XLTFormer: Regime-Adaptive Time–Frequency Expert Routing for Multi-Step Load and Energy Forecasting

Abstract

Multi-step load forecasting remains challenging because real-world demand series exhibit non-stationarity, multi-scale periodicity, and regime-dependent temporal structure. Performance degradation often arises not from insufficient model size alone, but from mismatch between the active operating regime and the inductive bias of a fixed forecasting backbone. This paper proposes XLTFormer, a regime-adaptive framework that treats multi-step prediction as descriptor-augmented routing over heterogeneous structural priors. Memory, attention, decomposition, cross-variable interaction, patch-based representation, and frequency-domain modeling are organized as route-level experts, while the methodological contribution centers on four mechanisms: regime-descriptor-guided sample-level routing, post-mixture peak-context correction, lightweight expert-preserving training, and validation-gated deployment selection. Across three real-load benchmarks, XLTFormer attains the lowest mean error on all five reported metrics under the ten-seed protocol and remains top-ranked overall under the controlled-capacity comparison. In the retained rerun, the final deployment repeatedly selects the cross-variable route on all three datasets, with WAPE/sMAPE values of 3.7274/3.7320 on Electricity, 3.9576/3.8643 on PJM, and 7.2619/7.5058 on GEFCom2014 Load. Relative to the strongest competing baseline, the WAPE improvement reaches 2.11% on Electricity, 1.11% on PJM, and 1.02% on GEFCom2014 Load, while the corresponding sMAPE reduction reaches 2.40%, 1.29%, and 1.08%, respectively. Seed-level paired tests, dependence-aware block bootstrap, and Friedman ranking analysis further show that this retained deployment advantage is stable rather than incidental. The temporal-routing audit nevertheless indicates that the sample-level gate still redistributes route responsibility across local regimes, especially when peak, ramp, or weather-linked stress increases. These results show that the practical forecasting gain comes from the full reliability-oriented framework, in which adaptive routing shapes route interaction and validation-gated selection determines the final retained output, rather than from attributing the deployed accuracy to raw adaptive fusion alone. Overall, the framework provides an effective direction for accurate, stable, and practically relevant multi-step load forecasting.