The key issue in SG is accurate load and price forecasting. A new enhanced classifier .... LSTM and RNN for load and price forecasting. These are known as the ...
Electricity Load And Price Forecasting using Enhanced Radial Basis Function Network Approach in Smart Grid (Technical Report for MSCS Course: Research Methodology for Information Technology RMIT CUI Fall 2018 16 ) Nosheen Nazar and Nadeem Javaid COMSATS University Islamabad, Pakistan Email: {noshinazar05, nadeemjavaidqau}@gmail.com
Abstract—In this paper, the main focus is the management of energy consumption and generation through smart grid (SG). The key issue in SG is accurate load and price forecasting. A new enhanced classifier technique is proposed to address the forecasting issue. The proposed model of enhanced technique consists of two stages: feature engineering and classification. Feature engineering contains feature selection and feature extraction. Feature selection is done by using Decision Tree (DT) and Recursive Feature Elimination (RFE) which improves the feature selection by reducing the feature redundancy. The next step of feature engineering, feature extraction is done by using ReliefF, which identifies the interaction of features. For classification, along with Support Vector Machine (SVM) another technique, Radial Basis Function Network (RBFN) is used. The comparison of proposed technique with the two existing techniques shows better results in terms of accuracy. In this paper, residential data is used. The simulation results shows load and price forecasting.
I. INTRODUCTION Smart grid (SG) is basically a power system that intelligently manage generation, consumption and distribution of energy. SG systems are flexible enough to exchange information in all directions which means it enable two-way communication between utility and customer. SG introduces new technologies and apply different techniques on power grid to manage the generation and consumption of energy. As we all know, the population of the world is increasing exponentially and the requirement of electricity is increasing even more than exponentially. The new generation is attracted towards SG due to the shortage of energy as it is our most valuable asset. The main goal of SG is to reduce the peak load and balance out the generation and consumption of energy which is not possible in traditional grids (TGs). In TG, electricity distribution is one-way process. In addition, more energy needs to be generated because of 30% to 40% dissipation and wastage of energy in TG or conventional grid. The usage of energy is not directly proportional to the population of human, even per human number of appliances are increasing so does the energy utilization. [?] SG is a modern form of TG. SG are saving one third of the energy as compared to TG. By energy preservation, energy cost can also be reduced in SG.
SG introduces new technologies such as smart metering, digital monitoring, sensors, various forms of distributed generation and solar energy to manage our quickly changing electric demand. SG ensures sustainability by providing efficient energy management. In SG, transmission of electricity is more efficient and faster in restoration after power disturbances as compared to TG. In SG, there are different domain areas which includes renewable energy sources, security and privacy for integration of SG with cloud computing however the most significant area in SG is data analytics. Data analytics is used in SG in terms of extracting useful information from actual data for guiding purposes like billing information and status of electricity. [?]. In SG, at each point there is data and there are different devices which can store and transmit data using data analytic techniques and estimate the electricity consumption like smart meters. In this paper, we are predicting load and price by using SG data. With the help of smart meter, a consumer takes part in SG operation to get enough information about the future energy generation and to manage energy consumption. Data analytics application will help to include data from the both internal and external systems like weather forecast data or demographic data of consumers of any product worldwide. Data analytics now provide extensive and vital role in SG with huge amount of data collections for future and provide vital roles in making future polices and plans. Advanced machine learning techniques are required to manage SG. The most common technique for machine learning is basically categorized as supervised and unsupervised learning through algorithms. There are different sources of data in SG like market data, residential data, commercial data, data of smart homes. This data can used to detect electricity thefts on wider networks spreading in different regions and also may help to deduce the development of people in different regions. It will differentiate the unit consumption of different regions along their statistics. Therefore, to make SG more efficient and productive, different models of data analytics are used. In data analytics, the most important methods are energy forecasting. SG can help to automate and manage the complexity and required needs of electricity in this century. It also empowers
