Statistical approaches to electric load forecasting

This dissertation studies statistical approaches to electric load forecasting. Electric load forecasts are used by electric utilities, energy generators, transmission companies, financial institutions, and other companies for important operational, financial, and infrastructure development decisions. More accurate load forecasts increase the efficiency of electric energy industry and reduce the energy used by the society.; We consider two problems: next year peak load forecasting and one day ahead electric load forecasting. For the next year peak load forecasting we present a statistical model that learns the load model parameters from historical load and weather data. We also describe the load pocket forecasting software that can be used to estimate and forecast the load growth in different service areas. This software builds statistical load models for various service areas (load pockets), estimates weather-normalized loads, normal weather conditions, estimates the ratios between the actual peak loads and the loads that would happen on designed days (weather normalized factors), and estimates the next year peak load. The estimation of the next year peak load is performed in the form of the probability distribution.; For one day ahead forecasting we investigate the applicability of Support Vector Machine (SVM) methods. SVM methods were originally introduced by V. Vapnik to solve pattern recognition problems. A generalization of the SVM algorithm to regression estimation is done by using Vapnik's e-insensitive loss function. Using historical hourly load, and hourly weather data we build an SVM model and apply it to load pockets, and to the entire system. We compare the results obtained by using the SVM models with the results obtained by using other methods such as neural networks and algorithms based on the statistical load modeling.