Department of Computer, Control and Management Engineering

Multidimensional times series prediction is a challenging task. Only recently the increased data availability has made it possible to tackle with such problems. In this work we devised a novel method to exploit the multiple correlated features in the time series. The recurrent neural networks and the wavelet transform have been important innovations in the fields of signal processing and time series prediction. This paper proposes a Wavelet Recurrent Network for multi-steps ahead prediction of multidimensional time series. The proposed model combines these two elements into a neural network that predicts multiple samples in the future that are multiple time steps ahead with respect to the input samples. This Wavelet Recurrent Network carries out a multiresolution decomposition of the input signal through the wavelet transform, predicts the future wavelet coefficients with the recurrent neural network and transforms the output back in the time domain. The proposed model is applied to the prediction of satellite telemetry data, that is composed of readings from multiple sensors which are highly correlated. The prediction of such telemetries can help the engineers to detect anomalies in the system, that, in the context of space missions, are particularly dangerous since they can compromise the entire mission if not handled properly. The results show that the proposed model outperforms the recurrent network without wavelet transform both in terms of accuracy and in the width of the forecast horizon.