A Combination of Pruning Algorithm and Parallel Networks Structure to Increase the Generalization of Neural Networks Used for Short-Term Load Forecasting of Iran Power System

In short term load forecasting using Neural Networks, when the training data is contaminated with high noise, this noise is mapped into network's weights, and causes increasing of forecasting error. This forecasting error make us apply some methods to increase accuracy in neural net. In this paper, load forecasting of such power systems is done based on employing two methods: Pruning algorithm and parallel networks structure. Considerable error reduction using these methods confirms that both methods improve the generalization of neural nets. Results of Tehran load forecasting whose training data is contaminated with high noise is a subsidiary of the ability of these methods in improving the generalization.