MCV-PatchTST: multi-scale cross-variate patch transformers for household short-term load forecasting

Abstract

Abstract Household short-term load forecasting is essential for demand response and real-time energy management, but it remains difficult due to volatile consumption, concurrent periodicities, and correlations among electrical variables. PatchTST is a strong patch-based Transformer baseline, yet its fixed patch size, strict channel-independent processing, and patch-level positional encoding limit its suitability for residential forecasting. This paper proposes MCV-PatchTST, a targeted extension of PatchTST that (i) performs multi-scale patch decomposition to capture temporal dynamics at multiple granularities, (ii) introduces a lightweight cross-variate attention module at the patch level to selectively exploit inter-variable dependencies, and (iii) restores absolute temporal awareness through learnable global positional encodings aligned to the patch sequence. Experiments on the UCI Individual Household Electric Power Consumption dataset across hourly, daily, and weekly resolutions show that MCV-PatchTST consistently improves accuracy over PatchTST, achieving 10–17% lower MAE and 8–15% lower RMSE while adding only modest computational overhead. Ablation studies confirm that each component contributes to the final performance and that their combination yields complementary gains, establishing MCV-PatchTST as an effective and efficient model for household load forecasting.