OPTIMIZATION OF TRANSFORMER LOAD FORECASTS IN SMART GRIDS THROUGH AI-DRIVEN REGRESSION AND WEATHER DATA FUSION

Abstract

This paper explores machine learning regression models for predicting maximum transformer load using historical and weather data. The growing energy demand and stress on infrastructure during peak periods motivate the need for accurate forecasting to enhance reliability and planning. Six models were evaluated: linear regression, Decision Tree, Random Forest, Support Vector Regression (SVR), K-nearest neighbors (KNN), and XGBoost. Three scenarios were tested. One year of historical data, one year of data plus weather variables, and ten years of synthetic data with weather fluctuations. Key features included connected load, date-based elements (day, month, year), and weather metrics like temperature, humidity, wind speed, and global horizontal irradiance (GHI). Data preprocessing involved merging transformers and weather datasets, feature engineering, and using Grid Search CV with Time Series Split for time-aware model tuning. Performance was evaluated using the root mean squared error (RMSE), the mean absolute error (MAE), and the R² coefficient. Scaled normalization facilitated visual comparison of models by plotting predicted versus actual line plots. In the one-year scenario without weather data, Linear Regression performed best (R² = 0.99), with Random Forest and KNN also performing well. When weather variables were added, Random Forest (R² = 0.90) and Linear Regression (R² = 0.99) remained strong, but SVR and KNN underperformed. With ten-year synthetic data, Random Forest (RMSE = 0.01, R² = 0.97) and XGBoost (RMSE = 0.02, R² = 0.98) outperformed others, capturing long-term and seasonal trends. Linear Regression and SVR struggled with extended forecasts. Correlation analysis revealed that transformer load had a moderate correlation with temperature (r = 0.34) and wind speed (r = 0.55), and a strong correlation with global horizontal irradiance (GHI) (r = 0.74). These findings validate the value of ensemble models and environmental variables in enhancing load forecasting accuracy. The study supports the integration of weather-aware machine learning for more intelligent energy grid management.