CN118114812A

CN118114812A - Shale gas yield prediction method, computer equipment and storage medium

Info

Publication number: CN118114812A
Application number: CN202410150800.5A
Authority: CN
Inventors: 游君昱; 黄小亮; 冯紫依; 戚志林; 李志强; 梁洪彬; 莫非
Original assignee: Chongqing University of Science and Technology
Current assignee: Chongqing University of Science and Technology
Priority date: 2024-02-02
Filing date: 2024-02-02
Publication date: 2024-05-31

Abstract

The application discloses a shale gas yield prediction method, computer equipment and a storage medium, wherein the method comprises the following steps of collecting historical daily gas production data of a shale gas well to be predicted, and preprocessing all the historical daily gas production data; establishing an ARIMA-EMD-LSTM yield prediction model, wherein the model comprises an ARIMA model, an EMD and an embedded LSTM model, the difference value between the ARIMA model prediction value and the original data is input data of the EMD, the EMD separates data inclusion modal components, n stable IMF functions are obtained, and the corresponding n embedded LSTM models are established based on the n stable IMF functions; dividing the history daily gas production data preprocessed in the first step into a training data set and a test data set; and predicting the productivity of the well to be predicted in a future time period by using the productivity prediction model. The application skillfully simulates the overall descending trend of production data, simultaneously accurately captures the local fluctuation in the data, and predicts faster and more accurate results.

Description

Shale gas yield prediction method, computer equipment and storage medium

Technical Field

The invention belongs to the technical field of shale gas exploitation, and particularly relates to a shale gas yield prediction method, computer equipment and a storage medium.

Background

Compared with the conventional oil and gas reservoirs, the shale gas reservoirs have the characteristics of extremely low permeability, strong reservoir heterogeneity, complex exploitation modes and the like. Unlike the production characteristics of conventional gas reservoirs, shale gas production mainly comprises three stages: free gas flow, adsorbed gas desorption, and dissolved gas diffusion. Development of shale gas reservoirs often accompanies hydraulic fracturing construction, and generated artificial cracks bring more uncertainty and greater complexity to yield prediction of shale gas, so that accurate prediction of the yield of shale gas wells is a precondition for reasonable and efficient development of shale gas reservoirs.

Conventional Decreasing Curve Analysis (DCA), well test analysis methods, numerical simulation methods, and the like all exhibit respective significant limitations in dealing with shale gas production predictions.

The assumption conditions established by the decreasing curve analysis are too simple and ideal, and geological and engineering factors influencing the productivity change are difficult to consider in practical application. Describing the production condition of the oil and gas reservoir by a well test analysis method through a mathematical model of a seepage equation, and obtaining relevant construction parameters (such as fracture conductivity and the like) by establishing a typical plate and fitting production data; however, the method has transition assumptions in the aspects of boundary conditions, crack morphology and the like of the model, and the application accuracy of the method in heterogeneous stratum is greatly limited. The gas reservoir numerical simulation requires a large amount of various data, and in actual production, the situation of data loss often exists to cause key data loss; and because of numerous input data, uncertainty in the data can cause errors of simulation results; meanwhile, the application range of the numerical simulation method is further limited due to the characteristics of high calculation cost and high cost of the numerical simulation method.

Compared with the respective limitations of the traditional method, the artificial intelligence and machine learning method has more application potential in the field of unconventional oil and gas reservoir development. Machine learning methods have also been used in recent years for yield prediction in hydrocarbon reservoirs due to their inherent information mining of data, the ability to perform problem analysis processing with limited data.

In the prior art, artificial neural networks have been used to predict oil production from discontinuous shale reservoirs. The seismic data, logging data, completion parameters and production data are mainly used to predict the production of a well after two years and to predict the well location of a new well. The key point of the research is that the existing data are utilized to generate the synthetic data of various data corresponding to coordinates in the range of the oil reservoir plane, the existing and synthetic data are utilized to predict the oil production, and meanwhile, the predicted oil production is used as an index of new well drilling.

