A graph-based Seq2Seq framework for causal discovery and out-of-distribution forecasting in building energy systems

In building energy systems, data-driven models have been widely applied to energy consumption forecasting and operational optimization. However, most existing approaches rely primarily on statistical correlations, making them difficult to maintain stable performance under out-of-distribution (OOD) scenarios, when data distribution shifts significantly. Moreover, these models generally fail to support counterfactual reasoning and other hypothetical analysis tasks that are critical for decision- making in real-world operations. To address these challenges, this paper proposes a causal graph-driven time series generative framework for building energy systems. The framework first applies the PCMCI causal discovery algorithm on high-dimensional multivariate time series data to automatically extract time-lagged causal graph structures that characterize the underlying system dynamics. Besides, a Graph Attention Network (GAT) is employed to encode the causal graph into continuous low-dimensional vector representations, which are subsequently processed by an optimized Sequence-to-Sequence (Seq2Seq) model to generate the final results. Experimental results on standard building energy forecasting scenarios show that, although causal learning incurs additional computational overhead, the proposed method achieves strong and stable predictive performance, with the RMSE of 1.21 kW and R2 of 0.98, outperforming an optimal ablation baseline from 36.9% to 72.3%. These findings highlight that combining causal discovery with a generative model provides an effective solution for counterfactual analysis and OOD forecasting in building energy systems.