The use of neural networks to determine the long-term impact of demand side management / by Werner Heinrich Kaiser
Kaiser, Wernich Heinrich
MetadataShow full item record
Eskom, South Africa national electricity provider, estimates that in the year 2006 South Africa will experience a capacity shortage. One way to address this problem is through the implementation of Demand-Side Management (DSM) in various sectors, ranging from the commercial to the industrial. For Eskom to succeed in their vision of implementing DSM, a tool has to be developed that can illustrate the long-term impact of various DSM options for a region. This tool could then be used to illustrate the various role players the advantages of implementing DSM. The purpose of this study was to test if Neural Networks (NN's) can be applied in the construction of an hourly baseline for a region. This baseline could then be used to illustrate the Long-term impact of various DSM options. In this study a technique was developed using hourly data to construct a baseline model for the calculation of the long-term impact of DSM. This technique was then tested and evaluated on a case study. To achieve this goal, an investigation was launched to determine which inputs have an influence on the energy use of a region. The different variables that influence the neural network topology were also investigated. This information was then used to develop a technique that models the energy use of the area. To determine accuracy of the simulated energy use, a verification procedure was developed based on an internationally accepted verification model, using data the NN did not "learn" on. Sufficient accurate results were obtained using the defined indices. Thus NN can be used to model the energy use of a region. The major disadvantage of this technique is that hourly data for the whole year was used to train the model on. The question arises into just how much information is needed to model the NN. Subsequently an investigation was launched to determine the minimum data set needed to model the energy use. It was found that a full factorial data set is the minimum set of data that a NN needs to train on. In choosing this data, a study has to be conducted on previous data to determine exactly when the best combination could be obtained. The results indicated that the data could not be minimised due to the configuration of input data. For this study, the months of the year were encoded. This aided the NN in learning the relationship between the various inputs and the energy use. It was found that this is a crucial step in aiding the NN. Thus the NN could not simulate accurate enough results without the encoded data. This results that the number of months cannot be minimised with the current technique used. The model then can be used to evaluate different DSM options by subtracting the hourly differences from the baseline. This information is then used to evaluate the options using various indices. The indices included monthly energy use, maximum demand, the energy use during the various time of use periods and the impact of greenhouse gasses. The concept was illustrated by an actual case study. The use of NN for modelling the baseline for the forecasting of the long-term impact of DSM is considerable faster than current techniques with a timesaving element of up to 90 %. The use of NN is thus a viable technique to model the baseline. The results indicate that NN can successfully be used for cases with a high diversity in the load, and with little or no knowledge of the underlying systems.
- ETD@PUK