Abstract: Towns and cities require wastewater treatment plants to ensure the ecological environment, human health, and sustainable development. It is necessary to predict the future values of the effluent quality indicators of wastewater plants in order to improve the efficiency of urban wastewater treatment plants and optimize energy use. However, due to large number of effluent quality indexes of wastewater plants, the extremely complex reactions involved in the wastewater treatment process, and the high degree of time-series nonlinearity of the data, the mechanistic model based on determining the biochemical reactions and the statistical-based prediction method cannot achieve the desired results. As a solution to this issue, this paper proposes a deep learning-based prediction method for wastewater plant effluent water quality, and a town in Jilin Province has used a CASS process wastewater plant to monitor water quality data at the influent and effluent as the source data, as well as the neural networks combined with the encoder-decoder structure for predicting wastewater plant effluent water quality. The model is also applied to complex environments by adding data on local environmental factors（temperature and precipitation） that affect wastewater treatment. The historical（20 steps） data as the model input to predict the future water quality data. At the same time, to better compare the model performance, this paper divides experiments into short-term prediction（single step） and long-term prediction（20 steps） of two different time dimensions for prediction. Experimental results show that the proposed structure improves the prediction ability of both LSTM and GRU networks, especially the improvement of long-term prediction ability. Based on the ED-GRU model, which has the best prediction effect, the root mean square error（RMSE） of the four effluent water quality indexes of COD, NH3-N, TP and TN in short-term prediction are 0.755 1, 0.219 7 and 0.073 4, 0.314 6, respectively. The goodness of fit（R2） is 0.901 3, 0.933 2, 0.916 7, 0.953 2, which can be used to make a better prediction of the local water quality trends. The RMSE of the four indicators in the long-term prediction is 1.720 4, 1.768 9, 0.447 8, 0.831 6, and the R2 is 0.484 9,0.550 7,0.450 2,0.759 5, which can predict the overall trend of future changes in effluent water quality. Compared with the sequential structure GRU, it is predicted that the short-term RMSE will decrease by more than 10%, R2 to increase by more than 2%, and the long-term RMSE will decrease by over 25% and the R2 will increase by more than 15%. Based on the results of this paper, the neural network based on the encoder-decoder structure can make accurate predictions of the effluent quality of wastewater plants and provide technical support for the next step of wastewater treatment process improvement.