Jan Kraljič (2011) Forecasting the electricity consumption data stream. EngD thesis.
Forecasting data streams of electricity consumption data is becoming more and more relevant for business risk management of electrical power distributors and traders. The forecasted values are used in electric market, load distribution, power plants load and power plants reserve management. As the numbers of measurement points are increasing the electricity consumption data is measured in increasingly shorter intervals. The data, read at equal width intervals generates data stream which we use for short term consumption forecast. Data mining of data streams has to be treated specially by machine learning algorithms. In this work forecasting problem has been split into two subproblems, one day ahead consumption forecast and hourly values for one day ahead. Data stream tests are performed on data for 1\% of Ljubljana's electricity consumption between years 2005 and 2008. Additionally, weather data and calender have been taken into account. Various combinations of data mining algorithms, discretizations and sliding windows are compared for both subproblems. Classical learning algorithms are used with sliding windows. Golden standard, ARIMA, linear model, naive Bayes classifier, k-nn, neural networks and random forest model are used in combination with equal frequency and equal width discretization and different sized sliding windows. Most of the mentioned models are commonly used on these type of data but not in combination with sliding windows and different discretizations. The error rate of selected combinations is bellow 5\%, which is already acceptable for practical use.
Actions (login required)