Multi-temporal PV power prediction using long short-term memory and wavelet packet decomposition

Abstract

The integration of photovoltaic (PV) systems into power grids presents operational challenges due to the inherent variability in solar power generation. Accurate PV power forecasting can help address these issues by enhancing grid reliability and energy management. This study introduces a novel hybrid deep learning approach that combines Wavelet Packet Decomposition (WPD) and Long Short-Term Memory (LSTM) networks to improve forecasting accuracy across multiple time horizons. The proposed model incorporates a dynamic weighting mechanism to optimally integrate the forecasts of decomposed subseries, effectively capturing both high- and low-frequency components of the power signal. Using real-world data from a solar parking site at the University of Twente, Netherlands, the proposed models are compared with standard LSTM, Linear Regression, and Persistence baselines across 15 min, 1-hour, and day-ahead horizons. The WPD-LSTM model with weight optimization reduces nRMSE by up to 72.5%, 52.9%, and 34.7% compared to Persistence, and by 68.6%, 36.1%, and 7.5% compared to standalone LSTM, respectively. These results highlight the effectiveness of the hybrid approach in delivering more accurate and robust PV power forecasts.