Coal-bed gas well yield prediction method based on hybrid model

Abstract

The invention discloses a coal-bed gas well yield prediction method based on a hybrid model, and relates to the technical field of coal-bed gas yield prediction.The coal-bed gas well yield prediction method based on the hybrid model mainly comprises the steps that historical dynamic production parameters are preprocessed and divided to obtain a training set and a test set; and training the preliminary prediction model by using the training set to obtain a trained preliminary prediction model, predicting the test set to obtain a preliminary prediction result, and further obtaining a daily gas production prediction result by using an XGBoost algorithm. By means of the coal-bed gas well yield prediction method based on the hybrid model, the accuracy of coal-bed gas well daily gas yield prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coalbed methane production prediction, and more particularly to a coalbed methane well production prediction method based on a hybrid model. BACKGROUND

[0002] Coalbed methane is a kind of unconventional natural gas occurring and storing in coal seams. As a kind of associated energy of coal, its exploration and development have been paid more and more attention in recent years. The research on the production prediction of coalbed methane can provide suggestions for the subsequent development of coalbed methane and is beneficial to the prediction of the economic benefits of coalbed methane, and plays an important role in the development process of coalbed methane.

[0003] The traditional coalbed methane production prediction usually adopts a numerical simulation method, but the method is complex to use, requires a large amount of reservoir and production data, and the calculation results are difficult to match the data of different production wells. The production data of coalbed methane is time series data that changes over time, and many scholars have proposed a long-short term memory (LSTM) network that can predict the time series to predict the production of coalbed methane: the near neighbor propagation algorithm is used to cluster the well data, and the LSTM model is used for prediction, and the results show that the accuracy of this method is higher than that of the traditional shallow neural network and numerical simulation method; a multivariate LSTM model is used to predict the production of coalbed methane, and good results are obtained, and transfer learning is introduced into the neural network to provide a basis for the production prediction when the production data set is insufficient. However, the accuracy of the prediction results of the traditional LSTM model of coalbed methane production can be further improved.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The purpose of the present application is to provide a coalbed methane well production prediction method based on a hybrid model, which can improve the accuracy of the daily gas production prediction of coalbed methane wells.

[0006] The present application provides a coalbed methane well production prediction method based on a hybrid model, comprising the following steps: S1: preprocessing and dividing the historical dynamic production parameters to obtain a training set and a test set; S2: constructing a preliminary prediction model, training the preliminary prediction model using the training set to obtain a trained preliminary prediction model; S3: predicting the test set using the trained preliminary prediction model to obtain a preliminary prediction result; S4: obtaining a daily gas production prediction result according to the preliminary prediction result using an XGBoost algorithm.

[0007] Further, the historical dynamic production parameters include oil pressure data, casing pressure data, orifice plate data, upflow pressure data, upflow temperature data, daily water production data, and daily gas production data.

[0008] Further, step S1 specifically comprises: S11: processing the abnormal values and missing values in the historical dynamic production parameters to obtain historical dynamic production data; S12: standardizing the historical dynamic production data to obtain standardized historical data; and S13: dividing the standardized historical data to obtain a training set and a test set.

[0009] Further, the preliminary prediction model comprises an input layer, a CNN layer, an LSTM layer, an attention layer and a full connection layer connected in sequence.

[0010] Further, the CNN layer comprises a convolution layer and a pooling layer connected in sequence; the convolution layer is used for feature extraction, and a weighted sum is calculated by sliding a convolution kernel on an input function; and the pooling layer is used for feature dimension reduction, and the size of a data space is reduced by aggregating elements in a pooling window.

[0011] Further, the update formula of the LSTM layer is as follows:

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] wherein, , respectively represent a forgetting gate, an input gate and an output gate; t represents a time slice; z represents an activation function; and respectively represent a weight and a bias; represents a hidden layer state of a neuron at an instant ; is an input vector of the instant ; represents an output of a candidate cell state; represents an output of a candidate cell state; represents a cell state before updating; represents a cell state after updating; represents a weight matrix of the cell state after updating; tanh is a hyperbolic tangent function; is a hidden layer state.

