To address the issues of low prediction accuracy and slow convergence rate of ammonia nitrogen concentration model under small sample data， a method of establishing ammonia nitrogen concentration prediction model by using long short-term memory （LSTM） and optimizing the model by using Bayesian optimization algorithm and Meta-learning mechanism was proposed. The Bayesian optimization algorithm was used to optimize the superparameters of the prediction model， and the initial values of the model parameters were given. Then the Meta-LSTM algorithm was used to learn the model gradient and allow the parameter sharing and updating among the optimizers， and finally the optimization of the prediction model of ammonia nitrogen concentration was realized. Compared with LSTM， GRU and RNN models， the result shows that the root-mean-square error， mean absolute error and mean square error of the proposed model are 0.027 6， 0.023 9 and 0.000 76， respectively， which are better than other prediction models. It is further indicated that the prediction model of ammonia nitrogen concentration based on Bayesian and Meta-learning is effective for modeling small sample data， and can achieve convergence in rapid training. The accuracy of the model meets the prediction requirements of ammonia nitrogen concentration for aquaculture.