Solar Power Prediction Using Deep Learning ... - onlinepresent.org

6 downloads 0 Views 934KB Size Report
Abstract. This paper presents solar power output prediction using long short- term memory (LSTM) network of artificial neural network. The neural network.
Advanced Science and Technology Letters Vol.146 (FGCN 2017), pp.148-151 http://dx.doi.org/10.14257/astl.2017.146.26

Solar Power Prediction Using Deep Learning Technique Prasis Poudel and Bongseog Jang Department of Multimedia Engineering Mokpo National University Korea [email protected]

Abstract. This paper presents solar power output prediction using long shortterm memory (LSTM) network of artificial neural network. The neural network is applied to model solar power data obtained from the experimental site in the Hae-Nam, Korea. LSTM neural network with one input node, two hidden layers and one output node is used to model a day solar power data. The results obtained from the comparison of LSTM neural network and moving average indicate that the LSTM neural network approach is reasonable for short term solar power prediction. Keywords: solar power output, long short-term memory neural network

1

Introduction

Solar energy obtained from solar power module is fluctuating in nature affected by external environment conditions like solar irradiance and cell temperature. This fluctuating nature has a great impact on the PV system planning, operation, and its economic analysis. Thus it is good if we have a prediction method for photovoltaic power system. Currently, a number of methods are being applied for photovoltaic power generation prediction. In terms of theory and methodology, they are classified into three categories; artificial neural network (ANN) based model [1], time series model [2], and time trend extrapolation model [3]. Among these models, ANN has a high prediction accuracy, however, it is complex modelling together with high requirements of data samples. In this paper, LSTM network [4] is used for short term PV power output prediction. LSTM network is a recurrent neural network, trained using backpropagation through time and overcomes the vanishing gradient problem. The objective of this study is to develop a real time solar power prediction system having acceptable accuracy with low complexity. This paper is organized as follows, description of LSTM network is presented in Section 2 and simulation results are described in Section 3. Finally, conclusion is stated in Section 4.

ISSN: 2287-1233 ASTL Copyright © 2017 SERSC

Advanced Science and Technology Letters Vol.146 (FGCN 2017)

2

LSTM Neural Network

LSTM is based on recurrent neural networks (RNN). Due to its recurrent connection between the neurons in RNN, it uses temporal information of the inputted data. An LSTM has a special neuron structure called memory cell. These memory cell will store information over the time. Each memory block consists of three gates: the input gate, forgot gate and output gate. These gate will control the information flow for each neuron. Furthermore, each gate has an activation function. For example, if the input gate takes a high activation, the input will be stored in the memory cell. If the output gate has a high activation, it will release the stored information to the next neurons. And if the forget gate notes a high activation, the memory cell will be cleared. The basic structure of LSTM network is shown in Figure 1.

Fig. 1. LSTM basic structure.

The LSTM neural network shown in above figure has xt input and output yt, where t is the prediction period. it, ft and ot are input, forget and output gate respectively.

3

Simulation Result

The inputs for LSTM neural network are real-time solar power output obtain from solar modules at Hae-Nam in Korea. In the LSTM neural network, we used the single input layer, two hidden layers, and one output layer. The collected solar power was

Copyright © 2017 SERSC

149

Advanced Science and Technology Letters Vol.146 (FGCN 2017)

divided into 90% for training and 10% for testing the data. Inside the neural network model, final output layer is activated with linear regression function. To accelerating the model training procedure, algorithms was implemented in Python with Anaconda Jupiter notebook [5]. Here, we used the 500 epochs to train the network. Loss value obtain from the LSTM model is shown in Figure 2.

Fig. 2. Loss value in LSTM model

From Figure 2, we can see the loss value is smoothly decreased as the number of epochs increased. After the completion of the training, the model predicts short-term power output. The overall accuracy of the performed neural network was 80.082% achieved with 8.679 root mean square error (RMSE). For comparison purpose, the solar power prediction by applying moving average (MA) of order 3 was conducted. Prediction result obtained from LSTM model and MA(3) are shown in Figure 3.

Fig. 3. Comparison of prediction result between the neural network and moving average

150

Copyright © 2017 SERSC

Advanced Science and Technology Letters Vol.146 (FGCN 2017)

Table 1. Comparison of RMSE Algorithm LSTM Neural network MA (3)

RMSE 8.679 10.6775

Table 1 shows the RMSE value of LSTM network and MA (3) predictions. The result shows that the better prediction was achieved by the LSTM model. This implies that, in this simple simulation case, the neural network is better than MA(3) for solar power output prediction in terms of the RMSE result.

4

Conclusion

In this paper, solar PV power data is used for short-term power prediction using LSTM neural network. The neural network predicted power is compared with the moving average modeling. From the result, it was found that the neural network is better for solar PV prediction in terms of the RMSE comparison results. This implies that neural network may be more effective tool for solar power output prediction. We are continuing storing the daily solar PV power output from Hae-Nam site to our lab of Mokpo national university. For the next phase of this work, we will train the neural network with one-year historical data for more accurate prediction results. Further, we will also use the meteorological data as an input to predict the solar power using the neural networks. Acknowledgments. This work is supported by Korea Electric Power Corporation (KEPCO) Smart Energy Campus Microgrid Project (2016-2019).

References 1. S. Alam, S. C. Kaushik and S. N. Garg, “Assessment of diffuse solar energy under general sky condition using artificial neural network”, Applied Energy, vol. 86, no. 4, (2009), pp. 554-564. 2. Y. Cui, Y. C. Sun and Z. L. Chang, “A review of short-term solar photovoltaic power generation prediction methods”, Resources Science, vol. 35, no. 7, (2013), pp. 1474-1481. 3. Y. Li, L. He and N. Niu, “Forecasting power generation of grid-connected solar PV system based on Markov chain”, Acta Energiae Solaris Sinica, vol. 35, no. 4, (2014), pp. 611-616. 4. S. Hasim, S. Andrew and B. Francoise, “Long Short- Term Memory Recurrent Neural Network Architecture for Large Scale Acoustic Modeling”, Proceedings of the Annual Conference of the International Speech Communication Association, (2014). 5. Python Software Foundation, https://www.python.org.

Copyright © 2017 SERSC

151