Research On The Prediction Of Remaining Oil Distribution Based On Deep Learning

Remaining oil distribution is one of the most important issues that oilfield developers have always paid attention to.After years of research,the predecessors have formed various methods for predicting remaining oil,such as comprehensive analysis method,measured data method and numerical simulation method,which play a huge role in improving oil recovery.In the process of oilfield development,a large amount of valuable data which can provide data sample support for machine learning has been accumulated.This paper intends to use the deep learning method to deeply analyze the main factors affecting the distribution of remaining oil,and then build a deep learning model and carry out training to form a new prediction method of remaining oil distribution.Based on the analysis of the factors affecting the distribution of remaining oil,this paper selects the reservoir physical property parameters distribution,reservoir fluid characteristics,well pattern and injection and production parameters as the key data of remaining oil distribution prediction and research on data processing methods.The numerical simulation method is used to establish nearly 500 sets of samples which have different physical property parameters,different fluid characteristics,different well patterns and different injection and production parameters for data calculation,after collecting the data calculation results,the data is cleaned and a learning sample library is generated.After researching and analyzing the deep learning algorithm,the support vector machine(SVM)and long-term and short-term memory network(LSTM)were used to establish the model which predicts remaining oil distribution after classification.The model prediction accuracy and model training time-consuming are used as evaluation indexes,and the model is repeatedly tested to determine the optimal parameters selection of the model.After learning the model with learning samples,a prediction model of remaining oil distribution based on deep learning is formed.The remaining oil distribution prediction model is used to predict the oil distribution of homogeneous conceptual model,heterogeneous conceptual model and actual reservoir model.Compared with the numerical simulation results of the reservoir,the prediction accuracy can reach over 95%,and the prediction time is much smaller than the reservoir numerical simulation method,the feasibility and efficiency of the model is shown.Applying the idea of residual oil distribution prediction,this paper attempts to predict the formation pressure distribution using LSTM network and the prediction effect also meets the mine requirements,which fully demonstrates the huge application potential of deep learning methods in oilfield development.