Development of models for short-term load forecasting using artificial neural networks

Optimal daily operation of electric power generating plants is very essential for any power utility organization to reduce input costs and possibly the prices of electricity in general. For a fossil fuel – fired power plant for example, the benefits of power generation optimalization (i.e. generate what is reasonably required) extends even to environmental issues such as the subsequent reduction in air pollution.Now to generate “what is reasonably required” one needs forecast the future electricity demands. Because power generation relies heavily on the electricity demand, the consumers are also practically speaking required to wisely manage their loads to consolidate the power utility’s optimal power generation efforts. Thus, for both cases, accurate and reliable electric load forecasting systems are absolutely required.To date, there are numerous forecasting methods developed primarily for electric load forecasting. Some of these forecasting techniques are conventional and often less favoured. To get a broad picture of the problem at hand, a literature survey was first conducted to identify possible drawbacks of the existing forecasting techniques including the conventional one. Artificial neural networks (ANNs) approach for short-term load forecasting (STLF) has been recently proposed by a majority of researchers. But there still is a need to find optimal neural network structures or convenient training approach that would possibly improve the forecasting accuracy. This thesis developed models for STLF using ANNs approach. The evolved models are intended to be a basis for real forecasting application. These models are tested using actual load data of the Cape Peninsula University of Technology (CPUT) Bellville campus reticulation network and weather data to predict the load of the campus for one week in advance.The models were divided into two classes: first, forecasting the load for a whole week at once was evaluated, and then hourly models were studied. In both cases, the inclusion of weather data was considered. The test results showed that the hour-by-hour approach is more suitable and efficient for a forecasting application. The work suggests that incremental training approach of a neural network model should be implemented for on-line testing application to acquire a universal final view on its applicability.