Formation pressure inversion method and apparatus based on multivariate time series

Abstract

The present invention discloses a formation pressure inversion method based on multivariate time series, including: collecting multivariate time series feature data of drilling, and performing convolution and mapping processing to obtain a high-level time series feature matrix, and then splicing it with the wellbore geometry constraint feature matrix to obtain an attribute feature matrix; using a graph adaptive learning algorithm to automatically learn the optimal adjacency matrix; using graph convolution to obtain the hidden associations between the nodes of the attribute feature matrix and the adjacency matrix, and obtain a deep attribute feature matrix; using a multilayer perceptron and combining the deep attribute feature matrix to predict formation pressure. The present invention also discloses a device for formation pressure inversion based on multivariate time series. Through the above scheme, the present invention has the advantages of simple logic, accuracy and reliability, and has high practical value and promotion value in the field of drilling technology.

Description

Technical Field

[0001] The invention relates to the technical field of drilling, and in particular to a formation pressure inversion method and device based on multivariate time series. Background Art

[0002] There is growing interest in cleaner underground energy sources around the world. For example, renewable geothermal resources and natural gas. In 2021, global geothermal power generation increased by 1.6GW, while natural gas consumption increased by 4.6%. In order to meet the growing demand, more wells need to be drilled. It is worth noting that the risks of well kicks and blowouts are fully considered during the drilling of geothermal resources and unconventional natural gas. A large part of geothermal resources is hydrothermal fluids. High-temperature, high-pressure water or steam stored in geothermal reservoirs often leads to abnormally high pore pressures, indicating a huge risk of well kicks. As the focus of natural gas extraction is on unconventional natural gas (such as shale gas and tight gas), the target reservoirs are deeper and more complex. In order to increase natural gas production, the length of horizontal wells is increasing. For example: the average depth of the horizontal wells recently drilled in the Sichuan Basin exceeds 5000m (1800m+horizontal section). The increase in depth, as well as the complex underground conditions, leads to a greater probability of formation fluids invading the wellbore during drilling, and overflow accidents occur from time to time, which is difficult to completely avoid.

[0003] Overflow refers to the phenomenon that when the formation pressure is greater than the bottom hole pressure, formation fluids such as oil, gas and water invade the wellbore, the drilling fluid outlet flow is greater than the inlet flow, and the fluid in the wellbore automatically overflows from the wellhead. Once overflow occurs, if well control measures are not taken or adjusted in time, it will evolve into serious accidents such as well kick and blowout. Rapid and accurate acquisition of formation abnormal forces can effectively improve well kick margin and reduce well control risks. Accurately obtaining abnormal formation pressure is the primary prerequisite for subsequent overflow treatment. After overflow, there are two traditional methods for obtaining abnormal formation pressure: the first is to determine the formation pressure by reading the casing pressure and riser pressure after shutting in the well, which is simple and reliable. However, during the overflow process, due to the decrease in formation pressure in the near-wellbore area, the bottom hole pressure is not equal to the formation pressure at the initial stage of overflow shut-in. After shutting in the well, it is necessary to wait for a period of time until the bottom hole pressure returns to the formation pressure. For formations with good permeability, it takes 10-15 minutes from shutting in the well to the formation pressure stabilizing, but for formations with poor permeability, it takes longer to wait. For overflow caused by gas invasion (most overflows are caused by gas invasion), long-term well shut-in waiting will cause the gas to slip upward to a greater height, further reducing the hydrostatic column pressure of the mud in the wellbore, and ultimately leading to excessive wellhead pressure. When dealing with overflow, the kick margin is small, which increases the risk of well control. Another method, numerical simulation, studies the characteristics of gas-liquid two-phase flow during gas invasion, establishes a gas-liquid two-phase flow model, and simulates the overflow process. This type of model is mostly composed of continuity equations and mixed momentum equations for each phase. With given initial conditions and boundary conditions, the model is solved by the finite interpolation method. In the actual formation pressure prediction, this method assumes the initial bottomhole negative pressure difference and gas invasion amount to fit the boundary conditions. The boundary conditions are mostly determined by the wellhead information. Finally, the formation pressure is determined based on the reasonable bottomhole negative pressure difference. This method has a large amount of calculation and an extremely complex calculation process. It needs to calculate a large number of intermediate processes and use a large number of intermediate parameters. The deviation of the intermediate process calculation and the value of the intermediate parameters have a great influence on the calculation results. The accuracy of numerical simulation methods depends greatly on the setting of initial conditions, boundary conditions and parameter selection in the model. After overflow occurs, rapid and accurate prediction of abnormal formation pressure is crucial to ensure well control safety. Therefore, it is necessary to study the method of rapid and accurate prediction of abnormal formation pressure in view of the problems of long time, complex calculation and low accuracy of obtaining abnormal formation pressure by traditional methods.

