networking systems at a largerthe scale government planned to build a digitalized digitalized nation. Nowadays, ideaand hasthe gained in prevalence as more buildings can benation. found Nowadays, the idea has gained in prevalence as more buildings can be found equipped with smart equipped with smart environments that can control heating or lights and are saving energy environments that can control heating or lights and are saving energy consumption, although continued consumption, although continued criticism suggests that there is still a long way to go to meet criticism suggests that there still a long way to gohigh to meet household requirements [6]. As there is household requirements [6].isAs there is relatively proportion of energy-demanding industries relatively high proportion of energy-demanding industries in Korea, as compared to other countries, in Korea, as compared to other countries, the recorded energy consumption was 244,626 × 106 kWh the recorded energy consumption wasin244,626 × 106[7]. kWh 2013,Korea, which is theof eighth in the in 2013, which is the eighth largest the world In in South 38% total largest electricity is world [7]. In South Korea, 38% of total electricity is consumed by residential and commercial buildings, consumed by residential and commercial buildings, 55% by the industrial sector, 6% by public 55% byand the 1% industrial sector, 6% by public1)sector sector by transportation (Figure [2]. and 1% by transportation (Figure 1) [2].

South Korea Electricity Consumption 1% 6%

38%

Residential/Commercial Industrial Public

55%

Transport

Figure 1. 1. Energy Energy consumption consumption distribution distribution [2]. [2]. Figure

As mentioned earlier, IoT-based smart homes can help in energy optimization through realAs mentioned earlier, IoT-based smart homes can help in energy optimization through real-time time monitoring. The primary goal of real-time energy monitoring should be exceptional monitoring. The primary goal of real-time energy monitoring should be exceptional predictions of predictions of energy usage. It also must help to implement equipment with accurate load balance. energy usage. It also must help to implement equipment with accurate load balance. Earlier, when IoT Earlier, when IoT systems were not in use, traditional energy management systems used to collect systems were not in use, traditional energy management systems used to collect energy consumption energy consumption data samples at an interval for monitoring purposes. Although, these data samples at an interval for monitoring purposes. Although, these traditional management systems traditional management systems were good at collecting energy consumption data, they failed to were good at collecting energy consumption data, they failed to predict the demand, extract usage predict the demand, extract usage patterns, to alarm the users in case of spikes, or to suggest patterns, to alarm the users in case of spikes, or to suggest effective settings. effective settings. IoT is helping us save energy through the use of energy-efficient appliances, smart meters, IoT is helping us save energy through the use of energy-efficient appliances, smart meters, transferring to off-peak power hours, and smart heating systems [8]. IoT-based energy management transferring to off-peak power hours, and smart heating systems [8]. IoT-based energy management systems are very effective in order to record energy consumptions as well as to predict consumption systems are very effective in order to record energy consumptions as well as to predict in future. Through the use of IoT devices, we are able to collect different data about the smart home consumption in future. Through the use of IoT devices, we are able to collect different data about environment e.g., its temperature, humidity level, user occupancy, the operational status of appliances the smart home environment e.g., its temperature, humidity level, user occupancy, the operational etc. This sort of data can be further used in different machine-learning approaches to train systems status of appliances etc. This sort of data can be further used in different machine-learning in order to predict future energy consumption. The importance of building energy consumption approaches to train systems in order to predict future energy consumption. The importance of predictions in order to make the most informed real-time decisions has been highlighted in many

Energies 2018, 11, 1289

3 of 21

recent studies; and some contributions have come up with prediction models through data-driven or machine-learning approaches [9,10]. According to [11], these predictions are equally important and effective for both the user of a residential building and power-generating companies. These predictions can also help us optimize electricity usage at peak hours, enabling the user to be aware of his energy usage patterns. A user can use these systems to take decisions e.g., to switch between energy sources (e.g., electrical to solar), according to his requirements. In [11], it has also been shown that in South Korea commercial and residential sectors cover around 40% of national energy consumption. Therefore, predicting the energy consumption will not just help a smart home user to effectively utilize resources, it will also have a growing impact on the national economy as well. This has led us to develop an IoT-based prediction model which will help users know their electricity consumption based on previous data also to effectively manage and optimize their energy consumption. This paper focuses on the prediction through the use of machine-learning algorithms along with optimization algorithms. For this we have proposed modification in the hybrid approach of particle swarm optimization (PSO) and neural networks (NN) which can be called the PSO-NN algorithm. As the name suggests, it combines the properties of NN and the PSO algorithm to come up with optimal weights and bias values for neural network training and hence the most accurate prediction results. For the purpose of experiments, we have used a data set of 4 buildings located in Seoul. The rest of the paper is structured as follows. A brief state-of-the-art overview is given in Sections 2 and 3 to present our prediction methodology using PSO and NN. Section 4 describes the data set and experimental set up. The results are presented in Section 5, and Section 6 is based on discussions about possible future directions which conclude the paper. 2. Literature Review Among all the energy sources, electricity plays a vital role. Therefore, in order to make a well-informed energy policy, it is also important to have knowledge about electricity consumption, its usage pattern and demand. There have been many studies focusing on electricity forecasting methods, and energy prediction in commercial and residential buildings has received an exceptional amount of consideration in the research community [12–15]. These forecasting methods for building energy consumption mainly include time-series analysis [16], computational intelligence methods [17–19], statistical methods and traditional multivariant regression [20] methods etc. In [13,21], a detailed review of prediction methods has been presented. As residential buildings are the substantial consumers of electricity in every country, therefore, they are being focused on for the consumption forecast. However, forecasting energy consumption for residential buildings is not simple, as the factors required for the prediction are complex and intertwined, and data unavailability for residential buildings still creates challenges for researchers leaving them with a requirement for comprehensive techniques [22]. This study has focused on providing a contemporary review of different modeling methods and techniques in order to model the energy consumption of residential buildings. Primarily, two different techniques have been identified i.e., bottom up and top down approaches. Each approach depends on various level of information used as an input, various calculations, and the output or results are generated with various applicability. Therefore, a comprehensive and critical analysis is presented along with a detailed explanation of each technique’s pros and cons and comparisons of all techniques. In order to curtail the complications of precise methods, some studies have also proposed simplified methods [23,24]. The key variables used as input information for the energy consumption prediction are very important. In a study conducted by [25], for the forecasting of electricity of the coming one month, a comprehensive list of input features is used. In total, the features amount to 20 significant variables including 14 variables related to weather, 5 variables related to social factors, and monthly electricity consumption. The proposed model, selects the significant input features by using the PSO algorithm along with support vector regression (SVR) and fuzzy rough feature selection. All the features which do not lead towards effective predictions are discarded. In order to validate the

Energies 2018, 11, 1289

4 of 21

prediction results, South Korea’s historical data from 1991–2012 was used. The data was divided into training and testing data sets. The first 20 years of data was used as training data and the remainder for testing. For evaluation along with other standard matrices, it was also analyzed against the results obtained from artificial NNs (ANNs), regression and other models proposed in previous studies. The method proposed in this study appears to have an advantage over previous models due to the automatic selection of significant features for safe predictions. The forecasting models for sensor-based energy are primarily based on machine-learning algorithms as these algorithms make use of historical data for training purpose in order to find the relationship between energy usage and key variables influencing the results e.g., peak hours etc. ANNs [20,26–28] and support vector machines (SVM) [29–31] are two basic algorithms comprehensively used in previous studies for sensor-based energy consumption predictions. ANNs have also being used for the forecasting of energy consumption of solar buildings. In another study [32], three modeling approaches were introduced for the forecasting of electricity consumption. These three approaches include the NNs, decision trees, along with the traditional regression analysis. Two possible cases are investigated by [33] for a building which has been insulated and for a building with at least one brick wall. The investigations were conducted for both winter and summer seasons with some experimental setups. The purpose of this work was to use ANNs to generate a simulation program in order to model the thermal behavior of the building. The learning algorithm was back propagation used in a multilayer recurrent architecture. A recent study [11] performed on energy consumption prediction also used residential buildings data from Seoul, South Korea. In this paper, the proposed model is based on the hidden Markov model based on an algorithm. The predicted results are evaluated against three main prediction algorithms namely, ANNs, SVM and classification and regression trees (CART). An improvement of 2.96%, 6.09%, and 9.03% is seen in the ANN, SVM and CART algorithms respectively. The proposed model was also analyzed on different aggregates of data, e.g., hourly, and weekly etc. We aim to apply PSO-based NNs for making predictions for smart building energy consumption. In PSO-NN the weights and bias values for neural network training are optimized using the PSO algorithm. PSO is very well known for solving optimization problems and its many modified versions have also been tailored for different applications depending on the scenario and problems. In [34], a modified PSO approach is proposed based on the fitness of the particles and the distance between the particles in the population. In [35], a modified approach for the social and cognitive learning factors of PSO is proposed by pre-defining a predicted velocity index. Some other modifications of PSO are presented in [36–40]. For prediction purposes, PSO comes in hybrid with some learning algorithms. One more focused approach in recent times is the PSO-NN, as it has been used for many [41–48] prediction problems. To the best of our knowledge, PSO-NN has not been used for predicting smart building energy consumption. Hence, we consider applying PSO-NN to our smart building data and also offer two modified approaches and make comparisons among them. 3. Particle Swarm Optimization-Based Neural Networks (PSO-NN) Prediction Methodology We introduce neural networks in Section 3.1, particle swarm optimization in Section 3.2, and then elaborate the prediction methodology of particle swarm optimization based neural networks (PSO-NN) for building energy consumption prediction data in Section 3.3. In Section 3.4, we present a modified approach for PSO-NN prediction learning to improve the accuracy of our implemented PSO-NN approach. 3.1. Neural Networks (NNs) The computational model (named as threshold logic) proposed in 1943 by McCulloch and Pitts led to the research of artificial intelligence-based neural networks [49]. Artificial neural networks


