Deterministic Multi-Energy Load Forecasting for Integrated Energy Management System Using a Random Search Transformer Network

Multi-energy load forecasting techniques form the foundation for the operation and scheduling of integrated energy management systems (EMS). This paper introduces a novel approach that integrates multi-task learning (MTL) with a multiple-decoder transformer (MDT) to forecast electricity and steam loads within the EMS framework. The MTL model is designed to capture the intricate correlated information among various loads, while the MDT model effectively identifies the correlation-based patterns between these loads. To enhance the model's performance, a grid search multi-head attention mechanism is proposed. This mechanism is optimized to capture the complex coupling relationships and uncertainties among loads by refining the distribution of network parameters. Furthermore, a multi-task balancing strategy based on random search is developed to support one-day-ahead load forecasting for scheduling purposes and intra-day rolling forecasting for real-time updates. The proposed methodology was validated through real-world simulations involving an industrial company located in Taoyuan, Taiwan. Experimental results demonstrate that the proposed approach significantly improves the prediction accuracy for both electricity and steam loads.