Research On Measuring Rod Pump Well's Working Liquid Level Modeling Based On Online Learning Gaussian Processes

For wells which have low formation energy and lose the flowing capacity, we have to pumps oil by using rod pump unit, and it is important to keep abreast of the various parameters and analyze all aspects of the oil wells working conditions. We can analysis underground condition to adjust the operating parameters of the well by the well's working liquid level, so that the working conditions for the fluid matching the fluid supply capacity. It is significant for ensuring a high yield and high pump efficiency. Further, when the well's working liquid level changes, we can make real-time adjustment of parameters of the well to make sure the security and stability of oil system and to ensure device's proper operation.Currently, the measurement method of the well's working liquid level is mainly echo method. Generate sound waves at the wellhead and let the waves transmit down alone well; when waves encounter the liquid level, they reflect back to determine the working liquid level. The measurement method requires stopping the device's operation, being complex and high cost, having a degree of risk and can not be operate. Therefore, the thesis introduces the soft sensor depended on online Gaussian processes regression. Online Gaussian process regression algorithm is a non-parametric incremental learning algorithm depended on Bayesian learning. The method is extensively applied to solve the high-dimensional, small-sample and nonlinear regression problems, and it can update the model to make the test result more accurate, with the time going on. Compared with the parametric learning approach, online Gaussian process regression algorithm has the advantage of implementing easily, self-adaptive to determinate hyper parameters and results with probability interpretation.The thesis is mainly depended on online learning Gaussian process algorithm and I improved its algorithm. Improvements are as follows. Firstly, clustering on training data, I can get the sparse basic vector and its size. Secondly, the expected improvement of each new input data is calculated, so that I can determine whether new data is added to the basic vector.Thirdly, using covariance functions seeks the information gain of the algorithm. Lastly, I use AFSA to optimize the hyper parameters. Through qualitative analysis of the parameters related with the working liquid level and ARD, I can determine the auxiliary variables. Then, I simulate the model, and compare the offline Gaussian process regression algorithm with the online Gaussian process algorithms. Finally, because the traditional conjugate gradient optimization method relies initial values and is easy to reach local optimum, I use AFSA optimize the hyper parameters. The online Gaussian processes regression algorithm using the AFSA can obtain much better prediction results.