5 of 21

3.1. Neural Networks (NNs)

Energies 2018, 11, 1289

5 of 21

The computational model (named as threshold logic) proposed in 1943 by McCulloch and Pitts led to the research of artificial intelligence-based neural networks [49]. Artificial neural networks started to flourish once the processing power of computers increased dramatically, as computation started to flourish once the processing power of computers increased dramatically, as computation power was one of the key issues faced in the progress of ANNs at the initial stages [50]. power was one of the key issues faced in the progress of ANNs at the initial stages [50]. Biologically inspired ANNs are known to produce most accurate prediction results [51]. Biologically inspired ANNs are known to produce most accurate prediction results [51]. ANN ANN learning has two operational modes of training and testing, the system has a set of inputs, learning has two operational modes of training and testing, the system has a set of inputs, weights weights associated with the inputs, hidden layers and a number of outputs. In training, the neuron associated with the inputs, hidden layers and a number of outputs. In training, the neuron learns to learns to decide whether to fire an output for a specific pattern or not, while in testing mode the decide whether to fire an output for a specific pattern or not, while in testing mode the accuracy of accuracy of the learned model is determined. the learned model is determined. The structure of a three-layer neural network is shown in the Figure 2, where we have five The structure of a three-layer neural network is shown in the Figure 2, where we have five inputs, six hidden layers and three outputs. The working of a simple neuron can be explained by inputs, six hidden layers and three outputs. The working of a simple neuron can be explained by Equation (1) [52], whereby a typical neuron computes the output in the following manner: Equation (1) [52], whereby a typical neuron computes the output in the following manner: n

ak = = f ((∑ wki xi )) i =0

(1) (1)

where, are the the inputs inputs to to the the neuron. the output output ofof kth kth neuron. neuron. x1 , x2,, . ...., . , x nare where, ak is the x0 input is bias ( input is bias ()assigning it bk )assigning +it 1 value, + 1 value, with w = b = 1. w , w , . . . , wkn weights are the with = k0 = 1. wk1k, wk2, …,k1 wknk2 are the weights associated eachf input. f is the activation function, which incorporates flexibility in the associated to each to input. is the activation function, which incorporates flexibility in the neural neural networks. networks.

Figure Figure 2. 2. Neural network (NN) layers.

3.2. Particle Swarm Optimization (PSO) In 1995, 1995, Kennedy Kennedy and and Eberhart Eberhart proposed proposed PSO, PSO, which which is is aa population-based population-based optimization optimization In technique inspired inspired by by bird bird flocking flocking and and fish fish schooling schooling theory theory and and also also has has strong strong ties ties to to genetic genetic technique algorithms and and artificial artificial life life [53]. [53]. In search for for food food by by flocking flocking birds, birds, the the bird bird algorithms In the the example example of of the the search closest to the food leads and others follow. As soon as some other bird thinks it has a food source closest to the food leads and others follow. As soon as some other bird thinks it has a food source close close it, it makes a sound and all start birdsfollowing start following it, changing the direction. particle in to it, ittomakes a sound and all birds it, changing the direction. Each Each particle in PSO PSO represents a bird in the flocking example, moving at a certain velocity looking for the optimal represents a bird in the flocking example, moving at a certain velocity looking for the optimal solution solution in the search space. in the search space. In PSO, PSO,first firsta apopulation population particles is initialized defining the number of particles thatcarry will In ofof particles is initialized defining the number of particles that will carry the search for the optimal solution. Each particle has velocity with which it moves through the the search for the optimal solution. Each particle has velocity with which it moves through the search searchand space and values. fitness values. The particles in thespace search by following the with particles space fitness The particles move inmove the search byspace following the particles the best solution so far. Each particle maintains the track of two values as particle’s best (pbest) and global

Energies 2018, 11, x FOR PEER REVIEW Energies 2018, 11, 1289

6 of 21 6 of 21

with the best solution so far. Each particle maintains the track of two values as particle’s best (pbest) and (gbest); is the best solution by itself, the particle itself, is while gbest is the best global (gbest);best pbest is thepbest best solution achieved byachieved the particle while gbest the best solution best solution byinany in the entire population. After finding the pbest and gbest, the found by anyfound particle the particle entire population. After finding the pbest and gbest, the particle updates particle updates its velocity andthe position using the following its velocity and position using following equations [54]. equations [54]. (2) )+ 2× = + 1× ×( − ×( − ) v = v + c1 × rand × ( pbest − present) + c2 × rand × ( gbest − present) (2) (3) = + present = present + v (3) where, is the particle’s velocity, is the current particle position (solution), is where, v is the particle’s velocity, present is the current particle position (solution), pbest is particle’s particle’s personal best solution found so far in the search process, is the global best solution personal the search process, gbest isnumber the global best solution found by any found bybest anysolution particle found so farso infar theinsearch, is a random generated between 0 and 1, particle far the in the search, rand isusually a random generated 0 and 1, and c1, c2 are the and 1, so 2 are learning factors; bothnumber c1 and c2 are keptbetween 2. learning factors; usually both c1 and c2 are kept 2. 3.3. Training NNs with PSO 3.3. Training NNs with PSO Recently, many researchers have been focusing on using PSO for learning ANNs. PSO Recently, many researchers haveprocess been focusing on using PSO for learning ANNs. PSOwork, improves improves the learning and training of an ANN in an efficient manner. In this we the learning and training process of an ANN in an efficient manner. In this work, we present present how PSO can be used to train a neural network and improve its results. The PSO algorithm howoptimize PSO canthe be neural used tonetwork train a by neural networktoand its results. algorithm can can attempting findimprove the optimal weight The and PSO bias values for the optimize the neural network by attempting to find the optimal weight and bias values for the neural neural network’s training. network’s Figuretraining. 3 presents the PSO-based NN model diagram. We are using neural networks for Figure 3 presents energy the PSO-based NN model We are usingPSO neural networks for training training our building consumption data diagram. to make predictions. is used in order to fine our building energy consumption data to make predictions. PSO is used in order to fine tune and tune and optimize weights and biases for the neural networks. It computes the particle’s positions optimize weights the neural networks. It computesPSO, the particle’s positionsofand passes and passes them and on biases for theforlearning process. In NN-based the population particles them on for the learning process. In NN-based PSO, the population of particles generated by PSO generated by PSO struggles its best to look for optimal weight and bias values for the neural struggles its best to look for optimal weight and bias values for the neural network’s training process. network’s training process. The local and global positions in the PSO are updated at each iteration The local and global positions in the updatedThe at each iteration untilwhether the best weights are found until the best weights are found forPSO NNaretraining. analyzer checks best outputs are for NN training. The analyzer checks whether best outputs are computed or more learning is needed. computed or more learning is needed.

Figure 3. Particle swarm optimization-based neural network (PSO-NN) model. Figure 3. Particle swarm optimization-based neural network (PSO-NN) model.

Figure 4 shows the configurations of the prediction system diagram for the PSO-based NN Figure 4 shows the configurations the prediction system for the NN model. The input is divided into test and of training data, training datadiagram to be trained byPSO-based the PSO-based model. networks, The input is divided into test and training data, training data to be trained by results. the PSO-based neural and test data for making predictions and computing accuracy In the neural networks, and test data for making predictions and computing accuracy results. In the processing unit, first the neural networks and particles of PSO are created. Then the initialization processing unit,random first thepositions neural networks and particles of PSO ins areundertaken. created. Then initialization for the weights, and velocities of the particles Thethe PSO algorithm

Energies 2018, 11, 1289

7 of 21


7 of 21

for the weights, random positions and velocities of the particles ins undertaken. The PSO algorithm shuffles and finds thethe new positions to be to neural networks for shufflesthe theparticles particlesatateach eachiteration, iteration, and finds new positions to passed be passed to neural networks training untiluntil the best positions are found. Neural networks use two methods such as tanh for training the best positions are found. Neural networks useactivation two activation methods such as (Equation (4), [55]) and softmax (Equation (5), [56]) for the training, and for computing the output tanh (Equation (4), [55]) and softmax (Equation (5), [56]) for the training, and for computing the values. mean square (6), [57]) is computed by neural by networks time and output The values. The mean(Equation square (Equation (6), [57]) is computed neural each networks eachsent timeback and to the PSO unit for comparing whether global best value is found or not. Once the total epochs have sent back to the PSO unit for comparing whether global best value is found or not. Once the total finished running, then the final weights biases are used makeare predictions for thepredictions test data and epochs have finished running, then theand final weights and to biases used to make for the accuracy of the test data is computed. the test data and the accuracy of the test data is computed.

Figure 4. Configurations diagram. Figure 4. Configurations diagram.

