Reservoir feature prediction method and model based on deep learning
A deep learning and model technology, applied in the field of reservoir transformation, can solve the problems of low prediction accuracy, and achieve the effect of reasonable and efficient reservoir transformation and production optimization management
Pending Publication Date: 2021-09-07
PETROCHINA CO LTD
0 Cites 2 Cited by
AI-Extracted Technical Summary
Problems solved by technology
[0007] The object of the present invention is to provide a method and model for predicting reservoir characteristics based on deep learning, so as to at least solve...
Method used
A kind of reservoir feature prediction method and model based on deep learning provided by the present invention combines convolutional neural network (Convolutional Neural Networks, CNN), two-way long-short-term memory (Bi-directional LongShort-TermMemory, Bi-LSTM ) and the attention mechanism (Attention Mechanism), using the advantages of two-way long-short-term memory to efficiently and accurately predict time series, coupled with the attention mechanism layer, to make up for the lack of convolutional neural network processing data with serial correlation, accurate Predict the reservoir characteristics such as porosity and permeability of reservoirs at different depths, so as to reasonably and efficiently carry out reservoir transformation and production optimization management, at least solve the problem of using convolutional neural networks in the prior art when predicting reservoir characteristics at different depths Technical issues with low prediction accuracy.
In summary, the reservoir feature prediction method based on CNN and Bi-LSTM and Attention can make full use of the spatial correlation and time series correlation of reservoir features, identify the correlation between features, and reduce the error; by Taking advantage of the large-scale parallel collaborative processing capability of artificial neural networks and utilizing the advantages of different neural networks, the generalization ability of the model has been effectively improved, and an efficient reservoir parameter prediction model has been established, which provides a new model for the prediction of characteristic parameters of reservoir reconstruction. train of thought.
Please refer to Fig. 4, after the pixel map is input to the multimode Bi-LSTM model, the double-layer convolutional neural network first carries out feature extraction to the pixel map, then the feature input of extraction is processed in the Bi-LSTM neural network, in Before outputting the results, the Attention layer is used to calculate the weight of each time series, extract the most important information, and use the weighted sum of all time series vectors as feature vectors to obtain the prediction results of reservoir characteristics. In the multi-mode Bi-LSTM model, the LSTM layer adopts a two-layer bidirectional structure to improve the efficiency and accuracy of time series correlation feature analysis.
The essence of the attention mechanism (Attention mechanism) is to imitate the human visual attention mechanism, learn a weight distribution to image features, and then apply this...
Abstract
The invention discloses a reservoir feature prediction method based on deep learning. The method comprises the following steps: acquiring a logging data training set; constructing a convolutional neural network and a forward-backward long-short term memory neural network added with an attention layer, and performing sequence modeling by combining the convolutional neural network and the forward-backward long-short term memory neural network to generate a multimode Bi-LSTM model; inputting the training set into a multi-mode Bi-LSTM model, training the multi-mode Bi-LSTM model by adopting a joint training mode, and performing joint optimization on parameters of a convolutional neural network and a forward-backward long-short term memory neural network added with an attention layer; and inputting actual logging data into the trained multimode Bi-LSTM model, and performing prediction through the model to obtain a prediction result of reservoir characteristics. According to the method, the advantage that bidirectional long and short time memory can efficiently and accurately perform time sequence prediction is utilized, and the attention mechanism layer is added, so that the defect that the convolutional neural network processes data with sequence correlation is made up, and reservoir characteristics such as porosity and permeability of reservoirs with different depths can be accurately predicted.
Application Domain
Neural architecturesNeural learning methods +1
Technology Topic
Sequence modelingData input +12
Image
Examples
- Experimental program(2)
Example Embodiment
[0065] Embodiment one
[0066] Please refer to figure 1 , image 3 and Figure 4 , the present invention provides a method for predicting reservoir characteristics based on deep learning, comprising:
[0067] S101. Obtain a training set of well logging data;
[0068] S102. Construct a convolutional neural network and an Attention-Based Bi-LSTM neural network with an Attention layer, and combine the two for sequence modeling to generate a multi-mode Bi-LSTM model;
[0069] S103. Input the training set into the multi-mode Bi-LSTM model, train the multi-mode Bi-LSTM model by means of combined training, and jointly optimize the parameters of the convolutional neural network and the Attention-Based Bi-LSTM neural network ;
[0070] S104. Input the actual logging data into the trained multi-mode Bi-LSTM model, and obtain the prediction result of the reservoir characteristics through the prediction of the trained multi-mode Bi-LSTM model.
