Research on Reservoir Hydrocarbon-Bearing Property Identification Method Based on Logging Data and Machine Learning

The hydrocarbon-bearing property of a reservoir is a crucial index for its evaluation. Although various evaluation methods based on well-logging data can reasonably interpret the hydrocarbon-bearing property of most reservoirs, these methods often exhibit significant randomness and ambiguity. This is due to various external influences, making it challenging to quickly and accurately evaluate the hydrocarbon-bearing property of a reservoir. To address this issue, this study investigates the identification of hydrocarbon-bearing properties in reservoirs based on well-logging data and machine learning techniques. Initially, 1731 sets of well-logging data with hydrocarbon-bearing property identification result labels from 356 wells in the Shahejie Formation of the Bohai Bay Basin’s Qikou Sag were collected. The distribution of different hydrocarbon-bearing property categories was analyzed on three types of well-logging data: gas logging, quantitative fluorescence logging, and Rock-Eval pyrolysis. Subsequently, seven model inputs were formed by combining these three types of well-logging data, and their performance was evaluated in combination with three machine learning techniques: K-nearest neighbor, random forest, and artificial neural networks. The influence of different inputs and models on classification performance was compared. Lastly, the importance of each input feature was analyzed. The results showed that the combination of quantitative fluorescence logging and Rock-Eval pyrolysis as inputs with the random forest model could achieve the best classification performance, with a macro F1 score of 95.36%. This suggests that this method has sufficient precision for the identification of hydrocarbon-bearing property categories in formations, providing a more efficient classification method for the hydrocarbon-bearing property of reservoirs compared to manual identification.