11−−e−2x tanhx tanh == 11++e−2x

(4) (4)

Exp( Xi ) ( ) Softmax( Xi ) = k Softmax( ) =∑ Exp j = 0 ∑ (X j )

(5) (5)

∑in=0 ( Xi − X 0 i ) MSE = ∑ ( n − ′ ) =

(6) (6)

3.4. Modified PSO-NN

3.4. Modified PSO-NN In this sub-section, we present two modified versions of the typical PSO-NN; the first is regeneration-based PSO-NN where we consider the distance and performance particles, In this sub-section, we(R-PSO-NN) present two modified versions of the typical PSO-NN; of the first is and the second is velocity boost based PSO-NN (VB-PSO-NN) where we consider the inertia weight regeneration-based PSO-NN (R-PSO-NN) where we consider the distance and performance of for updating thethe next velocity. particles, and second is velocity boost based PSO-NN (VB-PSO-NN) where we consider the inertia weight for updating the next velocity. 3.4.1. Re-Generation Based PSO-NN (R-PSO-NN) 3.4.1.InRe-Generation re-generation Based basedPSO-NN PSO-NN,(R-PSO-NN) the first step is cluster-based regeneration. In cluster-based regeneration, a close cluster of particles is found then half of the particles from the cluster are moved In re-generation based PSO-NN, the first step is cluster-based regeneration. In cluster-based to some other position where no such cluster exists (Figure 5). The distances between the particles regeneration, a close cluster of particles is found then half of the particles from the cluster are position arrays are calculated using Euclidean distance (Equation (7), [58]) to identify the clusters. moved to some other position where no such cluster exists (Figure 5). The distances between the particles position arrays are calculated using Euclidean distance (Equation (7), [58]) to identify the clusters.


8 of 21

Energies 2018, 11, 1289

8 of 21

( , ) = s

Euclidean Distance ( x, y) =

n

(

− )

∑ ( xi −

y i )2

(7) (7)

i =1

Firstly, Euclidean distance for all particles is calculated and then, based on Euclidean distance, the clusters are identified. A cluster is particles identifiedisifcalculated more thanand three particles found close to each Firstly, Euclidean distance for all then, based are on Euclidean distance, other with aare distance between themis less than aifthreshold The threshold distance the the clusters identified. A cluster identified more thandistance. three particles are found close toiseach minimum distancethem defined below: other with inter-particle a distance between lessas than a threshold distance. The threshold distance is the minimum inter-particle distance defined as below: (8) ( )= × Inter distance Particle Distance NumParticles ( IPD ) = c ×two where, is the minimum threshold between particles, is a constant value (8) for limiting inter-particle distance; c = 0.15; and is the total number of particles in the where, IPD is the minimum distance threshold between two particles, c is a constant value for limiting population. inter-particle distance; = 0.15; NumParticleswith is thevarying total number particleshaving in the population. The constant c is cset after and experimentation values.ofParticles less than the The constant c is set after experimentation with varying values. Particles having less than defined threshold distance between them will count as a cluster and be considered for the redefined threshold distance between them will count as a cluster and be considered for re-generation. generation.

Figure 5. 5. Particles’ Particles’ re-generation re-generation based based on on aa cluster. cluster. Figure

Before re-generating the particle forming a closed cluster, their past performance is checked. Before re-generating the particle forming a closed cluster, their past performance is checked. Particles in a cluster are not re-generated if they are all contributing to finding the global optimal Particles in a cluster are not re-generated if they are all contributing to finding the global optimal value; value; they are moved only when stuck in local optima. Figure 6 elaborates the steps to find particle they are moved only when stuck in local optima. Figure 6 elaborates the steps to find particle clusters clusters and re-generate particles. Clusters are made based on the distance between them and the and re-generate particles. Clusters are made based on the distance between them and the duplicate duplicate clusters (if any) are removed. For each particle Pi, its last updated pbest value is checked. clusters (if any) are removed. For each particle Pi , its last updated pbest value is checked. If the last If the last update in the pbest is found to be 15 iterations ago from current iteration, then it is added update in the pbest is found to be 15 iterations ago from current iteration, then it is added to the list of to the list of particles eligible to be re-generated (R_List). If the particles added to the re-generation particles eligible to or bemore re-generated (R_List). re-generation list are list are about 90% of the total, then If allthe theparticles particlesadded must to bethe re-generated as they areabout stuck 90% or more of the total, then all the particles must be re-generated as they are stuck in local optima. in local optima. Otherwise, half of the particles from the list are re-generated randomly to new Otherwise, positions. half of the particles from the list are re-generated randomly to new positions. If re-generation is made, thethen possibility for performance-based regeneration If nonocluster-based cluster-based re-generation is then made, the possibility for performance-based is evaluated; is forevaluated; which each pbest value’spbest trackvalue’s is maintained. A counter re-generation regeneration forparticle’s which each particle’s track is maintained. A counter rethreshold (RT) is maintained until the particle has no improvement in its pbest it must generation threshold (RT) is maintained until the particle has no improvement in then its pbest thenbeit re-generated to a new position. each epoch, it checks whether thewhether particle’sthe pbest value ispbest improved must be re-generated to a newAt position. At each epoch, it checks particle’s value or not. If pbest is improved with consecutive epochs, then it is not re-generated. Otherwise, if the is improved or not. If pbest is improved with consecutive epochs, then it is not re-generated. particle’s pbest value does not change a set iterations RT andofalso it is notRT theand current Otherwise, if the particle’s pbest valueover does notnumber change of over a set number iterations also global best, then the particle is believed to be stuck in local optima. it is not the current global best, then the particle is believed to be stuck in local optima.

Energies 2018, 11, 1289

9 of 21


9 of 21

Figure6.6.Re-generation Re-generation of particles Figure particlesin inaaclosed closedcluster. cluster.

3.4.2. Velocity Boost-BasedPSO-NN PSO-NN(VB-PSO-NN) (VB-PSO-NN) 3.4.2. Velocity Boost-Based In velocity boost-based PSO-NN we update the particle’s velocity using a new inertia weight In velocity boost-based PSO-NN we update the particle’s velocity using a new inertia weight value, value, occasionally. Initially, we keep our inertia weight constant (w = c = 0.729) as given in [37]. occasionally. Initially, we keep our inertia weight constant (w = c = 0.729) as given in [37]. After each After each epoch in the PSO-NN learning algorithm, we examine the change in the particle’s pbest, epoch in the PSO-NN learning algorithm, we examine the change in the particle’s pbest, maintain a maintain a counter for each particle and have a defined threshold as a velocity boost threshold counter for each particle and have a defined threshold as a velocity boost threshold (VBT); reaching (VBT); reaching VBT, if there’s no improvement in the particle’s pbest then the new inertia weight is VBT, if there’s no improvement in the particle’s pbest then the new inertia weight is used to update used to update the velocities in that epoch. The threshold VBT is further finalized after extensive theexperimentation velocities in that epoch. The values. threshold is furtherinertia finalized after[40] extensive experimentation with different WeVBT use constant weight shown in Equation (9) with different values. We use constant inertia weight [40] shown in Equation (9) and inertia and random inertia weight [41] shown in Equation (10) to tailor our new inertia weightrandom as shown in weight [41] (11): shown in Equation (10) to tailor our new inertia weight as shown in Equation (11): Equation ℎ = = c11 = = 0.7 Contant Inertia Weight 0.7

(9) (9)

() (10) ℎ = 0.5 + Rand() 2 Random Inertia Weight = 0.5 + (10) 2 () (11) ℎ ( ) = 1 + Rand() 3 New Inertia Weight (Wi ) = c1 + (11) 3 In our tailored inertia weight equation; c1 = 0.801, which is selected after playing around with our tailored inertia weight equation; = 0.801, which is selected after playing around with the theInvariations of constant and inertia weightc1combinations. variations of constant and inertia weight combinations. Figure 7 shows the detailed flowchart for our proposed modifications of the PSO-NN Figure 7 shows for our modifications of the PSO-NN algorithm. algorithm. Firstly, the thedetailed neural flowchart network and PSOproposed are generated and initialized. PSO initializes velocities, positions, pbest and gbest error for each particle in theinitializes population. The maximum Firstly, the neural network and PSO areand generated and initialized. PSO velocities, positions, number epochs is the total in number of iterationsThe for maximum finely optimizing theofneural pbest and of gbest and(MaxEpochs) error for each particle the population. number epochs network’s weights using PSO. In each epoch, first the cluster particle regeneration possibility is (MaxEpochs) is the total number of iterations for finely optimizing the neural network’s weights using checked. detailed of cluster-based particle regeneration in Figure 6. If clusterPSO. In eachThe epoch, first flow the cluster particle regeneration possibilityisisgiven checked. The detailed flow of based particle regeneration is not done, then it checks for the performance-based velocity boost cluster-based particle regeneration is given in Figure 6. If cluster-based particle regeneration is not first,then comparing of pbest countvelocity with VBT. If the last update count equals VBT,ofthen done, it checksthe forlast theupdate performance-based boost first, comparing the last update pbest for the current iteration the particle’s next velocity is calculated using new inertia weight (Equation count with VBT. If the last update count equals VBT, then for the current iteration the particle’s next (8)). After updating velocity using the new inertia weight, the inertia weight is reset for the next

Energies 2018, 11, 1289

10 of 21

velocity is calculated using new inertia weight (Equation (8)). After updating velocity using the new Energies 2018, 11, x FOR PEER REVIEW 10 of 21 inertia weight, the inertia weight is reset for the next iterations. The new inertia weight is only used wheniterations. any particle’s last inertia updateweight counter condition is met. the inertia updated too, is then it The new is only used when anyIfparticle’s lastweight updateisn’t counter condition met. the inertia weightparticle isn’t updated too, then it checks pbest for performance-based particle checks for If performance-based regeneration. The particle’s last update count is compared The particle’s pbestRT lastthen, update count isthe compared to be RT;stuck if theatlast updatewith count to RT;regeneration. if the last update count equals assuming particle to position further equals RT then, assuming the particle to be stuck at position with further improvement in pbest improvement in pbest possible, the particle is regenerated to a new position. Upon re-generation, possible, the particle is regenerated to a new position. Upon re-generation, the re-generation the re-generation counter for the particle is reset. counter for the particle is reset.

