Abstract

The electricity systems used in power distribution in the United States also consume a lot of electricity to warm up until it gets to homes, business or industries. These are technical losses which are caused by the physical resistance of the aging electrical equipment like transformers, conductors and distribution lines. With the aging of infrastructure, the resistance rises, and the amount of wasted energy increases. Although this is a large problem, most utility companies continue to use rough age estimations or customer complaints instead of using more direct and objective data-driven techniques to determine the concentration of the worst losses. This thesis fills that gap by creating and testing a completely open-source and reproducible data analytics pipeline to map energy loss hotspots in the United States power grid. The study design merges four publicly available datasets, i.e., the EPA eGRID database of over 12,634 active power generation nodes, the Global Power Plant Database of over 35,000 international facilities, the U.S. Department of Energy outage records of over 1.3 million events and the NOAA Storm Events Data of over 1.9 million weather events. A sequence of python analysis pipelines are used to process these datasets, comprising physics-based parameter estimation, machine learning classification, unsupervised clustering, spatial econometrics, anomaly detection and temporal trend analysis. IEEE 33-bus standard impedance parameters were determined to obtain the Aging Degree Index (ADI) of each plant node, where a model was used to degrade the resistance and reactance by 0.5 percent/year and 0.3 percent/year respectively to represent the physical effects of aging infrastructure cumulative. There were five stratified cross-validation folds with five machine learning classifiers. Random Forest had the best performance with ROC-AUC of 0.9167, better than XGBoost, Gradient Boosting, Support Vector machine and Logistic Regression which had a ROC-AUC of 0.9142, 0.8627 and 0.7759, respectively. The SHAP feature importance analysis indicated that the most important single predictor of high aging risk was the number of generator units per plant as older baseload plants with more than one generator unit have a higher topological complexity and are more susceptible to impedance decay. K-Means stratified the plant nodes into four groups of risk Low-Risk, Moderate-Risk, High-Risk and Critical-Risk with 734 Critical-Risk nodes, with an average infrastructure age of 58.1 years and an average ADI of 0.203. Isolation Forest anomaly detection detected 632 anomalous plants with a mean of 69.1 years of infrastructure age (compared to only 21.9 years of infrastructure age in normal plants), and 42.9 percent of K-Means Critical-Risk plants also were anomalies, meaning that both algorithms agree on the same highest-priority intervention targets. the Local Moran I spatial analysis showed that there were 1,027 statistically significant High-High hotspots node with p less than 0.05, and Getis-Ord G* showed that the fleet had 1,639 high-aging hotspots clusters and 2,463 low-aging cold spots, and a Global Moran I of 0.3189 indicates that there is significant spatial clustering in the fleet. The results confirm a Pareto-like energy loss distribution: 60-70 percent of grid nodes provide 15-20 percent of chronic technical losses. Hotspots were found to be of two types. Hotspots along the Gulf Coast and Atlantic Seaboard coastline are highly correlated with storms and Pearson r is large (0.78) along the coastline in the coastal states, but hotspots in Midwest, Northeast and Appalachian regions are smaller (r = 0.31) and mainly due to the chronic degradation of infrastructure. Ad hoc tests of augmented Dickey-Fuller stationarity have shown that the national outage time series is not randomly varying, but is actually increasing over time, i.e. grid stress is worsening over time and will not correct itself without an outlay of conscious large-scale modernisation. Aging degree, nameplate capacity, age of infrastructure, spatial hotspots status and anomaly detection status called Capacity-Weighted Vulnerability Index identified 2,527 plants in the Critical risk band which make up 20 percent of the national fleet. The highest percentage of plants belonged to the Critical band in New Hampshire with 70.5 percent and then Washington State with 31522 megawatts of capacity on the high-risk list and Alabama and Tennessee with over 44 percent plant critical ratios. The highest number of plants that required urgent upgrade were in Wisconsin at 54 plants. The 783 plants make up 6.2 percent of the total fleet, and are already above the IEEE 33-bus aging tolerance threshold, and require immediate replacement of infrastructure. Hurricanes and typhones were identified as the most devastating type of storms as the total property damage of 155.48 billion and customer-hours lost were highly correlated with the storm property damage at r = 0.642. The entire codebase is published on GitHub under an open-source licence such that the entire methodology can be re-executable at no cost by any utility, researcher or regulator. This thesis presents a proactive and scalable evidence-based model of data-driven grid modernisation planning, featuring direct avenues to extension into real-time SCADA integration, smart meter data fusion, and country-wide implementation of regulations.

Publication Date

5-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Ehsan Warriach

Campus

RIT Dubai

Download

Share

COinS