| Low-resistivity-low-contrast(LRLC)pay zones are widely distributed with huge reservation,and they have become the main targets for the recheck of old wells in most oilfields.The resistivity of LRLC pay zones is very similar to that of water layers,so the primary task for logging interpretation of LRLC pay zones is to identify LRLC pay zones.The commonly used LRLC pay zone identification methods can be divided into two types.The first type is based on conventional well logging data.According to the LRLC genetic mechanism in the study area,sensitive logging parameters,single or combined,are selected,and cross-plots of these parameters are used for LRLC pay zone identification.The second type of method is based on the data of new well logging methods.Array induction logging method,for example,can measure the resistivity of reservoir at different detection radius,and can be used to analyze the genetic mechanism caused by mud filtrate invasion.Nuclear magnetic resonance(NMR)well logging method can jump out of the trap of low resistivity contrast,and directly analyzes the hydrocarbon components of the reservoir by measuring the NMR signal of the hydrogen atoms in pore fluid.For the quantitative evaluation of reservoir parameters in LRLC pay zones,the reservoir parameter calculation methods based on conventional well logging data all try to establish certain parameter calculation models.However,the classification methods of LRLC pay zones based on conventional logging data are often ineffective and locally applicable.While the data of new well logging methods,such as the NMR well logging data,which can directly obtain reservoir fluid composition information and pore structure information,is not always available.Reservoir parameter evaluation of LRLC pay zones also faces the same dilemma.Therefore,this thesis proposes a globally applicable method based on machine learning algorithms to classify LRLC pay zones using only conventional well logging data,and designs a machine-learning-based integrated software to efficiently accomplish relative works.Finally,a variety of algorithms in this software have been applied to calculate reservoir parameters of LRLC pay zones.The genetic mechanism of LRLC pay zones of two study areas,F and J oilfields,is analyzed in detail.The genetic mechanisms of these two study areas have a lot in common,which are complex distribution of formation water salinity,high clay content with high volume of mixed-layer illite/smectite,and high irreducible water saturation.The genetic mechanisms of these two study areas,however,also have differences.The main genetic mechanism in the F oilfield is the complex distribution of formation water salinity,while the main genetic mechanism in the J oilfield is the high irreducible water saturation.In this thesis,two machine-learning-based methods are proposed for the identification of LRLC pay zones.The first one is to add the well logging data of the adjacent shale layer as reference data and combine it with the well logging data of the oil testing layer.Then,density-based spatial clustering algorithm with noise(DBSCAN)algorithm is applied to cluster the dataset above,and the dataset is projected to a higher dimension feature space by calculating the distance between the dataset samples and the centers of the feature clusters obtained from DBSCAN algorithm.Finally,the dataset in the high dimension feature space is treated as training data for a multi-layer perceptron(MLP).The proposed method was applied in two oilfields.The identification result of F oilfield has an 89.97% accuracy,and the identification result of J oilfield has an 95.94%accuracy.The second method is to pixelate the well logging data of all the oil testing layers,and resample the data with SMOTE-Tomek algorithm.A convolutional neural network with a long-short-term-memory layer(LSTM-CNN)is applied for LRLC pay zone identification.Good and reasonable results is obtained with a 90.40% accuracy.Since the number of parameters in the LSTM-CNN model is in very large amount,the model needs to be pruned by a power function pruning algorithm to avoid overfitting problem and reduce computational complexity.The number of parameters in the pruned model are reduced by 40% and the storage space is also reduced by 67% comparing with the unpruned model,but the identification result of the pruned model is almost unaffected with a 90.06% accuracy.The LSTM-CNN method mentioned above has a good identification results in F oilfield where the number of oil testing layers(i.e.,the number of samples)is large,but it is less effective for J oilfield whose identification accuracy is 76.92% due to the small number of oil testing layers and high risk of over-fitting problems.Since the two oilfields have similar genetic mechanisms and the same kinds of well logging curves in the input data,model parameters in the convolutional and LSTM layers trained in the F oilfield can be transfered to the neural network model in the J oilfield using the transfer learning algorithm.Only the full connected layers need to be retrained at a very small computing cost,and then the model can be fine-tuned.The transferred model can achieve better results in the identification of LRLC pay zones in J oilfield,and the identification accuracy is 85.38%.A software that integrates various machine learning algorithms is also designed to identify and quantitatively evaluate LRLC pay zones using well logging data.Four modules are included in this software,which are data preprocessing module,individual machine learning algorithm module,ensemble learning algorithm module,and other functions module.The data preprocessing module can convert well logging data into training data that can be directly used by other machine learning algorithms.The individual machine learning algorithm module and the ensemble learning algorithm module provide a large number of machine learning algorithms for application.The other functions module is a complement of the software functions,including algorithm optimization,model pruning,transfer learning,resampling,and simple figure output.Finally,the software is used to predict the pore structure parameters of LRLC pay zones in J and F oilfield using various algorithms,and the optimal algorithm is selected by results comparison,which reflects the easy-to-use and efficient features of the software. |