consumers with real time information system about their energy consumption. It also helps electricity providers to reduce outages and integrate renewable energy consumption through solar, wind and hydro plants. Above methods and models are used to predict load and price. For demand response,these models can describe the customer behavior. Data analytics will help you to keep, relate and evaluate historical data with the real time data.In this paper, data analytics is used for the sake of load and price forecasting. Machine learning consists of two categories of algorithms: classification which includes categorical values and regression used in case of continuous values. These algorithms are used for two types of prediction: temporal (hourly intervals) and spatial (building by unit) [?]. SVM is a supervised learning algorithm. Data handling is difficult in traditional grids which creates difficulty in training the data and its utilization. SVM reduces that difficulty by dividing the data into correct categories. The main disadvantage of SVM is that it is weak in processing the uncertain data and its key parameters needs to be set correctly for best classification results. In this paper, both SVM and RBFN based models are presented as classifiers to predict short-term electricity load and price. The enhanced version of RBFN is known as Enhanced Radial Basis Function Network (ERBFN). The main objective of this paper is to achieve higher accuracy for load and price forecasting. The terms prediction and forecasting are used alternatively. By using SVM and RBFN, proposed model has achieved better classification results. In general, each model effectiveness seems to depend on the dataset. The paper is organized in such a manner: Section 2 deals with related work, Section 3 covers the proposed model. Results are given in Section 4 whereas Section 5 deals with the performance metrics. Conclusion is given in Section 6. TABLE I: List of Abbreviations Abbreviation AEMO ANN ARIMA CNN DR DS DRN DTR EPEX ERBFN FFNN LSTM MISO MLP NLS-SVM NYISO PJM RBFN RFE SDA SG SVD SVM TG
Full Form Australia Electricity Market Operators Artificial Neural Network Auto Regressive Integrated Moving Average Convolutional Neural Network Demand Response Decision Tree Deep Residual Network Decision Tree Regression European Power Exchange Enhanced Radial Basis Function Network Feed Forward Neural Network Long Short Term Memory Midcontinent Independent System Operator Multi Layer Perceptron Nonlinear Least Square Support Vector Machine New York Independent System Operator Pennsylvania-New Jersey-Maryland Radial Basis Function Network Recursive Feature Elimination Stacked De-noising Autoencoders Smart Grid Singular Value Decomposition Support Vector Machine Traditional Grid
A. Motivation After reviewing the work done by the authors in [?] [?], we have following motivations for this paper: 1) Short term forecasting should be done for both load and price by using data analytic techniques. 2) A new technique should beat the existing technique in terms of accuracy and efficiency. B. Problem Statement As the population of the world and the requirement of electricity are increasing exponentially, the main issue is energy shortage. To overcome this issue, migration form TG towards SG is required.In TG, we need Energy Renewable Sources (ERS) for energy management. However, SG ensures sustainability by providing efficient energy management. By using latest techniques of SG, load and price could forecast more efficiently and it may help to save one third of the energy as compared to TG. The limitations of SVM are the tunning of parameters and computational complexity [?]. To overcome this limitation, we have proposed a new classifier technique and after combining these two techniques, better simulation results have been achieved. This new technique is known as SVM-ERBFN. C. Contribution In this paper, the new proposed classification technique called ERBFN shows better simulation results than the previous two techniques. This new technique is known as SVMERBFN. The main goal of this paper is to achieve higher accuracy. We have integrated selection, extraction and classification in our proposed system model. For the given dataset [?], there are five techniques which we have implemented for different purposes such as Relief-F is used for feature extraction. DT and RFE are used for feature selection and dimensionality reduction. For classification, we have used two techniques which are SVM and RBFN to forecast load and price. The proposed system model forecasting performance is better than before as we have performed extensive simulations to achieve better results. II. R ELATED WORK The purpose of this work is to understand the relationship between energy consumption of appliances and different predictors. There are many techniques which have been individually used in the past for load and price forecasting. These techniques are categorized as: data-driven, classical and artificially intelligent. Classical methods includes the mathematical models like ARIMA, Random Forest etc and artificially intelligent methods are more like biological neurons like LSTM, CNN, RBFN, FFNN etc. In this paper,data-driven predictive models [?] for appliances’ energy use has been discussed. Some widely used methods for energy prediction are ANN and SVM. The main objective of these models is to efficiently predict load and price. RBFN is an artificial neural network which trains rapidly that uses radial basis functions as activation functions. Linear combination of radial basis functions is the output of
TABLE II: Summary of the Related Work Technique ARIMA ARMAX Pooling Deep RNN CNN, LSTM DT Hilbertian ARMAX MLR, BaggedT, NN LSTM, DNN, GRU ARIMA, TV-ABC, NLS-SVM MI, ANN DWT-IR, SVM, Sperm Whale ANN, SVM, RBFNN QO-ABC, LS-SVM Relief-F, MI, CFS DRN DNN ELM, WNN