Figure 7. PSO-NN based on regeneration and velocity boost.

Figure 7. PSO-NN based on regeneration and velocity boost.

We chose to use the inertia weight on a specific condition only, as moving a particle’s position too frequently canthe also be a hurdle finding the optimal solution. we try to give a boost to too We chose to use inertia weightinon a specific condition only, Hence, as moving a particle’s position the velocity when a particle’s performance becomes relatively slow after a certain number of frequently can also be a hurdle in finding the optimal solution. Hence, we try to give a boost to the iterations and then reset it back to the old velocity pace. This procedure proved to be more effective. velocity when a particle’s performance becomes relatively slow after a certain number of iterations In our experimentation, we tested different values for the particle’s re-generation threshold (RT) and then reset it back to the old velocity pace. This procedure proved to be more effective. In our and inertia velocity boost threshold (VBT). These two thresholds are supposed to improve particle experimentation, tested to different for the particle’s re-generation threshold (RT) and inertia pbest once it iswe believed be stuckvalues in the local optima. After multiple comparisons among different

velocity boost threshold (VBT). These two thresholds are supposed to improve particle pbest once it is believed to be stuck in the local optima. After multiple comparisons among different input values, we

Energies 2018, 11, 1289

11 of 21


11 of 21

finalized VBT = 11 and RT = 25, as these two values brought maximum improvements in the prediction input values, we finalized VBT = 11 and RT = 25, as these two values brought maximum accuracies for building energy consumption. improvements in the prediction accuracies for building energy consumption. 4. Data Set and Experimental Setup 4. Data Set and Experimental Setup In In this section environment.We Welook look depth this sectionwe wepresent presentour ourdataset dataset and and experimental experimental environment. inin depth at at thethe data and try to find helpful trends in it also. data and try to find helpful trends in it also. 4.1. Data Set 4.1. Data Set Figure 8 elaborates The data dataset setisisgathered gatheredfrom fromfour fourresidential residential Figure 8 elaboratesthe thedata datacollection collection phase. phase. The buildings of Seoul, South Korea, from January to December 2010. buildings of Seoul, South Korea, from January to December 2010.

Figure 8. Data collection—Year 2010. Figure 8. Data collection—Year 2010.

We have energy consumption data of four residential buildings in Seoul, South Korea. The We have consumption data of relative four residential buildings in Seoul, South Korea. available dataenergy comprises hourly temperature, humidity and energy consumption readings. TheThe available comprises hourly temperature, relative and energy readings. energydata consumption readings are collected forhumidity each available floorconsumption and also overall consumption per hour for each building is added up. The buildings constructed ofconsumption reinforced The energy consumption readings are collected for each available floorare and also overall same as most of theup. residential buildings in South Korea. The first concrete, building has 33 perconcrete, hour forthe each building is added The buildings are constructed of reinforced the same floors (394 ft. tall), the second building has 15 floors (183 ft. tall), the third building has 36 (440 as most of the residential buildings in South Korea. The first building has 33 floors (394 ft. ft. tall), floors, and the has fourth building hasft.33tall), floors ft. tall). The primary source of energy used thetall) second building 15 floors (183 the(394 third building has 36 (440 ft. tall) floors, andinthe thesebuilding buildings is The usedprimary for lighting, cooling, drying, is fourth hasis33electricity, floors (394which ft. tall). sourcecooking, of energy used inwashing, these buildings entertainment etc. Space heating consumption is a combination of electricity via the use of electric electricity, which is used for lighting, cooking, cooling, washing, drying, entertainment etc. Space heaters and gas via floor heating. Our collected data only considers total electricity consumption heating consumption is a combination of electricity via the use of electric heaters and gas via floor per floor of each building. The energy consumption in many aspects is related to the weather of a heating. Our collected data only considers total electricity consumption per floor of each building. region. Seoul has wet and humid summers, dry and cold winters, a windy spring and mild autumn. The energy consumption in many aspects is related to the weather of a region. Seoul has wet and The average range of temperature in winter is [−3 °C, 5 °C], in summer is [12 °C, 30 °C], in spring is humid summers, dry and cold winters, a windy spring and mild autumn. The average range of [5 °C, 15 °C] and in autumn is [6 °C, 14 °C] [59]. temperature in winter is [−3 ◦ C, 5 ◦ C], in summer is [12 ◦ C, 30 ◦ C], in spring is [5 ◦ C, 15 ◦ C] and in Figure 9 shows an overview of the collected data. We have plotted the first 60 h (the first 2.5 autumn is [6 ◦ C, 14 ◦ C] [59]. days of January) sample readings in the graph. The x-axis shows the sample reading number and

Energies 2018, 11, 1289

12 of 21

Figure 9 shows an overview of the collected data. We have plotted the first 60 h (the first 2.5 days of January) sample readings in the graph. The x-axis shows the sample reading number and Energies 2018, 11, x FOR PEER REVIEW 12 of 21 the y-axis shows the values recorded for each hour. Figure 9a shows the energy consumption readings, the9by-axis shows the valuesfor recorded forhumidity each hour. Figure 9a 9c shows the the energy consumption Figure shows the readings relative and Figure shows sample readings for readings, We Figure shows humidity, the readings for relative and humidity Figure 9c shows the sample temperature. have9brelative temperature energyand consumption readings for the four readings for temperature. We have relative humidity, temperature and energy consumption buildings. The temperature range for the plotted sample days is between [−13, 0]. Since the plotted readings for the four buildings. The temperature range for the plotted sample days is between [−13, 0]. days are at the peak of winter, the temperature readings are all negative, reaching −13, but the overall Since the plotted days are at the peak of winter, the temperature readings are all negative, reaching average of winter-collected data is same as mentioned above. The relative humidity readings for the −13, but the overall average of winter-collected data is same as mentioned above. The relative plotted samplereadings are in the range of [30, 90]. are Energy levels for building 1 and building 2 humidity for the plotted sample in theconsumption range of [30, 90]. Energy consumption levels for are between 20 kW to 60 kW, while the energy consumption levels for building 3 and building 4 are building 1 and building 2 are between 20 kW to 60 kW, while the energy consumption levels for between 40 kW to 100 kW. 4 are between 40 kW to 100 kW. building 3 and building Energy Consumption 100

Values Recorded

90 80 70 60

Energy Building 1

50

Energy Building 2

40 30

Energy Building 3

20

Energy Building 4

10 0 1

6

11

16

21

26

31

36

41

46

51

56

Data Samples

(a)

Humidity

Temperature Data Samples

90 80 70 60 50 40 30 20 10 0 1

6 11 16 21 26 31 36 41 46 51 56 Data Samples

(b)

Temperature Readings (Celsius)

Relative Humidity Readings

100

0

1

6 11 16 21 26 31 36 41 46 51 56

-2 -4 -6 -8 -10 -12 -14

(c)

Figure 9. Collected data overview. (a) Sample data for energy consumption; (b) sample data for

Figure 9. Collected data overview. (a) Sample data for energy consumption; (b) sample data for relative relative humidity; (c) sample data for temperature. humidity; (c) sample data for temperature.

Figure 10 shows one day’s sample data for hourly energy consumption of all the four buildings. The graph energy consumption five classesoffor better Figure 10 shows one chart day’sdivides samplethe data for hourly energy into consumption all athe fourvisual buildings. understanding. We can observe that most of the hourly consumption is in the green class i.e., 40–60 The graph chart divides the energy consumption into five classes for a better visual understanding. kWh.

We can observe that most of the hourly consumption is in the green class i.e., 40–60 kWh.

Energies 2018, 11, 1289 Energies 2018, 11, x FOR PEER REVIEW

13 of 21 13 of 21

Energy Consumption (kW)

Buildings Hourly Energy Consumption 100 80-100

80

60-80

60

40-60

40

20-40

20

0-20

0

1 2 3 4 5 6 7 8 9 10 11

12 13 14 15 16 17 18 19 20 21 22 23 Hours 24

Energy Building 4 Energy Building 3 Energy Building 2 Energy Building 1

Figure buildings—1 day. day. Figure 10. 10. Hourly Hourly energy energy consumption consumption for for buildings—1