The method also has the advantages that the genetic algorithm is combined with the feedforward neural network to predict the capacity of the shale gas horizontal well in Changning region, the average error of the model for predicting the volume fracturing capacity of the shale gas horizontal well is found to be 8.76 according to the research, and the average error of the model relative to the multiple regression model is found to be 56.55%.

In addition, the initial yield, the progressive rate and the progressive index of shale oil and gas yield are predicted by using an artificial neural network, so that the oil and gas yield curve is fitted, and the applicant finds that the method has poor fitting effect on disturbance on the oil and gas yield curve caused by production measures and construction parameter changes, and can only fit large trend.

For example, patent number 202210771878.X, named "a method and system for predicting daily output of shale gas single well", wherein the output prediction is mainly performed by using an LSTM model, however, the shale gas output is affected by various factors, so that the historical production table shows significant nonlinearity and sudden fluctuation, that is, the shale gas output data should have two parts of linearity and nonlinearity, so that the applicant believes that all data are directly processed by adopting the LSTM model, and the error increase is caused by neglecting the linear expression of the data, and the prediction accuracy is still lacking.

Disclosure of Invention

In view of the above, the invention provides a shale gas yield prediction method, computer equipment and storage medium, which are used for solving the problems of complicated calculation, inaccurate prediction, slower prediction speed and the like of the shale gas yield prediction in the prior art.

The technical scheme is as follows:

The shale gas yield prediction method is characterized by comprising the following steps of:

S1, acquiring historical daily gas production data of a shale gas well to be predicted, and preprocessing all the historical daily gas production data;

S2, an ARIMA-EMD-LSTM yield prediction model is built, the ARIMA-EMD-LSTM yield prediction model comprises an ARIMA model, an EMD and an embedded LSTM model, wherein the difference value between the ARIMA model prediction value and original data is input data of the EMD, the EMD separates data inclusion modal components, n stable IMF functions are obtained, and based on the n stable IMF functions, corresponding n embedded LSTM models are built;

S3, dividing the historical daily gas production data preprocessed in the step S1 into a training data set and a test data set, wherein the training data set is used for training an ARIMA-EMD-LSTM yield prediction model, and the test data set is used for testing the established ARIMA-EMD-LSTM yield prediction model;

s4, predicting the productivity of the well to be predicted in a future time period by using an ARIMA-EMD-LSTM yield prediction model.

With the scheme, the method combines autoregressive integral moving average model (ARIMA), empirical Mode Decomposition (EMD) and long-term short-term memory network (LSTM) technology to meet the requirement of rapid and accurate shale gas yield prediction. The ARIMA model is first applied in the proposed model to separate linear and nonlinear features in the raw production data. ARIMA is used to model and predict linear components, while the residual between ARIMA prediction and the original dataset captures the nonlinearities present in the production data by EMD. An Empirical Mode Decomposition (EMD) is then used to extract eigenmode functions (IMFs) representing local feature signals of different time scales within the nonlinear component. Subsequently, a plurality of LSTM models are trained using the data of these IMFs. The super parameters of the LSTM model are carefully designed to optimize performance, and the LSTM prediction is combined with the ARIMA prediction to form a comprehensive production prediction model in order to obtain the final prediction.

As preferable: the preprocessing of the historical daily gas production data in the step S1 comprises the supplement of the missing values and the normalization of the data.