[0004] With the continuous development of machine learning, a large number of machine learning technologies have been successfully applied in the oil industry. The end-to-end computing characteristics of machine learning provide a new research direction for predicting abnormal formation pressure. At present, Yan Tie et al. have established a model for predicting abnormal formation pressure based on the LSTM (Long Short-Term Memory) algorithm to capture the law of changes in the mud pool increment over time after overflow, and have achieved good results. However, there are two key factors that hinder the application of this model in actual drilling. First, the data source of the training model is artificial data. The time series data of the mud pool increment over time under different formation pressures obtained through the numerical model is quite different from the field data. Secondly, only the mud pool increment (Pit gain) is used to predict abnormal formation pressure. In the actual drilling process, due to the complex formation conditions and poor on-site construction environment, certain errors will occur during tool measurement, which increases the difficulty of predicting abnormal formation pressure. The prediction of abnormal formation pressure based on a single factor is greatly affected by error noise. This technology takes into account that during the overflow period, parameters such as mud pool increment, wellhead riser pressure, and inlet and outlet flow rate difference will be affected. The sequence of these parameters changing over time contains information about abnormal formation pressure. At the same time, this technology also takes into account that the changes in these parameters are also constrained by factors such as wellbore trajectory and wellbore size. Therefore, it is not accurate enough to predict abnormal formation pressure only through time series characteristic parameters, and the nonlinear relationship between the geometric constraints of the wellbore itself and the time series characteristics must be added to the model. Therefore, after determining that an overflow has occurred, the time series characteristics of parameters such as mud pool increment and wellhead riser pressure after the overflow point can be extracted, combined with the geometric constraints of the wellbore itself, and the abnormal formation pressure can be predicted based on the machine learning algorithm.

[0005] There are two important challenges in predicting abnormal formation pressure based on machine learning algorithms: the extraction of multivariate time series features and how to train a reliable model under small sample conditions. The traditional RNN model and its variants (LSTM, GRU) have achieved great success in time series modeling. Shi Yu proposed a combined neural network of LSTM+MLP, which combines time series features with physical constraint features. LSTM is used to extract the sequence relationship of production data, and MLP is used to learn the nonlinear relationship between geothermal capacity and constraints. The model has achieved good results in predicting geothermal capacity. However, there are two main factors that hinder the application of RNN in formation pressure prediction. First, the RNN network has high requirements on the quality and quantity of data, but even in the middle and late stages of block development, this technology still has difficulty in providing a large number of high-quality overflow samples. Secondly, although the emergence of LSTM and GRU has improved the problem of gradient vanishing of the RNN model when processing longer time series, the problem still exists. With the advancement of technology, a set of data is obtained in about 2 seconds during the logging process, which means that this technology needs to extract features from multiple longer time series, and it is difficult to train a reliable model through RNN. CNN (Convolutional Neural Networks) has strong feature extraction and generalization capabilities due to its local connection, weight sharing and multi-level results, and has achieved great success in the field of computer vision. In recent years, some researchers have applied CNN to the field of NLP (Natural Language Processing), and have also achieved surprising results. For most NLP tasks, they are completed by processing sequence problems, with the input being sentences in the form of matrices, and each row of the matrix corresponds to a word vector. The task of extracting multivariate time series features in predicting abnormal formation pressure is very similar to the task of NLP. Therefore, CNN can also be used to extract time series of parameters such as mud pool increments.

