Improved SOM based method for short term load forecast of Iran power network

The paper presents an improved method for a 24-hour load forecasting in the power system by using self organizing map (SOM) of Kohonen neural network .In this paper two models are compared with each other. The main difference between these models is about determining the training patterns procedure. In the first basic model, the training of neural networks performs in similar patterns with the most common properties, but in the training of neural networks in the second improved model, a visual SOM network is used in order to achieve the most similar patterns in its winner neuron for training main SOM networks of the proposed model. Each of these models is sensitive to atmospheric factors such as temperature. In addition, these models are able to forecast normal and abnormal days of year such as holidays, ceremonies, religious events and etc, with high accuracy. Ten sub-models are considered for forecasting each day of week, special holidays, the days before special holidays and the days after special holidays. The numerical results of 24-hour power forecasting for Iran power system during the years 2000-2006 Shows that the improved model is very efficient in reduction of MAD and increasing the accuracy of forecasting loads.