Study On Reservoir Porosity Prediction Based On Deep Learning

Porosity is one of the key rock physical parameters,which can measure the storage capacity of petroleum fluids in the reservoir.Laboratory measurement methods for core sampling calculations are not only expensive and time-consuming,but also can only evaluate isolated points.Porosity prediction based on logging data has the advantages of data continuity and economic-time effectiveness.The situation of underground reservoir is complex,so the logging data obtained have strong non-linearity and non-uniformity.Although existing methods have achieved certain prediction results,they are insufficient in mining the correlation between sedimentary strata.Therefore,this thesis focuses on sample construction method and porosity prediction methods based on deep learning.The main research contents are as follows:(1)To mine transversal feature information among parameters and longitudinal depth deposition information in logging data,and make full use of this information for porosity prediction,a construction method of Imaged Sequence Samples(ISS)is proposed and the Sequence to Sequence(Seq2Seq)method is first introduced to establish a porosity prediction model.Firstly,the logging parameters are selected and standardized based on correlation coefficients;an ISS set is established by slidingly grouping sample points with depth,and the selected logging parameters are in a row.Secondly,due to its two sets of serial Bi-directional Gate Recurrent Units(BGRU),Seq2 Seq have good sequential data processing capabilities and flexible structures.As a result,a Seq2 Seq model for porosity prediction is established.Experiments show that the proposed ISS construction method can represent the transversal characteristics and longitudinal depth information of data;The Seq2 Seq method can effectively predict porosity and achieve high prediction accuracy.(2)To correct the cumulative bias and exposure bias that are prone to occur in the Seq2 Seq method and make the model more efficient in generating porosity prediction sequences,a Transcendental Learning(TL)module is designed and the TL-Seq2 Seq model is established for porosity prediction.Firstly,a Seq2 Seq framework based on BGRU is established.Then,the encoded information from the encoder and the real label information are fused in the TL module,and the transcendental learning knowledge containing the right prediction information was obtained by decisions from the Teacher Forcing strategy.Finally,transcendental learning knowledge is applied to guide the decoder in generating a porosity prediction sequence.Experiments with different research methods have been conducted both in two wells.The experimental results show that the proposed TL-Seq2 Seq model can correct cumulative bias and exposure bias,and has strong convergence and generalization capabilities.It manages training the ISS data for porosity prediction efficiently and achieves the best prediction accuracy among existing porosity prediction methods.(3)To fully extract transversal feature information from ISS data,and reduce network burden,a Semantic Query Transformer(SQ Transformer)method is proposed and the Semantic-based Parallel Network(SPNet)which is composed of parallel BGRU and SQ Transformer is established.Firstly,the concept of transcendental learning is combined with attention mechanisms in SQ-Transformer.The label semantics and Teacher Forcing strategies are used respectively to improve and control the proportion of correct information in Query.Secondly,the SQ-Transformer is pre-trained based on transfer learning technology to initialize network parameters.Finally,in the parallel network SPNet,the porosity feature is obtained by fusing the transversal feature information extracted by the upper SQ-Transformer and the longitudinal depth series information extracted by the lower BGRU.The experimental results show that the feature weights that have important contributions to porosity prediction can effectively improve in SPNet without trapping the training model into overfitting.TL-Seq2 Seq model achieves quite a high accuracy and relatively small parameters,flops and memory compared to the TL-Seq2 Seq,indicating that SPNet has good research and application prospects in actual porosity prediction tasks.