Transcripción del resumen del autor: The permanent downhole gauge (PDG) is a promising tool for reservoir testing but has yet to reach its full potential. Generally, conventional well-testing methods are most able to use small sections of constant-flow-rate data. However, data mining, a newly developed technique in computer science, is a tool that can reveal the relationship among variables from large volumes of data. The application of data-mining algorithms to synthetic and field data has been successful in extracting the reservoir model from variable-flow-rate and pressure-transient data. In fact, because of uncertainty in flow-rate measurements, this technique is one of the few ways to make use of the flow-rate data that are now available with some modern PDG tools. The application is conducted in two steps. First, the pressure and flow-rate data from the PDG are used to train a nonparametric data-mining algorithm. The reservoir model is obtained implicitly in the form of polynomials in a high-dimensional Hilbert space defined by kernel functions when the algorithm converges after being trained to the data. Next, a specific flow-rate input (for example, constant rate) is fed into the data-mining algorithm. The datamining algorithm will make a pressure prediction subject to the input flow rate. Because the data-mining algorithm has already obtained the reservoir model, the pressure prediction is expected to be the reservoir response given the constant flow rate. Therefore, the proposed constant flow rate and the predicted pressure reveal the reservoir model underlying the variable PDG data, without needing to specify in advance which reservoir model is to be used. Three methods (Methods A, B, and C), differing by input vectors, kernel functions, and presence of breakpoints detection, are proposed in the paper. Synthetic noisy data and real field data were used to test this approach. Methods B and C were able to satisfactorily recover the wellbore/reservoir model in most considered cases. Even in extreme cases when the flow-rate data are noisy and changing frequently, and in the absence of any shut-ins, the method was still able to extract the reservoir models.

1