LASH: A Leakage-Aware Sequential Hybrid Framework with Validation-Gated Residual Routing for Hourly Rolling Direct 24-Step Load Forecasting in Educational Facilities

Energy management in educational facilities requires reliable day-ahead electricity-demand trajectories that can be regenerated after each new hourly measurement without using information unavailable at the forecast origin. This study proposes LASH, a Leakage-Aware Sequential Hybrid framework with validation-gated residual routing for hourly rolling direct 1–24 h forecasting. LASH combines a no-year-over-year anchor, a multi-output Ridge residual expert, a causal temporal-convolutional expert, and a bounded horizon router calibrated on purged validation data. All benchmarks use historical-only weather proxies and exclude future demand and observed target-period weather. Four datasets and sixteen direct forecasting methods were evaluated under common chronological partitions, with stochastic models assessed across ten fixed-setting refits. LASH achieved the best numerical overall performance in both university-cluster reanalyses, while the router reverted to Ridge in the two BDG external-source checks. Dependence-aware inference supported improvements over Ridge in both university clusters, with the larger relative gain in Cluster 1. Reduced-architecture analyses showed that the preferred degree of nonlinear complexity was dataset-dependent. Under a common constrained storage-dispatch model, LASH also produced competitive scheduling performance. Overall, LASH combines leakage-aware rolling multihorizon forecasting with validation-controlled nonlinear capacity and an explicit fallback mechanism.