1 Neural Networks with Wavelet Based Denoising Layers for Time Series Prediction UROS LOTRIC 1 AND ANDREJ DOBNIKAR University of Ljubljana, Faculty of Computer and Information Science, Slovenia, e-mail: {uros.lotric, andrej.dobnikar}@fri.uni-lj.si Abstract To avoid preprocessing of noisy data, two special denoising layers based on wavelet multiresolution analysis are integrated into the layered neural networks. A gradient based learning algorithm is developed which uses the same cost function for setting both the neural network weights and the free parameters of denoising layers. The proposed layers, integrated into feedforward and recurrent neural networks, are validated on time series prediction problems: the Feigenbaum sequence, the rubber hardness time series and the yearly average sunspot number. It is shown that the introduced denoising layers improve the prediction accuracy in both cases. Keywords: feedforward and recurrent neural networks, wavelet multiresolution analysis, denoising, gradient based threshold adaptation, time series prediction 1 Corresponding author: Uros Lotric, University of Ljubljana, Faculty of Computer and Information Science, Trzaska 25, 1000 Ljubljana, Slovenia, e-mail: uros.lotric@fri.uni-lj.si, phone: +386 1 4768 874, fax: +386 1 4768 369