Intelligent Solar Power Forecasting Using a Neuro-Evolutionary DE-Optimized TCN–LSTM Hybrid Model

Abstract

Solar photovoltaic (PV) generation changes a lot, which makes it hard to keep the grid stable, trade energy, and manage storage. This shows how important it is to be able to accurately predict the short term. This research presents DE-TCN-LSTM, a hybrid deep learning framework that integrates Temporal Convolutional Networks (TCN), bidirectional Long Short-Term Memory (LSTM), and Differential Evolution (DE) for the automation of hyperparameter optimization. Rather than introducing novel model architectures, this study presents a regulated comparison of Differential Evolution (DE) with Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) for the optimization of Temporal Convolutional Networks-Long Short-Term Memory (TCN-LSTM) models in medium-scale photovoltaic output forecasting. The experimental evaluation uses one year of operational data from a 7.7 MW solar plant. A 6-fold rolling-origin cross-validation with 10 repetitions yields 60 runs per model. The DE-TCN-LSTM model achieves RMSE = 59.2 ± 3.0 kWh, MAE = 31.7 ± 1.8 kWh, and R² = 0.981 ± 0.003, with a normalized nMAE of 0.41%. Statistical analyses, encompassing ANOVA, Wilcoxon tests with Holm correction, Diebold-Mariano, and bootstrap confidence intervals, validate substantial enhancements compared to all baseline methodologies. RMSE is reduced by 32.1% compared to TCN-LSTM, 13.8% compared to PSO-TCN-LSTM, and 21.4% compared to GA-TCN-LSTM. Cross-dataset validation using SURFRAD and BSRN datasets shows robust generalization, with RMSE degradation limited to 5.4% or less across different climatic conditions. These findings highlight the importance of optimizer selection in improving deep hybrid models for large-scale PV forecasting.