Currently, models based on the application of artificial intelligence methods are actively developed and applied in solving a variety of problems, including in the practice of petroleum engineering. Evaluation of the accuracy and reliability of the developed models is usually reduced to defining standard statistical criteria, while the developers do not always use a separate examination sample. This article presents the results of the study, which are reduced to multivariate testing of the neural network model previously developed by the authors to determine the dynamic reservoir pressure in the selection zones of oil wells. The model is characterized by a number of advantageous characteristics, including minimal requirements for the amount of initial data, which determines its relevance and practical demand. However, the closed nature of computational algorithms related to the "black box" category does not allow us to reasonably formulate the conditions and criteria for applying the model, the reliability of the retro- and prospective forecast of the reservoir pressure. Three oil deposits of one field with different geological and physical conditions were selected as the object of study. The availability of a large number of actual reservoir pressure determinations by means of hydrodynamic well testing at the field allowed testing the model under a variety of scenarios, for each of which the forecast error was estimated and analyzed. As a result, high estimates of the model for retro- and prospective reservoir pressure reproduction were confirmed. It was found that forecast errors are reduced to zero in the presence of a large number of actual reservoir pressure determinations. However, to perform the calculation for each well, a single measurement for the entire history is sufficient. It was found that a sharp change in the well flow rate should also be accompanied by an actual determination of reservoir pressure with the entry of the obtained value into the model. In the absence of even a single reservoir pressure measurement for the wells, the model reliably reproduces its value using the kriging procedure used in the algorithms.