Session: 01-02: AI for Energy Sustainability II

Paper Number: 130477

130477 - A Deep Learning-Based Method for Non-Intrusive Load Monitoring and Load Disaggregation of 11kV/400V Electrical Substations 

Abstract: 

A deep learning-based method has been developed to achieve Non-Intrusive Load Monitoring (NILM) and load disaggregation at 11kV/400V electrical substations, accomplished using the aggregate load measured at the substation and separating it by property. This allows Utility & Power Distribution companies better electricity demand forecasts and aids demand-side planning due to understanding the routine of household electrical consumption. There has been extensive research of NILM at household-level using household “smart meter” data, however substation-level disaggregation offers a cost-effective alternative and streamlined approach as only the substation requires monitoring devices versus the alternative of every house on the distribution network being fitted with monitoring equipment.

Behavioural patterns in device usage were identified at current uptake levels of Low-Carbon Technology with analyses of devices such as Electric vehicles (EVs), Heat Pumps (HP) and Photo-Voltaic Solar (PV) with the relationship with substation load. Aggregate substation load data were captured by EA Technology’s VisNet Hub® units fitted to the Cheshire and Warrington Local Enterprise Partnership (LEP). Load and device data were obtained from ‘Pecan Street Dataport’, with datasets created within neighbourhoods in New York and Austin for validation, before applying the code method to the LEP fleet. The code was developed in Python, using the machine learning packages ‘Tensorflow’, ‘Keras’ and ‘scikit-learn’. An eventless NILM approach was taken, using Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers. The neural network was initially applied to NILM at household level, disaggregating the entire household load into individual components; and subsequently scaled to substation level, with a working proof-of-concept. The neural network was successfully able to disaggregate household loads for the Pecan St. Dataset including EVs, PV and Battery Energy Storage Systems (BESS). Furthermore, it was able to model two-way power flow, common with households with PV and BESS that export electricity back to the distribution grid.

Ongoing work is achieved on refining the model architecture and hyperparameters to ensure the optimal accuracy, compared to the empirical ground truth values. In parallel, the disaggregation method is being adapted to the VisNet Hub fleet, which requires creating a synthetic dataset which is an amalgamation of real data measured at the substation from the VisNet Hub fleet, empirical data and simulations that accurately portray device usage in each household. This ultimately allows for an evaluation of a scenario-based uptake in low-carbon technologies and an analysis of how the Power Distribution Network will need to adapt to support these loads as decarbonisation forces a paradigm shift in customer usage and introduction of new household devices.

Presenting Author: Elliott O'Malley Cardiff University

Presenting Author Biography: Elliott O'Malley is in his third academic year of his PhD at Cardiff University. He received a bachelor's degree in Mechanical Engineering (BEng) from Cardiff University in 2021 and has since joined the Resilient Decarbonised Fuel Energy Systems Centre for Doctoral Training (a partnership involving the universities of Cardiff, Sheffield, and Nottingham). His research interests include Non-intrusive load monitoring, artificial intelligence and deep learning applied to the energy networks. His industrial sponsor is EA Technology.

Authors:

Elliott O'Malley Cardiff University
Daniel Pugh Cardiff University
Jianzhong Wu Cardiff University
David Clements EA Technology