A benchmark study of optimizers for short-term solar PV power forecasting using neural networks under real-world constraints

Accurate short-term photovoltaic (PV) power forecasts are critical for efficient grid balancing, yet training optimizers, often overlooked compared to neural network architectures, significantly influence prediction accuracy and convergence speed. Prior research primarily focuses on adjusting network architectures, typically employing a single optimizer (commonly Adam), thus leaving optimizer selection underexplored, especially under noisy and incomplete real-world PV data. This study systematically benchmarks four optimizers—Adam, Adaptive Gradient (Adagrad), Rectified Adam (RAdam), and Lookahead—across three deep-learning architectures (Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN)-LSTM, and LSTM-Autoencoder) using data from two distinct PV sites. Unlike prior works, we assess optimizer effectiveness across a wide range of conditions, including varying training data lengths, sampling intervals, and missing data patterns (both random and block-wise). Using two real-world PV datasets representing semi-arid and desert climates, we analyze forecasting accuracy, convergence time, and robustness. Our empirical results demonstrate that RAdam consistently outperforms Adam by achieving up to 36% lower forecasting error under noisy and incomplete data conditions, while Lookahead offers up to 40% faster convergence in deep hybrid models. These gains translate into tighter reserve-margin planning and smoother inverter set-points, advancing state-of-the-art PV forecast pipelines. The paper concludes with optimizer-architecture recommendations for practitioners facing latency or compute constraints.