[0071] As a specific implementation, the logging data includes features and labels, wherein the features include acoustic wave, density, neutron and natural gamma, and the labels include porosity corresponding to the four features;
[0072] The actual logging data is input into the trained multi-mode Bi-LSTM model, and the predicted results of reservoir characteristics are obtained through the trained multi-mode Bi-LSTM model prediction, including:
[0073] Input the actual sound wave, density, neutron and natural gamma into the trained multi-mode Bi-LSTM model, and obtain the predicted result of porosity through the prediction of the trained multi-mode Bi-LSTM model; or,
[0074] The logging data includes features and labels, wherein the features include resistivity, induction logging, spontaneous potential and borehole diameter, and the labels include permeability corresponding to the four features;
[0075] The actual logging data is input into the trained multi-mode Bi-LSTM model, and the predicted results of reservoir characteristics are obtained through the trained multi-mode Bi-LSTM model prediction, including:
[0076] The actual resistivity, induction logging, spontaneous potential and borehole diameter are input into the trained multi-mode Bi-LSTM model, and the predicted results of permeability are obtained through the prediction of the trained multi-mode Bi-LSTM model.
[0077] As a specific implementation, the actual logging data is input into the trained multi-mode Bi-LSTM model, and the predicted results of the reservoir characteristics are obtained through the trained multi-mode Bi-LSTM model prediction, specifically including:
[0078] Randomly distribute the actual logging data to pixels of appropriate scale to form a pixel map;
[0079] The pixel map is input into the trained multi-mode Bi-LSTM model, and the prediction result of the reservoir characteristics is obtained through the prediction of the trained multi-mode Bi-LSTM model.
[0080] As a specific implementation, the pixel map is input into the trained multi-mode Bi-LSTM model, and the predicted result of the reservoir characteristics is obtained through the trained multi-mode Bi-LSTM model, specifically including:
[0081] Feature extraction of pixel images through convolutional neural network;
[0082] Input the extracted features to the Bi-LSTM neural network for processing;
[0083] The weight of each time series is calculated through the Attention layer, and the weighted sum of all time series vectors is used as the feature vector to obtain the prediction result of the reservoir characteristics.
[0084] Please refer to image 3 , the convolutional neural network adopts a two-layer convolutional neural network model. The model has a 4-layer structure, which are the pixel input layer, the first convolutional feature layer, the second convolutional feature layer, and the last traditional neural network output layer. Eight kinds of logging parameters are randomly distributed into the pixel input layer of the double-layer convolutional neural network model, and the size of the input layer is set to twice the number of parameter features, that is, 4*4 pixels. The first convolution feature layer is set with 4 convolution kernels and bias values of 3*3 parameters, and the second convolution feature layer is set with 6 convolution kernels and bias values of 4*2*2, two The feature layer uses the Sigmoid function as the activation function. The output layer takes combined features as input, sets weights and biases to get an output.
[0085] Please refer to Figure 4 , after the pixel image is input to the multi-mode Bi-LSTM model, the two-layer convolutional neural network first extracts the features of the pixel image, and then inputs the extracted features into the Bi-LSTM neural network for processing, and passes through the Attention layer before outputting the result. Calculate the weight of each time series, extract the most important information, and use the weighted sum of all time series vectors as feature vectors to obtain the prediction results of reservoir characteristics. In the multi-mode Bi-LSTM model, the LSTM layer adopts a two-layer bidirectional structure to improve the efficiency and accuracy of time series correlation feature analysis.
[0086] LSTM is an improved neural network to solve the problem that the gradient disappears during RNN backpropagation, which makes it difficult to process longer sequences. Compared with the general RNN model, the LSTM neural network model introduces the cell state, and uses three gates: input gate, forget gate, and output gate to maintain and control information. Bi-directional Long Short-Term Memory (Bi-LSTM) is a recurrent neural network (Recurrent Neural Networks, RNN). Bi-LSTM combines the information of the sequence in both forward and backward directions on the basis of LSTM. For the output at time t, the forward LSTM layer has information at time t in the input sequence and the time t1 before, while the backward LSTM layer has information at time t in the input sequence and the time t+1 after. The vectors output by the two LSTM layers before and after can be processed by adding, averaging or connecting, so as to obtain better sequence processing results.
