Search eceee proceedings

Estimating appliance presence and use through smart metres for effective demand-side management of residential cooling

Panel: 2. Future and innovative policies

This is a non-peer reviewed extended abstract. No paper available.

Authors:
Veena Shirsath, Ahmedabad University, India
Saket Sarraf

Abstract

The residential sector forms a quarter of total electricity demand in India, with space cooling accounting for 33-65% of total electricity consumption in residential apartments. With the rise of global temperatures and average incomes, residential cooling is expected to rise in Indian households, posing a challenge to low-carbon energy system transition. In fact, the projected residential AC demand for Inida is expected to cross the total current electricity consumption of Africa by 2030.

Air-conditioners contribute to peak demand, and hence DisComs have an incentive to affect cooling demand through targeted demand side measures. In order to introduce, monitor, and evaluate Demand Side Measures (DSMs), DisComs need to identify households contributing to the cooling load and know their AC usage.

This study aims to estimate appliance ownership, and consumption behaviour based on aggregated sub-hourly household level electricity consumption data. We use various machine learning models to predict AC ownership and usage patterns, commonly termed as Non intrusive Load Monitoring (NLIM) or energy signature analysis.

Most work in this field is based on very granular current and voltage data. However, this data is difficult to access for a DisCom looking to implement demand side measures using smart metre data. Here we attempt to achieve a similar output with more readily available data for DisComs i.e., through smart metres measured at a 15 minute interval. This also allows for achieving scalable results. The resultant information is useful for DISCOMs to design better policies, such as shifting demand from peak to non-peak hours within the day, managing smart thermostats, and identifying households with inefficient appliances. Our results provide an important use-case to support the deployment of the Smart Meter National Programme and recent discussions on grid interactivity.

The general structure of our model uses a binary variable for AC use (whether AC was used at a given time or not) as the independent variable. The features involve the sub hourly total household energy use along with temperature measures for the given day, time of day, and month. The paper uses data from detailed submetering of household energy use of 144 households and their appliances at 15-minute intervals from 2018 to mid-2020 for three Indian states. Climate data is collected from the Indian Meteorological Department.

We use successively complex classification models, like logistic regression, ANN, RNN, and LSTM, to estimate appliance usage patterns. For each model, two different data structures are used. The first structure takes the whole day load curve as an input (wide form dataset). We use the first difference between the adjacent time periods as there is a significant correlation between them. The second structure takes each 15-minute interval as the input in time series format (long form dataset). This allows for a better account of the temporal dynamics of appliance use.

We find that LSTM model with the wide form dataset performs the best with a 77% accuracy and F1 score of 0.62. When testing the same model on test datasets to which the model was not exposed to, we get an accuracy of 68% and an F1 score of 0.17.

The results may be used by DisComs to identify appropriate households for demand-side measures and also formulate user-specific measures. The information can also be relayed to households for them to improve their energy consumption. Our model closely matches the accuracy of other NILM models which use appliance specific energy signatures but with a much more granular energy data. Household level socio-demographic characteristics could further enrich the model, but it is difficult to access that data by the DisComs. Our next steps include (a) expanding the model to predict the duration of AC use (in minutes), (b) AC energy consumption, and (c) creating a model ensemble for more robust prediction.

Downloads

Download this display as pdf: 2-216-24_Shirsath_Display.pdf