A Machine Learning Framework to Infer Time-of-Use of Flexible Loads: Resident Behavior Learning for Demand Response

Milad Afzalan,Farrokh Jazizadeh

doi:10.1109/access.2020.3002155

Abstract

Load shapes obtained from smart meter data are commonly utilized to understand daily energy use patterns for adaptive operations in applications such as Demand Response (DR). However, they do not provide information on the underlying causes of specific energy use patterns - i.e., inference on appliances' time-of-use (ToU) as actionable information. In this paper, we investigated a scalable machine learning framework to infer the appliances' ToU from energy load shapes in a collection of residential buildings. A scalable and generalized inference model obviates the need for model training in each building to facilitate its adoption by relying on training data from a set of previously observed buildings with available appliance-level data. To this end, we demonstrated the feasibility of using load shape segmentation to boost ToU inference in buildings by learning from their nearest matches that share similar energy use patterns. To infer an appliance ToU for a building, classification models are trained for inference on subintervals of load shapes from matched buildings with known ToU. The framework was evaluated using real-world energy data from Pecan Street Dataport. The results for a case study on electric vehicles (EV) and dryers showed promising performance by using 15-min smart meter load shape data with 83% and 71% F-score values, respectively, and without in-situ training.

Highlights

In recent years, conventional centralized power systems are shifting to decentralized alternatives that integrate distributed energy resources (DER) such as solar panels, district resources, storage systems, and advanced technologies for smart metering and control
In this research, we have investigated the feasibility of an ML framework for inferring ToU of major flexible loads from smart meter data with the 15-minute resolution that accounts for the following features: (1) integrating a resident behavior learning component that leverages energy load shapes for identification of a training dataset to increase the efficiency of the machine learning training process, (2) inferring ToU for unseen buildings which have not contributed to the training process, and (3) investigation of appliance ToU is performed over the hourly basis with application to dynamic load management
We considered electric vehicles (EV) and dryer as two instances of flexible loads that are suitable for DER management and Demand Response (DR) applications

Summary

INTRODUCTION

Conventional centralized power systems are shifting to decentralized alternatives that integrate distributed energy resources (DER) such as solar panels, district resources, storage systems, and advanced technologies for smart metering and control. Jazizadeh: ML Framework to Infer ToU of Flexible Loads: RBL for Demand Response as supplementary information to the load shapes at the aggregate level, helps manage power systems more efficiently by engaging a subset of consumers that will result in a higher gain in efficient operations This could be achieved by using appliance-level energy use data and data-driven solutions that characterize energy consumption styles of users [1]. In this research, we have investigated the feasibility of an ML framework for inferring ToU of major flexible loads from smart meter data with the 15-minute resolution that accounts for the following features: (1) integrating a resident behavior learning component that leverages energy load shapes for identification of a training dataset to increase the efficiency of the machine learning training process, (2) inferring ToU for unseen buildings which have not contributed to the training process, and (3) investigation of appliance ToU is performed over the hourly basis with application to dynamic load management (e.g., demand-response). We primarily investigated 10 as the number of neighbors for the KNN, while comparing it against other values in the result section

APPLIANCE TIME OF USE INFERENCE

RESULTS AND DISCUSSION

CONCLUSION