Sliding Window-Based Deep Learning Approach for Solar Power Forecasting in Malaysia Utility-Scale PV Systems

2025-09-10T15:40:08+07:00

Accurate forecasting of solar power in utility-scale photovoltaic (USPV) systems is critical for grid stability but remains challenging due to meteorological variability and the large spatial scale of these systems. However, the choice of sliding window size in time-series forecasting remains underexplored. This study introduces a deep learning-based forecasting framework that systematically evaluates
the impact of sliding window size on forecasting accuracy using multivariate time-series data. The data collected from a 25 MWac USPV system in Malaysia between August 2022 and April 2023, comprises 5-minute interval measurements of solar irradiance, module temperature and solar power output. Multiple deep learning (DL) models, namely LSTM, CNN and GRU across window sizes ranging from 12 to 288 steps and forecasting horizons of 1 to 12 hours were investigated. Results show that a 144-step window consistently improves accuracy over conventional one-step input methods, with LSTM outperforming other models by achieving up to 23.1% RMSE reduction, 30.7% MAE reduction and a 8.6% increase in R² at 60 minutes forecasting horizon. This work emphasizes the importance of window size selection in optimizing forecasting accuracy for USPV systems and supporting renewable energy grid integration. By improving forecasting capabilities, this research is expected to provide critical insights to enhance renewable energy integration into the grid system.

Journal of Renewable Energy and Smart Grid Technology

Sliding Window-Based Deep Learning Approach for Solar Power Forecasting in Malaysia Utility-Scale PV Systems