Intelligent Discrimination Method for Reservoir and Formation Fluids Based on Deep Learning

In the process of oil and gas exploration, well logging and mud logging are the primary means for identifying underground reservoirs and distinguishing fluid types within them. Traditional interpretation of well logging and mud logging data mainly relies on expert experience and manual operations, leading to issues such as low interpretation efficiency, significant human factor influence, and difficulty in ensuring accuracy. This paper proposes a method for intelligent discrimination of reservoir and fluid types based on a bidirectional Gated Recurrent Unit (GRU) deep learning model. The specific procedure is as follows: (1) Collection, cleaning, and preprocessing of well logging and mud logging data; (2) Construction of a labeled well logging and mud logging dataset; (3) Construction of a deep learning model for reservoir and fluid discrimination; (4) Model training and optimization; (5) Utilization of the trained model for reservoir and fluid discrimination. By combining the sliding window sampling method with the bidirectional GRU deep learning model, the model is enabled to fully learn the geological information contained in the curve shape, curve amplitude, and background values of adjacent well intervals of the well logging curves, thereby improving the accuracy of reservoir and fluid discrimination. A weighted cross-entropy loss function is adopted to overcome the issue of imbalanced data distribution in actual production scenarios. Verification using real oilfield data demonstrates that this method achieves an accuracy rate of 92% in reservoir and formation fluid discrimination in the Neogene oilfields of Bohai Bay, significantly improving the efficiency of well logging and mud logging data interpretation. It provides technical support for the digitalization and intelligent transformation in the field of oilfield exploration and development.