Abstract: Due to complex characteristics and uneven correlation of water quality data, it is difficult to predict dissolved oxygen concentration.To improve the prediction accuracy of water quality dissolved oxygen concentration, a Feature Engineering and Northern Goshawk Optimization-Long Short Term Memory(FE-NGO-LSTM) hybrid model was proposed.Firstly, missing value imputation, feature screening, and feature polynomial construction were performed on the water quality dataset.Then, the model parameters were optimized based on the NGO-LSTM model to improve prediction performance.After analyzing the feature prediction performance under different polynomial orders, the model was compared with LSTM models based on grey wolf optimization algorithm, whale optimization algorithm, and particle swarm optimization algorithm.Finally, the model was validated with the dataset of the Chengnan monitoring section on east Tiaoxi River, and the prediction results of the FE-NGO-LSTM model were calculated for prediction periods of4,8,12,16,20,and24hours.The experimental results demonstrated that when the polynomial order was2nd, the model had the best prediction performance.Compared with LSTM models based on other optimization algorithms, the average absolute error, mean square error, and root mean square error of FE-NGO-LSTM model were reduced at least9.0%,12.9%,and6.3% respectively.Moreover, as the prediction period increased, the prediction error was still within an acceptable range, indicating that the FE-NGO-LSTM model has certain advantages and generalization in predicting dissolved oxygen concentration.