As preferable: the ARIMA model prediction comprises the following steps:

s2.1, original data stability identification;

s2.2, if the original data is non-stationarity data in the step S2.1, processing the original data by using d-order difference operation to obtain a stable sequence;

S2.3, if the original data are stable data in the step S2.1, calculating an autocorrelation function and a partial autocorrelation function of the stable sequence, selecting a corresponding model according to the two calculated correlation functions, and if the partial autocorrelation function of the stable sequence is truncated and the autocorrelation function is trailing, establishing an autoregressive model; if the partial autocorrelation function is trailing and the autocorrelation function is truncated, establishing a moving average model; if both the partial autocorrelation function and the autocorrelation function are trailing, the sequence fits the hybrid model;

S2.4, performing parameter estimation on the selected model parameters by using a red pool information quantity criterion and a Bayesian information criterion method to obtain an optimal order number p and an optimal order number q;

s2.5, according to the obtained parameters, carrying out yield prediction through an ARIMA detection model, and carrying out residual calculation.

As preferable: the complete prediction result of the ARIMA-EMD-LSTM yield prediction model is equal to the sum of the prediction values of the n LSTM models and the prediction values of the ARIMA model.

As preferable: and comparing the independent LSTM model prediction result with the ARIMA-EMD-LSTM yield prediction model result, and evaluating the performance of the model by using four indexes of average absolute error, average deviation error, root mean square error and decision coefficient.

Another aspect of the present application provides a computer device, which is characterized in that the electronic device includes:

A memory storing executable instructions;

And the processor runs the executable instructions in the memory to realize the shale gas yield prediction method.

In another aspect, the present application provides a computer readable storage medium, which is characterized in that the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the shale gas yield prediction method.

Compared with the prior art, the invention has the beneficial effects that:

By adopting the shale gas yield prediction method provided by the application, the data is processed by mainly using an ARIMA method so as to separate linear and nonlinear components. The nonlinearity represented by the difference between the ARIMA prediction and the raw data is then used as input data in the subsequent EMD step, and the EMD decomposed eigenmode functions (IMFs) are then input to the LSTM model and used as input. Since multiple IMFs are obtained for each well, an equal number of LSTM models are trained. These LSTM model generated predictions are essentially decomposed IMFs, which are derived from the residuals of ARIMA results. Thus, predictions of multiple LSTM models are summarized and combined with ARIMA predictions to ultimately produce final predictions of the raw production data.

The ARIMA-EMD-LSTM model skillfully simulates the overall descending trend of production data, simultaneously accurately captures local fluctuation in the data, and has more accurate prediction result.

Drawings

FIG. 1 is a schematic flow chart of the present invention;

FIG. 2 is historical daily gas production data (after pretreatment) of three shale gas wells to be predicted;

FIG. 3 is a predicted outcome for the ARIMA model for well number 1;

FIG. 4 is a graph of the resulting residuals calculated from the prediction of the ARIMA model;

FIG. 5 is an eigenmode function (first four) of EMD residual extraction according to the prediction result shown in FIG. 4;

FIG. 6 is a graphical illustration of the comparison of results of yield predictions using the ARIMA-EMD-LSTM yield prediction model of the present application with direct predictions using a conventional independent LSTM model for three shale gas wells to be predicted.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings.

The shale gas yield prediction method shown in reference to fig. 1 to 6 mainly comprises the following steps: firstly, collecting historical daily gas production data of a shale gas well to be predicted, and preprocessing all the collected historical daily gas production data, wherein in the actual implementation process, the method mainly comprises the steps of supplementing missing values and normalizing the data. The missing value is added by recording the missing data, which is not recorded at a certain time point, to zero, and the data normalization processing is performed according to the following formula (1).

Wherein T is the recorded value of the current day, T _min is the recorded minimum value in all the historical daily gas production data, T _max is the recorded maximum value in all the historical daily gas production data, and T _norm is the normalized product value.

And secondly, constructing an ARIMA-EMD-LSTM yield prediction model, wherein the ARIMA-EMD-LSTM yield prediction model comprises an ARIMA model, an EMD and an embedded LSTM model (embedded herein means that the LSTM model is used together with the EMD and is not an independent operation model), the difference value between the ARIMA model predicted value and the original data is input data of the EMD, the EMD separates data inclusion modal components, n stable IMF functions are obtained, and corresponding n embedded LSTM models are established based on the n stable IMF functions.

