Hybrid Statistical Machine Learning Framework for Demand Modeling and Intelligent Monitoring in Distribution Systems

Abstract

Abstract—Distribution utilities need high-resolution demand modeling and voltage-condition monitoring. In practice, traditional state estimation requires accurate topology and line parameters that are not always available. This study presents a hybrid statistical-machine learning workflow for the CDA– L273 distribution feeder in Tegucigalpa, Honduras. From 96,486 15-minute records, which resulted in 72,548 validated activepower measurements (Jan 2023-Dec 2025), with only 0.05% outliers. The feeder exhibits stable operation, with a mean demand 7,082 W (peak 11,485 W), voltage stability of 1%, and a consistently power factor 0.96. Time-series analysis reveals strong short-term autocorrelation (r=0.90 at one 15-minute step) and repeatable daily/seasonal patterns. For demand prediction, a multilayer perceptron trained on 23 engineered features (lags, rolling statistics, and cyclical encodings) achieved R^2 = 1.0000, RMSE of 6–8 W, and MAE of 5–6 W (< 0.1% of the mean load). In parallel, an MLP-based model-free mapping from P,Q,PF to phase voltages supports lightweight monitoring without network models, with typical test errors of 0.2–0.6%. Overall, the proposed framework provides a practical, low-cost basis for planning and operational decision-making in data-sparse feeders.