A cross-dataset benchmark for neural network-based wind power forecasting

Wind power generation is a critical and promising renewable energy source, and accurately forecasting its output can optimize energy management and yield substantial economic benefits. Despite the fact that a considerable number of the recent Wind Power Forecasting (WPF) studies employs neural network-based deep learning techniques, these studies are often conducted independently. There is an urgent need for a comprehensive benchmark to validate the effectiveness and robustness of neural networks in this domain and provide more valuable guidance for engineering practice. In this study, we first methodically delineate the task objectives of neural networks in WPF. Subsequently, we categorize neural network structures and task paradigms into autoregressive/non-autoregressive networks and deterministic/probabilistic predictions. Building on this, we establish a unified cross-dataset benchmark for neural networks in the WPF domain, which incorporates eight global wind power operation datasets at both turbine and farm scales. Finally, we conduct a series of neural network evaluation experiments based on this benchmark. The results indicate that neural networks excel in longer forecasting horizons, while autoregressive models show greater robustness in short-term forecasting, with non-autoregressive models progressively mitigate disparities in long-term forecasting.