[0006] From the perspective of machine learning, the prediction of abnormal formation pressure after overflow is essentially a regression problem. Due to the expressive power and arbitrary nonlinearity of neural networks, they are extremely competitive in regression tasks. However, due to the number of overflow samples, the limited amount of information is not enough to support the neural network to infer model parameters, and it is difficult to train a high-performance abnormal formation pressure prediction model. In recent years, due to its permutation invariance, local connectivity and combinatoriality, the graph neural network (GNN) has shown a strong ability in processing relational dependencies (graphs are a special form of data that describes the relationship between different entities). Graph neural networks are widely used. Based on the perspective of the graph, this technology can naturally combine multivariate time series features with geometric constraints and observe their relationship with abnormal formation pressure. At the same time, by propagating information in the graph structure, the graph neural network allows each node in the graph to know its neighborhood information. By aggregating the information of its neighboring nodes to the target node, the information richness of the data set and the diversity of samples can be effectively increased. Thanks to the full use of information, downstream tasks can be well completed even under the condition of small samples.

[0007] In view of the problem that the traditional method of obtaining abnormal formation pressure after overflow takes a long time and has low accuracy, it is urgent to propose a formation pressure inversion method based on multivariate time series with simple logic, accuracy and reliability. Summary of the invention

[0008] In view of the above problems, the purpose of the present invention is to provide a formation pressure inversion method based on multivariate time series. The technical solution adopted by the present invention is as follows:

[0009] In the first part, the present technology discloses a formation pressure inversion method based on multivariate time series, which includes the following steps:

[0010] Collect multivariate time series feature data of drilling, perform convolution and mapping processing to obtain an advanced time series feature matrix, and combine it with the wellbore geometric constraint features to obtain an attribute feature matrix;

[0011] Adopt graph adaptive learning algorithm to adaptively learn the adjacency matrix;

[0012] Use graph convolution to obtain the hidden associations between nodes in the attribute feature matrix and the adjacency matrix, and obtain the deep attribute feature matrix;

[0013] The formation pressure is predicted using a multi-layer perceptron combined with a deep attribute feature matrix.

[0014] In the second part, the present technology discloses a device for formation pressure inversion based on multivariate time series, which includes:

[0015] The time convolution module performs convolution and mapping processing based on the collected multivariate time series feature data to obtain an advanced time series feature matrix, and combines it with the wellbore geometric constraint features to obtain an attribute feature matrix;

[0016] Graph adaptation module, adaptively learning the adjacency matrix;

[0017] The graph convolution module is connected to the graph adaptation module to obtain the hidden associations between the nodes of the attribute feature matrix and the adjacency matrix, and output the deep attribute feature matrix;

[0018] And, the multi-layer perceptron output module is connected with the graph convolution module to predict the formation pressure according to the deep attribute feature matrix.

[0019] In the third part, the present technology provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a formation pressure inversion method based on a multivariate time series when executing the computer program.

[0020] In the fourth part, the present technology provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of a formation pressure inversion method based on a multivariate time series are implemented.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) The present invention ingeniously constructs a new model to invert abnormal formation pressure. The model consists of four modules: a time convolution module is used to extract the time series characteristics of six parameters such as mud pool increment, and then splice them with geometric constraint features to obtain an attribute feature matrix; a graph adaptive learning module is used to automatically learn the graph structure; a graph convolution module is used to extract hidden associations between nodes; and an output module is used for the final prediction task.

[0023] (2) The present invention ingeniously performs convolution and mapping processing on multivariate time series feature data to obtain a high-level time series feature matrix. The advantage is that the extracted high-level time series features are of higher quality and have stronger representation capabilities.

[0024] (3) Based on the perspective of graph, the present invention regards overflow as a node, and cleverly splices the multivariate time series features with the wellbore geometric constraint features as the attribute feature matrix of the node. The advantage of this is that the model can be trained based on the information transmission method of the graph.

[0025] (4) The present invention cleverly uses graph convolution to obtain the hidden associations between nodes in the attribute feature matrix and the adjacency matrix, and obtains the deep attribute feature matrix. The advantage is that by aggregating the information of the neighboring nodes of the target node, the information richness of the data set and the diversity of samples can be effectively increased. Thanks to the full utilization of information, even under the condition of small samples, downstream tasks can be well completed.

[0026] (5) The present invention uses a multi-layer perceptron and combines it with a depth attribute feature matrix to predict formation pressure. The advantage is that the model has high prediction accuracy and good robustness.

