A unified benchmark of deep learning and classical machine learning models for weather-based solar power forecasting: Balancing complexity and skill
Year: 2026
Type: article
Abstract: Numerical weather prediction (NWP) is the primary input for forecasting photovoltaic (PV) power generation for lead times ranging from a few hours to several days. Data-driven and hybrid irradiance-to-power methods are considered state-of-the-art; however, it remains unclear whether the latest improvements in machine learning (ML) hold significant potential for this domain. A comparison of 12 classical ML and deep learning (DL) models is presented to compare sequence-to-sequence architectures with feature-engineered tabular regression models from various aspects. All models are implemented and optimized for seven locations, using four years of measured PV power and ECMWF weather forecasts. Results reveal that after careful implementation and proper optimization, each considered ML model is capable of achieving a skill score of 46.7-47.7% over persistence, outperforming the 40.3% skill of reference physical model chains. The most accurate forecasts are created by recurrent neural networks (RNN), including long short-term memory and gated recurrent unit setups, closely followed by Pyraformer, Informer, and Autoformer transformer variants. However, classical multilayer perceptrons (MLP) and kernel ridge regression are also great alternatives, and even easy-to-use ensemble learning algorithms, such as ExtraTrees (ET), can catch up if expert inputs with physical hybridization and temporal feature engineering are provided. The similar forecast quality suggests that model selection should rely on the joint consideration of multiple factors, including the significant computational complexity of DL models and the large sensitivity of ensemble and transformer models to their hyperparameters. Overall, physical modeling, ET, MLP, and RNN are emerging as Pareto-optimal trade-offs of accuracy and runtime.
Source: Energy and AI
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,630
OpenAlex ID: W7166808708