Strategies for predictive power: Machine learning models in city-scale load forecasting

Abstract

This study focuses on enhancing machine learning (ML) algorithms' performance in predicting daily loads for Kirkuk, Iraq—an essential element in energy planning, resource allocation, and policymaking. We explore single and ensemble learning algorithms, including AdaBoost, Bagging, Support Vector Regression (SVR), and Decision Tree (DT). To assess accuracy, we employ critical metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (R2). MSE and RMSE gauge precision through squared differences between predictions and actual energy consumption. MAPE reveals percentage deviations, while R2 quantifies model fit, indicating its ability to capture energy consumption variance. Our results highlight Bagging's superiority, particularly in day-ahead forecasts, demonstrating superior accuracy over Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). These findings underscore the importance of feature reduction methods in enhancing performance. Bagging exhibits promise across various feature reduction scenarios, proficiently predicting energy consumption. Furthermore, the AdaBoost algorithm demonstrates commendable performance. The application of voting ensemble learning emerges as a particularly insightful approach, effectively reducing the squared differences and deviations in energy consumption forecasts. The significant implications of these findings suggest that the models, with their impressive performance, could serve as valuable tools for energy planners and policymakers in Kirkuk, playing a key role in optimizing resource allocation for efficient energy utilization.

Keywords

• Load forecasting
• Electricity planning
• Machine learning
• Ensemble learning
• Feature reduction
• Kirkuk city