Part 5. Artificial neural networks and machine intelligence

Combined application of recurrent neural networks and statistical methods for improved oceanographic data forecasting accuracy

V. Yu. Kuz'min

Lomonosov Moscow State University, Faculty of Space Research

Abstract: This paper is devoted to the application of statistical models to increase the prediction accuracy of oceanological data. The initial time series are modelled with mixtures of finite normal distributions. Statistical characteristics of constructed mixtures are used to pre- initialize the layers of a recurrent neural network. Forecasts made with the application of statistical models are compared with forecasts made using only the original data. It is demonstrated that a significant improvement in accuracy is observed for all analyzed series.

Keywords: EM, MSM, LSTM, neural networks, machine learning, feature enrichment, finite normal mixtures.