Time-series modeling and prediction of weather-driven system-level electrical load – Case of Abu Dhabi

摘要

Loadforecasting has long been used in operations and planning of the electric powersystem. In this study, weather variables were used for modeling and predictionof the system-level electrical load of the city of Abu Dhabi, UAE. A TransferFunction (TF) model was developed and its accuracy was compared to that of anAutoregressive Integrated Moving Average (ARIMA) model. We also tested anArtificial Neural Network (ANN) model based on the same weather variables thatwere used in the TF model. Assuming perfect knowledge of the weather variablesover the forecasting horizon, the TF model was more accurate for forecasthorizons of up to 48 hours. The ANN model, on the other hand, was more accuratefor one-week ahead forecasts. Assuming imperfect knowledge of the weathervariables (i.e., they are not known over the forecasting horizon and have to beforecasted first), the TF model was more accurate than the ANN model in allcases. Average accuracy of the best TF method does not exceed 1.5% for 24-hourhorizon, 2.5% for 48-hour horizon and 4% for 168-hour horizon. With the addeduncertainty of forecasted weather drivers, the accuracy of the proposed methoddegrades only slightly, while the ANN model is much less robust and becomesunusable beyond a two-day horizon.