[0027] In summary, the present invention has the advantages of simple logic, accuracy and reliability, and has high practical value and promotion value in the field of drilling technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope of protection. For those skilled in the art, other related drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 It is a logic flow chart of the present invention.

[0030] Figure 2 It is a schematic diagram of the working of the temporal convolution module of the present invention.

[0031] Figure 3 Schematic diagram of the graph adaptive learning module of the present invention.

[0032] Figure 4 Schematic diagram of the workflow of the graph convolution module and output module of the present invention.

[0033] Figure 5 This is a schematic diagram of the shale gas reservoir location of the present invention.

[0034] Figure 6 The following is a typical geological layering and well structure diagram of the target reservoir of the present invention.

[0035] Figure 7 This is a profile diagram of the formation pore pressure in the well section 3218-3696m of the well of the present invention.

[0036] Figure 8 This is the relative error distribution diagram of the TGP model test of the present invention.

[0037] Fig. 9 The RMS plot was tested for the TGP model of the present invention.

[0038] Fig.10 The MAPE plot was tested for the TGP model of the present invention.

[0039] Fig.11 It is the relative error diagram of the CNN+MLP model of the present invention.

[0040] Fig.12 It is the adjacency matrix graph of the present invention.

[0041] Fig.13 This is a diagram of the LSTM network architecture used in the test of the present invention.

[0042] Fig.14 Relative error graph for the LSTM network test of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of this application clearer, the present invention is further described below in conjunction with the accompanying drawings and embodiments, and the embodiments of the present invention include but are not limited to the following embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0044] Example

[0045] like Figures 1 to 14 As shown, the present embodiment provides a formation pressure inversion method and device based on multivariate time series. First, in order to capture the key information contained in the multivariate time series, the time convolution module performs convolution and mapping operations on the original multivariate time series feature data to obtain a high-level time series feature matrix, which will be used as the input of the graph adaptive module. Subsequently, in the graph adaptive learning module, the high-level time series features and the wellbore geometric constraint features are spliced ​​into an attribute feature matrix and then output, and an adaptive adjacency matrix is ​​output at the same time. The attribute feature matrix and the adjacency matrix are used as inputs to the graph convolution module, and the hidden associations between nodes are obtained through graph convolution. Finally, the output module composed of MLP (multi-layer perceptron) will predict the formation pressure based on the deep attribute feature matrix output by the graph convolution module.

[0046] The temporal convolution module of this embodiment applies multiple one-dimensional convolution kernels of different sizes to extract high-level temporal features. The difference between a convolutional neural network and an ordinary neural network is that a CNN includes a feature extractor consisting of a convolution layer and a subsampling layer (pooling layer). In the convolution layer of a CNN, a neuron is only connected to some neurons in the adjacent layers, and in a convolution layer, several feature maps (featureMap) are usually included, each feature map is composed of some rectangularly arranged neurons, and the neurons of the same feature map share weights, and the shared weights here are the convolution kernels. The convolution kernel is generally initialized in the form of a random decimal matrix, and the convolution kernel will learn to obtain reasonable weights during the training process of the network. The direct benefit of sharing weights (convolution kernels) is to reduce the connection between the layers of the network, while reducing the risk of overfitting. Subsampling is also called pooling, which usually has two forms: mean subsampling and maximum subsampling. Subsampling can be regarded as a special convolution process. Convolution and subsampling greatly simplify the complexity of the model and reduce the parameters of the model.

[0047] By using multiple convolution kernels of different sizes, it is possible to effectively process longer time series, avoid the problem of gradient vanishing, and capture time series features in different ranges. The workflow of the temporal convolution module constructed in this embodiment is as follows: Figure 2 shown.

[0048] Figure 2 In, x ij is the jth time series feature in the i-th sample. The total length of the time series is T, t i is the i-th time step. First, four convolution kernels with different lengths are used, and the lengths of the convolution kernels are f1=1×3, f2=1×5, f3=1×6, and f4=1×7, respectively, to convolve the original time series feature data. After the convolution is completed, time series features in different ranges are obtained. Subsequently, the convolution results of the four convolution kernels are spliced. This technology performs two splicing operations. The first time, the same time series feature x is concatenated. ij The convolution results of the corresponding different convolution kernels are spliced. After the first splicing is completed, the corresponding time t i The convolution results of different convolution kernels are averaged to get x ij [t i ] The second time, this technique will get the sample x ij The original time series samples are reassembled according to their order to obtain the convolved time series feature matrix. Finally, the new time series feature matrix is ​​mapped to the specified size to obtain the high-level time series feature matrix.

