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

Area of Science:

• Data Science
• Machine Learning
• Energy Systems

Background:

• Missing values in electricity consumption data hinder smart grid analytics, demand forecasting, and energy management.
• Existing studies lack comprehensive comparisons of imputation methods across diverse algorithmic paradigms.

Purpose of the Study:

• To systematically benchmark 30 machine learning models for imputing missing household electricity consumption time series data.
• To provide evidence-based guidance for selecting optimal imputation methods in energy data quality improvement.

Main Methods:

• A three-phase evaluation framework was employed, assessing individual models, ensemble combinations, and a hybrid cascaded framework.
• 30 machine learning models, including linear methods, tree-based ensembles, support vector machines, and deep neural networks, were compared.
• Performance was evaluated using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).

Main Results:

• Tree-based ensemble methods, particularly Extra Trees Regressor and Random Forest, demonstrated superior performance among individual models.
• Ensemble techniques (Bagging, Boosting, Stacking) further improved imputation accuracy.
• A cascaded hybrid framework achieved the best results with an MAE of 0.0468 kW.

Conclusions:

• Tree-based ensemble methods offer comparable or superior performance to deep learning models like LSTM for electricity data imputation.
• Model complexity does not inherently guarantee better performance for this specific task.
• The study provides a robust benchmark for future research in energy data imputation and quality enhancement.