4.2. Experimental Setup 4.2. Experimental Setup The implementation is performed on an Intel® Core™ i5-4570 CPU (Intel, Santa Clara, CA, The implementation is performed on an Intel® Core™ i5-4570 CPU (Intel, Santa Clara, CA, USA) USA) at 3.20 GHz with 8 GB installed memory (DDR3, Samsung, Seoul, South Korea) and 64-bit at 3.20 GHz with 8 GB installed memory (DDR3, Samsung, Seoul, South Korea) and 64-bit operating operating system (10.0.17134 Build 17134, Microsoft, Redmond, WA, USA). The implementation system (10.0.17134 Build 17134, Microsoft, Redmond, WA, USA). The implementation platform is platform is Visual Studio 2017 (Microsoft, Redmond, WA, USA) and the implementation language Visual Studio 2017 (Microsoft, Redmond, WA, USA) and the implementation language is C#. is C#. The available data as described in Section 4.1 is residential building data spanning a one-year The available data as described in Section 4.1 is residential building data spanning a one-year time period. We have divided our data into 16 classes for hours of the day, each containing two hours, time period. We have divided our data into 16 classes for hours of the day, each containing two and we aim to predict the energy consumption of a smart building every two hours. The data values hours, and we aim to predict the energy consumption of a smart building every two hours. The are normalized before feeding to the prediction algorithm as input. We have 5 inputs to our system as data values are normalized before feeding to the prediction algorithm as input. We have 5 inputs to temperature, humidity, energy consumption, hour of the day and day of the week. We aim to predict our system as temperature, humidity, energy consumption, hour of the day and day of the week. the hourly energy consumption of a building. The data is divided as 75% for the training phase and We aim to predict the hourly energy consumption of a building. The data is divided as 75% for the 25% for the testing phase. training phase and 25% for the testing phase. 5. Results Analysis 5. Results Analysis In this section, we present results analysis and comparisons between our implemented NN In this section, we present analysis and comparisons our of implemented NN and PSO-NN algorithms with ourresults proposed modified PSO-NN. Forbetween the purpose detailed results and PSO-NN algorithms with our proposed modified PSO-NN. For the purpose of detailed results analysis, we implemented our re-generation based (R-PSO-NN) and velocity boost-based (VB-PSO-NN) analysis, weseparately. implemented our we re-generation based (R-PSO-NN) and velocity boost-based (VB-PSOprocedures Hence, make our comparisons between NN, PSO-NN, and the proposed NN) procedures Hence, we ourthe comparisons betweeninNN, PSO-NN, the modified versions.separately. For the comparisons wemake consider prediction accuracy percentages in and contrast proposed modified versions. For the comparisons we consider the prediction accuracy in with number of iterations, number of PSO populations, and re-generation threshold (RT) and velocity percentages in contrast with number of iterations, number of PSO populations, and re-generation boost threshold (VBT) for the modified PSO-NN. threshold (RT) velocity boost threshold (VBT) for the the modified PSO-NN. Figure 11 and shows the output comparisons between prediction accuracy of NN, PSO-NN, Figure 11 shows the output comparisons between the prediction accuracy NN,initially PSO-NN, RR-PSO-NN and VB-PSO-NN. Neural networks start with an accuracy of around of 94.3% until PSO-NN and VB-PSO-NN. Neural networks start with an accuracy of around 94.3% initially until 300 iterations and then rise to 99.08% accuracy at 400 iterations, with a maximum accuracy of 99.32%. 300 iterations and the then rise to of 99.08% accuracy at 400 iterations, a maximum accuracy of PSO-NN achieves accuracy 98.04% at 100 iterations rising towith an accuracy of 99.42% finally. 99.32%. PSO-NN achieves the accuracy of 98.04% at 100 iterations rising to an accuracy of 99.42% R-PSO-NN achieves an accuracy of 99.13% at 100 iterations and reaches the maximum accuracy of finally. beating R-PSO-NN an accuracy of 99.13% at 100 and reaches atthe 99.45%, both achieves NN and PSO-NN. VB-PSO-NN starts withiterations an accuracy of 98.77%, 100maximum iterations accuracy of 99.45%, beating both NN and PSO-NN. VB-PSO-NN starts with an accuracy 98.77%, and achieves maximum accuracy of 99.45% before reaching 200 iterations. We can observe of that while at 100 iterations and achieves maximum accuracy of 99.45% before reaching 200 iterations. We can PSO-NN performs better than NN, achieving high accuracy in less iteration, the modified versions of observe that while PSO-NN performs better than NN, achieving high accuracy in less iteration, the modified versions of PSO-NN beat that too. Modified versions of PSO-NN with their functionality

Energies 2018, 11, 1289

14 of 21

Energies 2018, x FOR REVIEWversions of PSO-NN with their functionality of regenerating particles, 14 of 21 PSO-NN beat11, that too.PEER Modified Energies 2018, 11, x FOR PEER REVIEW 14 of 21 stopping them being stuck at local optima, not only achieved higher accuracy but also with fewer of particles, stopping them being stuck at local not achieved higher of regenerating regenerating particles, stopping being at accuracy local optima, optima, not only only achieved higher iterations. R-PSO-NN took 700 epochsthem to reach itsstuck highest of 99.45% whereas VB-PSO-NN accuracy but also with fewer iterations. R-PSO-NN took 700 epochs to reach its highest accuracy of accuracy but also with fewer iterations. R-PSO-NN took 700 epochs to reach its highest accuracy of achieved the same accuracy in less than 200 epochs. 99.45% whereas VB-PSO-NN achieved the same accuracy in less than 200 epochs. 99.45% whereas VB-PSO-NN achieved the same accuracy in less than 200 epochs.

PredictionAccuracy Accuracy(%age) (%age) Prediction

Building Building Energy Energy Consumption Consumption Prediction Prediction Accuracy Accuracy 100 100 99 99 98 98

NN NN PSO-NN PSO-NN R- PSO-NN R- PSO-NN VB-PSO-NN VB-PSO-NN

97 97 96 96 95 95 94 94

100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 Number of Epochs Number of Epochs Figure 11. Prediction accuracy results comparison. Figure11. 11.Prediction Prediction accuracy accuracy results Figure results comparison. comparison.

Figure Figure 12 12 shows shows aa close close up up view view of of the the comparison comparison between between R-PSO-NN R-PSO-NN and and VB-PSO-NN. VB-PSO-NN. In In Figure 12 shows a close up view of the comparison between R-PSO-NN and VB-PSO-NN. In the the the figure, figure, it it is is quite quite evident evident that that although although initially initially VB-PSO-NN VB-PSO-NN starts starts slowly slowly with with accuracy accuracy lower lower figure, it is quite evident that although initially VB-PSO-NN starts slowly with accuracymaximum lower than than than R-PSO-NN R-PSO-NN at at 100 100 epochs, epochs, after after that that it it drastically drastically increases increases accuracy accuracy and and crosses crosses the the maximum R-PSO-NN at 100 epochs, after that120 it drasticallyHence, increases accuracy and crosses the maximumenergy accuracy accuracy accuracy of of R-PSO-NN R-PSO-NN within within 120 epochs. epochs. Hence, for for our our given given data data of of smart smart building building energy of R-PSO-NN within 120 epochs.approach Hence, for our given dataand of smart building energy consumption, consumption, the VB-PSO-NN is more effective responsive for making the optimal consumption, the VB-PSO-NN approach is more effective and responsive for making the optimal thepredictions. VB-PSO-NN approach is more effective and responsive for making the optimal predictions. predictions.

PredictionAccuracy Accuracy(%age) (%age) Prediction

Velocity Velocity Boost Boost and and Re-generation Re-generation based based PSO-NN PSO-NN 99.6 99.6 99.4 99.4 99.2 99.2 99 99

VB-PSO-NN VB-PSO-NN R-PSO-NN R-PSO-NN

98.8 98.8 98.6 98.6 98.4 98.4

100 110 120 130 131 132 133 134 135 100 110 120 130 131 132 133 134 135 Number of Epochs Number of Epochs

Figure 12. Comparison of re-generation based PSO-NN (R-PSO-NN) and velocity boost based PSOFigure 12. Comparison of re-generation based PSO-NN (R-PSO-NN) and velocity boost based PSOFigure 12. Comparison of re-generation based PSO-NN (R-PSO-NN) and velocity boost based NN (VB-PSO-NN). NN (VB-PSO-NN). PSO-NN (VB-PSO-NN).

5.1. 5.1. Re-Generation Re-Generation Threshold Threshold (RT) (RT) 5.1. Re-Generation Threshold (RT) In In the the R-PSO-NN, R-PSO-NN, we we defined defined aa threshold threshold value value for for the the regeneration regeneration of of particles. particles. In In order order to to find the correct threshold value we played around with different thresholds in the range 15 In the R-PSO-NN, we defined a threshold value for the regeneration of particles. In order find find the correct threshold value we played around with different thresholds in the range 15 to toto30. 30. chose the window of [15, 30], considering that before 15 it is too soon to be changing a particle’s theWe correct threshold value we played around with different thresholds in the range 15 to 30. We chose We chose the window of [15, 30], considering that before 15 it is too soon to be changing a particle’s needs time around also to the number theposition windowas 30],some considering that before 15 and it is we too are soon to looking be changing a particle’s as it position asofit it[15, needs some time to to move move around and we are also looking to minimize minimize the position number of of

Energies 2018, 11, 1289

15 of 21

Energies 2018,time 11, x to FOR PEER around REVIEW and we are also looking to minimize the number of iterations, 15 of 21 needs some move which is why we consider 30 to be the ideal maximum value, after which the particle gets time to resettle in which is some why we considerepochs. 30 to be the ideal maximum value, after which the particle gets theiterations, new position for minimum time to resettle in the new position for some minimum epochs. Figure 13 shows the results of prediction accuracy of energy consumption data with the Figure 13 shows the results of prediction accuracy of energy consumption data with the rere-generation thresholds 15, 20, 25 and 30 in Figure 13a–d respectively. The results are calculated generation thresholds 15, 20, 25 and 30 in Figure 13a–d respectively. The results are calculated for for 100 iterations each. We observe that, at re-generation threshold 15, maximum accuracy is 98.87%. 100 iterations each. We observe that, at re-generation threshold 15, maximum accuracy is 98.87%. The accuracies for threshold 20 went down to around 94% while again rising at threshold 25. At 25, The accuracies for threshold 20 went down to around 94% while again rising at threshold 25. At 25, maximum accuracy is achieved of 99.13% while afterafter 25 the againagain tendstends to decrease. Hence, maximum accuracy is achieved of 99.13% while 25 accuracy the accuracy to decrease. from the comparisons, we could conclude that 25 iterations is a most suitable number for regenerating Hence, from the comparisons, we could conclude that 25 iterations is a most suitable number for theregenerating particles’ positions for a positions quick search an optimal solution. the particles’ for afor quick search for an optimal solution.

