Frequency-enhanced and fully activated expert-gated xLSTM for cooling-load prediction in industrial air-conditioning systems under composite operating conditions

In industrial processes such as tobacco conditioning and food manufacturing, accurate cooling-load forecasting is critical for maintaining ambient temperature and enabling efficient operation of central air-conditioning systems. However, cooling-load time series are highly nonlinear and volatile, which makes demand-side prediction challenging for conventional models. This article proposes an extended long short-term memory (xLSTM) model that combines an improved frequency-enhanced channel attention mechanism (iFECAM) with a fully activated mixture-of-experts (FMOE) gating structure. The iFECAM module extracts multiscale features from multisource inputs and performs frequency-domain reweighting to highlight informative components, while the xLSTM architecture employs augmented memory cells to better capture long-range temporal dependencies. The FMOE head further generates sample-adaptive mixing weights to aggregate specialist experts corresponding to different operating regimes, thereby explicitly representing composite operating conditions and improving robustness. Experiments on a real-world industrial tobacco plant dataset show that the proposed model consistently outperforms a range of deep learning baselines in terms of prediction accuracy and stability under varying operating conditions.