Technical Summary: Predictive Modelling for Individual Household Electric Power Consumption Using Machine Learning Approaches

1. Problem Statement

The study addresses the critical need to analyze and predict household electric power consumption amidst rising energy demands and costs. While existing literature often relies on synthetic data, short-duration datasets, or single-target predictions (e.g., active power only), these approaches frequently fail to capture the multifaceted, temporal variations inherent in real-world household consumption. Furthermore, many studies neglect multivariate relationships between electrical parameters (such as voltage, current, reactive power, and global intensity) and lack interpretability in their "black-box" models. This research aims to bridge these gaps by utilizing a highly granular, real-world, multivariate time-series dataset to develop accurate, interpretable, and robust machine learning models for household energy forecasting.

2. Methodology

Dataset

The study utilizes the Individual Household Electric Power Consumption (IHEPC) dataset, comprising over 2.07 million minute-level measurements from a single French household recorded between December 2006 and November 2010. Key variables include:

• Global Active Power: Total real power consumption (kW).
• Global Reactive Power: Unused power flowing between source and appliances.
• Voltage & Global Intensity: Electrical potential and current (Amperes).
• Sub-metering (1, 2, 3): Specific consumption for kitchen, laundry, and heating/cooling systems respectively.
• Derived Feature: Unmonitored consumption calculated by subtracting sub-metered loads from global active power.

Data Preprocessing

A systematic pipeline was implemented to ensure data quality:

• Handling Missing Values: Time-based linear interpolation was used to preserve temporal patterns without bias.
• Outlier Detection: Z-score analysis identified and removed noisy readings exceeding defined thresholds.
• Normalization: Min-Max scaling transformed numerical features into a 0–1 range to handle varying scales (e.g., voltage vs. current).
• Resampling: Data was aggregated into hourly, daily, and weekly granularities to capture both short-term fluctuations and long-term trends.
• Splitting: The dataset was divided into an 80% training set and a 20% testing set.

Machine Learning Models

Five regression models were trained and compared:

1. Decision Tree (DT): A single tree-based model using impurity metrics (Gini/Entropy) for splitting.
2. Random Forest (RF): An ensemble of decision trees using bootstrapped subsets and random feature selection to reduce overfitting.
3. Gradient Boosting (GB): A sequential ensemble method fitting trees to residuals of previous models.
4. LightGBM: A gradient boosting framework using leaf-wise tree growth for faster convergence and efficiency.
5. XGBoost: An optimized gradient boosting algorithm incorporating regularization (L1/L2) and sparsity-aware split finding.

Evaluation Metrics

Model performance was assessed using four standard regression metrics:

• R² (Coefficient of Determination): Percentage of variance explained.
• MSE (Mean Squared Error): Average squared difference between actual and predicted values.
• RMSE (Root Mean Squared Error): Square root of MSE.
• MAE (Mean Absolute Error): Average absolute difference.

Explainability

SHAP (SHapley Additive exPlanations) was integrated to interpret model predictions, quantifying the contribution of each feature to the output and bridging the gap between complex ensemble models and explainable analytics.

3. Key Results

Model Performance Comparison

All models demonstrated high predictive accuracy, with Test R² values ranging from 0.9980 to 0.9990. However, distinct differences emerged regarding generalization and overfitting:

• Decision Tree (DT): Achieved the highest training accuracy (Train R² = 0.9990, Train RMSE = 0.0080) but exhibited significant overfitting, evidenced by a large gap between training and testing errors (Test RMSE = 0.0440).
• Random Forest (RF): Demonstrated the most balanced performance. It achieved the highest Test R² (0.9990) and the lowest testing errors across all metrics (Test MAE = 0.0183, Test RMSE = 0.0340). The small gap between training and testing metrics indicates superior generalization and robustness.
• Gradient Boosting, LightGBM, and XGBoost: These models showed consistent and reproducible results with strong predictive power. XGBoost and LightGBM achieved very low testing errors (Test RMSE ≈ 0.0320) and minimal overfitting, though RF slightly outperformed them in overall test metrics.

Feature Importance and Correlation

• Correlation Analysis: Both Spearman and Pearson correlation heatmaps revealed a perfect positive correlation (1.00) between Global Active Power and Global Intensity. A strong negative correlation (-0.62 to -0.65) was observed between Global Active Power and Voltage.
• SHAP Analysis: The SHAP beeswarm and feature importance plots identified Global Intensity (Feature 2) as the dominant driver of prediction accuracy, with a mean absolute SHAP value of approximately 0.87. Other features like Global Reactive Power, Voltage, and Sub_metering_3 had negligible contributions, while Sub_metering_1 and Sub_metering_2 showed near-zero importance.

4. Key Contributions and Significance

The paper claims the following contributions to the field of residential energy analytics:

• Multivariate Real-World Approach: Unlike many studies using synthetic or short-term data, this work utilizes a high-resolution, real-world dataset spanning four years to model complex, multivariate relationships between electrical parameters.
• Comprehensive Model Comparison: The study provides a rigorous comparison of five distinct ML algorithms (DT, RF, GB, LightGBM, XGBoost) specifically for household power consumption, identifying Random Forest as the most robust model for this specific task due to its balance of accuracy and generalization.
• Interpretability: By integrating SHAP, the study moves beyond "black-box" predictions to provide actionable insights, specifically highlighting that Global Intensity is the primary determinant of power consumption in this context.
• Actionable Insights for Energy Management: The findings support the development of smart grid systems and household monitoring tools. The high accuracy and interpretability of the RF model suggest it is suitable for deployment in real-time applications (e.g., smart meters, residential dashboards) to facilitate proactive energy management, device scheduling, and sustainable living practices.

5. Limitations and Future Scope

The authors acknowledge that the study is based on data from a single household, which may limit generalizability to diverse demographic or behavioral patterns. Future work is suggested to expand the dataset to include diverse households and regional variations. Additionally, the authors propose investigating the trade-offs between prediction accuracy and computational overhead, as well as exploring hybrid or deep learning models for dynamic responsiveness.