An approach of machine learning was used to evaluate and predict the production of the heterogeneous carbonate gas reservoirs in the horizontal development wells of the late Precambrian Dengying Formation. The present data set of machine learning consists of gamma ray log, laterolog, high-resolution electrical image logs, and production rate data. The previous data set acquired the conventional openhole logs, including gamma ray log, neutron-density log, sonic log, laterolog, and dipole acoustic log. The challenge in the previous data set was that the training process for machine learning was not convergent. It was most likely that the conventional log responses did not fully correspond to the productivity of the heterogenous carbonate gas reservoirs. Forty-one wells associated with the present data set were used to set up the training sample data set for the machine learning to the productivity prediction of the carbonate gas reservoirs. The data set construction includes log depth shift, calibrated image log creation, classification of reservoir types from core and carbonate reservoir heterogeneity variables extraction from image logs. Core observation and core laboratory analysis indicate that the pore space of the carbonate gas reservoirs mainly consists of vugs, caves, and fractures. However, the vugs and caves are selectively developed and randomly distributed both laterally and vertically. This represents a complex heterogeneous carbonate reservoir in which the vugs and caves are key contributor to the total pore space of the carbonate gas reservoir. The attributes of the vugs and caves can be extracted from the electrical image logs, including connectedness, surface proportion, size, and thickness of vug, and cave zones. Six horizontal development wells were used to validate the machine learning approach. The predicted gas production rates in the four wells separately were 700,000 m3/d, 2,000,000 m3/d, 800,000 m3/d, 300,000 m3/d, 1,100,000 m3/d and 1,180,000 m3/d, and the respective actual gas production rates are 1,019,790 m3/d, 1,820,000 m3/d, 800,000 m3/d, 396,000 m3/d , 1,700,000 m3/d, and 1,411,900 m3/d. The machine learning workflow and approach provided satisfactory results in the six horizontal wells. Subsequently, the electrical image logs have run in the standard logging program in the more than 50 horizontal development wells.
Gamma-ray log derivative and volatility attributes assist facies characterization in clastic sedimentary sequences for formulaic and machine learning analysis
