Modeling and forecasting electric daily peak loads using abductive networks

Forecasting the daily peak load is important for secure and profitable operation of modern power utilities. Machine learning techniques including neural networks have been used for this purpose. This paper proposes the alternative modeling approach of abductive networks, which offers simpler and more automated model synthesis. Resulting analytical input-output models automatically select influential inputs, give better insight and explanations, and allow comparison with other empirical models. Developed using peak load and extreme temperature data for 5 years and evaluated on the sixth year, a model forecasts next-day peak loads with an overall mean absolute percentage error (MAPE) of 2.50%, outperforming neural network models and flat forecasting for the same data. Two methods are described for forecasting daily peak loads up to 1 week ahead through iterative use of the next-day model or using seven dedicated models. Effects of varying model complexity are considered, and simplified analytical expressions are derived for the peak load. Proposals are made for further improving the forecasting accuracy.