Prediction of Transformer Load-Rate and Analysis of Meteorological Feature Contributions for Sustainable Operation of Regional Substations: An Explainable Machine Learning Approach

Prediction of transformer load-rate at regional substations can provide information to support operating condition assessment, capacity margin analysis, and the identification of potential high-load risks. This study selected a 40 MVA main transformer at the Panzhou regional substation in Maoming City, Guangdong Province, China, as the research case. Based on hourly transformer load-rate and observed meteorological data from 2024, the forecasting performance of various time-series baselines and machine learning models was compared. The Random Forest (RF) model was then selected for input feature ablation experiments and explainability analysis. On this basis, Shapley Additive Explanations (SHAP) and Accumulated Local Effects (ALE) were employed to analyze the contributions of meteorological features to model predictions and their nonlinear response relationships. The model used observed meteorological data at the target time as inputs; therefore, the reported results reflect load-rate forecasting performance conditional on observed meteorological conditions. The results show that the RF model achieved a coefficient of determination (R2) of 0.9654, a mean absolute percentage error (MAPE) of 3.93%, a mean absolute error (MAE) of 0.97 percentage points, and a root mean square error (RMSE) of 1.46 percentage points on the independent test set for 2024. On the independent validation set from June to August 2025, the R2 and MAPE were 0.9296 and 3.72%, respectively, indicating that the model retained a certain degree of predictive capability on warm-season data from a different year at the same site. The ablation experiments showed that historical load features were the primary source of the model’s predictive performance, with an R2 of 0.9510 obtained using all historical load features alone. After incorporating temporal and meteorological features, the R2 further increased to 0.9654. SHAP analysis indicated that temperature and relative humidity had relatively high importance among the meteorological variables, while ALE analysis showed that the model exhibited different nonlinear responses to individual meteorological variables. The findings may provide data-driven support for reliable and sustainable operation and load management of regional substations.