Abstract: Abstract: Pipeline is a carrier of cascade pumping station with long distance water conveyance. Therefore, it is particularly important to keep the stable operation of pipeline structure. Because of the complexity and diversity of pipeline structure, it is difficult to measure vibration response signal of pipeline of pumping station. In order to minimize risks and ensure safe operation of pipeline, it is significant to search for some methods that use fewer unit monitoring data to forecast pipeline vibration state. Support vector machine (SVM) was designed as the core for the proposed prediction model considering its advantages in solving the small sample size, nonlinear and high dimensional pattern recognition, and so on. For the purpose of the improvement of data utilization efficiency, particle swarm optimization (PSO) algorithm was applied because of its advantage of special memory function. Combining advantages of PSO algorithm and SVM, a PSO-SVM prediction model was proposed in this paper. Global search tracking algorithm of PSO was used to optimize the kernel functions and penalty factors of SVM, which weakened the problem of low accuracy of prediction caused by SVM parameters optimization deficiency. The No.2 pipeline of Pumping Station 3 in Jindian River pumping irrigation was selected as the research object, which was connected with No.4 and No.5 units, and 3 points were set up at the top of the volute of the unit and on both sides of the tail of the volute respectively for these 2 units. First of all, based on the vibration monitoring data of units and pipeline, with the mathematical statistics theory and spectrum analysis, the dominant frequencies of pipeline structure were counted and the contribution rates of vibration sources were determined for pipeline vibration. At the same time, correlation coefficients of vibration between unit and pipeline were calculated. Except No.3 measuring point, the correlation coefficients of the other 5 measuring points were greater than 0.57, of which the correlation coefficients of No.1 and No.6 measuring points were relatively large. Strong coupling relationship between units and pipeline was determined. Selecting the unit monitoring vibration data in the different periods as input factors, and the pipeline vibration response data of vibration sensors #1, #2, #16 and #17 during corresponding periods as output factors, the PSO-SVM prediction model of pump station was established. In order to compare prediction accuracy, back propagation (BP) neural network was established with the same data for training and test. The results showed that the PSO-SVM prediction result coincided highly with actually measured data, and BP neural network only reflected the trend of pipeline vibration response. PSO-SVM prediction model had a fairly high promotion in prediction compared to BP neural network. Aiming to quantitatively compare 2 methods, mean relative error (MRE) and root mean square error (RMSE) were introduced as the evaluation indices. The maximum values of MRE and RMSE for PSO-SVM were 6.8% and 0.261, respectively, much lower than BP neural network. The research shows that, in this test condition, when the correlation coefficient between unit and pipeline is above 0.67, this proposed method can realize effectively vibration prediction of pipeline, which has stronger generalization ability so as to achieve the purpose of pipeline safe operation and online monitoring.