Lake water level changes show randomness and the complexity of basin hydrological simulation and lake water level response. We constructed a vine copula model to simulate and predict lake water level that incorporated rolling decisions and real-time correction of prediction results. The model was applied to predict the long- and short-term water levels in Erhai Lake on the Yun-gui Plateau, southwest China. The results showed that (1) the predicted daily water levels (with ME=0.02~0.09, RMSE=0.02~0.024, NSE=0.99, and IA=0.99) were more accurate than the predicted monthly water levels (with the ME=0.039~0.444, RMSE=0.194~0.279, NSE=0.913~0.958, and IA=0.977~0.989), and the accuracy of the predictions improved as the number of variables increased. (2) The vine copula model outperformed the back-propagation neural network and support vector regression models, and, of the three model types, gave the best estimate of the nonlinear relationships between the predicted water level and climatic factors, especially in the wet season (May to October). (3) The prediction accuracy of the vine copula model was lower for small sample sizes and when there was a lack of runoff data. By improving the analysis of the model’s errors, the percentages of the relative errors of the prediction accuracy less than 5%, 10%, 15%, and 20% increased to 70%, 83%, 95%, and 98%, respectively.