Abstract: Ammonia-nitrogen is one of the key parameters to evaluate aquaculture water quality and aquatic ecosystem security. When the ammonia-nitrogen content exceeds 0.2 mg/L, it will cause serious harm to the health of aquatic organisms, and even cause the disease and death of aquaculture objects. However, the pond culture environment is complex and variable. Many factors have affected ammonia-nitrogen concentration, such as dissolved oxygen, nitrite, salinity, blue-green algae, turbidity, pH value, oxidation reduction potential and so on. And these factors are related to each other and exhibit nonlinear changes. In order to master rules of the changes of ammonia nitrogen, a lot of information systems and models were developed to monitor the pond water. But the data acquired by internet of things for aquaculture has a lot of noise due to environmental weather factors and the sensor or network failure. Traditional mechanism methods are different from statistical model, and mainly depend on small amount of data. So traditional mechanism methods are difficult to achieve the desired accuracy using the time sequence data online, and can’t be used in the real time system because of the large amount of data. Some scholars have also carried out lots of research on statistical models such as traditional neural networks in the field of forecasting. But it is difficult to accurately predict the ammonia-nitrogen content because of the model’s local convergence and poor computational efficiency. In this paper, a method for predicting ammonia-nitrogen in pond aquaculture water based on multi-variate deep belief network model optimized by particle swarm optimization (PSO-MDBN) was proposed. The mothed constructs a non-linear deep-seated network structure through self-learning, and has strong data prediction ability and data classification and recognition ability. Firstly, because the ammonia-nitrogen content in pond culture water was affected by various environmental factors, the ammonia-nitrogen ecosystem of pond culture water was analyzed, and the data collection scheme of pond culture environment information was formulated according to the actual conditions of pond equipment and modeling needs. In this part, principal component analysis was used to screen the key influence factors of ammonia nitrogen change as auxiliary variables of the prediction model. Secondly, the wavelet threshold denoising method was used to reduce the data of the selected auxiliary variable data, and the data dimension and improve the reliability of the data. Thirdly, based on the deep belief network theory, a multi-parameter deep belief network model was constructed, and the optimal network structure was determined by experimental methods. Finally, the particle swarm optimization algorithm was used to optimize the deep belief network parameter setting to get the optimal combination of parameters. The experimental results showed that the average percentage error of the proposed method was 0.117 2. Compared with the traditional LSSVM(least squares support vector machine), BPNN (back propagation neural network) and DBNs(deep belief networks) methods, the RMSE of PSO-MDBN is reduced by 0.285 9, 0.214 6 and 0.013 9. In addition, if the number of samples increases, the model error would show a downward trend. Above all, the proposed method of this paper has high prediction accuracy and generalization performance. The research can provide data theoretical basis and parameter support for accurate prediction of ammonia nitrogen content in pond water.