Peak-aware short-term household electricity consumption forecasting using deep learning and transformer-based architectures

Abstract

Accurate short-term prediction of domestic electricity consumption is a prerequisite for smart grid management, demand response, and peak-load reduction. This work presents a detailed deep learning analysis of the UCI household power consumption dataset, which includes minute-level measurements from French households over four years. The data is resampled to hourly and undergoes an intensive feature engineering pipeline and then used with different neural architectures: a standard long short-term memory (LSTM), bidirectional attention LSTM, hybrid CNN-LSTM, temporal fusion transformer (TFT-Lite) lightweight, patch time series transformer (PatchTST), and hybrid CNN-transformer. To address the asymmetric operational cost of under-predicting electricity demand, the transformer-based models employ a custom asymmetric MSE loss that assigns twice the penalty to under-predictions compared with over-predictions. Model performance is evaluated in the original kW scale after inverse transformation using twelve evaluation metrics including RMSE, MAE, and sMAPE, together with a peak MAE metric for observations exceeding 2.0 kW to specifically assess peak-load forecasting capability. The comparative results demonstrate that the hybrid CNN-LSTM provide improved overall household load forecasting on RMSE (0.4766 ± 0.01 kW), MAE (0.3214 ± 0.01 kW), MASE (0.8511 ± 0.02), WAPE (32.99 ± 0.85%), sMAPE (35.72 ± 2.71%), peak precision (64.31 ± 6.11%), false alarm rate (3.54 ± 0.62%), forfcast bias (− 0.006 ± 0.02 kW), and peak timing error (4.25 ± 0.08 h). In addition, TFT-Lite is advantageous for peak-aware forecasting peak MAE (0.5882 ± 0.01 kW) with conventional recurrent baselines. In addition, a paired t-test for statistical significance and comprehensive ablation studies were conducted, and the results further strengthen the effectiveness and validity of the proposed approach.