Spatiotemporal data relevance analysis in short-term wind forecasting in LSTM networks

Abstract

The expansion of global wind capacity has made accurate short-term wind forecasting essential for power grid integration. Wind dynamics exhibit complex spatiotemporal characteristics determined by non-linear atmospheric processes and regional air mass interactions. This paper presents a comparative study of single-location wind forecasting models based on LSTM neural networks, using a three-hour prediction horizon. The influence of different spatiotemporal input selection strategies was investigated: a full-dataset model, a representative strategy using clustering, and a model relying strictly on local information. The results demonstrate that LSTM models consistently outperform a persistence baseline. The full-dataset model achieved the best overall performance, while the clustering strategy showed competitive results, validating dimensionality reduction through intelligent input selection. Furthermore, spatiotemporal-aware LSTM models exhibited greater robustness during critical time segments marked by abrupt changes, and their residual analysis revealed superior capture of temporal dependencies. These findings are crucial for optimizing training configurations in wind forecasting networks.