Multi-Step Load Forecasting with State-Guided Expert Routing for Energy Management Systems

Accurate load forecasting is crucial for operational planning and rolling decision-making in energy management systems (EMSs). However, the dynamic nature of temporal patterns and the inconsistent availability of covariates across the forecasting horizon complicate this task. While covariate-aware Transformer models distinguish between endogenous and exogenous inputs, their feed-forward modules typically employ homogeneous transformations across different operating conditions, and traditional Mixture-of-Experts (MoE) routing mechanisms usually rely solely on implicit token representations for decision-making. To address these limitations, we propose SA-TimeXer-MoHE, a scenario-aware TimeXer model with heterogeneous experts. First, the model partitions its inputs by forecast-time availability into the load target, historical-only measurements, deterministic future calendar variables, and operating-scenario covariates. Then, it replaces the homogeneous feed-forward module with MLP, depthwise-convolution, drift, and residual experts and conditions sparse Top-K routing jointly on token representations and nine EMS scenario variables. On a real EMS dataset, the full model achieves reported mean MAE, RMSE, P95_AE, and Ramp_MAE values that are 7.4%, 5.7%, 4.0%, and 6.0% lower, respectively, than those of TimeXer across 1, 2, 4, and 8 h forecasting horizons. These results highlight the potential of SA-TimeXer-MoHE as a forecasting component for rolling scheduling, peak regulation, and storage dispatch in practical EMS applications.