In order to improve the speed and accuracy of water level prediction, a prediction model based on improved echo state network (ESN) was proposed. Xavier method was introduced to optimize the weights to adapt to the water level prediction task. At the same time, the concept drift detection method (EDDM) was introduced to adapt to the actual water level environment, monitor the change of water level data distribution, and trigger the corresponding model update or adaptation strategy when the concept drift was detected, so as to improve the prediction effect of the real water level. The experimental results of data from nine water level stations show that compared to traditional sequence prediction models (SVM, RNN, GRU, LSTM, ESN, and XESN(Xavier-ESN)), the proposed model shows higher prediction accuracy for each water level station in terms of overall prediction performance and short -term, medium - term and long-term predictions, further improving the prediction accuracy of  inland waterway water level.