AbstractMost of the tools that are currently used for energy management fail to detect irregular behaviour when a dataset is measured in a time domain rather than a frequency domain. The objective of this research was to develop an analytical technique that can be used as an energy management tool in detecting irregular behaviour of a time series. Two electricity demand time series, two simulated half hourly series and an airline passenger time series were chosen as case studies for this research; however, the electricity demand series were heavily influenced by the presence of multiple seasonalities and heteroskedasticity. Most established time series methods were developed on the assumption that the errors are
homoskedastic, hence the prediction limits that are created from such established method will often fail in detecting irregular behaviours.
In this research, a recently developed time series method that models a time series explicitly was modified to accommodate the presence of multiple seasonal components, as well as the presence of heteroskedasticity. Upon incorporating multiple seasonal components and heteroskedastic components into the modified time series method, its forecast accuracy were comparable with established time series methods such as the double seasonal autoregressive integrated moving average (DSARlMA) and the double seasonal Holt-Winters exponential
smoothing, which have been used as benchmarks. The prediction limits of the benchmarks and the modified time series method were evaluated to examine if they are able to detect irregular electricity demand which have been simulated. However, the prediction limits of the modified time series method were adjusted by giving more weight to older observations and less weight to recent observations.
As part of detecting irregular consumption, a procedure was also developed to test for the difference between the number of observations outside the prediction limits before and after a change. The modified time series method proved to be a tool that can be of significant importance in the area of energy management as it is able to produce forecasts that are comparable with existing time series methods, as well as produce prediction limits that can be used for detecting changes in consumption pattern.
|Date of Award||2013|
|Supervisor||Hasan Al-Madfai (Supervisor), Steve Thomas (Supervisor), Steve Lloyd (Supervisor) & Steve Lakin (Supervisor)|