A horizontal well production prediction method based on an extended long short-term memory neural network

Abstract

The present invention discloses a horizontal well production prediction method based on an extended long short-term memory neural network, which relates to the technical field of production prediction. The method includes the following steps: obtaining the production data of the developed horizontal wells and the characteristic parameters composed of the data affecting the production; decomposing the characteristic parameters at two levels in the time domain and the frequency domain based on empirical mode decomposition and wavelet transform to obtain a three-dimensional feature tensor; respectively decomposing the three-dimensional feature tensor based on TCN and STL decomposition, and fusing the decomposition results through a dynamic feature selection gate to obtain a fused feature; establishing a prediction model, and training the prediction model based on the fused feature, and the trained model can perform production prediction; the main body of the prediction model is a bidirectional LSTM model, using a spatio-temporal attention mechanism to provide weights for the fused feature, and using an optimized sparrow search algorithm to select better hyperparameters for the bidirectional LSTM model. The method of the present invention can effectively predict the production of horizontal wells, and the prediction result has a high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of production prediction, and specifically to a horizontal well production prediction method based on an extended long short-term memory neural network. Background Art

[0002] In the field of oil and gas exploitation, the production prediction of horizontal wells has always been a challenging task because it is affected by various factors, such as reservoir characteristics, wellhead pressure, liquid flow, etc. With the continuous progress of data acquisition technology and information means, how to accurately predict the production of horizontal wells through historical data has become one of the keys to oilfield management and decision-making optimization. Traditional production prediction methods often rely on physical models or statistical methods and are difficult to deal with complex non-linear relationships and high-dimensional data problems. Especially when applied to the production prediction of tight oil, due to the great differences between tight oil reservoirs and conventional oil reservoirs, the application of traditional methods is more difficult and the accuracy is lower. Therefore, methods based on machine learning and deep learning have gradually been widely used.

[0003] As a deep learning model for processing time series data, the long short-term memory (LSTM) neural network has been widely used in various prediction tasks, but its application in oilfield production prediction still faces some challenges, such as the training efficiency of the model, overfitting problems, and limited ability to capture complex data features. Summary of the Invention

[0004] To solve at least one of the above problems, the present invention proposes a horizontal well production prediction method based on an extended long short-term memory neural network.

[0005] The technical solution of the present invention is as follows: A horizontal well production prediction method based on an extended long short-term memory neural network includes the following steps:

[0006] S1. Obtain the production data of the developed horizontal wells and the characteristic parameters composed of the data affecting the production.

[0007] S2. Based on empirical mode decomposition and wavelet transform, decompose the characteristic parameters at two levels in the time domain and the frequency domain to obtain a three-dimensional feature tensor.

[0008] S3. Decompose the three-dimensional feature tensor based on TCN and STL decomposition respectively, and fuse the decomposition results through a dynamic feature selection gate to obtain a fused feature.

[0009] S4. Establish a prediction model, and train the prediction model based on the fused feature. The trained model can be used for the production prediction of the remaining horizontal wells in this block; the main body of the prediction model is a bidirectional LSTM model, which uses a spatio-temporal attention mechanism to provide weights for the fused feature, and uses an optimized sparrow search algorithm to select better hyperparameters for the bidirectional LSTM model.

[0010] Beneficial effects: The method of the present invention can effectively predict the production of horizontal wells in tight reservoirs, and the prediction results are highly accurate, the calculation time required is short, and it can well guide actual production. Description of the drawings

[0011] Figure 1 It is a comparison chart of several common models and the model in this article for the production prediction in 100 days;

[0012] Figure 2 It is Figure 1 The enlarged view of the part marked with a red frame in Detailed implementation manners

[0013] Next, the specific implementation manners of the present invention will be clearly and completely described in conjunction with examples and the drawings. Obviously, the described examples are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0014] Example 1, a method for predicting the production of horizontal wells based on an extended long short-term memory neural network, comprising the following steps:

[0015] S1. Obtain the production data of the developed horizontal wells and the characteristic parameters composed of the data affecting the production data;

[0016] Specifically, in this embodiment, the production data can select the daily production of a single well, which is mainly to facilitate a more detailed analysis of the production change; the data affecting the production mainly include the flowing fluid level, bottom hole flowing pressure, casing pressure, stroke, pumping speed, and water cut, and these data are obtained by the inventor through a large number of experiments. Those skilled in the art can also select different data according to the different blocks where the oil and gas fields are located.

