Multiresolution process neural network and its learning algorithm

Abstract

A new model of multiresolution process neural network (MRPNN) which incorporates the characteristics of hierarchical, multiresolution and local learning capability is proposed based on the multiresolution analysis theory and process neural network model. This type of neural network facilitates in tackling with continuous input signals, which makes it possible to forecast time series problem. In addition, in order to approximate the nonlinear system, the hidden layer is used to deal with the nonlinear and complexity problems. A novel learning algorithm is given to expand the input functions and network weight functions based on the expansion of the orthogonal basis functions, subsequently The learning algorithm then builds the network by locating high error regions and adding nodes that get its activation function from the higher resolution space of the current local node, and its support falls within the high error region. Finally, the network is used to forecast the medium-term load of power system. Simulation results show that the network has good convergence and high accuracy. This method provides an effective solution to medium-term load forecasting in power system.