Study Of Water Quality Parameters Detection And Prediction Methods For Aquaculture Based On Machine Learning

The quality of aquaculture water is directly related to the safety of aquaculture.Therefore,it has tremendous economic value to study the changing trend of aquaculture water quality parameters and develop new monitoring methods,to achieve accurate prediction of aquaculture water quality parameters,to have an early warning,to realize scientific water quality management,and to solve the outbreak of fish diseases caused by deterioration of water quality.In this thesis,aquaculture water quality parameters are taken as the research subject.The methods of water quality parameter prediction and modeling,parameter optimization of the model,and the prediction accuracy improvement algorithm of water quality peak area,the development of water quality parameter measurement equipment and application of water quality parameter prediction model are mainly studied.The primary work and results are as follows:(1)The optimization algorithms of superparametric γ and σ for LSSVR model are studied.IPSO-LSSVR model is established by improving the particle swarm optimization algorithm.The distribution estimation algorithm is studied and improved by chaotic mutation.The combination of gamma is optimized by this algorithm.A CMEDA-LSSVR prediction model is established.The two models are used to predict dissolved oxygen and pH respectively.The prediction performance of the two models is compared.The comparison of the two models shows that the standard error and the average absolute percentage error of the former model are 10.62% and 8.79% less than those of the latter model respectively,which proves that the IPSO-LSSVR model is more suitable for the prediction of dissolved oxygen.By Comparing the predicted results of the two models,the standard error and the average absolute percentage error of the former model are respectively 7.27% and 9.84% higher than the latter model,which proves that the CMEDA-LSSVR model is more suitable for predicting the pH value.(2)The problem of improving the prediction accuracy of LSSVR model in the peak area of water quality parameters is studied.The LSSVR model training sample distribution impact on the performance of LSSVR model is analyzed.Based on this,a LSSVR prediction model weighted by the distribution density of training samples and the expected output amplitude of the model is studied.The improved weighted IPSO-LSSVR model is applied to the prediction of dissolved oxygen.And the results show that the standard error and the average absolute percentage error reductions of the improved model are 14.95% and 26.35% respectively.This proves that the improved algorithm of the standard IPSO-LSSVR model can significantly improve the prediction accuracy of dissolved oxygen in aquaculture water.The improved CMEDA-LSSVR model was used to predict the pH value of crab culture water.The results showed that the standard error and the average absolute percentage error reductions of the improved model are 17.23% and 22.19%,respectively.It is proved that the improved algorithm of standard CMEDA-LSSVR model can significantly improve the pH value prediction accuracy of the crab culture water.(3)The construction method of multi-core function of LSSVR model is studied.All input vectors of learning sample set are grouped according to homologous features.All input samples of homologous features are mapped to high-dimensional feature space using the same non-linear mapping function for regression fitting,and a grouped multi-core LSVR prediction model is established.The experimental results show that the standard error and the average absolute percentage error of the model for predicting dissolved oxygen are reduced by 14.9% and 23.6% respectively compared with those using single radial basis nucleus.The model was used to predict the pH value of crab culture water.The results show that the standard error and the average absolute percentage error of the model are 15.5% and 22.8% less than those of the single radial basis nucleus model.It proves that the grouped multi-core LSSVR model can significantly improve the dissolved oxygen and pH value prediction accuracy in aquaculture water.(4)A cruise water quality parameter measurement platform is developed to achieve the acquisition of large area water quality information in real time at low cost.The distribution principles of main water quality parameters is studied by using this measuring platform,and the correlation between water quality parameters and meteorological factors is studied,which provides guidance for the selection of input parameters for the establishment of water quality parameter prediction model.According to the weather forecast,a cloudy day is divided into five grades,which verifies the impacts of correcting the model according to different grades of the cloudy day.The results show that this method can effectively improve the dissolved oxygen prediction accuracy of the model in cloudy weather.Finally,the energy-saving effect of water quality prediction model applied to water quality early warning and wireless sensor nodes is verified.The results show that the average lifetime of wireless sensor nodes in the aquaculture IoT system is around twice as much as that of the direct data transmission,which greatly improves the timeliness of wireless sensor nodes,and achieve cost reducion and efficiency improvement.