And thirdly, dividing the historical daily gas production data preprocessed in the first step into a training data set (called a training set for short) and a test data set (called a test set for short), wherein the training data set is used for training an ARIMA-EMD-LSTM yield prediction model, and the test data set is used for testing the established ARIMA-EMD-LSTM yield prediction model.

And fourthly, predicting the productivity of the well to be predicted in a future time period by using an ARIMA-EMD-LSTM yield prediction model.

The second step mainly comprises the steps of constructing a main model framework, determining a combination mode of an ARIMA model, an EMD model and an embedded LSTM model and a data transmission mode, and training and perfecting the model according to the training data set in the third step, wherein in the implementation process, the second step is carried out after data division is directly advanced on the premise of existing data, and the model is constructed and trained and perfected simultaneously.

The ARIMA model in the ARIMA-EMD-LSTM yield prediction model is basically consistent with the traditional meaning, and the main difference is that the difference between the prediction result and the original data is used as the input data of EMD. The ARIMA model is one of the most widely used linear regression models for predicting stationary time series. The model is denoted ARIMA (p, D, q), the parameters p, D and q represent the structure of the prediction model, and three basic forms are actually included, namely an Autoregressive (AR) model, a Moving Average (MA) model and a hybrid model (AR & MA hybrid model), namely the ARIMA model combines an Autoregressive (AR) represented by AR (p), a Moving Average (MA) represented by MA (q) and a differential degree D.

The mathematical expression of ARIMA (p, D, q) can be described as follows (2):

Where L represents the hysteresis operator, phi _i represents the parameters of the autoregressive portion, theta _i represents the moving average model, and epsilon _t is the error term.

When the application is implemented, ARIMA model prediction mainly comprises the following steps:

S2.1, original data stability identification, wherein the ARIMA model algorithm is used for requiring that the processed time sequence is under a stable state, namely that the properties such as variance, covariance and the like of the sequence cannot change drastically with time, so that the stability identification is required for the data before calculation.

S2.2, if the original data is non-stationarity data in the step S2.1, the original data is processed by using d-order difference operation to obtain a stable sequence, and then the step S2.3 is performed.

S2.3, if the original data are stable data in the step S2.1, calculating an autocorrelation function and a partial autocorrelation function of the stable sequence, selecting a corresponding model according to the two calculated correlation functions, and if the partial autocorrelation function of the stable sequence is truncated and the autocorrelation function is trailing, establishing an autoregressive model (AR); if the partial autocorrelation function is trailing and the autocorrelation function is truncated, a moving average Model (MA) is built; if both their partial and autocorrelation functions are tailing, the sequence fits to the hybrid model.

S2.4, parameter estimation is carried out on the selected and established model parameters by mainly utilizing methods such as a least square method, a red pool information amount criterion, a Bayesian information criterion and the like, so as to obtain an optimal order layer number p and an order number q;

And carrying out residual diagnosis analysis by using the obtained D, q and p coefficient values, and carrying out model detection on the selected model.

S2.5, according to the obtained parameters, carrying out yield prediction through an ARIMA model, and carrying out residual calculation, wherein the difference value between the predicted value of the ARIMA model and the original data is the residual value.

Because the result of ARIMA model prediction only represents the linear part of the original data, the difference (residual) between the predicted result and the original data is the nonlinear part of the original data, and the part of data needs to be processed by EMD and then is predicted by combining with an embedded LSTM model, wherein the residual data is calculated by using an EMD algorithm, and the data connotation modal components are separated. The method mainly comprises the following steps:

S2.6, positioning all maximum points on the original data, and using cubic spline interpolation fitting to form an upper envelope curve of all maximum values in series;

s2.7, positioning all minimum value points on the original data, and forming a lower envelope curve of all minimum values in series by using cubic spline interpolation fitting;

s2.8, calculating the average value of the upper envelope line and the lower envelope line, wherein the average value line is the average envelope line;

s2.9, subtracting the average envelope curve from the original data curve to obtain a new time sequence;

s2.10, subtracting new data after envelope average from the original data, and repeating the steps S2.6 to S2.9 until no further decomposition can be performed if negative local maxima and positive local minima exist, which means that more than one eigenmode function (called IMF) exists.

