New neuron models for simulating rotating electrical machines and load forecasting problems

The existing neuron structure has an aggregation function (usually summatio n) and its transformation through non-linear filter or squashing or thresho lding functions. Such structure of neural networks has a number of disadvan tages like large number of neurons, hidden layers and huge training data re quired for complex function approximations. The present paper proposes new neuron models to overcome the above problems in the existing neural network s. The model has been developed and tested for modelling of electrical mach ines like DC motor, induction motor and synchronous generator and load fore casting problems using different new neuron models in the neural network li ke Sigma neuron in all the layers (hidden and output layers) as in the exis ting neural networks, Pi neurons in all the layers, Sigma neurons in hidden layer and Pi neurons in the output layer and finally Pi neurons in hidden layer and Sigma neurons in output layer. After simulating the above mention ed models for mapping the starting transient characteristics of induction m otor, it was found that, the Pi neurons in the hidden layer and Sigma neuro ns in output layer neural network model requires least training time and al so giving least rms error as compared to the other models. Hence, it is qui te clear that the existing Sigma neurons backprop neural network models can be replaced by some other efficient neural network which will incorporate all the properties of the simple existing neural network as well as the hig her order neural networks. For exploring these possibilities various compen satory neuron models have been proposed in these paper. The above mentioned models have also been compared with the existing model to highlight their simplicity and accuracy. (C) 1999 Elsevier Science S.A. All rights reserved .