Ultrashort-term prediction of photovoltaic power based on a bidirectional LSTM model

Abstract

The rapid increase in the use of solar PV, a climate-dependent energy source, presents a challenge for the management of power grids. This intermittency makes the development of reliable generation forecasting models critical for maintaining system balance and operational integrity. In this work, an improved long short-term-memory (LSTM) network model called Bi-LSTM, which consists of two concurrent LSTM layers, is proposed to predict subseries with different minute–hour–day intervals. The bidirectional nature of the Bi-LSTM allows it to integrate information from preceding and succeeding points in a sequence. This comprehensive context awareness underpins the success of the model in applications such as natural language processing, audio recognition, and time series analysis. Finally, a comparative study with the LSTM, MLP, SVR, Bi-LSTM, CNN, GRU, KNN and BPNN models is performed by using an actual dataset collected from the National Renewable Energy Laboratory (NREL) in the United States. The simulation results show that the proposed prediction model outperforms the other 7 prediction models with different statistical matrix methods regarding PV power forecasting in every time interval and season. According to the R² statistical metric results, the Bi-LSTM models exhibit high consistency for the training set overall, with scores approaching 0.99.