**Research on Well Log Completion and Generation Based on Machine Learning**

Author:Chen Yun Tian

Supervisor:zhang dong xiao

Database:Doctor

Degree Year:2020

Download:102

Pages:138Size:8790K

Keyword:Gradient-free，neural network，Physical constraint，Small data size，uncertainty quantification，Well log

Machine learning algorithms,especially neural networks,have become powerful tools for modeling in the engineering field.These methods can fit the highly nonlinear mapping relationship between different variables of a higher dimension.Neural networks are particularly suitable for the engineering problems that have observation data,but the mapping relationship between variables is too complicated,which results in the problem that traditional physical models or empirical models cannot be effectively solved.An important application area of neural networks in engineering is petroleum engineering,especially for well logging which is critical to exploration and development of petroleum resuorce.Well logging is a physical measurement method used to describe and analyze underground conditions,and is of great significance for oil and gas exploration and development.Geologists and engineers can build accurate geological models based on well logs and design exploration and development strategies.However,the acquisition of well logs is often expensive and time consuming.In the actual measurement process,well logs are often missing or incomplete due to various objective reasons,and it is also possible to give up measuring some of the entire well logs considering the high cost.Therefore,the completion and generation of well logs is a study of academic and engineering value.However,because the formation is often complex and anisotropic,the mapping relationship between different well logs is extremely complicated.Whether it is a traditional physical model or an empirical model,it is difficult to accurately describe the relationship between the well logs.These existing models are also difficult to complete the missing logs or generate unmeasured logs.In this thesis,we propose an efficient solution to the use of machine learning methods for the problem of well log completion and generation.At the same time,we find that there are four problems in the process of directly applying machine learning algorithms to solving engineering problems.In order to solve these problems,this thesis introduces physical constraints as prior knowledge into the model,and integrates the algorithms in the history matching into the neural networks,and further proposes a new type of neural network called ensemble neural networks(ENN)and ensemble long short-term memory network(En LSTM).Using the newly proposed models,we successfully integrated the domain knowledge into the machine learning algorithm,making the model more in line with the physical mechanism and further improving the prediction accuracy of the model.Specifically,the first problem is that the current application of neural networks to practical engineering problems is simple and naive and lacks the consideration of specific domain knowledge.In fact,the introduction of domain knowledge into neural networks is equivalent to providing valuable prior knowledge to the model,which is conducive to constructing models that are more in line with the physical mechanism,breaking the bottleneck of model performance and further improving the prediction accuracy of the model.Simple and direct application of neural networks does not guarantee that the model has good prediction accuracy.The combination of neural network and engineering problems needs to fully consider the domain knowledge,and the neural networks should not be simply and directly applied to solving problems.Instead,the domain knowledge should be organically integrated and coupled with the neural networks.On the one hand,the ability of neural networks should be fully utilize to describe complex mapping relationships.On the other hand,using the characteristics of the problem itself to feed back the algorithm can make the model and the problem more effectively combined and improve the performance of the model.The second problem is the lack of quantitative analysis of the uncertainty of the predictions.In engineering applications,because the predictions tend to have large economic and social impacts,it is extremely important to analyze the uncertainty of the predictions.Uncertainty of the predictions is unavoidable in any model,and the prediction uncertainty includs data uncertainty due to noise data,and model uncertainty from model parameters and model structure.Therefore,the predictions of the neural network are also unlikely to be always accurate.If the model can provide uncertainty analysis on the predictions,the problem can be handed over to the manual judgment when the prediction is highly uncertain.By this means,it is possible to effectively reduce the loss due to the inaccurate predictions.Therefore,for applications with large economic value or impact on life safety,it is extremely important to provide uncertainty analysis of predictions.The third problem is the availability of data and the amount of useful data.High quality data is extremely important for training machine learning models.However,for most engineering problems,there are often two characteristics in the data: in some cases,the amount of data is huge,but the data is unstructured and the ratio of missing values and outliers is high,and there is not much actually useful data;in other cases,the data quality is high,but the acquicition of data is often time consuming and expensive.This problem greatly restricts the practical application of machine learning in engineering problems.The last problem is that the activation and loss functions of the current neural network must be easy to derive,otherwise the iterative optimization cannot be performed by the backpropagation algorithm.Although many loss functions are provided in the field of machine learning,in engineering applications,many effective loss functions are often combined with domain knowledge and difficult to derive.In addition,this problem also constrains the choice of neuron structure and limits the exploration of new networks with stronger predictive capabilities.In order to solve the above four problems,this study completed the following work:1.Apply machine learning algorithms to energy engineering to solve the problem of well log completion and generation.Considering the geological continuity of the reservoir,long short-term memory neural network(LSTM),which is good at processing sequence data and can learn long-term correlation,is chosen as the prediction model.Besides,a cascaded long short-term memory neural network(Cascaded LSTM)is constructed based on the LSTM and used to generate artificial well logs.2.Introduce physical constraints and domain knowledge into machine learning algorithms,and construct a physics constrained long short-term memory neural network(PC-LSTM)by adding mechanism mimic network architecture and formation adjusted stratified normalization to the traditional LSTM as physical constraints.The geomechanical well logs are predicted based on conventional well logs using the PC-LSTM,and the 3D geomechanical fields are successfully constructed.This method is conducive to the construction of accurate geological fields based on easily available conventional well logs,which is of great significance for actual petroleum exploration and development.3.The ensemble randomized maximum likelihood(En RML)method in the field of energy engineering is used to improve the fully connected neural network(FCNN)and LSTM.By combining the feedforward process of traditional networks(FCNN and LSTM)with En RML,the ensemble neural network(ENN)and the ensemble long short-term memory network(En LSTM)are constructed and their iterative optimization is based on covariances.These networks are constructed based on Bayes’ theorem.They can provide uncertainty quantification,can be trained based on small data set,and do not depend on derivative calculation,and they are more suitable for practical engineering problems.This research proves that energy engineering algorithms can be applied to machine learning algorithms and optimize and improve machine learning models.In the En LSTM,the problem of excessive convergence is effectively solved by introducing the model parameter perturbation method,and the proportional perturbation method of normalized observations is introduced as well.Finally,the En LSTM is applied to the well log generation problem and achieved good results.