Predictive model for PV power generation using RNN (LSTM)

• Published in: Journal of mechanical science and technology Vol. 35; no. 2; pp. 795 - 803
• Main Authors: Park, Min Kyeong, Lee, Jong Man, Kang, Won Hee, Choi, Jong Min, Lee, Kwang Ho
• Format: Journal Article
• Language: English
• Published: Seoul Korean Society of Mechanical Engineers 01.02.2021; Springer Nature B.V; 대한기계학회
• Subjects: Alternative energy sources; Artificial neural networks; Coefficient of variation; Control; Dynamical Systems; Electric power generation; Energy management systems; Energy storage; Engineering; Error reduction; Industrial and Production Engineering; Machine learning; Mechanical Engineering; Meteorological data; Neural networks; Optimization; Original Article; Photovoltaic cells; Prediction models; Predictive control; Recurrent neural networks; Renewable energy sources; Renewable resources; Root-mean-square errors; Vibration; 기계공학; Deep learning; PV power generation; Model predictive control; Recurrent neural network; Renewable energy
• ISSN: 1738-494X; 1976-3824
• DOI: 10.1007/s12206-021-0140-0

In recent years, advanced information technologies, such as deep learning and big data, have been actively applied in building energy management systems to improve energy efficiency. Various studies have been conducted on the prediction of renewable energy performance using machine learning techniques. In this study, a recurrent neural network (RNN) was utilized in predicting photovoltaic (PV) power generation. An RNN is an artificial neural network in which the connection between units is composed of a cyclic structure that can reflect the characteristics of time series. Therefore, to eventually incorporate a model predictive control technique for energy demand and supply matching, this study uses previously measured weather data and PV power generation data to predict the future PV power generation. Various optimization processes, such as normalization, classification of learning data, and setting of layer options, are performed to create a predictive model. Furthermore, 500 hidden neurons and 1 and 3 hidden layers are created and compared. The initial learning rate for both single and multiple-layer options was set to 0.05. As a result, the coefficient of variation of the root mean squared error, Cv(RMSE), which is a statistical index to assess the overall accuracy of the predicted values, was 13.8 % for a single layer and 13.2 % for a multiple-layer long short term memory model. Although no significant difference was observed between the two models, the multiple-layer model showed a slightly reduced error rate than the single-layer model, and the accurate prediction was confirmed to be achievable. Based on this study, a stable energy supply system can be established through linkage with energy storage control by predicting the amount of energy production through renewable energy sources and the amount of energy demand at urban scale.