Does Causal Decomposition Improve Residential Load Forecasting? A Leakage-Aware, Multi-Dataset Reproducibility Study of a Decomposition-Augmented Transformer

Title: Decomposition-Augmented Transformer for Direct Multi-Horizon Residential Load Forecasting Without Climate Features: A Cross-Dataset Benchmark Study Description / Summary: This paper introduces the Decomposition-Augmented Transformer (DAT), a framework designed for direct multi-horizon short-term residential load forecasting using historical consumption data alone. Residential electricity demand is non-stationary, intermittent, and characterized by sharp appliance peaks, making it difficult to predict without accumulating error or relying on external weather feeds. DAT addresses these challenges by applying Variational Mode Decomposition (VMD) or CEEMDAN in a strictly causal, sample-wise manner within rolling lookback windows, effectively eliminating lookahead bias and temporal information leakage. Boundary end-effect mitigation (BES) techniques—including mirror padding and historical similarity matching—are integrated to suppress spectral drift at window edges. High-frequency stochastic noise is adaptively filtered using fuzzy entropy, and the resulting mode components and dynamic center frequencies are embedded into a multi-head self-attention Transformer encoder. The architecture produces forecasts for all horizons up to 24 hours in a single forward pass, completely bypassing recursive error propagation. Evaluated across three benchmark datasets (UCI Household Power Consumption, London Smart Meter, and MORED), the proposed VMD-Transformer achieves superior accuracy over recurrent networks (LSTM), linear baselines (DLinear), modern attention architectures (PatchTST, iTransformer), and a zero-shot foundation model (TimesFM). It attains the lowest 24-hour RMSE, MAE, and MAPE and the highest R² across all datasets while preserving peak volatility, recording standard-deviation ratios (StdR) between 0.964 and 0.981. Pairwise Diebold–Mariano tests, Wilcoxon signed-rank tests, and Model Confidence Set analysis confirm the statistical significance of the performance gains. The study demonstrates that robust residential load forecasting is achievable from smart-meter consumption histories alone, offering a simplified and practical solution for distribution grids and demand-side management where meteorological data are unavailable or costly.