The Research On The Soft-sensing Technique Of Dynamic Liquid Level For Sucker Rod Pumping System Based On Multiple Models

In the exploitation of crude oil, the sucker rod pumping system is used commonly in many oil fields. In the production process of the sucker rod pumping, the measurement of liquid level in the annulus space is a very important work. Dynamic liquid level is the dynamic data of the actual level when the well is in work. The dynamic liquid level, on the one hand, is a direct reflection of the liquid supply ability of the well, on the other hand, can be used to help us to adjust the depth of the sucker rod pump. There are some defects in the accuracy, real-time property, safety and other aspects in the traditional measurement methods. So it is necessary to research new testing technology. The application of online soft measurement technology is not only economic and reliable but also overcomes the defects of the traditional measurement methods.In order to have a better reflection of the actual production process characteristics, we usually have to collect more sample data in the establishment of soft measurement model. But the sample data for different regions and time has different amount of disturbance and characteristics. The application of single model will result in that the process characteristics is of mismatching, and have a bad accuracy and generalization ability etc. So this thesis apply a soft measurement modeling method using multiple model integration. Firstly, we should get the sub-model based on the RBF neural network, LS-SVM and ELM method for the sample data. Secondly, on the process of predicting, the prediction of three sub-models should be integrated based on the Gauss-Markov estimate method to get the final result.Since there are a large number of the sample data and have a great difference of them, and which may cause that the prediction is not accurate, so this thesis apply a method of multiple model integration based on affinity propagation clustering method. First of all, the sample data should be classified using affinity propagation clustering method, secondly, we should get the models for each of the cluster, then get three sub-models for each of the cluster, at last, the prediction of three sub-models should be fused based on the Gauss-Markov estimate algorithm. The final predicted value is the forecasting result of the model of the cluster which the test sample data belongs to.In the practical application, the system have the problem like when the sample data is near the working point, the correlation between the sample data and the working point is larger, but when the sample data is far from the working point, the correlation between the sample data and the working point is smaller, and the problem that the dynamic changes is lack of considering in the running of the system. On account of these problems, a kind of online soft-sensing technique of multiple models fusion is applied. This method can be adaptive to the changes of the working conditions, and can update the model dynamically, and further improve the accuracy and robustness of the model.At last, we applied the method to the prediction of the dynamic liquid level, the result also verifies that the proposed algorithm has better prediction accuracy and robustness.