PNNL

Abstract

Abstract —An accurate representation of the voltage- dependent, time-varying energy consumption of end-use electric loads is essential for the operation of many modern distribution automation (DA) schemes. One such scheme, volt-var optimization (VVO), has been widely deployed because of its ability to decrease energy consumption and peak demand. Many modern VVO schemes leverage electric network models and power flow results to inform control decisions, and are sensitive to errors in the end-use electric load models. Load modeling for utility VVO systems is typi- cally performed based on a load allocation algorithm, which uses a sparse set of measurements to estimate demand for each electric customer. End-use load modeling can potentially be improved using additional measurements, such as from advanced metering infrastructure (AMI). This paper presents two independent and novel machine learning algorithms for creating accurate, data-driven, time-varying load models for use with DA technologies such as VVO. The first algorithm uses historic AMI data, k -means clustering, and least-squares optimization to create a predictive load model for individual electric customers. The second algorithm uses deep learning (via a convolution-based recurrent neural network) to incor- porate additional data and increase model accuracy for each electric customer. The improved accuracy of the load models for both algorithms is validated through simulation.