[0017] S2. Based on empirical mode decomposition and wavelet transform, decompose the characteristic parameters at two levels in the time domain and the frequency domain to obtain a three-dimensional characteristic tensor;

[0018] As a powerful time-frequency analysis tool, wavelet transform can decompose signals in both the time and frequency dimensions simultaneously, has strong localization characteristics, and can extract features of different scales in time series signals. The empirical mode decomposition (EMD) method decomposes complex signals into a series of intrinsic mode functions (IMFs) and residual terms in an adaptive manner, enabling us to better capture the local and global features of the signals. Therefore, in this embodiment, combining EMD and wavelet transform will help enhance the feature extraction ability of the historical data of the oil field and further improve the effect of the prediction model. The inventor calls it the MD-Wavelet hybrid decomposition method, which combines empirical mode decomposition and wavelet transform, decomposes the characteristic parameters at two levels in the time domain and the frequency domain, and obtains a richer and more detailed feature expression.

[0019] The method of this step includes the following sub-steps:

[0020] S21. Perform empirical mode decomposition on the characteristic parameters: , where x (t) represents t the characteristic parameters at time IMF k ( t ) is t the k th IMF component at time k , representing the frequency components of the signal, Res ( t ) is t the residual term at time

[0021] , representing the long-term trend of the signal; through this decomposition, the periodic components, noise, and trend components in the data are separately extracted, which can help the model better understand the non-linear and multi-scale characteristics in the data.

[0021] S22. Apply wavelet transform to the intrinsic mode functions to extract time-frequency features , where W n represents the time-frequency feature of the n th IMF component at time t after wavelet transform, is t the improved Meyer wavelet basis function at time a and b are the scale and position parameters of the wavelet function;

[0022] S23. Concatenate the residual term of S21 and the time-frequency features of S22 to form a three-dimensional feature tensor: , where represents the concatenation operation, which combines each IMF component and its corresponding wavelet transform result to form a three-dimensional feature representation. This three-dimensional feature tensor contains the time-frequency information of the signal at different time steps, different frequency scales, and different feature dimensions, and can comprehensively describe the multi-scale characteristics of the signal; IMF n represents the IMF th Res global component.

[0023] The final obtained feature tensor X is a three-dimensional tensor, which contains the features of the oilfield data at different frequencies and time scales, and helps the model better understand the complex patterns in the oilfield production data.

[0024] S3. Decompose the three-dimensional feature tensor based on TCN and STL decomposition respectively, and fuse the decomposition results through a dynamic feature selection gate to obtain fused features;

[0025] The production data of oil and gas fields is usually affected by multiple factors, which will fluctuate on different time scales. For example, seasonal changes (such as climate, production cycle, etc.) will affect the production of oil fields on a longer time scale, while equipment failures, production strategy adjustments, etc. in the short term may cause fluctuations in a shorter cycle. To solve these problems, in this embodiment, the three-dimensional feature tensor is processed based on TCN and STL decomposition, and the processing results are fused through a dynamic feature selection gate. Finally, fused features are obtained. These fused features express both the short-term changes and long-term trends of the features, which is more in line with the actual situation of oil and gas field production.

