Interpretable Deep Learning for Short‐Term Appliance Load Forecasting: A Comparative Study of LSTM–SHAP and TFT Attention Models
IET Smart Grid · 2026-01 · 9 (1)
DOI: 10.1049/stg2.70110 ↗Kurz gefasst
Fachartikel in der Zeitschrift IET Smart Grid, online erschienen am 28.07.2026. Verfasst von Yankai Yang und 4 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Accurate short‐term residential appliance load forecasting is critical for modern power system operation. Household electricity demand shows strong variability, nonstationarity and sudden fluctuations, which make accurate and interpretable forecasting challenging. This paper investigates interpretable deep learning approaches for short‐term residential appliance load forecasting using a long short‐term memory (LSTM) network explained by SHapley Additive exPlanations (SHAP) and a Temporal Fusion Transformer (TFT) explained by built‐in variable‐selection and attention mechanisms. To ensure an objective‐consistent point‐forecasting comparison, an MSE‐trained LSTM and an MSE‐trained TFT are evaluated using the same data partition, feature‐engineering procedure, feature‐selection setting and evaluation protocol. Their point‐forecasting performance is assessed using RMSE, MAPE and R 2 . In parallel, a quantile‐trained TFT is evaluated separately for probabilistic forecasting using the mean pinball loss, approximate CRPS, prediction interval coverage probability and mean prediction interval width. Results show that the MSE‐trained TFT achieves lower point‐forecasting errors than the LSTM under the loss‐aligned setting. The quantile‐trained TFT further provides reasonably calibrated uncertainty estimates, with an 80% prediction interval coverage probability of 0.815 and an absolute coverage error of 0.015. In addition, the TFT provides feature‐level and temporal explanations through its variable‐selection and attention mechanisms, whereas SHAP offers complementary post‐hoc explanations for the LSTM. The study provides a task‐consistent comparison of recurrent and Transformer‐based forecasting models and demonstrates how point accuracy, uncertainty quantification and interpretability can be jointly considered in residential load forecasting.
Zitieren
APA
Yang, Y., Musleh, A. S., Liu, B., Wang, G., & Ren, J. (2026). Interpretable Deep Learning for Short‐Term Appliance Load Forecasting: A Comparative Study of LSTM–SHAP and TFT Attention Models. IET Smart Grid, 9(1). https://doi.org/10.1049/stg2.70110
Vancouver
Yang Y, Musleh AS, Liu B, Wang G, Ren J. Interpretable Deep Learning for Short‐Term Appliance Load Forecasting: A Comparative Study of LSTM–SHAP and TFT Attention Models. IET Smart Grid. 2026;9(1). doi:10.1049/stg2.70110
BibTeX
@article{Yang2026,
author = {Yang, Yankai and Musleh, Ahmed S. and Liu, Boyu and Wang, Guanyu and Ren, Junhao},
title = {Interpretable Deep Learning for Short‐Term Appliance Load Forecasting: A Comparative Study of LSTM–SHAP and TFT Attention Models},
journal = {IET Smart Grid},
year = {2026},
volume = {9},
number = {1},
doi = {10.1049/stg2.70110}
}
EUVIMED