A Multi-Timescale Data-Driven Approach to Enhance Distribution System Observability

TL;DR

This paper introduces a multi-timescale, data-driven approach that estimates unmetered customer consumption patterns to improve distribution system observability, utilizing machine learning models and real utility data.

Contribution

It develops a novel multi-stage machine learning framework combining spectral clustering, multi-timescale learning, and Bayesian methods to estimate unmetered customer loads.

Findings

Effective estimation of unmetered customer loads demonstrated

Improved distribution system observability verified with real data

Method outperforms traditional approaches in accuracy

Abstract

This paper presents a novel data-driven method that determines the daily consumption patterns of customers without smart meters (SMs) to enhance the observability of distribution systems. Using the proposed method, the daily consumption of unobserved customers is extracted from their monthly billing data based on three machine learning models: first, a spectral clustering (SC) algorithm is used to infer the typical daily load profiles of customers with SMs. Each typical daily load behavior represents a distinct class of customer behavior. In the second module, a multi-timescale learning (MTSL) model is trained to estimate the hourly consumption using monthly energy data for the customers of each class. The third stage leverages a recursive Bayesian learning (RBL) method and branch current state estimation (BCSE) residuals to estimate the daily load profiles of unobserved customers without SMs. The proposed data-driven method has been tested and verified using real utility data.