An Attention-Based Photovoltaic Forecasting Scheme Combined with LSTM Model

Abstract

With the development of new energy around the world, the proportion of photovoltaic energy used as a clean energy in the distribution network is gradually rising. The forecast of photovoltaic power plants is vital to many energy providers for their marketing analysis. Thus, photovoltaic forecasting has become an important research direction at present. However, owing to the high volatility and intermittent characteristics of photovoltaic power generation, it is still a challenge to predict photovoltaic power accurately. As far as traditional photovoltaic forecasting methods are concerned, SVM and ARIMA, as machine learning methods can solve the timing prediction problem of certain scenarios, but they are often not appropriate for some time series closely related to features. To address this problem, this paper proposes a short-term photovoltaic load forecasting model based on the Attention mechanism and LSTM model. Firstly, the correlation coefficient and LASSO regression are used for feature selection to filter out redundant features. Secondly, a long short-term memory network (LSTM) is used to make predictions to solve the problem of gradient disappearance during model training. Finally, the Attention mechanism is added to better capture feature weights and further improve the prediction accuracy of the training model. The proposed method can predict the change trend well. Comparative results confirm that the proposed method with feature selection can has better effect than ARIMA, SVM, and ELM.