A Study of Accuracy of Neural Networks, Support Vector Machines and the Bayesian Regression in Load Demand and Price Forecasting (T-0409) (MFN 5061)

Abstract

For optimal power system operation, electrical generation must follow electrical load demand. The generation, transmission, and distribution utilities require some means to forecast the electrical load and its pricing so they can utilize their electrical infrastructure efficiently, securely, and economically. The short-term load and pricing forecast represents the electric load and cost forecast for a time interval of a few hours to a few days. This thesis will use three methods for forecasting: Neural networks, Support vector machines and the Bayesian regression. All these methods will use the same database obtained from the electrical company of Sydney, Australia. These regression models can be created and trained to receive historical load, price and future weather forecasts as inputs to produce a load and price forecast as its output. All the results from these methods will be recorded and compared to find which one gives us the best result with least MSE and NRMS.