[0049] In this embodiment, the graph adaptive learning module aims to automatically learn the connection relationship between nodes. Specifically, instead of using a predefined graph structure, the relationship between nodes in the calculation graph is calculated by embedding nodes to represent the graph structure, and the node representation is optimized by backpropagation based on the loss of the predicted target, thereby changing the graph structure and realizing a data-driven graph structure definition. Compared with the traditional graph structure given in advance after measuring the similarity between nodes by distance (dot product, Euclidean distance, etc.), graph adaptive learning can obtain a better graph structure.

[0050] At present, a variety of graph adaptive learning strategies have been proposed in the prior art. The commonly used methods can be summarized as follows: initialize node embedding representation, calculate the relationship between nodes through vectors between nodes, obtain a parameterized adjacency matrix, and continuously "adjust" the node representation in the back propagation based on gradient descent optimization, thereby changing the adjacency matrix and finally converging it to a stable state, such as Figure 3 shown.

[0051] In this study on predicting abnormal formation pressure, overflow events are regarded as nodes, and the connections between nodes are undirected. Therefore, the adaptive adjacency matrix calculation formula of this technology is:

[0052] A adp =soft max (ReLU(EE T ))

[0053] Among them, A adp is the adaptively learned adjacency matrix, and E is the embedding of randomly initialized nodes. By replacing E and E T Multiply them together to get the weight between the source node and the target node. Then use the ReLU activation function to eliminate weak connections. Apply the SoftMax function to normalize the adaptive adjacency matrix.

[0054] In this embodiment, the graph convolution module aims to converge the node information with the information of its neighbors, obtain the hidden associations between nodes, and make full use of the information in the feature matrix. Although, this technology can directly predict the labels of unknown nodes through MLP after splicing the attribute feature matrix. But at this time, each node only contains its own information, and the potential associations between nodes are lost. In other words, the nodes adjacent to the unknown nodes, relatively speaking, have a greater contribution to the prediction of the labels of the unknown nodes, but are still regarded as equally important as other non-adjacent nodes, which is obviously unreasonable. The purpose of the graph convolution module can be understood as further extracting features and increasing the abundance of data information and the diversity of samples in the feature matrix. This embodiment uses GCN to implement this idea.

[0055] GCN (Graph Convolutional Networks) is a graph convolutional neural network based on the spectral domain. After considering the attributes of the node itself and the adjacent nodes in the graph structure data, GCN performs weighted summation on the nodes around the node to obtain the new feature vector of the node, and then completes the feature extraction after passing the activation function. The input of each convolution (hidden layer) of GCN is the adjacency matrix A of the graph and the characteristic attribute H of the node. (l) , the output is the new node feature attribute H (l+1) , in the initial input layer feature attribute H 0 It is the feature matrix X. After integrating the characteristic attributes of the central node itself and the node normalization and smoothing, the final GCN convolution kernel expression is:

[0056]

[0057] In the formula, A N =A+I, I is the unit matrix (add self-loop), D N A N , H is the feature of each layer, W is the weight of each layer, σ is the nonlinear activation function. In practical applications, this technology uses the ReLU function.

[0058] This embodiment does not directly use GCN to complete the prediction of unknown node labels, but constructs an MLP separately for the final abnormal formation pressure prediction, such as Figure 4 shown.

[0059] In this example, the on-site drilling data of a shale gas reservoir in eastern Sichuan ( Figure 5 ) as the original data. Geologically, it is located in the compression-slip zone. The overall strata are relatively flat, mainly Jurassic and Cretaceous strata. Typical geological stratification and well structure of the target reservoir are as follows: Figure 6 As shown. The buried depth is 3750-3950 meters (Longma Creek). In order to increase the contact area with the reservoir, all deployed wells are horizontal wells. As of June 2021, 32 horizontal wells have been drilled, and more wells are planned in the next few years. So far, the average MD is about 5500m, and the average length of the horizontal section is about 1800m.

