Abstract
As the energy industry faces increasingly complex changes in demand and external environmental impacts, traditional forecasting models struggle to effectively integrate multi-source and multimodal data while maintaining interpretability. Therefore, a forecasting model capable of processing multimodal data simultaneously and capturing long-term dependencies is needed. Mamba-ECIS enhances the ability to capture both short-term fluctuations and long-term trends by incorporating multimodal data. It combines a dual architecture with a causal-consequence attention module, thereby improving model interpretability. Experimental results show that Mamba-ECIS outperforms existing benchmark models on both the UCI energy consumption dataset and the Pecan Street Dataport dataset, achieving improvements of approximately 3-12% across various evaluation metrics. Ablation studies further confirm the synergistic effects among the modules and demonstrate the unique contribution of each module to overall model performance. These findings indicate that Mamba-ECIS provides a more accurate, stable, and interpretable solution for power demand forecasting, demonstrating high application potential.