Renmin Zhujiang (Jan 2023)
Gaussian Process Regression Total Nitrogen Prediction Based on Data Decomposition Technology and Several Intelligent Algorithms
Abstract
Total nitrogen (TN) is one of the important indicators to reflect the degree of water pollution and measure the eutrophication status of lakes and reservoirs.To improve the accuracy of TN prediction,based on the empirical wavelet transform (EWT) and wavelet packet transform (WPT) decomposition technology,this paper proposes a Gaussian process regression (GPR) prediction model optimized by osprey optimization algorithm (OOA),rime optimization algorithm (ROA),bald eagle search (BES) and black widow optimization algorithm (BWOA) respectively.Firstly,the TN time series is decomposed into several more regular subsequence components by EWT and WPT respectively.Then,the paper briefly introduces the principles of OOA,ROA,BES,and BWOA algorithms and applies OOA,ROA,BES,and BWOA to optimize GPR hyperparameters.Finally,EWT-OOA-GPR,EWT-ROA-GPR,EWT-BES-GPR,EWT-BWOA-GPR,WPT-OOA-GPR,WPT-ROA-GPR,WPT-BES-GPR,WPT-BWOA-GPR models (EWT-OOA-GPR and other eight models for short) are established to predict the components of TN by the optimized super-parameters.The final prediction results are obtained after reconstruction,and WT-OOA-GPR,WT-ROA-GPR,WT-BES-GPR and WT-BWOA-GPR models based on wavelet transform (WT) are built.Eight models,including EWT-OOA-SVM based on support vector machine (SVM),the paper compares the unoptimized EWT-GPR,WPT-GPR models,and the uncomposed OOA-GPR,ROA-GPR,BES-GPR,and BWOA-GPR models.The models were verified by the monitoring TN concentration time series data of Mudihe Reservoir,an important drinking water source in China,from 2008 to 2022.The results are as follows.① The average absolute percentage error of eight models such as EWT-OOA-GPR for TN prediction is between 0.161% and 0.219%,and the coefficient of determination is 0.999 9,which is superior to other comparison models,with higher prediction accuracy and better generalization ability.② EWT takes into account the advantages of WT and EMD.WPT can decompose low-frequency and high-frequency signals at the same time.Both of them can decompose TN time series data into more regular modal components,significantly improving the accuracy of model prediction,and the decomposition effect is better than that of the WT method.③ OOA,ROA,BES,and BWOA can effectively optimize GPR hyperparameters and improve GPR prediction performance.
