Dynamical recurrent neural networks--towards environmental time series prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Dynamical Recurrent Neural Networks (DRNN) model synapses as autoregressive filters.
• DRNNs approximate time series laws using nonlinear difference equations, enabling history-sensitive forecasts without external memory.
• The training method is temporal-recurrent-backpropagation, efficient due to exponential gradient decay.

Purpose of the Study:

• To assess the predictive ability of DRNN models for meteorological and astronomical time series.
• To enable preset telescope instrumental modes hours in advance using reliable environmental forecasts.
• To compare DRNN performance against traditional methods for complex time series prediction.

Main Methods:

• DRNN model applied to meteorological and astronomical time series.
• Comparison with nonlinear Autoregressive (AR) and Autoregressive Moving Average (ARMA) models using feedforward networks.
• Fuzzy coding and fuzzy correspondence analysis for astronomical seeing prediction.
• Nonlinear multiple regression (nowcasting) on fuzzily coded seeing records.

Main Results:

• DRNNs provide history-sensitive forecasts by approximating time series dynamics.
• The DRNN model demonstrated superior performance in predicting astronomical seeing compared to the fuzzy k-nearest neighbors method.
• Efficiency of temporal-recurrent-backpropagation is enhanced by exponential decay of backpropagated gradients.

Conclusions:

• DRNNs are effective for time series forecasting, particularly for complex and erratic data like astronomical seeing.
• The model's ability to learn internal dynamics allows for accurate predictions without explicit external memory.
• DRNNs show promise for applications requiring reliable environmental forecasts, such as optimizing astronomical telescope operations.