| Petroleum geological big data analysis has become one of the important topics in hydrocarbon exploration and development and geological sciences since petroleum geological big data contains abundant petroleum geological knowledge.Machine learning(ML)refers to the automatic recognition of hidden patterns,structures and relationships in big data through the computer simulation of human learning behavior.It is argued that ML will play a key role in petroleum geological big data mining and knowledge discovery.However,among the foremost challenges facing the effective analysis of petroleum geological big data by using ML techniques:(1)how to fuse and organize multi-source heterogeneous and multi-scale petroleum geological big data to provide direct data sources,and(2)how to fully integrate the known geological information and domain knowledge of geologist and geophysicists in ML.Therefor this paper studies these two problems systematically to form a set of novel approaches for appling ML techniques in petroleum geological big data analysis,which provides methodologies and technologies for intelligent oilfield and intelligent geology.The major contributions of this paper are highlighted as follows:(1)This paper designs a relationship fusion strategy of petroleum geological big data based on data lake architecture and metadata,aiming at the problems of decentralized storage and multi-source heterogeneity of petroleum geological big data.This strategy allows petroleum geological big data to be stored in a variety of patterns and structures under the data lake architecture.From the perspective of metadata,the identifiers for data source metadata and business metadata are arranged to establish the hierarchical relationship,and the relationship fusion of petroleum geological big data is achieved from the metadata level.(2)This paper designs the ML-oriented methods of constructing 3D geological model and knowledge ontology.3D geological model is used as the direct data source of ML,and the organization method of petroleum geological big data based on 3D geological modeling is studied.The modeling strategy of 3D surface model is designed by using triangle patch as data organization unit and triangulated irregular network(TIN)as data organization mode;Moreover,the multi-scale 3D geological solid modeling strategy is formulated,taking voxel as organization unit and octree as data organization mode;The modeling example shows that the method can realize the organization and visual expression of multi-source and multi-scale petroleum geological big data.In addition,to implement the organization and reuse of business knowledge required by ML mthods,five basic elements of knowledge ontology are defined,including the concept,attribute,content,relationship under the business topic,and the operation on the 3D geological model;On this basis,the construction method of knowledge ontology is studied,and business knowledge and business data are extracted from3 D geological model by taking advantage of knowledge ontology.(3)In order to apply the mapping relationship between geological variables to predict the target variable,this paper studies prediction-oriented ML methods.Considering the key problems of ML prediction methods and demand of quantitative prediction of geological outcrop porosity,this paper proposes a method of geological outcrop porosity prediction based on ML.Compared to traditional methods,this method employs 3D outcrop surface model as data source,exploits spectral data to predict porosity and instantiates geological outcrop ontology accordingly.On the basis of discussion on the spectral response characteristics of porosity and prediction feasibility,the porosity prediction model is constructed based on ML methods.The application in sandstone outcrop of Yanchang Formation in Shanxi Province shows that this method can quantitatively predict sandstone porosity,thus providing a basis for accurate inversion of geological outcrop geological porosity.Then the application strategies of prediction-oriented ML methods are designed.(4)In order to deeply mine the multiple geological information of small samples to obtain reliable comprehensive evaluation results,this paper studies ML methods for multi-factor comprehensive evaluation task with small sample size.To assign reliable weights to evaluation factors,this paper utilizes expert knowledge and reduces its subjectivity and uncertainty in the weighting procedure,and develops a knowledge driven method called fuzzy analytic hierarchy process-grey relational analysis method(Fuzzy AHP-GRA).In view of the lack of known information in the early stage of marine hydrocarbon exploration,an evaluation method of marine hydrocarbon resources is presented based on ML.Different from traditional methods,this method uses 3D geological model as data source,determines and extracts evaluation factors according to knowledge ontology.Then the ML methods are applied to conduct comprehensive evaluation of marine hydrocarbon resources.The evaluation method of marine hydrocarbon resources based on ML is applied in Laoshan uplift in South Yellow Sea Basin and obtains reliable results compared with known geological information.Then this paper presents the application strategies of ML methods for multi-factor comprehensive evaluation task with small sample size. |
