W-LSTMix: A Hybrid Modular Forecasting Framework for Trend and Pattern Learning in Short-Term Load Forecasting

Motivated by the need for a lightweight, gener-alizable forecasting model in intelligent build-ing energy management and reliable microgridcontrol. We propose W-LSTMix—a modularand lightweight hybrid architecture designed togeneralize across diverse building types, withonly ∼ 0.13 M parameters integrates wavelet-based signal decomposition, neural basis expan-sion, and patch-based temporal mixing for effi-cient building-level load forecasting. The modelseparately forecasts decomposed time series com-ponents: long-term trends are modeled using anLSTM-enhanced N-BEATS stack, while resid-ual and seasonal patterns are captured throughan MLP-Mixer-enhanced N-BEATS structure.We compare our model against state-of-the-artTSFMs such as Lag-Llama, Moirai, Chronos, andTiny Time Mixers in zero-shot and fine-tunedsettings, under domain-specific training and test-ing. In both comparisons, our model consistentlyoutperforms the baselines, demonstrating robustgeneralization capabilities and suitability for real-world intelligent energy management.