Mix-BEATS: Mixer-enhanced Basis Expansion Analysis for Load Forecasting

Short-term load forecasting (STLF) is crucial for energy management and integrating renewables, but traditional models often fail to generalize across diverse buildings.While Time Series Foundation Models (TSFM) offer potential, most are transformer-based, making them computationally heavy and unsuitable for edge deployment.We introduce Mix-BEATS, a lightweight and accurate model combining N-BEATS and TSMixer architectures for buildinglevel energy prediction.Adapting TSMixer originated from MLP-Mixer, originally designed for vision tasks using pure MLP blocks for patchwise and spatial mixing, we leverage its principles for temporal forecasting.Mix-BEATS is trained on hourly smart-meter data from 38,956 buildings and evaluated on a held-out set of 1,000 buildings using the NRMSE metric.We compare Mix-BEATS against state-of-the-art TSFMs such as Lag-Llama, Moirai, Chronos, and Tiny Time Mixers in zero-shot and fine-tuned settings, as well as generic models including Autoformer, N-BEATS, TS-Mixer, etc., under domain-specific training and testing.In both comparisons, Mix-BEATS consistently outperforms the baselines, demonstrating robust generalization and efficiency for real-world STLF tasks.