Neural-network-based demand forecasting in a deregulated environment

The traditional approach to load forecasting is based on processing time series of load and weather factors recorded in the past. In the dynamic environment of the deregulated power industry, historical load data may not always be available. This paper explores the possibility of an alternative approach toward load forecasting based on indirect demand estimation from available customer data. This approach requires utilization of demand models for different customer categories. This paper presents a neural-network-based method of demand modeling. Neural networks are designed and trained based on the aggregate demands of the groups of surveyed customers of different categories. The performance of such models depends on the neural network design and representativeness of the training data. The forecast accuracy is also affected by the forecasted group size, customer characteristics, customer classification system, and the extent of demand survey. This paper discusses the issues of neural network design and illustrates the proposed method by its application to forecasting demand of residential customers.