Comparative analysis of optimized LSTM architectures for long-term solar power forecasting

Accurate long-term forecasting of photovoltaic (PV) power output is vital for ensuring the reliability and efficiency of renewable energy systems integrated into smart grids. This study proposes a systematic deep learning framework using Long Short-Term Memory (LSTM) networks, exploring the combined effects of network depth, activation functions (ReLU vs. Leaky ReLU), and optimization algorithms (Adam, Nadam, RMSprop, SGD) on predictive accuracy. A real-world dataset, recorded over one year in Van, Turkey, was used to evaluate the performance of 24 distinct model configurations. The results reveal that deeper LSTM architectures significantly improve generalization, particularly when paired with Leaky ReLU activation. Among all configurations, the four-layer LSTM model optimized with Nadam and activated with Leaky ReLU achieved the best forecasting performance, reaching a minimum SMAPE of 15.4%. This result represents a substantial improvement over conventional setups and demonstrates the effectiveness of adaptive optimization and nonlinear activation in capturing complex temporal patterns. The findings offer practical implications for enhancing grid stability, optimizing solar energy dispatch, and informing energy policy planning. The proposed model serves as a robust and scalable tool for energy stakeholders aiming to improve long-term solar forecasting in dynamically changing environments.