Improving Residential Load Forecasting Accuracy Through Semi-Supervised Multi-Binary Classifier Framework

The increasing integration of distributed energy resources into smart grids highlights the need for accurate household-level load forecasting. However, the high variability of single-household consumption poses challenges for conventional models. This study proposes a semi-supervised multi-binary classifier framework for improving forecasting accuracy by extracting structural features from daily load curves. Specifically, load shapes are categorized into five interpretable types (one-peak, two-peak, three-peak, four-peak, and high variation) and transformed into binary feature arrays that serve as supplementary inputs for forecasting models. With these features, the CNN+BiLSTM model on the CER dataset reduced RMSE from 0.2811 to 0.2209, while on the LCL dataset, a Transformer+MLP model achieved an SMAPE reduction from 40.17% to 37.66% together with consistent MAE improvements. Moreover, compared with clustering-based methods such as K-means, the proposed framework achieved understandable predictive results with substantially lower computational cost. These results demonstrate that the framework provides interpretable, scalable, and practical enhancements for residential load forecasting.