City Research Online

Abstract

This dissertation proposes and implements the inclusion of model structure in combining forecasts. Empirical investigations are conducted with an emphasis on neural networks and seasonal exponential smoothing models using synthetic data and real time series, from the electricity sector. It starts with a literature review on combining forecasts and ensembles of neural networks, and highlights their use in forecasting within the energy sector. Research gaps are identified and the questions to be addressed in this research are set, thus leading to
three empirical studies.

The first study provides a detailed sensitivity analysis of the goodness-of-fit and forecasting performance of feed-forward neural networks on time series with different characteristics. It expands existing literature by increasing the number and variety of time series and by using graphical and statistical diagnostics to objectively judge the influence of model specification on forecasting performance. Having identified conditions for achieving stable model performance, this study facilitated the identification of suitable models for different time series characteristics, which are then useful in developing combinations (ensembles) of feed forward neural networks.

The second study proposes structural combination methods based on clustering (CB) and genetic algorithms (GA) for forecasting time series. Clustering of neural networks using their parameter space is performed to identify a pool of forecasts to be combined. Three synthetic time series and two real time series (electricity demand and wind power production) were used to assess the performance of the two proposals against several benchmarks in univariate and multivariate forecasting problems. Structural combinations with GA were more competitive than those with CB for non-seasonal time series and the multivariate wind power forecasting application, whereas for the seasonal series, the CB tended to be more competitive.

The third study focused on forecasting univariate time series with seasonality, by structurally combining, in separate applications, multiplicative Holt-Winters and multiplicative Holt-Winters-Taylor models. Noise addition and block swapping were applied to the original time series in order to generate structurally diverse individual models. Applications were conducted using a seasonal daily peak electricity demand time series, an hourly double-seasonal electricity demand series and a half-hourly double-seasonal electricity demand series. Structural combinations worked better for the peak electricity demand and half-hourly demand time series when model variation was induced via noise addition. For the double-seasonal hourly electricity demand, block swapping, as a means for diversity in models, resulted in better forecasts.

Finally, in the last chapter of this dissertation, conclusions are drawn from this research. The contribution to the literature is assessed and a future research agenda is proposed.

Publication Type:	Thesis (Doctoral)
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD28 Management. Industrial Management
Departments:	Bayes Business School > Management Doctoral Theses Bayes Business School > Bayes Business School Doctoral Theses

Export

Downloads