Multivariate time series modeling and forecasting of Winnipeg's electrical load-temperature relationship
Date
2005
Authors
Klippenstein, Scott
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Manitoba Hydro has approximately 5000 megawatts of hydroelectric generating capacity. It generates, transmits and distributes almost all the electricity consumed in Manitoba, and also sells and purchases electricity under agreements with neighboring systems in Canada and the United States. Power demand, especially in space heating and cooling, is linked to several weather variables, mainly the air temperature. This process involves fitting univariate time series ARIMA models to the temperature data and then fitting multivariate models to load and temperature . In this practicum, we look at the average temperature models and its extremes at monthly and daily time intervals. Next, we model hourly temperature and examine its behaviour throughout the day with comparison to average load's behaviour. Later, we extend this into the multivariate case for load and temperature. In the problems that were investigated, it is shown that the temperature and load relationship are fairly consistent throughout the day on average for each month. It was shown for a given week, there was no relationship that existed between the two series. On the monthly time intervals, load was strongly dependent on temperature, but decreased as we moved to daily time intervals. This was due to a drop in demand for electrical load on the weekends than during the week. Because of this, partitioning of the daily model was required. Further investigation on other weather variables may provide a better understanding on the noise factors. It was also shown that ARIMA models gave fairly accurate results on all different time intervals. For smaller time units, only a small fraction of temperature data was needed to achieve optimal results. Future research on this area using state space models may provide a more up-to-date forecasting on load and weather or Bayesian VAR models to reduce overparameterization.