Abstract: Dissolved oxygen is one of the essential indexes reflecting the degree of water pollution, and thus, it can accurately, efficiently, and reliably predict the state of water quality. Obtaining efficient and accurate prediction models in the aquatic system is difficult because of the real-time variation and complexity of the water quality environment and the biases in data collection. Therefore, principal component analysis is first used to determine the number of variables that affect dissolved oxygen in water quality, where the data's dimension is reduced. To solve the problem of nonlinearity and non-stationarity among the variables, mutual information is proposed to select the factors that strongly influence this relationship as the input variables of the prediction model. Next, a nonlinear decreasing weight particle swarm optimization algorithm with Gaussian function is used to optimize the parameter selection process of a support vector regression(GNIPSO-SVR) model, which effectively overcomes the parameter selection limitations of the traditional SVR prediction model.To consider the effect of air pollutants, a GNIPSO-SVR model is constructed. This model is then employed to predict dissolved oxygen in the water quality of Shanghai. Comparing and analyzing the GNIPSO-SVR model results with those obtained from the BP neural network, the SVR, and PSO-SVR models, the proposed model effectively eliminated the redundancy and correlation problems among dissolved oxygen variables and improved the prediction accuracy and operating speed.