[0060] As of June 2021, a total of 112 overflow accidents occurred in the YX block, with a high frequency of overflows. The overflow layers are mainly concentrated in the Jialingjiang Formation, Feixianguan Formation and Longmaxi Formation, causing a cumulative loss of more than 227.9 days of drilling time. Figure 7 The figure shows the formation pressure profile of the 3218-3696 m section of Well L202 calculated by the Eaton method.

[0061] Figure 8The first track is the natural gamma curve, the second track is the P-wave time difference curve, the third track is the well inclination curve, the fourth track is the well diameter curve, and the fifth track is the formation pressure profile. It can be seen that the formation pore pressure of the 3218-3696m well section is between 57 and 77MPa, and the formation pore pressure gradient is between 1.80 and 2.15g / cm3, which is a typical abnormally high-pressure formation.

[0062] This technology constructs the original time series feature dataset Dataset_T and geometric constraint dataset Dataset_P based on 100 groups of overflow samples in the YX block (the remaining overflow sample data are too missing) for training and testing the model.

[0063] In this embodiment, Relative Error (RE), RMSE (Root mean square error) and MAPE (Mean absolute percentage error) are used to evaluate the performance of the model.

[0064] In this embodiment, considering that the number of overflow samples is small, this embodiment simplifies the structure of each module as much as possible when constructing the TGP model to avoid the phenomenon of underfitting of the model. In this test, the structure of the TGP model is shown in Table 1.

[0065] Table 1 TGP model structure and some parameters used for testing

[0066]

[0067]

[0068] In the temporal convolution module, 8 convolution kernels of each of the four lengths are used. After two convolution operations, a maximum pooling operation is performed. In the graph convolution layer, two graph convolutions are performed to aggregate the information of one-hop and two-hop neighbors to each node. In the output module, a 3-layer fully connected neural network is used, and the number of neurons in the hidden layer is set to 16. This structure is the optimal structure obtained after multiple tests.

[0069] In this embodiment, 80% of the data is used for training, and 20% of the data is used for testing. Since the data set is randomly divided in each test, the error generated by the model each time is different. In order to fully test the performance of the model, this technology has been tested 10 times. The test results are shown in Table 2.

[0070] Table 2 TGP model test results

[0071]

[0072]

[0073] According to the prediction results of the model, the changes of various indicators are as follows Figure 8-Figure 10 shown.

[0074] Depend on Figure 8 It can be seen that in the 10 tests, the maximum relative error produced by the TGP model is 9.89%, and the maximum relative errors produced in the 10 tests are all less than 10%. Most of the relative errors in the 10 tests are concentrated between 2% and 6%. Fig.10 , Fig.12 It can be seen that in 10 tests, the minimum RMSE is 0.063, the maximum is 0.11, and the average is 0.09. The minimum MAPE is 2.5, the maximum is 4.7, and the average is 3.9. The test results show that the TGP model has extremely high prediction accuracy, good robustness, and strong generalization ability. In general, the TGP model shows satisfactory performance and can be applied to the actual prediction of abnormal formation pressure.

[0075] Through the ablation experiment of the TGP model, the individual contributions of graph adaptive learning and graph convolution were evaluated. When the graph structure adaptive learning module is not included in the model, the structure of the graph is calculated by Euclidean distance. When the calculation result is less than 1.14 (average value), there are edge connections between nodes. When the model lacks the GCN module, the GAL module does not exist, and the high-level time series features and wellbore geometric constraint features are spliced ​​after the convolution is completed. Similarly, this technology was tested 10 times, and the calculation results of each model are shown in Table 3.

[0076] Table 3 Ablation experiment results

[0077]

[0078] As shown in Table 3, when the model lacks the GAL module, the maximum relative error of the model increases from 9.89% to 18.5%, and both RMSE and MAPE increase to a certain extent. When the TGP model lacks the GCN module, the error generated by the model increases significantly, and the maximum relative error increases from 9.89% to 46.1%, and both RMSE and MAPE increase significantly. Next, this technology analyzes the reasons for these phenomena.

[0079] When the GCN model is missing from the model, the overall performance of the model drops significantly, but interestingly, the test results of some groups are not much different from the TGP model. The relative error of CNN+MLP in 10 tests is shown in Figure 11.