[0026] This step includes the following sub-steps:

[0027] S31. Capture the features of the three-dimensional features on different time scales through TCN to obtain ; decompose the production data through STL decomposition to obtain ;

[0028] TCN is a convolutional operation used to capture long-term and short-term dependencies. Compared with ordinary convolution, TCN can capture dependencies in long time series by expanding the receptive field while maintaining the causality of the data, that is, ensuring that the model only uses past time steps for prediction and avoiding interference from future data. The main advantage of TCN is that it can effectively expand the receptive field of the convolutional kernel when processing time series data, thereby obtaining dependency features with a longer time span. In this step, when processing all the data in the three-dimensional feature tensor through TCN, the convolutional operation formula is as follows: , where represents t the feature captured by TCN at time represents the l th k convolutional kernel weight of the layer; t-k is the value of the input at time point t , representing the input that is delayed k steps backward in time from d is the dynamic dilation coefficient; b l is the bias term of the l layer; σIt is the ReLU activation function. At the same time, in this step, the size of the convolutional kernel can be adjusted through the dynamic dilation coefficient to expand its receptive field to capture features at different time scales.

[0029] STL decomposition is a time series decomposition method based on locally weighted regression, which can decompose time series data into a trend component ( T t ), a seasonal component ( S t ) and a residual component ( R t ). The advantage of STL decomposition is that it can adaptively decompose data in the case of significant non-linearity and seasonal variations, so as to extract the main patterns of the data. In this step, after using STL decomposition to decompose the oilfield production data, is expressed as follows: = T t + S t + R t , where represents the production data after STL decomposition at time T t ; S t is the trend component, reflecting the long-term change trend of the data; R t is the seasonal component, capturing periodic fluctuations; R t represents the residual component after STL decomposition, indicating the remaining part that is not explained by the trend and seasonality, representing noise or random fluctuations; the subscript t represents the time point.

[0030] S32. Through the dynamic feature selection gate, fuse the TCN output features and Y t : , where H f represents the fused features; g represents the dynamic gating fusion function.

[0031] Since TCN and STL decomposition extract features from different perspectives. TCN focuses on local short-term changes, while STL decomposition focuses on long-term trends and seasonal fluctuations. Therefore, how to effectively fuse these two types of features is the key to this module. The inventor found through a large number of experiments that introducing a dynamic feature selection gate can well fuse the two. The fused features express both the short-term changes and long-term trends of the features, which is more in line with the actual situation of oil and gas field production, and adaptively weights and fuses these two types of features through a gating mechanism.

[0032] Among them, the calculation formula of g is as follows: , where represents the weight parameter in the dynamic control model; represents the bias term of the dynamic feature selection gate; [ ] combines these two types of features into a new feature matrix.

[0033] S4. Establish a prediction model and train the prediction model based on the fused features. The trained model can be used for the production prediction of the remaining horizontal wells in this block; the main body of the prediction model is a bidirectional LSTM model (BiLSTM), which uses a spatio-temporal attention mechanism to provide weights for the fused features and uses an optimized sparrow search algorithm to select optimal hyperparameters for the bidirectional LSTM model.

[0034] BiLSTM is an extension of the traditional LSTM, which can handle the forward and backward dependencies of sequential data simultaneously. In this way, not only can past information be captured, but also future context information can be utilized, thereby enhancing the model's ability to understand time-series data.

[0035] The bidirectional LSTM network performs forward and backward propagation on the input t at each time step x t to generate t the forward output at time and t the backward output at time . The formula is as follows:

[0036]

[0037] Among them, and respectively represent the outputs of forward and backward propagation at time t; LSTM represents LSTM algorithm; and respectively represent the output of forward propagation at time t - 1 and the output of backward propagation at time t + 1.

[0038] However, in this embodiment, since the input features include time-related information and space-related information, therefore, in order to reflect its role, the inventor introduces a spatio-temporal attention mechanism and sets a spatio-temporal gate function to fuse the output of the bidirectional LSTM with the weighted information of the spatio-temporal attention mechanism, and finally forms a spatio-temporal feature representation.

[0039] Among them, the spatio-temporal attention mechanism can simultaneously focus on time and space information through a joint attention mechanism. In the process of weighting spatio-temporal features, we combine temporal information and spatial position encoding, and through querying Q and keyK Calculate the spatio-temporal weights dynamically, which enables the model to automatically allocate attention to different time steps and spatial positions according to the importance of the data.

