Sensitivity analysis of a neural network used for the forecasting of electric power load

Abstract

A feedforward multilayer neural network has been used for the estimation of the four-hour-ahead electric load in Power Plant in the island of Crete. An attempt was made to use few variables for the input vector, while keeping the degree of accuracy to acceptable levels. To this effect a sensitivity analysis of the input parameters was performed. The parameters investigated were both environmental (weather condition, minimum and maximum temperature) and seasonal (Julian day, holiday classification). The architecture of the network was a multi-slab feedforward structure using backpropagation. This served as the selected platform for the comparisons. The network was trained with data that were pruned in both size and content. The correlation coefficient between actual and predicted power load was 0.987 when all the parameters were used for the training of the network. The network has also been compared to a multiple linear regression analysis. The correlation coefficient for this technique was 0.983.