Building Household Electricity Load Profile Using Machine Learning

Abstract

The objective is to generate household electricity load-profiles by predicting device-wise electricity usage patterns using the time-series data, to trace the disutility of electricity. Further, the emphasis lies on how the socio-economic parameters of a household can combine with the time-series data to aid in better prediction. This is a comparative study representing a bottom-up model, with input granularity set to 10-min cycle power of 5 everyday household devices along with their associated timestamps as the building blocks and predicts, whether a device would be switched on or off at a given point in time. The model is trained and validated on REFIT dataset, comprising of 20 houses along with the socio-economic features of each house. For comparison purpose, two datasets are created, with and without the socio-economic parameters. Results point towards the impact of socio-economic features and how they improved the prediction accuracy by a fine margin for each device, leading towards promising high-resolution electricity load profiles. Using the socio-economic features, we were able to predict the state of a device up to an accuracy of 97%, whereas without these features the accuracy was 76%.