Comparative Evaluation of Ensemble Learning Methods for Capital Expenditure Deviation Prediction in FERC Rate-Base Transmission Projects

Abstract

Accurate capital expenditure (CapEx) prediction for large-scale transmission projects regulated under the Federal Energy Regulatory Commission (FERC) rate-base mechanism remains a persistent challenge, as cost overruns directly inflate the rates borne by electricity consumers. This study presents a comparative evaluation of ensemble learning methods—XGBoost, LightGBM, CatBoost, and Random Forest—for predicting CapEx deviations in FERC-jurisdictional transmission projects. Drawing on publicly available FERC Form 1 financial data (1994–2019) and supplementary macroeconomic indicators from the Bureau of Labor Statistics (BLS) and the Producer Price Index (PPI) for electrical equipment, we construct a tabular dataset of 1,247 utility-year records encompassing annual transmission capital additions, operations and maintenance expenditures, regional labor indices, and commodity price signals. Each record is labeled with a deviation indicator reflecting whether actual annual capital additions exceeded the prior three-year rolling average by more than fifteen percent. Experimental results across five-fold cross-validation indicate that LightGBM achieves the strongest predictive performance (AUC = 0.841, F1 = 0.762), with XGBoost following closely (AUC = 0.833, F1 = 0.748). SHAP-based feature importance analysis reveals that the construction cost index, year-over-year transmission plant growth rate, and regional wage differential are the three most influential predictors. These findings offer empirical evidence supporting the viability of gradient boosting approaches for early-stage cost deviation screening in regulated transmission investments.