In view of the poor prediction performance of conventional statistical models of concrete crack opening due to the insufficient consideration of temperature nonlinear factors and residual errors, an improved statistical model for the prediction of concrete crack opening was established incorporated with the temperature nonlinear factors. Then the residual error time series of the model was analyzed using chaos theory and its phase-space was reconstructed, meanwhile the residual error was predicted using BP artificial neural network optimized by genetic algorithm. Finally, an improved hybrid prediction model was established to predict concrete crack opening. Based on the monitored crack opening data of a drainage gate reservoir, the prediction results of the conventional statistical model, the improved statistical model and the improved hybrid prediction model were compared and analyzed. The result shows that the improved hybrid prediction model has smaller prediction error and can effectively improve the prediction accuracy.