Forecast And Early Warning System And Algorithm For Oil And Gas Wells Safety Production Monitoring Data

Along with our country economic developing into the normal growth path, and the industrial system reform deepening greatly, the energy demand is steadily increasing. In recent years, in the production process of the oil and gas, the representative of the fossil energy,security issues in mining, purification, transportation and other production processes have become increasingly prominent. Using data driven approach for anomaly detection and early warning has become the focus of the current research. In view of the characteristics of the process industry monitoring data, such as isolation, nonlinearity, non completeness, etc, the statistical process monitoring theory and kernel method are introduced to establish the model of overall characterization and the abnormal diagnosis for the oil and gas wells, in order to improve the accuracy of the algorithm. The main work of this thesis includes following aspects:1. A kernel principal component analysis algorithm based on clustering is used to solve the problem of oil and gas wells system status. The kernel mapping method is used to solve the problem of the nonlinear information loss in the data reduction process by traditional mehod.In high dimensional linear space conversion by the inner product computation, the feature decomposition is performed by using the cluster center of sample kernel mapping,the statistic distribution of squared prediction error is used for the overall characterization of the system state. Examples show that the algorithm can reduce the false alarm rate effectively.2. Based on the independent component analysis, the kernel decoupling fisher discriminant algorithm is proposed to solve the problem of the fault diagnosis and classification in the oil and gas wells under the complete sample. Independent component analysis is implemented to decouple the independence of samples. In the high dimensional linear space,the classification methods based on cosine angle are presented to build the optimal discriminant vector library of the complete historical data sample. The diagnosis and classification of real time abnormity in oil and gas wells safety production are realized. Examples show that the fault diagnosis accuracy of oil and gas wells is improved.3. A support vector description algorithm based on clustering method and ensemble learning approach is used to solve the problem of abnormal classification of oil and gas wells safety production monitoring samples under the condition of incomplete data. In view of the limitation of the traditional support vector description algorithm, the KMSVDD algorithm expands the application field to multi-class abnormal samples space. The incomplete samples are replaced by the sample cluster center vectors as classification input, to calculate the high-dimensional sphere radius and other classification description thresholds of the new algorithm,and classify the abnormal samples under the condition of multiple incomplete samples. The ensemble learning approach is introduced to solve the classification accuracy problem of traditional method by adjusting the training samples' weights of weak classifiers.Numerical examples show that the new algorithm can effectively classify the new isolated fault types.4. The integrated algorithm of the overall characterization and anomaly analysis is realized,and the system framework of oil and gas wells safety production monitoring based on OSA-CBM standard system is presented. Abnormal optimal discriminant vector library in the offline space is established to realize the overall characterization and anomaly analysis in online real-time spatial data space. Through optimal discriminant vector or abnormal samples being updated by the feedback of analysis, the closed loop of the data monitoring algorithm is formed,and the algorithm integration is realized. The functions of the monitoring and early warning system included data storage, calculation, analysis and interactive presentation. The dynamic management and elastic expansion functions of the model are realized to characterize the overall state of the system quickly, and update the abnormal sample library dynamically, and realizate the online fault diagnosis and early warning.An empirical analysis of the desulfurization and dehydration processes in this thesis is carried on by analyzing the monitoring parameters data of a natural gas purification plant in Southwest Oil and Gas Field. In the overall characterization, the number of false alarms in normal operating conditions is reduced from 27 to 6 with the new algorithm, compared to the traditional algorithm. In the decoupling discriminant analysis of complete samples, the dimension of the sample is reduced from 1050 dimensions to 149 dimensions, and Fisher's value is reduced from 3.45×108 to 5.87×104 with the new algorithm. In the analysis of multi class anomaly incomplete samples, the new algorithm can identify the abnormal samples which can not be identified by the traditional SVDD algorithm, and only 5 of the 50 introduced test exception samples failed to be recognized. Improved ensemble learning algorithm raises the false sample cluster classification accuracy up to 100%. The empirical results show that the proposed algorithm is effective.