An important part in supplying electricity to residential areas are companies selling them the electricity. Accurate daily forecasts alleviate some of the concern related to providing the right amount of electricity without any surplus or shortages. The predictions are made after using a wide array of methods to analyze time series data (autocorrelations, time series decomposition, motifs and discords, etc.) and algorithms to build forecasting models (tree based methods, deep learning, statistical methods, etc.). Recently a method was presented which efficiently calculates the matrix profile of a time series. Matrix profile enables us simple discovery of motifs and discords. We use matrix profiles to analyse our dataset (residential energy load in Maribor) and recognize the areas from which we should draw our features. We focused on residential energy consumption for which we built models, which forecast hourly energy demand. We compare different models and analyse the forecasting accuracy using different features. We apply the most successfull feature set to another data set and compare the forecasting accuracy. We also develop and present a method, that uses matrix profiles to generate features. The result indicate that we can use a similar set of features, which work well on Maribor dataset, and apply it to another residential consumption dataset and expect similarly good results.
|
