Abstract
Developing a model that can accurately predict internal fractured reservoirs in the context of the ultra-low physical properties of carbonate rocks by only employing conventional mathematical methods can be very challenging. This process is challenging because the relationship between basic fracture parameters and the logging response in carbonate reservoirs has not been studied, and the traditional method lacks adaptability due to the complex relationship between basic fracture parameters and the logging response. However, data-driven approaches supplemented by machine learning algorithms based on multi-layer perceptrons (MLP) provide a more reliable solution to this challenge. In this paper, a classical fracture parameter evaluation data set is established using fracture porosity, fracture density, fracture length, and fracture width data that can be identified by resistivity and acoustic imaging logging. Another data set can be composed of different types of logs, and it can be used to identify reservoirs. Two different data sets were validated by regression task evaluation indicators in machine learning, and the correlation coefficient R2 is greater than 0.82. This means that the model accuracy of the algorithm can reach 82%. Combined with the comparison results of eight conventional machine learning algorithms, the reliability and application validity of the MLP model are verified. This method’s accuracy is also verified by oil test data, which show that the MLP machine-learning algorithm can effectively simulate the relationship between lithology and fracture development. In addition, it can be used to predict key exploration horizons before drilling. The relationship between lithology and fracture development degree is well-simulated by the MLP machine learning algorithm, which shows that the degree of fracture development is mainly affected by fractures, indicating that the method can be used to predict key exploration horizons before drilling.