| In the process of aquaculture,water quality is the most important factor that affects the profit of aquaculture.With the continuous development of aquaculture technology,aquaculture water quality environment has become a focus of research in various countries.As one of the important parameters to measure water quality,ammonia nitrogen concentration is very important in the monitoring of aquaculture water quality.The high concentration of ammonia nitrogen will have a bad impact on the living environment of aquatic organisms,and damage the surrounding water quality and ecology.Therefore,real-time monitoring of ammonia nitrogen concentration in water quality is particularly important.The prediction of water quality data can solve the defects of existing detection methods.The prediction method can be used to analyze and model the data of ammonia nitrogen concentration in aquaculture water,establish the relationship between ammonia nitrogen concentration and other factors,and obtain the ammonia nitrogen concentration at the current moment,effectively solving a series of aquaculture problems caused by the failure of traditional detection methods to obtain ammonia nitrogen concentration in time.With the development of information technology,a large amount of data can be obtained.However,due to the relatively few related sensors in the breeding process,the digitalization degree is low,resulting in insufficient data.In the case of less data labels,the traditional prediction model will have difficulties in gradient updating,slow convergence speed and low prediction accuracy.This thesis puts forward the following two methods to solve the above problems:1.In order to improve the precision of small sample modeling,a prediction Model of ammonia nitrogen concentration based on meta-learning and Long Short-Term Memory(LSTM)was proposed.Model-agnostic Meta Learning algorithm was used.MAML)divides the model into LSTM base learner(base-learner)and MAML mate-learner,learning the complex nonlinear change process of ammonia nitrogen concentration,improving the adaptive ability of MAMLLSTM model,improving the generalization ability of the model,and realizing the prediction of small sample data set.The experimental results show that the MAML-LSTM model is effective for modeling small sample data sets,and can achieve convergence through fast training.The model error is reduced by 11.9% compared with the LSTM model,which can realize the prediction of ammonia nitrogen concentration in aquaculture,and provide a certain reference for the subsequent water quality control.2.In order to obtain the optimal initialization parameters for LSTM based learner in the meta learning training stage,an improved Bayesian optimization algorithm based on electron cloud is proposed.By adding kernel function,the probability of Bayesian optimization is increased according to the three-dimensional motion view of electron cloud,and the optimal hyperparameters are screened by calculating the optimal solution of matrix during output.This method can effectively reduce the error caused by artificial parameter adjustment and improve the training accuracy of the model.The improved Bayes algorithm is used to optimize the LSTM based learner,so that it has good initial parameters in the Mam L-LSTm-based ammonia concentration prediction model.The results show that the improved Bayesian optimization method can improve the accuracy of the prediction model,and the model error is reduced by 4.5%,which verifies the feasibility of the improved Bayesian optimization method. |
