Methodology And Application Research On Surface Water Quality Prediction Using Neural Network And Transfer Learning

Sustainable management of surface water resources under changing climatic conditions requires trend analysis and prediction of surface water quality.In recent years,relying on the accumulation of surface water monitoring data and the development of computing power and algorithm,the neural network(NN)-based surface water quality prediction technology has become a hot research topic.However,the practical application of the technology still faces the following challenges,which severely restricted the reliability of the prediction results: i)it is common to determine the neural network hyperparameters manually through trial and error in the field of water environment,which may introduce much randomness and require a considerable computational time,limiting the model prediction performance;ii)due to the lack of sufficient historical information as a support,it is difficult to utilize the neural network for surface water quality at data-deficient sites.iii)due to the heterogeneous distribution of surface water quality data in different time periods,it is difficult for the neural network model to maintain high prediction accuracy in long-term prediction.This study aims to seek solutions to the above challenges based on neural network hyperparameter optimization(HPO)methods and transfer learning(TL)theory.Specifically,first,an overall development approach of neural network models for surface water quality prediction was proposed,focusing on the neural network hyperparameter optimization problem to ensure the training efficiency and prediction performance.On this basis,from the perspective of practical applications,the study systematically explored how to i)improve the performance of neural network models in surface water quality prediction tasks at data-deficient sites based on TL,and ii)improve the performance of neural network models in long-term surface water quality prediction tasks.The thesis selected five case sites for experimental studies to verify the feasibility and effectiveness of the methodology framework.Finally,based on the methodology framework,a surface water quality prediction Web system was developed.The system architecture and the interaction logic of each module were designed according to the functional requirements of the system,then each module was developed and implemented separately.The main research results and conclusions of the thesis are as follows:(1)Five representative automatic hyperparameter optimization methods were implemented,including Grid Sampling(GS),Random Search(RS),Genetic Algorithm(GA),Bayesian optimization(BO)algorithms based on Gaussian Process(GP),and Tree Parzen Estimator(TPE).For evaluating the methods,this study proposed an approach that the “optimal hyperparameter value sets” achieved by the GS were regarded as the benchmarks,then the other HPO methods were evaluated and compared from three dimensions,i)convergence,ii)optimization orientation,and iii)consistency of the optimized values.The case study results indicate that the TPE-based BO is the recommended HPO method for the surface water quality prediction NN models in this study based on its satisfied performance of i)fast and stable convergence,ii)reasonable and efficient optimization orientation,iii)high consistency rates of the optimized hyperparameter values with the benchmarks,and iv)slight randomness of the HPO process.(2)The TL-based approach of NN model development for data-deficient sites was proposed.The case study results indicate that i)TL can significantly improve neural network models’ prediction performance with low computing cost in data-deficient sites when the source domain and TL hyperparameters are rightly selected;ii)the PRMSER and DISTANCE proposed in the study are two effective similarity measurement indexes(SMIs)for the source domain selection;iii)TL hyperparameters have a significant impact on TL performance,so the TL hyperparameters selection using a validation set from the target domain is recommended.(3)The TL-based approach to improve the long-term prediction performance of NN models was proposed.The case study results indicate that i)TL can significantly improve neural network models’ performance with low computing cost for long-term surface water quality prediction when the TL hyperparameters are rightly selected;ii)TL hyperparameters have a significant impact on TL performance,so the TL hyperparameters selection is recommended.(4)Based on the aforementioned surface water quality prediction methodology framework,a surface water quality prediction Web system was designed and developed.Through the six system modules,the following functions were realized: i)data collection,ii)data query and analysis,iii)NN model training(including HPO),iv)NN model recall and prediction,v)cold start of NN for a new site,and vi)fine-tuning of NN model in long-term prediction.The system can effectively reduce the technical difficulties of the relevant researchers to share the research results.Under the current background of smart water construction,this study points out the problems and challenges faced by the practical application of the neural network in surface water quality prediction tasks,discusses the solutions in light of the theory of HPO and TL,and realizes the instantiation application of the methodology through case studies and system development.The study can promote further research and application of neural network model in surface water quality prediction,and provide powerful help for water environment planning and management.