[0040] Specifically, the spatio-temporal attention mechanism first calculates the spatio-temporal joint attention weights a i,j , which represent the correlation between spatial positions and time steps. The formula for calculating the spatio-temporal joint attention weights is:

[0041]

[0042] where a i,j represents the weight at time step i and spatial position j ; Q and K are the query and key matrices respectively, d is the dimension of the query and the key. The spatio-temporal position encoding matrix P will incorporate the spatial position information into the calculation to ensure the influence of spatial features on spatio-temporal weighting; softmax is the expression form of the softmax function.

[0043] The process of fusing the output of the bidirectional LSTM with the weighted information of the spatio-temporal attention mechanism is as follows:

[0044]

[0045] where Γ is a fusion function that adjusts the output of the bidirectional LSTM according to the weight a i,j calculated by the spatio-temporal attention mechanism, so that the final spatio-temporal features output can better reflect the mutual relationship between time and spatial features.

[0046] At the same time, in the BiLSTM model of this embodiment, the selection of hyperparameters has a crucial impact on the training process and the final prediction performance of the model, such as network architecture parameters (such as the number of layers, the number of units in each layer), training parameters (such as the learning rate, batch size), and parameters in the optimization process (such as momentum, weight decay, etc.).

[0047] Manually selecting these hyperparameters is often time-consuming and laborious, and may fall into local optimal solutions. To overcome these problems, this module introduces an improvement to the Sparrow Search Algorithm (SSA), adopting a dynamic search space and a quantization search mechanism to optimize and expand the hyperparameters in the BiLSTM network. Through this optimization strategy, the training process of the model can be more efficient, and the optimization process can also be more comprehensive and flexible, thereby improving the accuracy and stability of the oilfield production prediction task.

[0048] Dynamic search space: In traditional optimization algorithms, a fixed search range is usually adopted for optimization. In our solution, a dynamic search space is used to dynamically adjust the search range for each hyperparameter. The core of this method is to determine the search range of a certain hyperparameter according to the mean value μ k and standard deviation σ k of this hyperparameter in the historical training process. The formula is:

[0049]

[0050] where ， is the candidate value of the s th hyperparameter at iteration t+ 1, represents the mean value of the s th hyperparameter in historical iterations, represents the standard deviation of the s th hyperparameter in historical iterations, α represents the dynamic adjustment coefficient.

[0051] This adjustment of the dynamic search space enables the optimization process to be carried out within a more targeted range, enhancing the global search ability and avoiding falling into local optimal solutions.

[0052] Quantization search mechanism: To balance exploration and exploitation in the search process, we introduce a quantization search mechanism. The quantization search mechanism guides the search process through the quantum state, enabling the optimization process to fully explore the solution space and quickly develop to the potential optimal solution region. The quantum state is a mixed state, composed of an exploration state (explore) and an exploitation state (exploit) mixed in proportion:

[0053]

[0054] where, α and β are the coefficients of the exploration state and the exploitation state respectively; Indicates the state. By controlling these two coefficients, the quantization search mechanism can dynamically adjust the balance between exploration and exploitation during the training process, avoiding premature convergence.

[0055] Fitness function design: In the optimized sparrow search algorithm, the design of the fitness function determines the quality of the search process. To improve the optimization efficiency, we designed a composite fitness function that not only considers the error of the model (such as the mean square error MSE), but also introduces gradient stability and feature entropy constraints. The gradient stability constraint can prevent the problems of gradient explosion or gradient disappearance, while the feature entropy constraint ensures the diversity of features and avoids overfitting.

[0056] The design of the fitness function is as follows:

[0057]

[0058] where, MSE is the mean square error, is the norm of the gradient, Entropy represents the feature entropy, ω 1 and ω 2 and ω 3 represent respectively MSE and and Entropy weights.

[0059] After obtaining the above model, the prediction output module converts the feature output of the model into specific oilfield production prediction values through the regression layer. At the same time, to improve the interpretability of the model, we added an interpretability enhancement layer during the prediction process. By quantifying the feature contribution degree, decision-makers can clearly understand the contribution of each feature to the final prediction result. In addition, a post-processing mechanism is introduced to smooth the prediction fluctuations and improve the stability and consistency of the model output.

