CLUSTERING-BASED PREDICTION OF RESIDENTIAL ELECTRICITY CONSUMPTION

The EU-funded project HESTIA (Holistic demand response services for European residential communities) is striving to develop modern ICT tools for the next generation of Demand Response (DR) services for residential consumers and prosumers. For the electric load monitoring, two demo sites were set up in the Netherlands and Italy. For the energy providers, grouping/clustering of households with respect to their electric consumption profile is of great importance. It can help them to predict electricity usage, find faulty devices, and help to adjust Time of Use tariffs. Electric load forecasting has financial benefits when trading on electricity markets and offers advantages for reliable operation
of electricity networks. Our present work focuses on the clustering, prediction and subsequent analysis of household electricity load profiles collected within the HESTIA project.
In 30 households both on the Italian and Dutch demo sites, the electric grid import/export data from smart meters were recorded over a one-year period with a resolution of 15 and 5 minutes respectively.
The households were equipped with PV panels and home batteries, and their electricity production were monitored. The total household consumption was calculated from the grid, PV and battery values. The data quality issues required considerable efforts and preliminary analysis.
Clustering the timeseries of single household daily consumption was performed by standard clustering methods such as K-Means. This was deployed as a software service running once a day on an AIT server on real-time updated monitoring data. The clustering training is performed once a month. Many
clustering methods such as K-Means involve non-deterministic steps, which makes the matching of the clusters between subsequent runs an important task. We have solved the cluster matching with the
Hungarian method, where we converted the task to a linear sum assignment problem. This enables to map cluster labels, and track cluster assignments across subsequent training runs. For predicting the cluster labels, we engineered a set of features from aggregated daily time series data, intraday data, measured and forecasted weather data, as well as calendar indicators. The cluster prediction task is interpreted as a classification problem, where the classifier is trained on historical data. Various machine learning algorithms (decision tree-based classifiers, support vector machines,
neural networks) were tested for classification performance. As the input data before the clustering was normalized, the centroid of the predicted cluster represents a forecast for the shape of the next day consumption. The accuracy of this classification-based 24 h ahead electrical load curve prediction is
evaluated in the presented study.