相比于传统的渠道水动力学方程求解站前水位，机器学习模型因其响应快、稳态误差小等特点，被广泛应用于渠道水位预测。然而，输入数据庞杂、时序预测困难等原因直接限制和影响了预测精度，如何准确评价输入数据对预测结果的影响，是水位预测模型应用亟待解决的难题之一。以南水北调东线两湖段典型工程为研究对象，提出一种测试输入数据对预测结果影响分析的框架，构建基于长短时记忆网络（long short-term memory, LSTM）模型的泵站站前水位实时预测模型，通过对各输入因子的敏感性分析，实现预测模型数据输入集的优化，并基于该数据组合对不同预见期下的邓楼泵站站前水位进行预测分析。结果表明：以长沟站前水位、长沟泵站流量、邓楼站前水位指标作为输入参数的组合最优，平均绝对误差（mean absolute error, EMA）、均方误差（mean squared error, EMS）、均方根误差（root mean square error, ERMS）和平均绝对百分比误差（mean absolute percentage error, EMAP）4项指标均小于0.05，相关系数（r）和决定系数（R2）均大于0.95，剔除突变数据后最大水位误差不超过0.02 m。

The step pumping station system is a very complex system that consists of the pumping station, barrage and canal section facilities, for which accurate water level prediction is recommended for the safe operation of water transfer projects. Machine learning models are widely used for water level prediction because of their fast response and small overshoot as compared with the traditional hydrodynamic equations. However, the complicated input data and the difficulty of time series prediction directly limit and affect the prediction accuracy, and raises the question as how to accurately evaluate the influence of input data on the prediction results to be solve in water level prediction model. The establishment of machine learning model helps to accurately predict the water level in front of the pump, realize intelligent dispatching, improve the operation efficiency of the pumping station, and is of great significance to the construction of water network. Long Short-Term Memory (LSTM), as a special recurrent neural network structure, has both the nonlinear properties of neural networks, and memory and long-term dependence, which facilitates its processing of time series data. The Changgou pumping station to Denglou pumping station in the Nansi-Dongping Lake section of the South-to-North Water Transfer East Route is taken as the research object. A framework for analyzing the influence of test input data on prediction results is proposed to construct a real-time prediction model of water level in front of pumping station based on LSTM model, and the optimization of the data input set of the prediction model is realized through the sensitivity analysis of each input factor, while the prediction analysis of the water level in front of Denglou Pumping Station under different foresight periods is carried out based on the above data combination. Four input factors, namely water level in front of Changgou station, flow rate of Changgou pumping station, water level in front of Denglou station and flow rate of Denglou pumping station, were initially determined as model inputs by the correlation coefficient method. The water level prediction model in front of Denglou pump station with a foresight period of 2 h is taken as an example, and the combination of input factors is further analyzed, with a Mean Absolute Error of 0.017 3, Mean Squared Error of 0.001 1, Root Mean Square Error of 0.032 7, and Mean Absolute Percentage Error of 0.000 5. The inputs for the pre-pumping water level prediction model were ultimately established as the water level in front of Denglou station, the flow rate at Changgou pumping station, and the water level in front of Changgou station. Following this, the accuracy of the prediction model for foresight periods of 2 hours, 4 hours, and 6 hours were analyzed, along with an error assessment based on these results. The results show that the prediction model with a foresight period of 6 h has higher accuracy. For the data anomalies existing in the model, most of them are caused by the manual regulation process, after removing the anomalies, the overall error of the model was reduced, and the prediction model with a foresight period of 4 h had the highest accuracy. The LSTM neural network real-time model constructed can accurately predict the water level in front of Denglou pumping station, and it fits well with the measured value, with high prediction accuracy, with the absolute value of the relative water depth error of less than 0.02 m except in the case of the abnormal value. The input parameters are optimized through sensitivity analysis, and when the input parameters are optimal combinations, the four indexes of the model were less than 0.05, and the correlation coefficient and the coefficient of determination were greater than 0.95. The LSTM model can reduce the complexity of the combination of input factors and ensure the accuracy of prediction. In summary, the model prediction results have a high degree of fit with the measured values, and their errors are within a reasonable range, and the LSTM prediction model had a stable prediction effect, which can be used for real-time prediction of the water level in front of the station of the terrace pumping station system.