Features Price Price Load Price Load Price Load Load and price Load and weather Load Load and price Load and price Load and price Load Load and temperature Load and price Price
Data Set PJM AEMO IRISH International Exchange Appliance Energy Prediction EPEX Beijing EPEX PJM, NYISO PJM NYISO NSW NYISO, PJM NSW ISONE EPEX AEMO
the network and neuron parameters. It uses as a non-linear classifier. Neural networks [?] are considered better than other techniques such as ARIMA, Random Forest and Naive Bayes because of its automatic feature and extraction training processes. It classifies by measuring the input similarities. In [?], two datasets are used for load forecasting. For price prediction, ARIMA model is used in PJM region containing residential data. [?]. ARIMA model needs stable electricity market to perform efficiently. [?]. In papers [?] and [?], artificial intelligent models are being used such as CNN, LSTM and RNN for load and price forecasting. These are known as the complex models in terms of time and space because of their hidden layers. In paper [?], comparison was done between different techniques to choose the best model for prediction. Similarly in paper [?], three techniques were compared for load and price forecasting. The results show that GRU outperforms DNN and LSTM in terms of accuracy. In [?], ARIMA and NLS-SVM are used in combination for load and price forecasting. In paper [?], feature selection is done by using the MI technique then selected features are fed to ANN model for load prediction using PJM dataset. The structure of SVM is commonly used in many research fields. In papers [?], [?], [?], SVM is used for feature selection process. However, it lacks to discuss about the redundancy of features. As stated earlier, the limitations of SVM are the tunning of parameters and computational complexity. Therefore, we have used another technique called RFE to remove redundancy of features to achieve better results. To optimize predictive performance,we combined some of the above mentioned techniques. Some widely used machine learning methods are SVM and ANN. By using such network architectures, we could achieve best overall performance. In [?], combination of DNN and LSTM is done for day-ahead price forecasting. DNN is popular for extracting complex patterns. In [?], two variants of DNN are being used. In paper [?], ELM is used in order to make the model efficient. In paper [?], different types of feature selection algorithms are used such as MI, Relief-F and CFS. For better accuracy achievement, this paper proposed a model in which two techniques are combined as a classifier and used for prediction.
Region US Australia Ireland Belgium Belgium Spain, Germany Beijing Belgium US, Australia US Australia Australia US Austrailia US France Australia
Limitations Computationally expensive Cannot deal with non-linearity Time complexity Computationally expensive Over fitting Cannot deal with extreme non linearity Not suitable for long term forecasting Over fitting Computationally expensive Time complexity Slow convergence scheme Hard to tune parameters Time complexity Time Complexity, Over fitting Space complexity Over fitting
Feature engineering is one of the application of classifiers. There are two main processes in feature engineering which includes feature selection and feature extraction. There are several techniques used for feature selection and extraction. However, it has been mentioned that DT faces over fit problem and it performs better in training process. For load and price forecasting, preprocessing is the most important element in order to get high accuracy. For preprocessing, data has to be in normalized form. III. PROPOSED SYSTEM MODEL To demonstrate the working of proposed model, house appliances’ energy consumption has been discussed along with other features including measurements of house temperature and humidity conditions with a 10-min interval for a period of 4.5 months. Home energy usage is also correlated with weather conditions. It is desirable to select the most relevant features and also find out the irrelevant features which ones do not improve the prediction of appliances’ energy consumption. Furthermore, load and price forecasting has been calculated with the help of several features discussed earlier by applying three different techniques including DT, RFE, Relief-F and SVM. Load and price are predicted by setting a target feature. In the proposed system, the enhanced version of RBFN classifier is implemented with SVM classifier to enhance the accuracy and to get good predictions. RBFN use radial functions as their activation function, resulting in low error rate in the predictions. After prediction, the predicted electric load and price values are plotted on the graphs for easy visualization. The purpose of this implementation is also to understand the relationships between appliances’ energy consumption and different parameters. According to the given dataset, 75 percent data has been considered as training data and 25 percent as testing data. Since the dataset contains several features or parameters. For prediction, model has considered different parameters which includes temperature, humidity, weather conditions, atmospheric pressure. To filter out non-predictive parameters, different methods are implemented to extract the irrelevant features and reserved the most important ones.
