Improving Short-Term Solar Power Prediction through a Hybrid Deep Learning Model

Abstract

Reliable solar energy is fundamental to sustaining clean energy systems; however, its inherent variability poses significant challenges to the stability and operational reliability of modern power grids. Accurate short-term Photovoltaic (PV) power forecasting is therefore essential for effective energy management and the smooth operation of smart grid infrastructures. In this work, we introduce an adapted hybrid Long Short-Term Memory–Temporal Convolutional Network (LSTM-TCN) architecture that represents the core methodological contribution of the study. This hybrid design leverages the ability of LSTM networks to capture long-term temporal dependencies while exploiting the expanded receptive field and efficient parallelization offered by TCNs. This architecture is applied for the first time to PV power forecasting under Moroccan climatic conditions. The proposed model is trained and validated using a decade-long real-world dataset (2013–2023) collected in Dakhla, Morocco, a region characterized by distinctive meteorological patterns that enhance the robustness and relevance of the evaluation. Comparative analyses against standalone LSTM and TCN architectures show that the hybrid model achieves the highest predictive accuracy, yielding the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE)values and a coefficient of determination (R²) of 0.9975. These results demonstrate the effectiveness of the proposed hybrid framework in delivering reliable PV power forecasts and supporting improved integration of solar energy into smart grid systems.