The effectiveness ofthe multiscale neural network (NN) architecture for time series prediction of nonlinear dynamic systems has been investigated. The prediction task is simplified by decomposing the time series into separate scales of wavelets, and predicting each scale by a separate multilayer perceptron NN. The different scales of the wavelet transform provides an interpretation of the series structures and information about the history of the series, using fewer coefficients than other methods. In the next stage, the predictions of all the scales are combined, applying another perceptron NN, in order to predict the original time series. Each network is trained by the backpropagation algorithm using the Levenberg-Marquadt method. The weights and biases are initialized by new clustering methods, which improved the prediction results compared to random initialization. Three sets of data were analyzed: the sunspots benchmark, fluctuations in a far-infrared laser and a numerically generated series (set A and D in the Santa Fe competition). Taking the ultimate goal to be the accuracy of the prediction, we find that our suggested architecture outperforms traditional nonlinear statistical approaches.
展开▼