Fig. 1: Proposed System Model
Dataset splitting into testing and training datasets has been done on two stages;before normalization process and after the selection of normalized features for the classifier. In general, the proposed model consists of following parts: 1) Splitting the dataset into training and testing sets on two stages. 2) Feature selection is done by using DT and RFE techniques. 3) Relief-F is used for feature extraction. 4) For prediction of load and price, SVM along with RBFN and ERBFN is used. 5) Four performance metrics are used for the measurement of accuracy. IV. R ESULTS A ND S IMULATIONS Five methods are implemented in the proposed system model.These five methods include Relief-F for feature extraction process, after the extraction of irrelevant features DT and RFE are considered together for feature selection process. SVM and ERBFN are the two techniques which are jointly implemented and work as a new classifier (SVM-ERBFN) in the proposed system model. For implementation, simulations are performed in Spyder (Python 3.6). Simulations are run on the system Intel Core i7, having 16GB RAM. The results and description of all four stages of the system model are divided as follows: A. Data Description In the given dataset, house appliances’ energy consumption has been discussed along with other features including measurements of house temperature and humidity conditions. Home energy usage is also correlated with weather conditions. The data is taken for 4.5 months. The dataset is divided into 75% and 25% i.e., 75 percent data has been considered as training data and 25 percent as testing data. Table 3 shows the features which are used in the dataset.
B. Feature Selection using DT and RFE Feature Selection is a process of selecting subset of relevant features for the purpose of model construction. DT correctly identifies the importance of all the features of given dataset and make a list to demonstrate it. Splitting of the dataset has been done before implementing DT. It eliminates the worst performing features on a particular model one after another until the best features are obtained.It is used to select the most important features near to the target feature. After applying DT on the given dataset, the most important number of features are selected and features with less importance are rejected. TABLE III: Feature Description and Units Short Form A L T1 RH1 T2 RH2 T3 RH3 T4 RH4 T5 RH5 T6 RH6 T7 RH7 T8 RH8 T9 RH9 To Ph
Features Appliances energy consumption Light energy consumption Kitchen Temperature Kitchen Humidity living room Temperature Living room Humidity Laundry room Temperature Laundry room Humidity Office Temperature in Humidity in office room Bathroom Temperature Humidity in bathroom Outside Temperature of north side building Humidity of the building Ironing room Temperature Humidity of ironing room Teenager room 2 Temperature Humidity of teenager room 2 Parents’ room Temperature Humidity of parents’ room Outside Temperature (from Chi`evres weather station) Pressure from weather-station
Units Wh Wh C % C % C % C % C % C % C % C % C % C mm Hg
C. Feature Extraction using Relief-F Feature Extraction is one of the key steps in the implementation of model. For efficient feature extraction, Relief-F is used.
Algorithm 1 Algorithm of RFE Take Input Features (F): Training set S Tune the model on the training set Set of features F = (f1, f2, .... fn) Rank the features Calculating variable importance Subset size of i (f1 go fn) Set rank of F f* - last ranked feature in F R(f-i + 1) - f* Recalculate the rankings F - F-f* Output Best features
D. Prediction After feature selection and extraction, both SVM and RBFN classifiers predicts the load and price by using selected features. 1) Support Vector Machine (SVM): SVM is a type of supervised machine learning classification algorithm. In the proposed model, SVM is implemented as a classifier which gives the prediction about the load and price. in this proposed model, SVM concept, known as Kernel SVM is used. Load and Price forecasting has shown in the figures below. 2) Radial Basis Function Network (RBFN): In the proposed system model, RBFN is implemented as a combined classifier. The RBFN model is composed of three layers and it uses Radial Function as its activation function. The predicted load of the RBFN model can be viewed from the figure below. E. Forecasting
Algorithm 2 Algorithm of Relief-F 1) 2) 3) 4) 5) 6) 7) 8) 9) 10) 11)
Input features Weights assigned to all features (F): W[B]=0 Select a random instance, Ri Find k-nearest hits, Hj Find k-nearest miss for all classes , Mj(C) Update weights of each attribute end Calculate the average of k-nearest hits and miss. Determine the number of important variables Use model correlate to the optimal S(i) output
Figures 2-6 show the load and price prediction comparison of two different techniques. The first graph is for SVM; the second graph is for RBFN. Actual SVM
1000
Load (MW)
Relief-F uses filter approach for feature extraction and dimension reduction. It is implemented for feature extraction as it finds the score between features and select top scoring features. This feature scoring based on the nearest neighbor instance pairs. According to the dataset, it has selected the fifteen features.
800 600 400 200 0
0
50
100
150
200
Hours
250
300
350
Fig. 2: Support Vector Machine Load Forecasting
10000
Algorithm 3 Algorithm of SVM 1) 2) 3) 4) 5) 6) 7) 8) 9) 10) 11) 12) 13) 14)
Input Dataset Splitting Train set T Tune SVM on training set for each class Ap PA Ai (Xi ) ← 1/AN i=1 N (yˇi − y)2 ifAi (Xi )
