In order to guide the intensive aquaculture effectively and improve the accuracy and stability of water temperature prediction, based on the analysis of water temperature factors, a prediction model (EMD-IGA-SELM) was proposed with the combination of empirical mode decomposition (EMD), improved genetic algorithm (IGA) and improved extreme learning machine (SELM). Firstly, the outlier and missing data were corrected with the calculation of composite meteorological index. Secondly, the Pearson correlation was utilized to explore the relationships between affecting factors and water temperature, and construct the input and output of prediction model. Then, Softplus function was used as activation function of SELM to replace Sigmoid. The best weight and threshold of SELM were obtained from the IGA, which introduced the chaotic sequence to traditional GA. Finally, EMD algorithm was applied to decompose the original water temperature time series into a series of intrinsic mode function (IMF). IGA-SELM prediction models were trained in each IMF sequence, and the predicted values were calculated by the sum of predicted value in each IMF sequence. The experimental results showed that EMD-IGA-SELM had better prediction accuracy, and the mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) of GA-SELM were 0.1233℃, 0.0043 and 0.1478℃, respectively. Research results met the practical needs of the aquaculture and provided decision support for water quality management and control.