Global-local attention-enabled multiple decoder Transformer for multi-energy load forecasting in user-level integrated energy system

Accurate multi-energy load forecasting (MELF) is crucial for reliable and cost-effective operation of integrated energy system (IES). Despite advancements, existing MELF models naively aggregate all input features without explicit global modeling mechanisms, leading to suboptimal MELF results. Additionally, they are primarily designed for regional-level IES, limiting their ability to capture local contexts and subtle fluctuations in complex user-level IES (UIES). To overcome these limitations, this study proposes a novel global-local attention-enabled multiple decoder Transformer for MELF in UIES. First, a new feature-temporal attention mechanism is proposed to jointly capture coupling relationships and temporal dependencies in a global context. Second, as a local modeling method, a residual bi-directional temporal convolution-based attention mechanism is incorporated into each decoder to efficiently extract subtle variations. Furthermore, a day-type tendency network is integrated into each decoder to address critical external factors. Case studies demonstrate that the proposed model outperforms state-of-the-art MELF models, achieving the highest average R2 of 0.9676 across multiple loads while requiring only 5.39 % more training time than the shortest training model. These results highlight the model's practical engineering value for robust UIES operation and optimal dispatch, with future extensions into anomaly detection for enhanced fault diagnosis, system resilience, and proactive energy management.