[0080] Depend on Fig.11It can be seen that the performance of CNN+MLP in Test 2, Test 4 and Test 8 is not much different from that of the TGP model, but in the tests of other groups, the relative error is significantly increased. This shows that the robustness of the model is poor, that is, the generalization ability is insufficient. The fundamental reason for this result is that the number of overflow samples is insufficient, and the model is underfitting. This shows that the information propagation method based on the graph can increase the information abundance and sample diversity in the data set by convolution, thereby improving the performance of the classifier.

[0081] When the GAL module is missing from the model, the performance of the model deteriorates. This technique compares the graph adjacency matrix (with self-loops) calculated based on Euclidean distance with the graph structure (with self-loops) learned by GAL. Fig.12 shown.

[0082] As can be seen from the figure, the values ​​in the critical matrix calculated by Euclidean distance are 0 or 1, while the values ​​in the critical matrix calculated by GAL are between [0,1]. In other words, the adjacency matrix calculated by GAL can assign weights to edges. Compared with the strict connection relationship between nodes calculated by Euclidean distance, GAL can assign more information to nodes, further increasing the diversity of samples in the feature matrix.

[0083] In this embodiment, several methods are also listed for comparison:

[0084] In order to verify the superiority of this technology, this embodiment is compared with the mainstream LSTM method. This embodiment constructs four LSTM networks and sets the time steps to 3, 5, 7, and 9 respectively to capture multivariate time series features in different ranges. After obtaining the time series features, a single LSTM network will be mapped to a 1×6 vector. Then the four mapped vectors are merged to obtain the final time series feature sample, such as Fig.13 shown.

[0085] The LSTM model was tested 10 times, and the test results are compared with the TGP model in Table 4.

[0086] Table 4 Comparison between TGP model and LSTM

[0087]

[0088] As shown in Table 4, the maximum relative error of the LSTM model in the test reached an astonishing 94.7%, which is 9.5 times the maximum relative error of the TGP model. The average RMSE and MAPE of the 10 tests were 0.5 and 21.74, respectively, which are 5.5 and 5.6 times that of the TGP model. This shows that the performance of the TGP model is significantly better than that of the LSTM model.

[0089] The relative error of the LSTM model in 10 tests is as follows Fig.14 As shown. Fig.14 It can be seen that, similar to the CNN+MLP model, the LSTM model also has relatively good prediction accuracy in some groups of tests. For example, in Test 4, the minimum relative error is 5.23, the maximum relative error is 19.6, and the average relative error is 13.3. However, the robustness of the model is very poor and the generalization ability is insufficient. The main reason for this phenomenon is that the amount of data is insufficient, resulting in underfitting of the model. On the other hand, by comparing with the CNN+MLP model, it can be seen that the LSTM model is not as good as CNN in extracting multivariate time series features.

[0090] After the overflow occurs, the rapid and accurate prediction of abnormal formation pressure is the basis for formulating well control measures and is crucial to ensuring well control safety. In this work, this technology constructs an intelligent model for end-to-end rapid inversion of abnormal formation pressure. This technology first extracts the high-level time series features of the riser pressure, mud pool increment, drilling time, inlet and outlet temperature difference, inlet and outlet flow difference, and inlet and outlet conductivity difference over time after the overflow through one-dimensional convolution. This technology uses multiple convolution kernels of different lengths to specifically complete this task, and the extracted high-level time series features contain time series features in different ranges. Then, based on the perspective of the graph, the overflow event is regarded as a node, and the high-level time series features extracted by convolution are concatenated with the wellbore constraints as the attribute vector of the node. The optimal connection relationship between nodes is obtained through the adaptive graph structure learning module. Based on the information transmission method of the graph, the information of the two-hop neighbors is aggregated to each node in the form of graph convolution to obtain a deep feature matrix. Finally, a 3-layer MLP predicts the abnormal formation pressure of the unknown node based on the deep feature matrix.