[0060] Regression output layer: The regression layer is responsible for mapping the feature tensor H obtained from the spatio-temporal fusion module to the prediction value y pred of the oilfield production. Through the fully connected layer, the model maps the spatio-temporal features to specific prediction values. The output prediction value represents the oilfield production prediction result and optimizes the regression coefficients through the training process.

[0061] The specific formula is:

[0062]

[0063] where, W reg is the weight matrix of the regression layer,b reg is the bias term H is the spatio-temporal feature

[0064] Interpretability Enhancement Layer: To improve the interpretability of the model, a feature contribution quantification method is designed. This layer helps users understand the reasons for the model's predictions by quantifying the contribution of each input feature to the prediction result. Specifically, we evaluate the influence of each feature on the prediction result by calculating its partial derivative.

[0065] The formula for feature contribution measurement is:

[0066]

[0067] In the formula, I ( x i ) represents x i the contribution to the prediction result, measuring the influence degree of this feature on the prediction result; x i represents the i th feature in the fused feature; y represents the prediction result of the model; represents the prediction result x i gradient with respect to the fused feature.

[0068] This method can generate importance scores for each input feature and provide a clear explanation behind the prediction for decision-makers.

[0069] Post-processing Mechanism: To correct possible abnormal fluctuations and prediction errors, the post-processing mechanism smooths the predicted values through methods such as moving average to ensure that the final prediction result is more stable and reasonable. Moving average is a common time series smoothing technique that can effectively remove small fluctuations and make the prediction result more stable.

[0070] The moving average formula is:

[0071]

[0072] where, y smooth ( t ) is the smoothed predicted value at time t, N is the window size, y pred ( t ) represents t the original predicted value at time

[0073] To further illustrate the superiority of the method of the embodiment of the present invention, specific test examples are given below.

[0074] Take the relevant data of a certain oilfield for the whole year of 2023. Since there is a large amount of data, only part of the data is given below, as shown in Table 1.

[0075] Table 1 Partial original data table

[0076]

[0077] Compare the method of the embodiment of the present invention with existing methods, such as ARIMA, LSTM, and TCN. The final results are as Figure 1 、 Figure 2 and shown in Table 2.

[0078] Table 2 Comparison of the performance of different models

[0079]

[0080] From Figure 1 、 Figure 2 and Table 2, it can be seen that compared with the existing models, the model of the present invention has higher accuracy and shorter training time.

[0081] To illustrate the rationality of each step in the embodiment of the present invention, an ablation experiment model comparison is carried out, that is, removing a certain module or step in the embodiment of the present invention, or replacing the module of the embodiment of the present invention with a conventional module, and forming a new method, and comparing the prediction accuracy of the new method with that of the embodiment of the present invention. The final results are shown in Table 3.

[0082] Table 3 Comparison of ablation experiment models

[0083]

[0084] As can be seen from Table 3, for the method of the embodiment of the present invention, any module or step is necessary, and only when all modules exist, the method of the embodiment of the present invention has better effects.

