Enhancement of Stochastic Gradient Descent Generalization for PV Power Prediction

The goal of this study is to increase the precision and effectiveness of photovoltaic (PV) power forecasting, which is essential for grid stability and the cost-effective integration of renewable energy. Dual-Enhanced Stochastic Gradient Descent (DESGD), a novel optimization technique, is presented. It improves the generalization capabilities of the conventional Stochastic Gradient Descent with Momentum (SGDM). The study contrasts DESGD with other optimizers such as Adam and Levenberg-Marquardt (LM) using the Photovoltaic Power Output Dataset (PVOD) and a Multilayer Perceptron (MLP) architecture. The results demonstrate that DESGD provides a compromise between accuracy and computational economy, achieving superior generalization over a range of climatic variables and PV stations. DESGD is a possible substitute for PV power forecast because it is substantially faster than LM, even though LM often offers higher accuracy. The study also investigates the effect of MLP depth and concludes that DESGD remains robust as model complexity rises. In order to solve other issues in PV forecasting, future study will assess DESGD on various deep learning architectures and datasets.