[0087] The essence of the attention mechanism is to imitate the human visual attention mechanism, learn a weight distribution of image features, and then apply this weight distribution to the original features to provide future tasks such as image classification, image recognition, etc. The influence of different features makes the task mainly focus on some key features, ignoring unimportant features, and improving task efficiency. The attention mechanism has a huge role in improving the sequence learning task. In the framework of the codec, by adding the Attention model in the encoding segment, the data weighted transformation is performed on the source data sequence, or the Attention model is introduced in the decoding end, and the target data is Weighting changes can effectively improve system performance in a sequence-to-sequence natural manner. The purpose of applying the Attention mechanism to the Bi-LSTM neural network in the present invention is to weight the results of the Bi-LTSM and focus on the most important information, thereby making the prediction results more accurate.
[0088]The present invention adopts CNN and Bi-LSTM network combined with the Attention mechanism method, makes full use of the advantages of CNN in identifying the spatial correlation of features, performs convolution and pooling processing on the logging data after image processing, and extracts the data hidden in the data. The feature relationship in the model, and then the Bi-LSTM network makes full use of the deep sequence correlation of the features to make predictions. During the prediction process, Attention integrates the information of multiple feature vectors through weighted summation to make the prediction results more accurate.
[0089] The prediction result of the invention is accurate and has better model generalization ability. Compared with the method of only using CNN for prediction, the advantages of Bi-LSTM in the time correlation processing of features in the method of the present invention make up for the defect of CNN's insufficient accuracy of logging prediction at different depths. At the same time, the present invention fully utilizes the large-scale parallel collaborative processing capability of the artificial neural network, integrates the advantages of different artificial neural networks, and makes up for each other's deficiencies. The CNN model and the multi-mode Bi-LSTM model in this method can adjust the structure of the model itself by adjusting the number of convolutional layers or the number of cycles of the model when the geological conditions of the predicted target change uncertainly, so as to adapt to the specific situation , which enhances the generalization ability of the model. In addition, the CNN and Bi-LSTM networks used in the present invention have the characteristics of fast training speed, and can handle complex geological conditions and a large amount of data with ease.
[0090] A deep learning-based reservoir feature prediction method and model provided by the present invention combines convolutional neural networks (Convolutional Neural Networks, CNN), bidirectional long-short-term memory (Bi-directional LongShort-TermMemory, Bi-LSTM) and attention Attention Mechanism, using the advantages of bidirectional long-short-term memory to efficiently and accurately predict time series, coupled with the attention mechanism layer, makes up for the insufficiency of convolutional neural networks in processing data with serial correlation, and accurately predicts different depths Reservoir characteristics such as porosity and permeability of the reservoir, so as to reasonably and efficiently carry out reservoir transformation and production optimization management, at least solve the problem of low prediction accuracy when using convolutional neural network to predict reservoir characteristics at different depths in the existing technology Low technical issues.
[0091] In summary, the reservoir feature prediction method based on CNN, Bi-LSTM and Attention can make full use of the spatial correlation and time series correlation of reservoir features, identify the correlation between features, and reduce the error; The large-scale parallel collaborative processing capability of the network, taking advantage of the advantages of different neural networks, effectively improves the model generalization ability, establishes an efficient reservoir parameter prediction model, and provides a new idea for the prediction of characteristic parameters of reservoir reconstruction.
[0092] As a specific implementation, before the training set is input into the multimode Bi-LSTM model, it also includes:
[0093] Normalize and normalize the data in the training set. Among them, normalized parameter value = (minimum value of sample parameter of original parameter value)/(maximum value of sample parameter-minimum value of sample parameter), the data can be normalized and standardized by Yeo-Johnson transformation, and its expression is :
[0094]
[0095] As a specific implementation, when obtaining the well logging data training set, the well logging data testing set will also be obtained;
[0096] After obtaining the well logging data test set, it also includes:
[0097] Input the test set into the trained multi-mode Bi-LSTM model, and obtain the prediction result of the reservoir characteristics through the prediction of the trained multi-mode Bi-LSTM model;
[0098] According to the prediction results of the reservoir characteristics and the prediction results of the reservoir characteristics in the test set, the root mean square error and Pearson correlation coefficient are used to evaluate the prediction effect.
