Enhancing LSTM Algorithms for Photovoltaic Power Forecasting

Photovoltaic (PV) energy, a cornerstone of renewable power generation, is instrumental in driving the shift toward eco-friendly industrial practices. By harnessing solar power, it fosters a balanced and sustainable progression that benefits economic growth, social advancement, and environmental preservation alike. Accurate PV power generation forecasting can help power system operators plan ahead, reduce the use of backup power, lower operational costs, optimize power generation schedules, and enhance the grid integration capability and generation efficiency of new energy microgrids. In practical applications, methods from the field of deep learning are commonly used for PV forecasting. Among these, Long Short-Term Memory (LSTM) networks excel at short-term photovoltaic power prediction, offering distinct benefits when dealing with long-term dependencies in time series data while maintaining remarkable flexibility. Their architecture proves particularly adept at capturing complex, time-dependent relationships that simpler models often miss. However, LSTM's high computational complexity, sensitivity to hyperparameter selection, and high data quality requirements make it less suitable for long-term forecasting. To improve forecasting accuracy, a combination of Variational Mode Decomposition (VMD) and an improved Grey Wolf Optimization (IGWO) algorithm is used to optimize the LSTM forecasting network, along with a self-attention mechanism, to establish a short-term PV forecasting model based on VMD-IGWO-LSTM-ATTENTION. This model is compared experimentally with LSTM, VMD-LSTM, and VMD-IGWO-LSTM, and the model accuracy is assessed using multiple error indicators. The findings demonstrate the proposed composite model effectively enhances PV forecasting accuracy.