An Improved Informer-Based Load Forecasting Method Incorporating Electricity Consumption Profiling

Accurate demand-side load forecasting is essential for reliable power system operation, energy management, and demand-side decision-making. However, heterogeneous electricity-use behavior among residential users and multiscale temporal variations remain major challenges for conventional forecasting models. To address these issues, this study develops an electricity-use-aware load forecasting method that integrates user behavior profiling with an improved Informer model. First, symbolic aggregate approximation (SAX) and hierarchical clustering are employed to extract representative load patterns from residential electricity consumption data. LightGBM is then used to identify informative questionnaire attributes associated with different load patterns, and the clustering results and selected user attributes are jointly used to construct multidimensional electricity-use behavior profiles. These profiles provide additional behavioral information for subsequent load forecasting. On this basis, the Informer architecture is improved through three strategies: a temporal convolutional network (TCN)-based embedding for multiscale periodic feature extraction, probabilistic block-sampling attention for local temporal dependency modeling, and a multilayer perceptron (MLP)-based cross-self-attention decoder for nonlinear feature interaction and fusion. Experiments on data from 3207 residential users show that the proposed model achieves an MAE of 0.0237, an RMSE of 0.0258, and an R2 of 0.945. Compared with the original Informer, MAE and RMSE are reduced by 17.4% and 19.8%, respectively, while R2 is increased by 6.65%. The model also outperforms Transformer and LSTM baselines, and the ablation results demonstrate the complementary contributions of the three architectural improvements. These results indicate that combining electricity-use behavior profiling with multiscale temporal modeling can effectively improve residential demand-side load forecasting.