A location-centric transformer framework for multi-location short-term wind speed forecasting

Abstract

Accurate spatiotemporal wind speed forecasting plays a vital role in power system optimization and renewable energy efficiency. However, conventional models often combine historical wind speed data from multiple locations into uniform feature channels, which compromises their ability to capture spatial correlations and impairs forecasting accuracy. This study proposes that maintaining location-specific distinctions while modeling spatiotemporal relationships can enhance forecasting performance. Based on this premise, this study develops a novel Transformer-based framework with a location-centric architecture that introduces several key innovations: (1) a spatiotemporal gated fusion unit that dynamically integrates geographical coordinates with temporal wind speed data while preserving location-specific information, (2) a restructured Transformer that employs self-attention for modeling spatial correlations across locations while using feedforward networks to capture temporal dependencies, and (3) a dual-enhancement mechanism combining reversible instance normalization to address concept drift and a frequency channel attention mechanism to leverage frequency-domain characteristics. Experimental results show significant improvements in multi-location, multi-step wind speed forecasting accuracy. This enhanced precision directly supports more efficient renewable energy management and grid integration.