[0085] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the equivalent embodiments by using the above-disclosed technical content within the scope of the technical solution of the present invention. However, as long as the content does not depart from the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting horizontal well production based on an extended long short-term memory neural network, characterized in that: The following steps are included: S1. Obtaining the production data of developed horizontal wells and characteristic parameters that affect the production data composition; S2. Based on empirical mode decomposition and wavelet transform, the feature parameters are decomposed in both time domain and frequency domain to obtain a three-dimensional feature tensor; The following steps are included: S21. Perform empirical mode decomposition on characteristic parameters: , where x (t) means t Characteristic parameters of the moment; IMF k ( t )for t Moment k indivual IMF Components, representing the frequency content of the signal, k =1,2,…,n, where n represents the maximum number of intrinsic mode functions; Res ( t )for t The residual term at time t represents the long-term trend in the signal; S22. Apply wavelet transform to the intrinsic mode function to extract time-frequency features: , where W n Indicates n indivual IMF The time-frequency characteristics of the component at time t after wavelet transformation, yes t The improved Meyer wavelet basis function of time, a and b are the scale and location parameters of the wavelet function; S23, concatenate the residual term of S21 and the time-frequency features of S22 to form a three-dimensional feature tensor: , where represents the concatenation operation, combining each IMF component with its corresponding wavelet transform result to form a three-dimensional feature representation; IMF n Indicates the nth IMF Quantity; Res global Represents the residual that forms a three-dimensional tensor, which is the set of residuals at all time steps; S3, decompose the three-dimensional feature tensor based on TCN, decompose the production data based on STL decomposition, and fuse the decomposition results through a dynamic feature selection gate to obtain fused features; S4. Establish a prediction model and train the prediction model based on the fusion features. The trained model can be used for production prediction of other horizontal wells in this block. The prediction model is mainly a bidirectional LSTM model. The spatiotemporal attention mechanism is used to provide weights for the fusion features. The optimized sparrow search algorithm is used to select better hyperparameters for the bidirectional LSTM model. The optimization method of the optimized sparrow search algorithm is to introduce a quantized search mechanism and a dynamic search space, and design a fitness function. The formula of the quantized search mechanism is as follows: , where represents the quantum state, explore and exploit Represent the exploration state and the development state respectively; α and β are the coefficients of the exploration state and the development state respectively; Represents the state; the fitness function is as follows: , where MSE is the mean square error, is the norm of the gradient, Entropy represents the feature entropy, ω 1. ω 2 and ω 3 respectively represent MSE , and Entropy The weight of .

2. The method according to claim 1, characterized in that The characteristic parameters include dynamic liquid level, bottom hole flow pressure, casing pressure, stroke, stroke frequency and water content.

3. The method according to claim 1, characterized in that S3 includes the following sub-steps: S31, through TCN to capture the characteristics of three-dimensional features at different time scales ; Decompose the production data through STL decomposition ; S32, through the dynamic feature selection gate, and Fusion: , where H f Indicates fusion features; g represents the dynamic gating fusion function; express t Features captured by TCN at all times; express t Production data decomposed by STL at all times.

4. The method according to claim 3, characterized in that In S31, The calculation formula is as follows: , where Indicates l Tier k convolution kernel weights; Is input at time point tk The value of indicates that the input is from t Backward Delay k Step input; d is the dynamic expansion coefficient; b l It is l The bias term of the layer; σ is the activation function ReLU; The calculation formula is as follows: , where T t It is the trend component, which reflects the long-term trend of data; S t is the seasonal component, capturing cyclical fluctuations; R t represents the residual component after STL decomposition, which represents the remaining part not explained by trend and seasonality, and represents noise or random fluctuations; t Indicates time; The calculation formula is as follows: , where θ g represents the weight parameter in the dynamic control model; b g Represents the bias term of the dynamic feature selection gate; [ ] combines these two types of features into a new feature matrix.

5. The method according to claim 1, characterized in that In S4, when the spatiotemporal attention mechanism is used to provide weights for fusion features, the weight calculation formula is: , where a i,j Indicates that at time step i and spatial location j The weight on Q is the query matrix, express K The transposed matrix of the key, d are the dimensions of the query and key, P is the spatiotemporal position encoding matrix; softmax is the expression of the softmax function.

6. The method according to claim 1, characterized in that In S4, an interpretability enhancement layer for evaluating the influence of features on the prediction results and a post-processing mechanism for increasing the stability of the prediction values ​​are also added to the prediction model; In the interpretability enhancement layer, the feature contribution measurement formula is used to calculate the influence of features on the prediction results: , where I ( x i )express x i Contribution to the prediction results, measuring the influence of the feature on the prediction results; x i Indicates the first i Features y Represents the prediction results of the model; Represents the prediction results x i Gradient of fused features; In the post-processing mechanism, the smoothness of the prediction results is increased by the sliding average formula: , where is the smoothed forecast value at time t; N is the window size; express t The original predicted value at time .

Images