S2.11, carrying out similarity test on n IMFs by using Dynamic Time Warping (DTW), and regarding the IMFs with higher similarity after new data are added, regarding the IMFs as stable connotation modal components, and then, applying the IMFs to subsequent research steps. The above steps result in n stable IMFs, designated IMF ₁、IMF₂、……、IMF_n.

S2.12, inputting different IMF data based on the IMF data obtained in the steps, so that the LSTM model prediction result also corresponds to different IMFs. Corresponding to the obtained n IMFs, corresponding n LSTM models are also established and are marked as LSTM ₁、LSTM₂、……、LSTM_n.

In this embodiment, the embedded LSTM model is in a form of "sequential input, and a number of columns output" is input as current and past data, and output is a prediction result of a next time point, and the structure of the embedded LSTM model is similar to that of a typical LSTM model, and mainly includes 1 input layer, 1 LSTM layer, 1 dropout layer, 1 LSTM layer, one full connection layer, and 1 regression layer.

Wherein the main superparameter settings are shown in the following table (1):

Main super parameter	Selecting/taking value
		Optimization algorithm	adam
Maximum training times	500
		Initial learning rate	0.005
Learning rate decline factor	0.1
		Learning rate decline period	800

Watch (1)

The sum of the predicted values of the n LSTM models and the predicted value of the ARIMA model is the complete predicted result of the ARIMA-EMD-LSTM yield predicted model.

To further demonstrate the accuracy of the ARIMA-EMD-LSTM yield prediction model of the present application, it was compared with the conventional independent LSTM model predictions and its performance was evaluated with four indices, mean Absolute Error (MAE), mean deviation error (MBE), root Mean Square Error (RMSE) and determinant coefficient (R ²), as shown in the calculations (3) - (6):

The prediction process of the present application is mainly calculated by a computer, and thus the present application also provides a computer device including a memory and a processor, wherein the memory stores executable instructions; the processor can run executable instructions in the memory to realize the shale gas yield prediction method, specifically mainly realize the calculation process of the model and output results, and for the convenience of visual observation, the output results are in a graphic form in the embodiment.

In another aspect, the present application provides a computer readable storage medium, where a computer program is stored, where the computer program when executed by a processor implements the shale gas yield prediction method described above, and specifically mainly implements a calculation process of the calculation model described above, and outputs a result.

Referring to fig. 1 to 6, in this embodiment, taking the number 1 to-be-pre-log as an example, the data obtained after preprocessing the historical daily output data are shown in the following table (2):

Watch (2)

The data of the daily gas production after pretreatment is divided into 90% of training data sets and 10% of training data sets, wherein the training data sets are used for training the ARIMA-EMD-LSTM yield prediction model provided by the application, and the testing data sets are used for testing the established yield prediction model.

The results of the prediction were first performed by the ARIMA model in the ARIMA-EMD-LSTM yield prediction model, as shown in FIG. 3, with the emphasis that the ARIMA model was developed specifically using training data sets, and the test data was used for verification and testing purposes only. Although the ARIMA model captures the linear trend in the raw data, it is clear that ARIMA predictions do not closely match actual production data. The upper and lower dashed lines in the figure represent the upper and lower limits of the 95% confidence interval, respectively, and the residual calculation is performed according to the prediction result obtained in fig. 3, and the result is shown in fig. 4.