Prediction Accuracy (%age)

100 99 98 97 96 95 94 93 92 91 90

Re-Generation Threshold = 20 100 99 98 97 96 95 94 93 92 91 90

11 12 13 14 15 16 17 18 19 20

11 12 13 14 15 16 17 18 19 20

Number of Particles

Number of Particles

(a)

(b)

Re-Generation Threshold = 25


100 99 98 97 96 95 94 93 92 91 90





100 99 98 97 96 95 94 93 92 91 90

11 12 13 14 15 16 17 18 19 20

11 12 13 14 15 16 17 18 19 20

Number of Particles

Number of Particles

(c)

(d)

Figure 13. Searching optimal re-generation threshold (RT) for R-PSO-NN (a) with re-generation Figure 13. Searching optimal re-generation threshold (RT) for R-PSO-NN (a) with re-generation threshold 15; (b) with re-generation threshold 20; (c) with re-generation threshold 25; (d) with rethreshold 15; (b) with re-generation threshold 20; (c) with re-generation threshold 25; (d) with generation threshold 30. re-generation threshold 30.

5.2. Velocity Boost Threshold (VBT) 5.2. Velocity Boost Threshold (VBT) In order to get the optimal VBT value, we ran experiments with values ranging from 5 to 20. In to getthe the optimal VBT value, we ran experiments with 5 to 20. Figureorder 14 shows output accuracies for VB-PSO-NN. From 5 to 10, wevalues get an ranging accuracy from of 94.57%, Figure the output accuracies for VB-PSO-NN. From to 10,11we an accuracy at 11 14 weshows get a sudden rise to the accuracy of 99.45% while again5 after theget accuracy drops of to 94.57%, 99.45 at 11 we get a sudden rise to the accuracy of 99.45% while again after 11 the accuracy drops and keeps fluctuating a little bit until 20. After testing the VBT for these values, we can say thatto1199.45 is and a little until After testing the VBT foraccuracy. these values, we can say that 11 is thekeeps right fluctuating threshold value forbit VBT, as 20. it gives maximum prediction the right threshold value for VBT, as it gives maximum prediction accuracy.

Energies 2018, 11, 1289

16 of 21


16 of 21


16 of 21

Prediction Accuracy (%) Prediction Accuracy (%)

Searching Optimal VBT for VB-PSO-NN

100 99 100 98 99 97 98 96 97 95 96 94 95 93 94 92 93 91 92 91

Searching Optimal VBT for VB-PSO-NN

VB-PSO-NN VB-PSO-NN 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Velocity Boost Threshold (VBT 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Velocity Boost Threshold (VBT Figure 14. Energy consumption prediction. Figure 14. Energy consumption prediction.

5.3. Number of Particles and Epochs Comparison Figure 14. Energy consumption prediction.

5.3. Number of Particles and Epochs Comparison

We also compare our results to the number of particles in PSO. In Figure 13 above, we also 5.3. Number of Particles and Epochs Comparison

observed that accuracy thenumber change in of in particles in the PSO 13 population. In also We also compare our changes results with to the of number particles PSO. In Figure above, we Figure 13, we noticed that higher accuracy is achieved with a particle population size of 12. Now We also compare our results to the number of particles in PSO. In Figure 13 above, we also observed that accuracy changes with the change in number of particles in the PSO population. we run three PSO-NN, R-PSO-NN andchange VB-PSO-NN for a a varying number particles seeNow the observed that accuracy with the in number ofparticle particles in theofPSO population. In we In Figure 13, all we noticed thatchanges higher accuracy is achieved with population size ofto12. effects on prediction accuracy (Figure 15). In the previous experiment, we observed a of population Figure 13, we noticed that higher accuracy is achieved with a particle population size 12. Now run all three PSO-NN, R-PSO-NN and VB-PSO-NN for a varying number of particles to see the effects size of 12 be most productive for PSO-NN and R-PSO-NN. In Figure 15, we that we run allto three PSO-NN, R-PSO-NN and VB-PSO-NN for a varying number of notice particles to for seeVBthe on prediction accuracy (Figure 15). In the previous experiment, we observed a population size of PSO-NN, is accuracy higher with 17 numbers of particles the PSO population, as athe accuracy effects on accuracy prediction (Figure 15). In the previous in experiment, we observed population 12 to rose be most productive for PSO-NN and R-PSO-NN. In Figure 15, we notice that for VB-PSO-NN, from to 99.50%. size of 12 99.45% to be most productive for PSO-NN and R-PSO-NN. In Figure 15, we notice that for VBaccuracy is higher with 17 numbers of numbers particlesofinparticles the PSO as the accuracy rose from PSO-NN, accuracy is higher with 17 in population, the PSO population, as the accuracy 99.45% to from 99.50%. rose 99.45% to 99.50%.

Varrying Number of Particles

Prediction Accuracy (%) Prediction Accuracy (%)

100

Varrying Number of Particles

99 100 98 99 97 98 96

PSO-NN (%)

97 95 96 94

R-PSO-NN (%) PSO-NN (%) VB-PSO-NN R-PSO-NN (%)

95 93 94 92 93 91 92

VB-PSO-NN 11

12

13

11

12

13

91

14

15

16

17

18

19

20

Number of Particles 14 15 16 17 18

19

20

Number of Particles Figure 15. Change in accuracy with varying number of particles.

In Table 1, we present maximum achieved for each of the NN, PSO-NN, R-PSOFigure our 15. Change in accuracy withaccuracies varying number of particles. Figure Change earlier, in accuracy with varying of particles. NN and VB-PSO-NN. As15. discussed the results changenumber with such parameters as the number of particles population and the total number of epochs to PSO-NN, run the training. In Tablein1,the we PSO present our maximum achieved accuracies for eachprovided of the NN, R-PSOHence in the table below, we give our best accuracies with the number of particles and number of NN and VB-PSO-NN. As discussed earlier, the results change with such parameters as the number In Table 1, we present our maximum achieved accuracies for each of the NN, PSO-NN, R-PSO-NN epochs required to achieve them. The table shows that VB-PSO-NN proves to be the best approach, of particles in the PSO population and the total number of epochs provided to run the training. and VB-PSO-NN. As discussed earlier, the results change with such parameters as the number of Hence in the table below, we and givethe ourtotal best accuracies the number of particles andtraining. number of particles in the PSO population number ofwith epochs provided to run the Hence epochs required to achieve them. The table shows that VB-PSO-NN proves to be the best approach,

in the table below, we give our best accuracies with the number of particles and number of epochs required to achieve them. The table shows that VB-PSO-NN proves to be the best approach, as it not

Energies 2018, 11, 1289

17 of 21

only gives the maximum accuracy but also consumes the minimum number of epochs. While for the accuracy of 99.50%, VB-PSO-NN took 17 particles in the PSO population, even with a lesser PSO population 1211, it xgives same maximum accuracy as R-PSO-NN, taking 133 epochs only. 17 of 21 Energiesof 2018, FOR PEER REVIEW Table 1. Best accuracy comparison. as it not only gives the maximum accuracy butparameters also consumes the minimum number of epochs. While for the accuracy of 99.50%, VB-PSO-NN took 17 particles in the PSO population, even with a Number of Epochs Numberaccuracy of Particles Best Accuracy lesser PSO Algorithm population of 12 it gives same maximum as R-PSO-NN, taking 133 epochs only. ANN 600 99.32% PSO-NN R-PSO-NN VB-PSO-NN

6. Discussion

800

12

12 comparison. Table 1.700 Best accuracy parameters 133

17

99.42% 99.45% 99.50%

Algorithm Number of Epochs Number of Particles Best Accuracy ANN 600 99.32% PSO-NN 800 12 99.42% R-PSO-NN 700 12 99.45% 133 17 99.50% PSO-NN proposedVB-PSO-NN a particle re-generation and velocity boost-based

