Enhancing short-term solar power forecasting using a hybrid deep learning approach and a comparative analysis of machine learning models

Photovoltaic (PV) energy is essential for sustainable urban infrastructure but remains constrained by weather-induced variability, particularly under tropical climatic conditions. Accurate solar power forecasting is critical for grid stability and resilient energy systems. This study proposes a hybrid deep learning model that integrates Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures within a sequential configuration designed to enhance temporal feature learning. In the proposed model, the GRU functions as the primary connector, extracting short-term temporal dependencies and transferring them to the LSTM layer for long-term contextual representation. The model employs a recursive forecasting strategy using time-indexed data to effectively capture the nonlinear dynamics of PV power generation. Two PV technologies, Poly-crystalline (Array 1) and Mono-crystalline (Array 2), are evaluated. Model accuracy is assessed using statistical metrics supported by residual analyses to examine correlations between observed and predicted outputs. The proposed hybrid demonstrates superior predictive capability, achieving root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) values of 18.55 W, 19.43 W, and 0.9961 for Array 1, and 14.82 W, 11.85 W, and 0.9952 for Array 2, respectively. Incorporating GRU as the primary hybrid component yields performance improvements of 13.97% (RMSE) and 16.03% (MAE) for Array 1, and 16.41% (RMSE) and 12.44% (MAE) for Array 2. Beyond these quantitative enhancements, the study explores the functional advantages and limitations of employing GRU as the dominant connection module. Overall, the findings confirm that the proposed model, supported by recursive forecasting, provides a scalable solution for improving PV power prediction in variable tropical environments.