Abstract: Dissolved oxygen concentration is a key indicator in the evaluation of lake ecological health. Due to the unique hydrology of shallow dish-shaped lakes, its dissolved oxygen (DO) has obvious instability and nonlinear characteristics. To accurately predict the DO concentration in the dish-shaped lake, we proposed a prediction model for DO concentrations combining long and short-term memory neural network (LSTM) with principal component analysis (PCA) and maximum information coefficient (MIC) based on the typical monitoring data set of the dish-shaped lake in Poyang Lake. The results showed that compared with the support vector regression (SVR) and LSTM models, the prediction accuracy of the PCA-MIC-LSTM model was significantly improved, with a determination coefficient of over 0.99, a root mean square error of 0.039 mg/L, and an average absolute error rate of 0.301%. Among them, the PCA noise reduction treatment affected the LSTM model prediction effect more than the MIC feature extraction, and can significantly reduce the error rate. The PCA-MIC-LSTM model in this study can provide a reference for the protection of water body in dish-shaped lakes.