MIRA-Mamba: Multi-Scale Interactive Recalibrated Mamba for Multi-Step Indoor Temperature Forecasting in HVAC Systems

For multivariate-input-driven multi-step temperature forecasting in air-conditioning systems, existing models often rely on a single temporal receptive field, limiting the amount of temporal information captured and hindering the effective characterization of multi-scale dynamic features. Moreover, accurately capturing condition-dependent variations in the contributions of different variables to the prediction target remains challenging, resulting in persistent error accumulation over long forecasting horizons. To address these issues, this paper proposes a Multi-scale Interactive Recalibrated Mamba (MIRA-Mamba) forecasting framework. The multi-scale temporal interaction embedding module extracts and fuses dynamic features from different temporal receptive fields, enhancing the representation of multivariate coupling relationships and alleviating information loss caused by single-scale representations. The bidirectional state–space recalibration encoder employs bidirectional Mamba to capture long-range temporal dependencies and incorporates a feature contribution recalibration mechanism to adaptively adjust variable weights, thereby strengthening high-contribution variables and suppressing interference from low-contribution variables. In addition, a private dataset, Hisense-AC, is constructed from real air-conditioning operation logs collected from 15 cities over half a year. Experimental results show that MIRA-Mamba achieves the lowest forecasting errors across all four operating modes on the Hisense-AC dataset, with a maximum MSE reduction of 5.4% compared with the best-performing baseline. On six public time-series datasets, MIRA-Mamba obtains the best MSE in 17 out of 24 forecasting settings, demonstrating its effectiveness in real HVAC scenarios and its generalization capability on public benchmarks.