[0099] The present invention uses root mean square error RMSE and Pearson correlation coefficient (R) to evaluate prediction effect. Among them, R is a method to measure the correlation between two variables, and its value range is [-1, 1], and the formula is
[0100]
[0101] In the formula: Y and P are the original parameter value and the predicted parameter value respectively; D is the variance; COV(Y, P) is the covariance function, which characterizes the characteristics of the relationship between variable Y and variable P.
[0102] RMSE reflects the deviation between the target porosity and the predicted actual porosity and permeability, and the formula is:
[0103]
[0104] Where: y i and p i are the values of actual porosity, permeability and predicted porosity and permeability, respectively; N represents the number of samples.
Example Embodiment
[0105] Embodiment two
[0106] Please refer to Figure 1 to Figure 4 , the present invention also provides a reservoir feature prediction model based on deep learning, comprising:
[0107] The training set acquisition module is used to acquire the logging data training set;
[0108] The multi-mode Bi-LSTM model generation module is used to construct a convolutional neural network and an Attention-Based Bi-LSTM neural network with an Attention layer, and combine the two for sequence modeling to generate a multi-mode Bi-LSTM model;
[0109] The joint optimization module is used to input the training set into the multi-mode Bi-LSTM model, and the multi-mode Bi-LSTM model is trained by means of joint training, and the convolutional neural network and the Attention-Based Bi-LSTM neural network are trained. The parameters are jointly optimized;
[0110] The prediction module is used to input the actual logging data into the trained multi-mode Bi-LSTM model, and obtain the prediction result of the reservoir characteristics through the prediction of the trained multi-mode Bi-LSTM model.
[0111] As a specific implementation, the logging data includes features and labels, wherein the features include acoustic wave, density, neutron and natural gamma, and the labels include porosity corresponding to the four features;
[0112] The prediction module is specifically used for:
[0113] Input the actual sound wave, density, neutron and natural gamma into the trained multi-mode Bi-LSTM model, and obtain the predicted result of porosity through the prediction of the trained multi-mode Bi-LSTM model; or,
[0114] The logging data includes features and labels, wherein the features include resistivity, induction logging, spontaneous potential and borehole diameter, and the labels include permeability corresponding to the four features;
[0115] The prediction module is specifically used for:
[0116] The actual resistivity, induction logging, spontaneous potential and borehole diameter are input into the trained multi-mode Bi-LSTM model, and the predicted results of permeability are obtained through the prediction of the trained multi-mode Bi-LSTM model.
[0117] As a specific implementation, the prediction module is specifically used for:
[0118] Randomly distribute the actual logging data to pixels of appropriate scale to form a pixel map;
[0119] The pixel map is input into the trained multi-mode Bi-LSTM model, and the prediction result of the reservoir characteristics is obtained through the prediction of the trained multi-mode Bi-LSTM model.
[0120] As a specific implementation, the pixel map is input into the trained multi-mode Bi-LSTM model, and the predicted result of the reservoir characteristics is obtained through the trained multi-mode Bi-LSTM model, specifically including:
[0121] Feature extraction of pixel images through convolutional neural network;
[0122] Input the extracted features to the Bi-LSTM neural network for processing;
[0123] The weight of each time series is calculated through the Attention layer, and the weighted sum of all time series vectors is used as a feature vector, and the prediction result of the reservoir characteristics is obtained through the multi-mode Bi-LSTM model prediction.
[0124] As a specific implementation, before the training set is input into the multimode Bi-LSTM model, it also includes:
[0125] Normalize and normalize the data in the training set.
[0126] As a specific implementation, when obtaining the well logging data training set, the well logging data testing set will also be obtained;
[0127] After obtaining the well logging data test set, it also includes:
[0128] Input the test set into the trained multi-mode Bi-LSTM model, and obtain the prediction result of the reservoir characteristics through the prediction of the trained multi-mode Bi-LSTM model;
[0129] According to the prediction results of the reservoir characteristics and the prediction results of the reservoir characteristics in the test set, the root mean square error and Pearson correlation coefficient are used to evaluate the prediction effect.
[0130] As for the specific implementation process of the second embodiment, since the method of the first embodiment has been described in detail, it will not be repeated here.
[0131] Those of ordinary skill in the art can understand that all or part of the steps in the realization of the above facts and methods can be completed by instructing related hardware through programs, and the related programs or the programs can be stored in a computer-readable storage medium , when the program is executed, it includes the following steps: at this time, the corresponding method steps are drawn, and the storage medium can be ROM/RAM, magnetic disk, optical disk, etc.
PUM


Description & Claims & Application Information
We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.