Artificial Intelligence based Power System Load Forecasting for Smart Grid

Abstract

Load forecasting is very crucial for economic and reliable operation of power systems. This type of forecast is required in daily operations of power plants and is normally carried by utility forecasters. Though a reasonable state of performance has been achieved in this field, many new challenges have emerged with smart grid initiative. Hence, there is an acute need for more advanced and more accurate load forecasting tools, which has motivated me to take further research in this field. The motivation and contribution of the research work is discussed further. Integrated Architecture for Short Term Load Forecasting: Usually, the models developed for forecasting the next day load use a single technique. Though, some hybrid approaches are developed, they are mainly meant for pre-processing and then forecasting the load. This forms an exciting area of research where an integrated architecture for an accurate load forecasting for the next day is developed. Clustering:As for the short term load forecasting, the selection of the training set is crucial. The criterion for selecting the training set is that there should be some correlation between the training and the testing data set. This forms another motivation to further explore the idea of the clustering of training patterns based on daily average load and peak load. Anomalous Day Modelling: Load forecast for holidays is always difficult to process, due to their dissimilar load behaviour. This forms an interesting field of research for their load forecast. Various case studies for generating enough training samples for anomalous days and their results when applied for integrated architecture and clustering approaches have been discussed. Long Term Load Forecasting: A time series data of long term load forecasting can be considered as comprising linear and non-linear components. Most long term load forecasting methods use one of the statistical techniques or an artificial intelligence to forecast the long term load.In the present work, a hybrid approach is proposed