Interpretation Methods Of Tight Sandstone Reservoir With Seismic Data And Well Logs Based On Machine Learning Method And Multi-Information Fusion

The tight sandstone reservoirs are characterized by the low porosity, low permeability and strong heterogeneity. It is difficult to carry out the reservoirs prediction with sesimic data and well logs data. The methods of machine learning and multi-information probability fusion provide new ideas and methods for the prediction and fine characterization of tight sandstone reservoirs, which is of great value for the prediction of tight sandstone reservoirs.Taking the tight sandstone reservoir of Zhao30 blocks in Sulige Gasfield as the study target in this paper, the general idea of reservoirs prediction is gradually realized as follows. Based on the petrophysical analysis and seismic forward modeling, a variety of machine learning methods are used by the way of qualitative, semi-quantitative and then quantitative analysis from multi-perspectives. Firstly, the seismic forward modeling and AVO forward modeling of the tight sandstone reservoirs are realized,which provides a qualitative analysis process. Second, the data mining method of frequent itemsets is used to extract the frequent information of sand and gas, and then the result is confirmed with the result of seismic forward modeling, which provides a preliminary reservoirs target. On the basis, the semi-supervised fuzzy C-means method is used to analyze the characteristics of sedimentary and gas distribution with different group of seismic attributes and different paramters. And further, the self-organized neural network method is also used to study the regularities of distribution of sand body and gas layer thickness, which provides a quantitative analysis. Finally, the three-dimensional geologic model is established under the constraint of multi-information probability fusion, which can achieve a higher accuracy of prediction.According to the study, the following conclusions can be obtained. Firstly, Based on the study of seismic forward modeling of tight sandstone reservoirs, the results show that the response characteristics are evidently different for the different thickness and gas-bearing reservoirs and AVO type are mainly class III, which are in accordance with the analysis of prestack seismic data. Secondly, the FP-Growth method in the frequent itemset is used to excavate the frequent combination model with the attribute of prestack and poststack, and the results are consist with the AVO forward modeling,which further verify the accuracy of the method. Thirdly, based on the combination of different seismic attributes and the horizontal well data, the semi-supervised fuzzy C-means method is used to analyze the characteristics of sand body thickness distribution,which improves the traditional fuzzy C-means method and is consist with the drilling data. Fourthly, with the analysis of seismic wave and seismic attribute, the interpretation of sand body and gas thickness are carried out by the self-organizing neural network method based on the four types of seismic reflection modes, which can obtain a detailed distribution compared with the other two methods and the priori knowledge. At last, we can combine the information achieved by the different machine learning methods using multi-information probability fusion method, and then the three-dimension geological model is built constrained by the probability field. In order to verify the accuracy of prediction, the cross validation method is used and the root mean square error value of sand body thickness is 3.74 m, and the gas thickness is 1.34 m.The study shows that the machine learning methods have achieved obvious effect in the process of reservoirs prediction and the three-dimensional geological model obtained by the multi-information probability provides the theoretical basis for the drilling deployment.