A State-Based Prediction Model for Energy Demand Forecasting

In this thesis we introduce the state-based prediction method that is applied on the domain of Power TAC for performing short-term energy demand forecasts. Power TAC is an agent-based competition that simulates an energy market in which energy broker agents compete against each other for the goal of pro_t maximization. One of the tasks of a Power TAC broker agent is to predict the short-term imbalance of energy supply and demand in its portfolio. The statebased method is designed to perform this kind of forecasting. Its main feature is that it uses states that each represents unique combinations of data features to acquire data relevant for predictions. Subsequently, a weighted method as well as a simple linear regression model is used to determine future energy imbalance. The state-based method is compared to the CART regression tree model in terms of prediction performance and time performance. We found by conducting experiments that overall the state-based method obtains better prediction performance, but that it is less robust to noise. In terms of time performance, the state-based method shows a considerable improvement