Optimization of Photovoltaic Power Forecasting: A Comparative Study of Deep Learning Architectures, Optimization Techniques, and Evaluation Metrics

This study presents a systematic evaluation of deep learning architectures for photovoltaic (PV) power forecasting, comparing nine model configurations across three architectures (MLP, LSTM, CNN) and three optimizers (Adam, RMSprop, Adagrad). Using six years of hourly data from a 350kWp grid-connected PV system in Scotland, we demonstrate that architectural choice and optimizer selection significantly impact forecasting accuracy. The LSTM-RMSprop configuration achieved superior performance with RMSE of 2.651 kWh and MAE of 1.197 kWh, showing a 90.29% coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>). This outperforms both CNN (RMSE: 2.767-2.902 kWh) and MLP architectures (RMSE: 3.104-3.115 kWh) across all optimizers. Our main contributions include: (1) comprehensive optimization of model architectures through hyperparameter evaluation, revealing optimal configurations for each model type; (2) systematic evaluation of optimizer impact, demonstrating RMSprop's superiority for LSTM with improved accuracy across architectures; (3) detailed error analysis showing model stability across different conditions, with NRMSE ranging from 31.15% to 39.07%; and (4) practical insights into computational requirements, where CNN architectures achieve fastest training times (2.3-3.8 minutes/epoch) compared to LSTM (3.9-5.9 minutes/epoch) and MLP (36.54-46.72 minutes/epoch). Results demonstrate that LSTM architectures with appropriate optimization can outperform simpler models in PV power forecasting, providing valuable guidance for practical implementations.