SolarSG-Net: A Noise-Resilient Chemical–Environmental Hybrid Framework for Solar Power Maximum Forecasting

Abstract

Improving power forecasting of solar photovoltaic (PV) is essential for sustainability and grid stabilization efforts. Nevertheless, the solar generation is sensitive to atmospheric conditions and environmental changes as well as chemical agents that reduce solar radiation. The SolarSG-Net is a new hybrid forecasting framework proposed in this study, which consists of Savitzky–Golay (SG), Principal Component Analysis (PCA), and Bidirectional Long Short-Term Memory (BiLSTM). The SG filter significantly improves data quality and removes high-frequency sensor noise while keeping irradiance peaks. PCA eliminates the redundancies among the parameters ambient air PM2.5, PM10, NO2, SO2, CO and aerosol optical depth (AOD). The BiLSTM model is capable of capturing forward and backward dependencies. The experimental results show that SolarSG-Net outperforms earlier methods, including LSTM, CNN-LSTM, Random Forest (RF), and Support Vector Regression (SVR), with RMSE = 2.84%, MAE = 2.17%, and R² = 0.985. The findings confirm that pollution-aware preprocessing can enhance prediction stability and generalization under varying atmospheres.