[0091] This technology uses the real overflow data of the YX block to train and test the TGP model, selects relative error, RMSE and MAPE to evaluate the model performance, and compares it with the combined neural network model of LSTM+MLP. In 10 tests, the maximum relative error generated by the TGP model was 9.89%, the average RMSE was 0.09, and the average MAPE was 3.9. Compared with the LSTM+MLP model, the maximum relative error was reduced by 93.7%, the RMSE was reduced by 82%, and the MAPE was reduced by 82%. The TGP model has good prediction accuracy and robustness in the task of abnormal formation pressure prediction, and shows strong competitiveness compared with other methods.

[0092] This technology introduces one-dimensional convolution and graph learning into the inversion of abnormal formation pressure, providing a new direction for predicting abnormal formation pressure and achieving good results. However, there are still several issues that need further research. First, in this study, the number of overflow samples is fixed, that is, in the graph structure adaptive learning, the connection relationship between nodes is static learning. This technology has not yet considered how to update the existing model and make predictions for new nodes when new nodes are added. This will be the next research direction of this technology. Another point is that the optimal architecture of the TGP model is obtained after multiple tests. This technology does not analyze the parameters of the model because it is not clear about the specific connection between the optimal architecture and the task of predicting abnormal formation pressure. In further research, this technology is prepared to collect data from more blocks and analyze the optimal architecture of the TGP model through the structure of different block data sets.

[0093] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any changes that adopt the design principles of the present invention and are made through non-creative work on this basis should fall within the protection scope of the present invention.

Claims

1. A formation pressure inversion method based on multivariate time series, characterized in that: The following steps are involved: Collect multivariate time series feature data of drilling, perform convolution and mapping processing to obtain a high-level time series feature matrix, and then combine it with the wellbore geometry constraint feature matrix to obtain an attribute feature matrix; Adopt graph adaptive learning algorithm to automatically learn the optimal adjacency matrix; Use graph convolution to obtain the hidden associations between nodes in the attribute feature matrix and the adjacency matrix, and obtain the deep attribute feature matrix; Formation pressure prediction is performed using a multi-layer perceptron combined with a deep attribute feature matrix; The wellbore geometric constraint characteristics include well depth, well inclination, azimuth and drill bit size; The multivariate time series characteristics include high-level time series characteristics of riser pressure, mud pool increment, drilling time, inlet and outlet temperature difference, inlet and outlet flow difference and inlet and outlet conductivity difference changing with time after overflow.

2. The formation pressure inversion method based on multivariate time series according to claim 1, characterized in that: The multivariate time series feature data is convolved and mapped using a time convolution module; the time convolution module contains a feature extractor consisting of a convolution layer and a subsampling layer; the subsampling layer is a pooling layer.

3. The formation pressure inversion method based on multivariate time series according to claim 1, characterized in that: The expression of the adjacency matrix is: YOUR adp =softmax(ReLU(EE T )) Among them, Aadp represents the adaptively learned adjacency matrix, E represents the embedding of randomly initialized nodes, and T represents matrix transpose.

4. The formation pressure inversion method based on multivariate time series according to claim 3 is characterized in that: The graph convolution adopts a graph convolutional neural network based on the spectral domain, and its convolution kernel expression is: A N =A+I Among them, I represents the identity matrix, DN represents the degree matrix of AN, H represents the features of each layer, W represents the weight of each layer, and σ is the nonlinear activation function.

5. The formation pressure inversion method based on multivariate time series according to claim 1, characterized in that: Graph convolution is used to obtain the hidden associations between the nodes of the attribute feature matrix and the adjacency matrix, including: based on the information transmission method of the graph, the information of the two-hop neighbors is aggregated to each node in the form of graph convolution, and the attribute feature vector of each node is updated.

6. A device for formation pressure inversion based on multivariate time series, characterized in that: include: The time convolution module performs convolution and mapping processing based on the collected multivariate time series feature data to obtain an advanced time series feature matrix, and combines it with the wellbore geometric constraint features to obtain an attribute feature matrix; Graph adaptation module, adaptively learning the adjacency matrix; The graph convolution module is connected to the graph adaptation module to obtain the hidden associations between the nodes of the attribute feature matrix and the adjacency matrix, and output the deep attribute feature matrix; And, the multi-layer perceptron output module is connected with the graph convolution module to predict the formation pressure according to the deep attribute feature matrix.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the formation pressure inversion method based on multivariate time series described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the formation pressure inversion method based on multivariate time series described in any one of claims 1 to 6 are implemented.

Images