Dual-axis attention network for granular load forecasting in industrial and residential demand response

Smart grids require precise demand response management through accurate granular-level load forecasting capabilities. This paper introduces a novel Dual-Axis Attention Network (DAAN) that addresses the critical challenge of short-term load forecasting at both user and appliance levels in industrial and residential environments. The proliferation of smart meter infrastructure has enabled fine-grained data collection. However, existing forecasting methods struggle to capture the complex temporal and feature dependencies inherent in granular consumption patterns. DAAN leverages a sophisticated dual-axis attention mechanism that simultaneously processes temporal correlations and inter-feature relationships through parallel horizontal and vertical attention blocks. This architecture enables enhanced learning of complex consumption behaviors while maintaining computational efficiency. Comprehensive evaluation across three datasets—including a newly released industrial appliance power consumption dataset containing minute-level data from five industrial sectors—demonstrates DAAN's superior performance. The proposed method achieves significant improvements over nine benchmark approaches, with mean arctangent absolute percentage errors not exceeding 0.25 across all test scenarios. These results validate DAAN's capability to provide accurate minute-level forecasting essential for effective demand response programs, load aggregation strategies, and distribution system optimization in smart grids.