Prediction of energy consumption and risk of excess demand in a distribution system

Abstract

An empirical model for prediction of energy consumption in a distribution system is described. The model resembles a normalized radial basis function neural network whose neurons contain prototype joint data about the consumption process and the environment. A set of prototype patterns of consumption and environmental variables is formed from a record of a multi-component time series by a self-organized process. Prediction of energy consumption is performed by a conditional average estimator based upon known prototype patterns and given future values of environmental variables. Importance of these variables for the prediction is determined by a genetic algorithm. Prediction performance of the model is tested on a one-year-long consumption record of a gas distribution system. Prediction error is determined by the difference between predicted and actually observed consumption. Its value depends on time and amounts to a few percent of the actual consumption. The probability distribution of prediction error is estimated from a properly selected time interval of prediction. This distribution can be used to estimate the risk of energy demand beyond some prescribed value. For an optimization of the distribution process, a cost function that includes operation and control costs of a distribution system as well as penalties related to excess energy demand is proposed. Its minimum corresponds to an economically optimal energy distribution.