Jonas Ströjby

Modelling non-linear stochastic dynamic systems with recurrent neural networks and
non-linear filters with applications in energy trading

Abstract:

In this Masters project we study time series modeling and forecasting by using
non-linear Kalman filters and recurrent neural networks. Neural networks are
universal approximators and the inclusion of internal recurrence allow for
temporal structures and memory. They are used to approximate the dynamics of
the studied system. The parameters of the neural network are estimated using
non-linear Kalman filters. The noise on both output and input is studied and
handled. A method to handle noisy inputs based on the unscented Kalman filter
is presented. Also, an approximative method to estimate conditional variance
is suggested. We present a couple of synthetic examples intended to show
various properties of the method and also show it's performance on two real
applications relevant in energy trading.