Abstract: In recent years, water quality prediction has become a hotspot in the field of water environment management. However, the complexity and dynamic nature of the water environment itself lead to low prediction accuracy and poor model stability during water quality prediction. To address these issues, a new water quality prediction model were proposed based on Optimality Variational Mode Decomposition(OVMD), Temporal Convolutional Network(TCN), and Autoregression(AR). First, OVMD was used to decompose the original data to obtain several sub-sequences. Then, the decomposed sub-sequences were used as inputs for TCN and AR models for water quality prediction, and the prediction results of the two models were stacked and reconstructed to obtain the final prediction result.Finally, the total phosphorus data from Longhua Creek monitoring station was used for experimental verification. The results showed that the OVMD-TCN-AR water quality prediction model significantly outperforms Long Short Term Memory networks(LSTM) and Longand Short-term Time-series network(LSTNet). The average absolute error of the OVMD-TCN-AR water quality prediction model was 0.00660, the root mean square error was 0.01166, the MAPE was 0.0494, and the fitting degree was 0.97, indicating that the OVMD-TCNAR water quality prediction model had high reliability and application value.