We have for residential electricity consumption prediction. The data we have used is from January to December 2010; one 6. Discussion of the vital questions is the validity of the data gathered 8 years ago. In order to make sure that data We have particle re-generation velocity boost-based PSO-NN residential is not outdated, weproposed examinea residential electricityand trends from 2010 onwards. Anforincrease of 7% is electricity consumption prediction. The data we have used is from January to December 2010; one seen in the total electricity consumption of South Korea from January 2010 to January 2018 [60] and an of the vital questions is the validity of the data gathered 8 years ago. In order to make sure that data increase of 1.2% is seen in the residential electricity consumption of South Korea from 2010 to 2016 [2]. is not outdated, we examine residential electricity trends from 2010 onwards. An increase of 7% is Hence, we can consider our provided data still to be in context. seen in the total electricity consumption of South Korea from January 2010 to January 2018 [60] and Inanorder to validate model’s efficiency, we chose [61] and increase of 1.2% is our seenproposed in the residential electricity consumption of ANN South Korea fromSVM 2010 [62] to for comparisons; as the literature review indicates ANNs and SVMs to be the most widely used algorithms 2016 [2]. Hence, we can consider our provided data still to be in context. in buildingInenergy consumption [14,15]. In this work, we implemented hybrid order to validate our predictions proposed model’s efficiency, we chose ANN [61] and aSVM [62] PSO-NN for comparisons; as theenergy literature review indicates ANNs and to be the most widely used algorithm for building predictions and proposed twoSVMs modifications in it for improving its algorithms in building consumption [14,15]. In this work, implemented performance. PSO-NN is anenergy optimized versionpredictions of ANN. The performance of we PSO-NN will beabetter hybrid PSO-NN algorithm for building energy predictions and proposed two modifications in it for than ANN as PSO-NN optimizes the weights of ANN. Hence, the results achieved by our two proposed improving its performance. PSO-NN is an optimized version of ANN. The performance of PSO-NN modifications of PSO-NN improve the prediction accuracy even more, outperforming the performance will be better than ANN as PSO-NN optimizes the weights of ANN. Hence, the results achieved by of ANNs 1). For further comparisons, we run the SVM our input data. We have our (Table two proposed modifications of PSO-NN improve thealgorithm predictiononaccuracy even more, used MATLAB R2018a (The MathWorks, Inc., Natick, MA, USA) for training our data on the outperforming the performance of ANNs (Table 1). For further comparisons, we run the SVM SVM model.algorithm The implementation of SVM in MATLAB automatically data into ofMA, [−1, +1], on our input data. We have used MATLAB R2018a (Thescales MathWorks, Inc.,range Natick, USA) for training our data on the SVM model. The implementation of SVM in MATLAB thus we do not need to pre-process data any further. We trained our data on six different kernels automatically scales data intoGaussian, range of [−1, +1], thusGaussian we do not and needcoarse to pre-process dataSVM. any further. as linear, cubic, quadratic, fine medium Gaussian The results We trained our data on six different kernels as linear, cubic, quadratic, fine Gaussian, medium with the quadratic kernel (97.89%) and cubic kernel (97.89%) were best in comparison to all six SVM Gaussian and coarse Gaussian SVM. The results with the quadratic kernel (97.89%) and cubic kernel implementations but still lower in comparison to ANN, PSO-NN and our proposed R-PSO-NN and (97.89%) were best in comparison to all six SVM implementations but still lower in comparison to VB-PSO-NN (Figure 16). ANN, PSO-NN and our proposed R-PSO-NN and VB-PSO-NN (Figure 16).

Energy Prediction Accuracy (%)

Energy Prediction Accuracy 99.6 99.1 98.6 98.1 97.6 97.1 96.6 Cubic/Quadratic SVM

ANN

PSO-NN

R-PSO-NN

Prediction Algorithm

Figure 16. Comparison of final prediction accuracies.

Figure 16. Comparison of final prediction accuracies.

VB-PSO-NN

Energies 2018, 11, 1289

18 of 21

The R-PSO-NN approach, while it increases accuracy and decreases the number of epochs, still is not as efficient as V-PSO-NN (Table 1 and Figure 16). One of the reasons could be the far re-generation of particles from possible solution positions most of the time. The particle re-generation involves a random factor and we are not fully controlling the re-generation position of the particle. While in V-PSO-NN the accuracy is higher with a drastic decrease in the number of epochs, we have carefully developed an equation for velocity boost after using multiple combinations of the equation constants and have fine-tuned the VBT value after experimenting with different threshold values. Hence, for V-PSO-NN, we can conclude that our fine-tuned parameter values allow the particle to make a quick move towards the right direction before it moves towards a non-solution direction or towards getting trapped in local optima. Our main motivation in this work is to further improve the performance of PSO-NN for domains where ANN prediction algorithm is best suited. In the context of previous implementations of ANN/PSO-NN, the proposed modifications will bring an improvement in the results and can easily be incorporated into any ANN/PSO-NN algorithm without changing the core of the algorithm, resulting in a more efficient hybrid approach. The conducted study and proposed mechanisms for accurate and efficient energy prediction will be of benefit in terms of timely predicting upcoming energy demands for smart buildings and managing energy consumption schedules accordingly. Moreover, accurate energy consumption also plays a vital role in load-shifting from peak hours to semi-peak or off-peak hours. One of the key factors in managing and optimizing the energy consumption of smart buildings is the detailed availability of multi-dimensional data. True user behavior observations and energy consumption trends can help us focus and set the right direction for finding solutions and optimizing energy. Hence, in future, we strive to collect timely and advanced data for smart buildings and explore various aspects of smart buildings’ energy consumption such as user activity-based consumption trends, peak and off-peak hours consumption patterns etc. 7. Conclusions In this work, we aim to address the issue of efficient energy predictions in order to optimize energy consumption in the smart buildings sector. In order to achieve this, we focused on prediction algorithms and their respective accuracy in different areas. We applied PSO-NN, a combined approach of using neural networks for learning along with one of the widely used optimization algorithms i.e., the particle swarm optimization (PSO) algorithm. The proposed algorithms were implemented on a dataset of smart buildings in Seoul, South Korea. This data was collected for a period of one year. Jn recent years neural networks and advanced machine learning techniques have been used extensively for prediction related problems, and so it is high time for advanced and hybrid approaches to be used in order to produce more accurate and optimized results. In this study, we have proposed two variants of PSO-NN, re-generation based PSO-NN (R-PSO-NN) and velocity boost based PSO-NN (VB-PSO-NN). Both the proposed approaches are experimented on using the aforementioned data set. The VB-PSO-NN approach showed higher performance in terms of the accuracy of results as compared with NN, PSO-NN and R-PSO-NN. Although the accuracy achieved through R-PSO-NN was slightly lower than VB-PSO-NN, it outperformed NN and PSO-NN. The results obtained in the form of prediction accuracy of energy consumption for smart buildings has given us the confidence to explore more dimensions of data and continue to build prediction models for smart buildings using PSO-NN and its variants. Author Contributions: S.M. designed the PSO-NN-based prediction methodology for energy prediction in residential smart buildings, performed the experiments, analyzed the data and wrote the paper. D.K. managed the data curation process, conceived the overall idea of energy prediction in residential smart buildings, and supervised this work. Both authors contributed to this paper. Acknowledgments: This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (2014-1-00743) supervised by the IITP (Institute for Information & communications Technology Promotion), and this work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government

Energies 2018, 11, 1289

19 of 21

(MSIT) (No. 2017-0-00756, Development of interoperability and management technology of IoT system with heterogeneous ID mechanism). Any correspondence related to this paper should be addressed to DoHyeun Kim; [email protected]. Conflicts of Interest: The authors declare no conflict of interest.

References 1.

2. 3. 4. 5.

6. 7. 8. 9. 10. 11. 12. 13. 14.

15. 16. 17. 18. 19. 20. 21.

Global Demand for Energy Will Peak in 2030, Says World Energy Council. Available online: https://www.theguardian.com/business/2016/oct/10/global-demand-for-energy-will-peak-in2030-says-world-energy-council (accessed on 5 April 2018). Energy Information Korea. 2016. Available online: https://www.keei.re.kr/web_keei/en_publish.nsf/XML_ Portal_New/B9C4089854585599492580BF0009503E/$file/EnergyInfo2016.pdf (accessed on 4 May 2018). International Energy Outlook. 2017. Available online: https://www.eia.gov/outlooks/ieo/pdf/0484(2017) .pdf (accessed on 5 April 2018). Deru, M.; Torcellini, P. Source Energy and Emission Factors for Energy Use in Buildings (Revised); No. NREL/TP-550-38617; National Renewable Energy Laboratory (NREL): Golden, CO, USA, 2007. The Internet of Things: Making the Most of the Second Digital Revolution. Available online: https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/409774/14-1230internet-of-things-review.pdf (accessed on 5 April 2018). Korea’s Smart Homes Need New Lessons. Available online: http://koreajoongangdaily.joins.com/news/ article/article.aspx?aid=3029201 (accessed on 5 April 2018). Korea Energy Agency. 2014. Available online: http://www.energy.or.kr/web/kem_home_new/energy_ issue/mail_vol22/pdf/publish_05_201507.pdf (accessed on 5 April 2018). 4 Ways the Internet of Things Is Saving Energy, Water & Money. Available online: http://dadolabs.com/ energyefficiency/ (accessed on 5 April 2018). Amasyali, K.; El-Gohary, N. Building lighting energy consumption prediction for supporting energy data analytics. Procedia Eng. 2016, 145, 511–517. [CrossRef] Hu, Y.-C. Electricity consumption prediction using a neural-network-based grey forecasting approach. J. Oper. Res. Soc. 2017, 68, 1259–1264. [CrossRef] Israr, U.; Ahmad, R.; Kim, D. A Prediction Mechanism of Energy Consumption in Residential Buildings Using Hidden Markov Model. Energies 2018, 11, 358. [CrossRef] Perez-Lombard, L.; Ortiz, J.; Pout, C. A review on buildings energy consumption information. Energy Build. 2008, 40, 394–398. [CrossRef] Zhao, H.-X.; Magoules, F. A review on the prediction of building energy consumption. Renew. Sustain. Energy Rev. 2012, 16, 3586–3592. [CrossRef] Ahmad, A.S.; Hassan, M.Y.; Abdullah, M.P.; Rahman, H.A.; Hussin, F.; Abdullah, H.; Saidur, R. A review on applications of ANN and SVM for building electrical energy consumption forecasting. Renew. Sustain. Energy Rev. 2014, 33, 102–109. [CrossRef] Fumo, N. A review on the basics of building energy estimation. Renew. Sustain. Energy Rev. 2014, 31, 53–60. [CrossRef] Nogales, F.J.; Contreras, J.; Conejo, A.J.; Espínola, R. Forecasting next-day electricity prices by time series models. IEEE Trans. Power Syst. 2002, 17, 342–348. [CrossRef] Massimiliano, M.; Caputo, P.; Costa, G. Paradigm shift in urban energy systems through distributed generation: Methods and models. Appl. Energy 2011, 88, 1032–1048. Adel, A.; Hemmati, M.; Abdoos, A.A. Short term load forecasting using a hybrid intelligent method. Knowl. Based Syst. 2015, 76, 139–147. Gallego, C.; Costa, A.; Cuerva, A. Improving short-term forecasting during ramp events by means of regime-switching artificial neural networks. Adv. Sci. Res. 2011, 6, 55–58. [CrossRef] Tutun, S.; Chou, C.-A.; Erdal, C. A new forecasting framework for volatile behavior in net electricity consumption: A case study in Turkey. Energy 2015, 93, 2406–2422. [CrossRef] Lu, X.; Lu, T.; Kibert, C.J.; Viljanen, M. Modeling and forecasting energy consumption for heterogeneous buildings using a physical-statistical approach. Appl. Energy 2015, 144, 261–275. [CrossRef]

