Modeling The Dynamics Of Dissolved Oxygen In Aquaculture

Dissolved oxygen(DO)is an important parameter in the process of water quality change,reflecting the self-purification ability of water bodies,because most processes of water quality purification require the participation of oxygen.At the same time,DO as the main demand for lives in water,plays an important role in aquaculture production,and suitable DO concentration range is beneficial to the growth of aquatic products.In order to promote the development of aquaculture industry,it is necessary to fully understand the regularity of DO dynamics.In this thesis,modeling methods based on data driven and mechanism-driven were studied in the modeling of the dynamics of DO in fishery ponds.1.A short-term prediction model based on Echo State Networks(ESN)was proposed for DO.In view of the lack of historical input information memory ability and slow training speed of traditional modeling methods such as Support Vector Machine(SVM)and Feed-Forward Neural Networks,ESN is introduced into DO prediction modeling.The two-way construction of training samples and ESN model integration strategies are proposed to solve the problem of generalization performance deterioration of the traditional ESN model in the case of network freedom parameter stoking and in the case of large reserve pool.The test results show that the optimized ESN has good parameter robustness,and in the case of large-scale reserve pool,it can effectively reduce the overfitting phenomenon of traditional ESN and guarantee or improve the generalization performance of the model.And the optimized ESN improved the test evaluation(by Mean Absolute Error and Root Mean Square Error)by 3% and 1.7% compared to the Least Square-Support Vector Machine(LSSVM),and 2.84% and 2.25% higher than the traditional ESN.The model of predicting short-term DO,which needs to update quickly with the single monitoring parameter in the complex water environment,can be realized by the optimized ESN.2.A medium-term predictive modeling method of DO based on the multivariate input and Gate Recurrent Unit(GRU)was studied.This method considers the effect of environmental information(water temperature,pH,turbidity and ammonia nitrogen concentration)on DO changes,combines environmental information with DO as a model input,and constructs a medium-term prediction model with GRU and Long Short-Term Memory(LSTM)that have good memory ability for historical input data.Both LSTM and GRU have obtained good prediction accuracy for the prediction of DO concentration after about 2 hours.The four evaluation indicator values of the GRU(Mean Absolute Error,Root Mean Square Error,Mean Absolute Percentage Error and Coefficient of Determination)were 0.450mg/L,0.641mg/L,5.4 mg/L,5.4% and 0.994 respectively,and LSTM achieved accuracy of 0.407mg/L,0.542mg/L,5.9% and 0.970 respectively.Although GRU predictive performance is similar to LSTM,GRU has better performance than LSTM due to its less time cost and fewer number of parameters than LSTM.This method is useful to solve the large time lag problem in artificial oxygenation.3.A mechanism-based method of DO modeling in cultured waters was studied.This method analyzes the key ingredients of the dynamic change process of DO,covers mathematical models including photosynthesis,atmospheric reoxygenation,artificial oxygenation,respiratory action and sediment oxygen consumption,and uses the Runge-Kutta method and the Levenberg-Marquardt algorithm to compute and adjust the model parameters.Using the data of the actual sampling,the thesis verifies the accuracy of the model,and shows that the Mean Absolute Error is about 0.447,the Root Mean Square Error is about 0.552,and the Mean Absolute Percentage Error is about 6.4%.The result proves that the mechanism-based model can describe the dynamics process of DO and explain the influence of water environmental factors on DO dynamics change.