EMD calculates according to the residual error shown in fig. 4, and separates the connotation modal components, mainly obtains the eigen mode function shown in fig. 5 (note: only the first four IMFs are shown in the figure), and then establishes a corresponding LSTM model for prediction, which can be understood as the residual error prediction process.

And finally, adding the predicted values of the n LSTM models and the predicted values (shown in figure 3) of the ARIMA model to obtain a final yield predicted result, wherein the result is a normalized yield representation, inverting the value according to a normalization processing mode to obtain actual single-day yield, inputting the existing yield data during prediction, obtaining future daily gas yield data through model calculation, and performing prediction test on the historical daily gas yield data, wherein the result is shown in a third column of a table (2), and the predicted result is very similar to the actual value.

To further demonstrate the accuracy of the ARIMA-EMD-LSTM yield prediction model of the present application, the present application compares the prediction results of the test data and the total data with the prediction results of the independent LSTM model (which is a conventional model and is not described in detail herein), the results of which are shown in FIG. 6 and the following tables (3) and (4), the left column in FIG. 6 shows the final prediction results of the training dataset and the test dataset, the right column shows the direct prediction results of the independent LSTM model, and the values shown at the top of each panel represent the overall RSME (root mean square error, corresponding to the bolded data in Table 4), which is a measure of the model prediction accuracy. Table 3 shows MAE, MBE, RMSE and R ² values for three well test datasets, and the ARIMA-EMD-LSTM model predictions based on test data compared to independent LSTM model direct predictions; table 4 shows the MAE, MBE, RMSE and R ² values for the total data (including training data set and test data set) for three wells, and the ARIMA-EMD-LSTM model predictions based on the total data compared to the independent LSTM model direct predictions.

As can be seen by combining the table (3) and the table (4), compared with the traditional LSTM model, the ARIMA-EMD-LSTM model provided by the application skillfully simulates the overall descending trend of production data, and simultaneously accurately captures local fluctuation in the data. In contrast, the traditional independent LSTM model is less effective in predicting these local changes than the ARIMA-EMD-LSTM model. ARIMA-EMD-LSTM model consistently produced lower MAE, MBE and RMSE values in all three wells. In addition, the R ² value of the ARIMA-EMD-LSTM model is closer to 1 than the R ² value of the LSTM model, and the evaluation indexes together confirm that the ARIMA-EMD-LSTM model has excellent prediction precision compared with the traditional LSTM model. Therefore, the ARIMA-EMD-LSTM model prediction result of the application is accurate to be better than the direct prediction result of the independent LSTM model.

Table (3) test data prediction vs. schematic table

Table (4) Total data prediction vs. schematic Table

Finally, it should be noted that the above description is only a preferred embodiment of the present invention, and that many similar changes can be made by those skilled in the art without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. The shale gas yield prediction method is characterized by comprising the following steps of:

2. The shale gas production prediction method as claimed in claim 1, wherein: the preprocessing of the historical daily gas production data in the step S1 comprises the supplement of the missing values and the normalization of the data.

3. The shale gas production prediction method as claimed in claim 2, wherein: the ARIMA model prediction comprises the following steps:

s2.1, original data stability identification;

4. The shale gas production prediction method as claimed in claim 1, wherein: the complete prediction result of the ARIMA-EMD-LSTM yield prediction model is equal to the sum of the prediction values of the n LSTM models and the prediction values of the ARIMA model.

5. The shale gas production prediction method according to any of claims 1 to 4, characterized in that: and comparing the independent LSTM model prediction result with the ARIMA-EMD-LSTM yield prediction model result, and evaluating the performance of the model by using four indexes of average absolute error, average deviation error, root mean square error and decision coefficient.

6. A computer device, comprising:

A memory storing executable instructions;

A processor executing the executable instructions in the memory to implement the shale gas production prediction method of any of claims 1-5.

7. A computer readable storage medium, characterized in that it stores a computer program which, when executed by a processor, implements the shale gas production prediction method of any of claims 1 to 5.