Research On Ocean Environment Data Prediction Based On Deep Learning

Ocean provides important resources for national life.It has become a significant part to support the sustainable development of national economy and society.However,with unlimited exploitation and utilization of ocean resources,ocean environment in China is deteriorated gradually.Accurate and efficient ocean environment data prediction plays an important role in preventing ocean pollution.It is conductive to the modernization of ocean governance and the construction of powerful ocean country.Recently,the development of deep learning provides technical support for data prediction.Given this,taking the ecological environment of offshore sea in China as the research object,the prediction methods for ocean environment data obtained by four-dimensional monitoring system are studied based on deep learning theory.The major contributions of this paper are as follows:Firstly,for outliers and noises in ocean environment data,a smoothed deep belief echostate network(SDBEN)is proposed for single factor prediction in ocean scenario.This model consists of three functional components: smooth preprocessing,feature extraction and nonlinear regression.The smoothing methods are used to weaken the influence of outliers on data prediction and reduce potential noise,and the predictor based on the deep belief echo-state network(DBEN)is used to extract the feature and achieve nonlinear regression.Experimental results show that compared with the classical machine learning models,our model has a smaller normalized root mean squared error(NRMSE),i.e.,better nonlinear approximation performance.Secondly,activation function(AF)has an important effect on the nonlinear approximation performance of deep prediction model.Actually,the choice of AF is taskrelated.For this problem,a self-adaptive selection DBEN(SADBEN)is proposed for oceanrelated multi-factor time series prediction.In this model,the optimal AF can be determined through a self-adaptive activation mechanism for a given ocean environment data prediction task.Experimental results show that the NRMSE of our model is less than 0.06.Hence,SADBEN can be applied to the ocean environment multi-factor prediction task effectively.Finally,for the problem that online prediction based on incremental learning affects the training efficiency,an online ocean environment data prediction based on equivalent-sample selection is proposed.When new samples are coming,equivalent-sample selection algorithm eliminates the farthest samples.In this way,the pressure brought to the model by incremental samples can be reduced.It can ensure that the proposed model can perform online prediction continuously and effectively.Experimental results show that our model can meet the requirements of both training efficiency and prediction accuracy in our ocean-related online prediction tasks.Figure 30;Table 12;Reference 56...