Energies 2018, 11, 1289

22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34.

35. 36. 37.

38.

39. 40. 41. 42.

43. 44. 45.

20 of 21

Swan, L.G.; Ugursal, V.I. Modeling of end-use energy consumption in the residential sector: A review of modeling techniques. Renew. Sustain. Energy Rev. 2009, 13, 1819–1835. [CrossRef] Al-Homoud, M.S. Computer-aided building energy analysis techniques. Build. Environ. 2001, 36, 421–433. [CrossRef] Wang, S.W.; Xu, X.H. Simplified building model for transient thermal performance estimation using GA-based parameter identification. Int. J. Therm. Sci. 2006, 45, 419–432. [CrossRef] Hyojoo, S.; Kim, C. Short-term forecasting of electricity demand for the residential sector using weather and social variables. Resour. Conserv. Recycl. 2017, 123, 200–207. Yang, J.; Rivard, H.; Zmeureanu, R. On-line building energy prediction using adaptive artificial neural networks. Energy Build. 2005, 37, 1250–1259. [CrossRef] Karatasou, S.; Santamouris, M.; Geros, V. Modeling and predicting building’s energy use with artificial neural networks: Methods and results. Energy Build. 2006, 38, 949–958. [CrossRef] Wong, S.L.; Wan, K.K.W.; Lam, T.N.T. Artificial neural networks for energy analysis of office buildings with daylighting. Appl. Energy 2010, 87, 551–557. [CrossRef] Dong, B.; Cao, C.; Lee, S.E. Applying support vector machines to predict building energy consumption in tropical region. Energy Build. 2005, 37, 545–553. [CrossRef] Kavaklioglu, K. Modeling and prediction of Turkey’s electricity consumption using support vector regression. Appl. Energy 2011, 88, 368–375. [CrossRef] Zhao, H.-X.; Magoulès, F. Feature selection for predicting building energy consumption based on statistical learning method. J. Algorithms Comput. Technol. 2012, 6, 59–78. [CrossRef] Tso, G.K.; Yau, K.K. Predicting electricity energy consumption: A comparison of regression analysis, decision tree and neural networks. Energy 2007, 32, 1761–1768. [CrossRef] Kalogirou, S.A.; Bojic, M. Artificial neural networks for the prediction of the energy consumption of a passive solar building. Energy 2000, 25, 479–491. [CrossRef] Peram, T.; Veeramachaneni, K.; Mohan, C.K. Fitness-distance-ratio based particle swarm optimization. In Proceedings of the 2003 IEEE Swarm Intelligence Symposium, SIS’03, Indianapolis, IN, USA, 24–26 April 2003; pp. 174–181. Cai, X.; Cui, Y.; Tan, Y. Predicted modified PSO with time-varying accelerator coefficients. Int. J. Bio-Inspired Comput. 2009, 1, 50–60. [CrossRef] Chen, X.; Li, Y. A modified PSO structure resulting in high exploration ability with convergence guaranteed. IEEE Trans. Syst. Man Cybern. Part B 2007, 37, 1271–1289. [CrossRef] Sedghi, M.; Aliakbar-Golkar, M.; Haghifam, M.R. Distribution network expansion considering distributed generation and storage units using modified PSO algorithm. Int. J. Electr. Power Energy Syst. 2013, 52, 221–230. [CrossRef] Ghomsheh, V.S.; Shoorehdeli, M.A.; Teshnehlab, M. Training ANFIS structure with modified PSO algorithm. In Proceedings of the IEEE Mediterranean Conference on Control & Automation, MED’07, Athens, Greece, 27–29 June 2007; pp. 1–6. Wang, S.; Watada, J. A hybrid modified PSO approach to VaR-based facility location problems with variable capacity in fuzzy random uncertainty. Inform. Sci. 2012, 192, 3–18. [CrossRef] Cui, G.; Qin, L.; Liu, S.; Wang, Y.; Zhang, X.; Cao, X. Modified PSO algorithm for solving planar graph coloring problem. Prog. Nat. Sci. 2008, 18, 353–357. [CrossRef] Chau, K.W. Particle swarm optimization training algorithm for ANNs in stage prediction of Shing Mun River. J. Hydrol. 2006, 329, 363–367. [CrossRef] Chatterjee, S.; Sarkar, S.; Hore, S.; Dey, N.; Ashour, A.S.; Balas, V.E. Particle swarm optimization trained neural network for structural failure prediction of multistoried RC buildings. Neural Comput. Appl. 2017, 28, 2005–2016. [CrossRef] Song, Y.; Chen, Z.; Yuan, Z. New chaotic PSO-based neural network predictive control for nonlinear process. IEEE Trans. Neural Netw. 2007, 18, 595–601. [CrossRef] [PubMed] Chau, K.W. Application of a PSO-based neural network in analysis of outcomes of construction claims. Autom. Constr. 2007, 16, 642–646. [CrossRef] Lu, W.Z.; Fan, H.Y.; Lo, S.M. Application of evolutionary neural network method in predicting pollutant levels in downtown area of Hong Kong. Neurocomputing 2003, 51, 387–400. [CrossRef]

Energies 2018, 11, 1289

46. 47. 48. 49. 50. 51.

52. 53. 54. 55. 56. 57. 58. 59. 60. 61. 62.

21 of 21

Ahmadi, M.A. Neural network based unified particle swarm optimization for prediction of asphaltene precipitation. Fluid Phase Equilib. 2012, 314, 46–51. [CrossRef] Gordan, B.; Armaghani, D.J.; Hajihassani, M.; Monjezi, M. Prediction of seismic slope stability through combination of particle swarm optimization and neural network. Eng. Comput. 2016, 32, 85–97. [CrossRef] Sheikhan, M.; Mohammadi, N. Time series prediction using PSO-optimized neural network and hybrid feature selection algorithm for IEEE load data. Neural Comput. Appl. 2013, 23, 1185–1194. [CrossRef] McCulloch, W.; Walter, P. A Logical Calculus of Ideas Immanent in Nervous Activity. Bull. Math. Biophys. 1943, 5, 115–133. [CrossRef] Minsky, M.; Papert, S. Perceptrons: An Introduction to Computational Geometry; MIT Press: Cambridge, MA, USA, 1969. Artificial Neural Networks as Models of Neural Information Processing|Frontiers Research Topic. Available online: https://www.frontiersin.org/research-topics/4817/artificial-neural-networks-as-models-of-neuralinformation-processing (accessed on 30 March 2018). Artificial Neuron Output. Available online: https://en.wikipedia.org/wiki/Artificial_neuron (accessed on 4 May 2018). Kennedy, J.; Eberhart, R. Particle Swarm Optimization. In Proceedings of the IEEE International Conference on Neural Networks, Perth, Australia, 27 November–1 December 1995; pp. 1942–1948. Poli, R.; Kennedy, J.; Blackwell, T. Particle swarm optimization. Swarm Intell. 2007, 1, 33–57. [CrossRef] Hyperbolic Function. Available online: https://en.wikipedia.org/wiki/Hyperbolic_function (accessed on 4 May 2018). Softmax Function. Available online: https://en.wikipedia.org/wiki/Softmax_function (accessed on 4 May 2018). Mean Squared Error Function. Available online: https://en.wikipedia.org/wiki/Mean_squared_error (accessed on 4 May 2018). Euclidean Distance. Available online: https://en.wikipedia.org/wiki/Euclidean_distance (accessed on 4 May 2018). Seoul Average Weather. Available online: https://www.ncdc.noaa.gov/ (accessed on 4 May 2018). Korea Electricity Consumption. Available online: https://www.ceicdata.com/en/korea/electricitygeneration-and-consumption/electricity-consumption (accessed on 4 May 2018). Bebis, G.; Georgiopoulos, M. Feed-forward neural networks. IEEE Potentials 1994, 13, 27–31. [CrossRef] Krammer, K.; Singer, Y. On the algorithmic implementation of multi-class SVMs. Proc. JMLR 2001, 2, 265–292. © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).