Transfer learning-based multi-energy load forecasting method for integrated energy system with zero-shot

In the planning and capacity design of integrated energy system (IES), the critical reliance on multi-energy load data faces a paradoxical dilemma: a scarcity or complete absence of historical operating data. This “fundamental demand vs. data scarcity” contradiction challenges optimal design. This paper systematically proposes a transfer learning (TL)-based forecasting framework designed for zero-shot scenarios, which addresses this challenge through a three-stage innovative approach: First, a novel algorithm, Tnet, is designed based on probabilistic generalization assessment. By decomposing temporal features and incorporating weighted mutual information entropy, a source domain selection paradigm guided by probabilistic judgment is constructed. This paradigm identifies source domain groups from multiple candidates with the highest generalization value for a given target domain. Second, an improved meta-learning strategy, Metas, is developed to optimize cross-domain parameter transfer by adapting task weights dynamically, significantly enhancing the modeling accuracy of temporal features. Third, an encoder-decoder model integrated with a multi-head attention mechanism is constructed to enable the coordinated forecasting of electricity, heating, gas, and cooling loads. Experimental results show that under zero-shot conditions, the proposed method reduces mean absolute percentage error by more than 42 % compared to benchmark models while improving the coefficient of determination by over 50 %. Further validation through few-shot fine-tuning (FSFT) demonstrates that when the target domain gradually acquires a small amount of real data, the model can achieve rapid correction within a few iterations and maintain high forecasting robustness. Its performance in the “cold-start” phase, where data is scarce, far exceeds that of direct training. This highlights the core role of the FSFT strategy in bridging the performance gap during the critical transition from zero-shot scenarios to those with sufficient data. It provides a complete, feasible, and efficient forecasting paradigm for IES that have not yet been commissioned or lack comprehensive historical data. This paradigm covers the entire process from a zero-shot start-up to few-shot optimization, offering valuable insights for energy planning and operational scheduling in real-world applications.