A systematic comparison of machine learning models for missing value imputation in household electricity consumption data

Missing values in electricity consumption datasets compromise the reliability of smart grid analytics, demand forecasting, and energy management systems. Although numerous imputation methods have been proposed, prior studies typically evaluate a limited subset of models using a narrow range of metrics, making it difficult to draw generalizable conclusions about relative performance across algorithmic paradigms. This paper addresses this gap through a systematic benchmarking study comparing 30 machine learning models spanning linear methods, tree-based ensembles, support vector machines, and deep neural networks for imputing missing values in household electricity consumption time series. We conducted a three-phase evaluation. In the first phase, individual model assessment revealed that tree-based ensemble methods achieved the strongest performance, with Extra Trees Regressor and Random Forest attaining Mean Absolute Error (MAE) values of 0.0474 kW and 0.0480 kW, respectively. In the second phase, ensemble combinations using Bagging, Boosting, and Stacking techniques demonstrated improved accuracy over single-model implementations. In the third phase, a three-stage cascaded hybrid framework employing error curve learning achieved a Mean Absolute Error of 0.0468 kW, Root Mean Square Error (RMSE) of 0.1477 kW, and Mean Absolute Percentage Error of 4.38%. A main finding is that tree-based ensemble methods performed comparably to or better than deep learning architectures, including Long Short-Term Memory (LSTM) networks for this imputation task, suggesting that model complexity does not necessarily translate to improved performance in this domain. We also provide a comprehensive cross-model ranking that offers practitioners evidence-based guidance for selecting imputation methods. The multi-metric evaluation framework established here can serve as a reusable benchmark for future research in energy data quality improvement.