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

SHIVAM DWIVEDI, Anuj Kumar, Harish Kumar Saravanan, Pandarasamy Arjunan

Keywords: Energy Informatics, Smart Grid, Energy Forecasting, Short-term Load Forecasting (STLF), Demand-side Load Management, Time Series Foundation Models (TSFM), Machine Learning

TL;DR: W-LSTMix uses wavelets with enhanced N-BEATS to learn trends and patterns, boosting time series forecasting accuracy.

Abstract: Recently, there has been growing research interest in developing domain-specific Time Series Foundation Models (TSFMs), particularly in areas such as energy. We present W-LSTMix - a modular, lightweight hybrid architecture with only 0.13M parameters - designed to generalize across diverse building types for efficient building-level load forecasting. W-LSTMix integrates wavelet-based signal decomposition, neural basis expansion, and patch-based temporal mixing. It forecasts decomposed time series components separately: long-term trends via an LSTM-enhanced N-BEATS stack, and residual plus seasonal patterns through an MLP-Mixer-augmented N-BEATS structure. We benchmark W-LSTMix against state-of-the-art TSFMs such as Lag-Llama, Moirai, Chronos, and Tiny Time Mixers, as well as N-BEATS, under both zero-shot and fine-tuned settings. Our model consistently outperforms these baselines, demonstrating robust generalization and suitability for real-world intelligent energy management.

Submission Number: 45