Developing a novel hybrid formulation for short-term solar-irradiance prediction by leveraging feature extraction algorithm

Accurate forecasting of short-term solar irradiance is critical for energy generation, transmission, and grid integration with a deep focus on renewable energy sources. A precise estimation of irradiance is challenging primarily due to the nonlinear and nonstationary characteristics influenced by meteorological conditions and geographical positions. In this regard, this study proposes a novel hybrid model, namely, the TVF-EMD-MRMR-Seq2Seq-LSTM for hourly solar irradiance prediction. It is based on (i) a time-varying filter empirical mode decomposition (TVF-EMD) for feature extraction, (ii) the minimum redundancy and maximum relevance (MRMR) technique for feature selection, and (iii) a bidirectional long short-term memory network infused with encoder–decoder architecture (Seq2Seq-LSTM) for regression. The proposed methodology is rigorously validated through hourly solar irradiance data encompassing an eight-year temporal period (2015–2022) across four distinct climatic regions from India, namely, Hamirpur (Himachal Pradesh), Jafarabad (Gujarat), Bhopal (Madhya Pradesh), and Thiruvananthapuram (Kerala). The experimental results based on the values of root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), R2, and correlation coefficient (ρ) reveal that the TVF-EMD-MRMR-Seq2Seq-LSTM model achieves 80%–93% accuracy improvement in terms of relative RMSE reduction in comparison to the standalone baseline MLP model. Moreover, the Nash–Sutcliffe efficiency coefficient, paired t-test values, and a graphical analysis of the prediction results validate the robustness of the methodological framework. Therefore, the findings strongly indicate the adaptability and efficacy of the proposed hybrid model.