Short term load forecasting based on hybrid artificial neural networks and particle swarm optimisation

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

Short term load forecasting (STLF) is the prediction of electrical load for a period that ranges from the next minute to a week. The main objectives of the STLF function are to predict future load for the generation scheduling at power stations; assessment of the security of the power system as well as for timely dispatching of electrical power. STLF is primarily required to determine the most economic manner in which an electrical utility can schedule generation resources without compromising on the reliability requirements, operational constraints, policies and physical environmental and equipment limitations. Another application of the STLF is for predictive assessment of the power system security. This system load forecast is an essential data requirement for off-line network analysis in order to determine conditions under which a system may become vulnerable. This information allows the dispatcher to prepare the necessary corrective actions. The third application of STLF is to provide the system dispatcher with more recent information i.e., the most recent forecast with the latest weather prediction and random behaviour taken into account. The dispatcher needs this information to operate the system economically and reliably. Due to the sensitivities surrounding a load forecast, it thus becomes crucial that the forecasting error is minimised. There are various methods that are used for short term load forecasting, namely; statistical methods and computational intelligence methods. Statistical methods are known as the regression methods which forecast the future electrical load based on historic time series load information. These methods have been in use for many years however due to the dynamic changes in the power system today such as the introduction of Independent Power Producers (IPPs) onto the grid; it becomes difficult to use these methods because they are very static and inflexible i.e. they cannot be manipulated by including rules or expert knowledge in order to counter the effect of any sudden changes in the power system. Their inability to adapt to the changing behaviour of the power system thus leads to high forecasting errors. Computational intelligence (CI) methods however are dynamic and are able to learn by experience. Short term load forecasts have been conducted by using various CI methods such as Artificial Neural Networks (ANNs), Genetic Algorithms (GAs), Fuzzy Logic (FL), Expert Systems (ES), and Particle Swarm Optimisation (PSO). Hybrid versions of these methods, where two or more CI methods are amalgamated in a process to forecast future load, have also been used. iv In this research, a traditional forecasting technique, Multiple Linear Regression (MLR), was compared with a CI technique, Artificial Neural Networks. ANN was also compared with another neural network method namely Elman Recurrent Neural Network (ERNN) to determine whether a more neural network method with memory yields better results as compared to ANN.