Day-ahead power forecasting in a large-scale photovoltaic plant based on weather classification using LSTM

Photovoltaic (PV) solar power generation is always associated with uncertainties due to weather parameters intermittency. This poses difficulties in grid management as solar penetration rate rise continuously. Thus, accurate Photovoltaic (PV) power prediction is required for the successful integration of solar energy into the power grid, and short-term forecasting (minutes-1 day ahead) is significant for real-time power dispatching. Day-ahead power output time-series forecasting methods are proposed in this paper, in which ideal weather type and non-ideal weather types have been separately discussed. For ideal weather conditions, a forecasting method is proposed based on meteorology data of next day for ideal weather condition, using long short term memory (LSTM) networks. For non-ideal weather conditions, time-series relevance and specific non-ideal weather type characteristic are considered in LSTM model by introducing adjacent day time-series and typical weather type information. Specifically, daily total power, which is obtained by discrete grey model (DGM), is regarded as input variables and applied to correct power output time-series prediction. Prediction performance comparison between proposed methods with traditional algorithms reveal that the RMSE accuracy of forecasting methods based on LSTM networks can reach 4.62% for ideal weather condition. For non-ideal weather condition, the dynamic characteristic is effectively described by proposed methods and the proposed methods obtained superior prediction accuracy.