Predictive Modeling of Urban Residential Power Consumption: An Explainable AI Approach to the Kenyan Prepaid Grid

Technical Summary: Predictive Modeling of Urban Residential Power Consumption in the Kenyan Prepaid Grid

Problem Statement
Accurate electricity load forecasting is critical for grid stability, revenue planning, and energy management, particularly in rapidly urbanizing regions where prepaid metering has become the dominant billing system. However, existing forecasting models often rely on assumptions derived from Western markets—specifically, high sensitivity to weather fluctuations (due to HVAC usage) and significant short-term price responsiveness. These assumptions may not hold in sub-Saharan African contexts, where residential consumption is driven by essential baseload activities and constrained by prepaid credit balances rather than climate control or immediate tariff elasticity. Furthermore, the "black-box" nature of advanced machine learning (ML) models often limits their utility for policymakers who require transparent insights into the drivers of demand. This study addresses the gap in understanding the true drivers of residential demand within the Kenyan prepaid sector and evaluates whether complex ML architectures offer meaningful advantages over simpler statistical approaches in this specific environment.

Methodology
The study employs a longitudinal dataset spanning January 2017 to April 2026, integrating historical consumption records, meteorological variables, and tariff information from the Energy and Petroleum Regulatory Authority (EPRA).

• Feature Engineering: The framework utilizes autoregressive features (Lag 1, Lag 2, and Lag 12 kWh) to capture short-term persistence and annual seasonality. Exogenous variables include monthly average temperature, Cooling Degree Days (CDD), and tariff parameters (Lifeline Rate and Limit).
• Modeling Framework: A comparative approach benchmarks an Extreme Gradient Boosting (XGBoost) model against a traditional Ordinary Least Squares (OLS) regression model.
• Validation Strategy: To prevent information leakage, a strict chronological validation strategy was used. Data from 2017–2024 served for training and hyperparameter optimization, while 2025–2026 data were reserved exclusively for out-of-sample testing.
• Explainability (XAI): The study integrates SHapley Additive exPlanations (SHAP) to quantify feature contributions and interpret the XGBoost model.
• Diagnostics: The analysis includes Variance Inflation Factor (VIF) checks for multicollinearity, autocorrelation/partial autocorrelation analysis (ACF/PACF), seasonal decomposition, and ablation studies to isolate the contribution of different predictor groups.

Key Results

1. Dominance of Historical Behavior: Correlation analysis and SHAP values reveal that historical consumption behavior is the primary determinant of future demand. Lag-based autoregressive variables (particularly Lag 1 and Lag 2) exhibit the highest predictive influence, far exceeding the impact of weather or tariff variables.
2. Model Performance Parity: Despite its greater complexity, the XGBoost model achieved forecasting performance broadly comparable to the OLS model. The XGBoost model showed a marginally lower Mean Absolute Error (MAE), while OLS yielded a slightly lower Root Mean Square Error (RMSE). This suggests that the predictive information in the dataset is largely linear and can be captured by simple autoregressive relationships.
3. Limited Impact of Exogenous Variables: Weather (CDD) and tariff variables contributed only marginal improvements to forecasting accuracy. Ablation studies confirmed that adding these variables to a "Lag Only" model resulted in negligible reductions in RMSE.
4. Behavioral Persistence: Autocorrelation diagnostics and seasonal decomposition indicate that residential demand is governed by strong temporal persistence and recurring consumption patterns rather than immediate responses to price changes or weather variability.

Significance and Claims
The paper claims that residential electricity demand within the Kenyan prepaid sector is governed primarily by behavioral persistence and recurring consumption patterns, rather than the weather sensitivity or price elasticity often assumed in Western literature.

• Methodological Contribution: The study demonstrates that computationally efficient, transparent statistical models (like OLS) can provide reliable demand estimates for utility planning in prepaid environments, challenging the assumption that increasingly complex ML architectures are always necessary for superior performance.
• Policy Implications: The findings suggest that short-term residential demand is relatively inelastic to moderate tariff adjustments. Policymakers should not expect immediate demand reductions following tariff revisions, as households prioritize essential services and maintain stable usage habits based on available credit.
• Operational Value: The research highlights the value of interpretable forecasting frameworks. By identifying that historical consumption is the dominant driver, utilities can prioritize accurate record-keeping of past usage over the deployment of complex, opaque models, thereby enhancing operational decision-making and stakeholder trust.

In conclusion, the study posits that for the Kenyan prepaid grid, the "black box" of advanced AI offers limited additional forecasting skill over simple autoregressive models, and that understanding the behavioral inertia of prepaid consumers is more critical than modeling external economic or climatic variables for short-term load forecasting.