An approach of improved dynamic deep belief nets modeling for algae bloom prediction

Algae bloom outbreak is a dynamic nonlinear process with time-varying characteristics and it is difficult for existing algal bloom prediction method to consider the complex characteristics, which leads to low accuracy prediction. For the problem, a dynamic deep belief nets model that combines time series analysis with deep learning methods is proposed by analyzing algal bloom outbreak mechanism. The model introduces historical moment in input layer, increases connection between input layer and hidden layer, uses contrastive divergence algorithm to introduce historical moment in hidden layer and weight and bias algorithms are given timing characteristic in pre-training stage. At the same time, the model adopts dynamic learning rate to complete pre-training and the back-propagation algorithm is used to fine tune network parameters to complete the whole model training. The instance validation results show that the method can more accurately describe dynamic nonlinear process than other prediction methods and further improve prediction accuracy.