A novel temporal domain adaptation framework for residential electricity consumption forecasting under incomplete information

With the development of smart grids and electricity markets, accurately predicting residential electricity consumption (REC) has become increasingly important. Data-driven techniques have proven to be a promising approach in electricity forecasting tasks. However, household data is often collected from intermittent household surveys rather than intelligent monitoring devices, leading to insufficient continuity, detail, and completeness, thus limiting the applicability of data-driven approaches. To address this issue, a stacked interval type-2 fuzzy perceptron is put forward to infer incomplete REC data and thereby ensure the feature space consistency between the historical and contemporary domains. Subsequently, a subspace-guided composite ensemble learning network is developed to model the correlations between household features and electricity consumption as transferable features. By integrating these developed components, a novel temporal domain adaptation framework is formed, facilitating domain adaptation between historical and contemporary domains and thereby enhancing contemporary REC forecasting performance. Comparisons with classical and state-of-the-art models indicate that the proposed framework can reduce prediction errors by 5.81% to 20.72% and showcase robustness under experimental scenarios. By combining neuro-fuzzy mechanisms, ensemble learning, and transfer learning, the proposed approach constructs a promising framework, facilitating REC forecasting under incomplete information scenarios.