A dynamic relearning neural network model for time series analysis of online marine data

In order to effectively predict the time series marine data obtained from online monitoring, contribution hypothesis and evolution hypothesis are proposed. A mathematical model describing the relationship between the number of samples and the weight of relearning is developed on the basis of the two hypotheses. Conventional neural networks with static parameters are modified to have dynamic parameters, which can “evolve” repeatedly in the process of online monitoring. The algorithm flow of dynamic relearning neural networks is established, which consists of two phases, named sample training phase and dynamic computation phase. In the first phase, proper number of samples and learning weight are obtained; in the second phase, dynamic computations are carried out with conditional relearning. A linear neural network is chosen and current velocity data is selected for experiment, in all of the three chronologically selected groups, the relearning neural network outperforms the conventional static neural networks, and the mean absolute errors (MAE) of the three groups are respectively reduced 3.40 percent, 6.67 percent and 7.93 percent. Experiment results show that MAE will be reduced more and more as time goes on, which verify the contribution hypothesis and evolution hypothesis. Focusing on improving the work flow of neural networks, the proposed method could be widely applied to various types of other geographic data as well as marine monitoring data.