A Bayesian Regression Prediction Model For Marine Chemical Data

Marine chemical data has the characteristics of disorder and nonlinearity.In order to make more accurate predictions of marine chemical data,this thesis proposes a prediction model for marine chemical data based on Bayesian theory.Since the numerical dispersion of marine chemical data is large,using the raw data directly as the input of the model will lead to too little training for each numerical sample,which in turn affects the prediction accuracy of the model.In this thesis,the differential series dataset of marine chemical data is used as the input samples of the model to reduce the number of training set categories and improve the training accuracy.Using the sliding window method from the network communication protocol,the difference series dataset is divided to generate the features and labels of the training set.Based on Bayesian theory,a naive Bayesian model and a semi-naive Bayesian model are proposed for marine chemical data prediction.The Bayesian model suffers from the problem of lacking a hyperparameter setting algorithm.To solve this problem,an optimized and improved algorithm was proposed in this thesis by combining linear regression theory.According to the characteristics of different marine ranch chemical data,the adaptive setting of model hyperparameters is realized,which further improves the prediction accuracy of the model.The specific studies in this thesis are as follows.(1)Naive Bayesian-based prediction model for marine chemical data.In order to make accurate prediction of marine chemical data,the values to be predicted are used as categories based on the assumption of naive Bayes.And the values corresponding to the maximum of posterior probabilities of all training samples in the training set are selected as the predicted values to achieve the accurate prediction of dissolved oxygen parameters in marine chemical data.Finally,the model was used to train and predict the dissolved oxygen data of shellfish culture marine ranches in Yantai,Shandong Province from February 18,2016 to January 31,2020,and the effect of dissolved oxygen prediction was analyzed based on different feature lengths to determine the optimal values of model parameters.Comparing the naive Bayesian model with the advanced algorithms such as multilayer perceptron regressor,the comparative experimental results show that the naive Bayesian algorithm can make good predictions of dissolved oxygen data from marine ranches.(2)A semi-naive Bayesian-based model for predicting marine chemical data.The Bayesian model is based on the ideal assumption of "conditional independence of attributes",which needs to be adjusted because of the interaction between different marine chemical data in the actual marine water environment.A semi-naive Bayesian model for predicting marine chemical data is proposed based on the analysis of marine chemical data of the northeastern China marine ranches in the past three years.Based on the semi-naive Bayesian theory,the first-order difference series of the original marine chemical data were used as categories,and the frequency of occurrence of different categories was counted and the first-order difference series of each category corresponding to the previous time interval were recorded.For the sequence of marine chemical data to be predicted,the category that makes the greatest posterior probability at the next moment is selected as the prediction value,so as to achieve the prediction of marine chemical data.By comparing with the advanced models such as LSTM,RBF,ARIMAX and SVR,the semi-naive Bayesian-based marine chemical data prediction model provides an effective marine chemical data prediction method with higher accuracy.(3)Optimization algorithm of Bayesian marine chemical model based on linear regression.Bayesian models need to be set manually for two hyperparameters,feature length and frequency threshold,before establishment.And the value size of hyperparameter settings will directly affect the prediction accuracy of the model,and there is a lack of effective hyperparameter setting algorithms for Bayesian prediction models.In order to further improve the prediction accuracy of Bayesian model for marine chemical data,an optimization algorithm for hyperparameter setting in marine chemical data prediction model is designed based on linear regression theory,which can set the model parameters adaptively for the chemical parameter characteristics of different marine ranches.The results of the comparison experiments show that the prediction error is further reduced after model optimization,which can meet the practical requirements of real-time accurate prediction of marine chemical data.