Study On Dynamical Prediction Methods Based On Grey System And Kernel Method

With many oil fields in China and abroad stepping into the mid-late period of exploitation in recent years, the oil companies are now facing the challenges such as the high water cut and the high-speed declining of the production, etc. It is important to make efficient schemas for exploitation or adjusting the existing schedules for the oil companies over the world. To this end, not only improving the technology and craft, but also accurate prediction of the reservoir performance is needed. Precise forecasts can reflect the current status of the reservoir performance, which can help make the schemas of exploitation and adjustment. However, the reservoir performance is effected by a lot of natural and anthropic factors, and the relationship between the indicators of reservoir performance and the reliance factors is often nonlinear, which make it difficult to predict the reservoir performance accurately. On the other hand, it becomes inevitable to deal with the large scale data when predicting the reservoir performance as the exploitation time becomes longer and the development scale becomes larger, which brings about another challenge of accomplishing the forecasts efficiently and time-saving.This dissertation is aiming at solving the main problems in the reservoir exploitation currently. In order to achieve the accurate and efficient forecasts, several dynamical prediction models have been built based on the Grey System Theory and the kernel methods considering different conditions, and the fast algorithms have been designed to train the kernel based prediction models based on the optimization methods. The main research of this dissertation is based on the philosophies above, and the main contents are outlined as follows:(1) The way of deriving the univariate grey discrete prediction models has been analyzed and summarized theoretically, based on which two multivariate grey discrete models have been proposed. The relationship between the discrete models and the traditional models has been discussed theoretically, and the advantages of the discrete models have been presented in the numerical examples. Theoretical analysis between the grey models and the Arps decline models has been presented, which specified the connection between the grey models and the production decline, and also provided theoretical basis for the reservoir performance prediction by the grey models.(2) The main philosophy and general modelling procedures of the kernel method have been abstracted based on the research of the modelling procedures for the traditional kernel based prediction models. The kernel method is then been used to build the nonlinear extension model of the Arps decline model, which contains the law of decline and the nonlinear relationship between the production and its reliance factors. Numerical examples have been presented to show the difference between the linear model and the nonlinear model, and also to show the advantages of the nonlinear model.(3) The kernel method has been used to make the traditional multivariate grey prediction model nonlinear, and then a new prediction model has been proposed. The ability of small sample modelling of the grey model and that of dealing with the nonlinearities have been coupled in the novel model. The advantages of the novel model in dynamical prediction with small samples and nonlinearities have been shown in the numerical studies.(4) The training algorithms for the kernel based prediction models proposed above have been designed based on the Sequential Minimum Optimization(SMO) and the Conjugate Gradient algorithm. The convergence, the rate of convergence and the accuracy of these algorithms have been discussed in the numerical studies.(5) The four proposed forecasting models have been used to building the dynamical prediction models with the real world production data, and the accuracy of these models have been compared. The applicable conditions have been summarized based on the analysis of the advantages and disadvantages for each model in the conditions of linearity and nonlinearity, small samples and large samples, respectively. Meanwhile, the efficiency and accuracy of the SMO and Conjugate Gradient algorithm have been compared, the superior algorithm for the real world prediction has been suggested.