Machine learning approach to determine the energy sharing potential in a neighbourhood.
Last Updated: 9-2019
The application of energy storage is crucial for the efficient use of renewable energy. In order to adopt and operate energy storage systems in the energy grid, accurate short-term forecasting of the energy load is essential. In this study, two years of (fifteen-minute interval) electricity consumption data of seventy dwellings inside a residential district are analyzed. To determine the best suitable electricity demand prediction model for the given data, a literature study together with an empirical test of multiple machine learning models is conducted. Prior to the prediction algorithm, a clustering model is run to group households with similar demand profiles in a two week period preceding the prediction. A random forest regressor is used to predict hourly electricity demand for each cluster, one day ahead. Over ten validation days, the proposed model achieves, on average, an R2 score of 0.77 and a cumulative variation of root mean squared error (cv-RMSE) of 0.41. The k-means clustering, together with random forest, is able to predict the trends of hourly electricity loads one day ahead, and determine the indirect sharing potential. The most important prediction feature is the electricity consumption of the previous day. Other important features are humidity, outside temperature and electricity consumption of two and three days previous. Based on the validations, it is estimated that the neighborhood can be completely self-sufficient for seven out of twelve months with a storage system of 660kWh, which is 9.43kWh per dwelling. When applying long-term energy storage, the self-sufficiency could be increased even more. This study covers the prediction of the short-term energy demand of a small residential district with on-site renewables and uses clustering in the pre-processing of the data, which is a new approach in this field of research. Currently, this study used only individual days for validation. Validating over longer periods of time, at least multiple days, could increase the understanding of the actual sharing potential between dwellings. Also, by experimenting with additional user-specific prediction features such as; occupancy, appliances ownerships, and solar radiation, improved model fitting might be achieved.