Accurate Short-term Forecasting for Photovoltaic Power Method Using RBM Combined LSTM-RNN Structure with Weather Factors Quantification

Abstract

After the operation of the photovoltaic system, there is a problem in grid connection and power sales which is the prediction of photovoltaic output power. The power forecasting problem of the photovoltaic systems is strongly influenced by the weather factors, and there is continuity of time variation within the input sequence. Most forecasting methods have volatility problems for short-time scale predictions, seldom consider the changes of weather-affecting factors and their influence modeling. When it comes to a newly implemented photovoltaic power plant, the scale of samples is usually not enough to finish the training process. In order to solve the above problems, this paper proposes a predictive network with cascaded network structure, the pre-stage network is used to quantify the weather influence factors, and the post-stage network is used to give the prediction of the photovoltaic power output. The proposed method considers the quantification of weather influencing factors, explores the time series intrinsic links contained in the power output sequence and uses the unique stacked organization of LSTM structure to reduce the dependence on the scale of training datasets. This paper uses a Tianjin local photovoltaic power plant as certain instance, and finally achieve the goal to make an accurate prediction of photovoltaic power output by hourly scale, which verifies the validity of the proposed method.