[0018] Further, the specific expression of the attention layer is as follows:

[0019]

[0020]

[0021]

[0022] where, denotes the un-normalized attention score, measuring the importance of the current time step t ; V denotes the weight vector, mapping the attention space to a scalar score, finally generating the attention score (scalar); denotes the activation function, introducing a non-linear function; the hidden layer state of the current time step, i.e., the output of the CNN-LSTM; is the hidden layer state of the historical time step, i.e., the hidden layer state of the time step i ; is the weight matrix, transforming the current hidden layer state to map to the attention space; is the weight matrix, transforming the historical hidden layer state to map to the same space; b denotes the bias variable, increasing the flexibility of the model; exp denotes the exponential function, converting the score to a positive number, which is convenient for Softmax normalization; denotes the normalized attention weight, indicating the contribution proportion; F denotes the context vector, focusing on the information of all historical time steps, and the weight is dynamically allocated by the attention mechanism; denotes the updated hidden layer state; f denotes the fusion function.

[0023] Further, the objective function of the above XGBoost algorithm is: , where, i represents the sequence number of the i th sample; is the total number of samples in the dataset; denotes the loss function, and denote the actual true value and the model prediction value of the th sample, respectively, measuring the difference between the two. represents the serial number of the tree in XGboost, i.e. k the first tree; K represents the total number of decision trees; represents the model of the first k decision tree, including tree structure and parameters (leaf weight); represents the regularization term for the first k tree, which controls the model complexity.

[0024] Further, the coalbed methane well production prediction method based on the mixed model further comprises: according to the daily gas production prediction result, using the determination coefficient and the mean square error to evaluate the prediction performance.

[0025] The application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the coalbed methane well production prediction method based on the mixed model.

[0026] The coalbed methane well production prediction method based on the mixed model has the following beneficial effects: The application solves the problem of ignoring time sequence in traditional coalbed methane production prediction, better extracts effective information of coalbed methane production influencing factors, collects and integrates coalbed methane well production dynamic parameters including oil pressure (MPa) and casing pressure (MPa), orifice (mm), upflow pressure (MPa), upflow temperature (℃), daily water production (m 3 ), and daily gas production (10 4 m 3 ), combines the efficient feature extraction capability of convolutional neural network (CNN) and the ability of long short-term memory network to process multiple time series data, and fuses the attention mechanism (Attention) to allocate the time step output of the LSTM layer to perform weight summation, and finally combines the tree structure of XGBoost to capture the nonlinear feature interaction of data to dynamically predict the coalbed methane production of the coalbed methane well, thereby realizing high-precision prediction of the coalbed methane well production.

[0027] The mixed model of the convolutional neural network based on the attention mechanism combined with the long short-term memory network and the XGBoost algorithm has higher accuracy and better generalization ability, and by combining the advantages of the two neural networks, the shortcomings of a single model in the prediction process are compensated, the attention mechanism can focus on the influence of important information on the coalbed methane production, thereby improving the production prediction accuracy. The model established by the random gradient descent method is not easy to fall into a local minimum value, which is beneficial to find all optimal solutions and realize high-precision output prediction of the coalbed methane production. BRIEF DESCRIPTION OF DRAWINGS

[0028] The application will be further described below in conjunction with the accompanying drawings and embodiments, wherein: Figure 1 is a flow chart of a coalbed methane well production prediction method based on a hybrid model provided by the application; Figure 2 is a schematic diagram of a coalbed methane well production prediction method based on a hybrid model provided by the application; Figure 3 is a schematic diagram of a CNN-LSTM-Attention combined XGBoost hybrid model provided by the application; Figure 4 is a CNN-LSTM-Attention combined XGBoost hybrid model for predicting coalbed methane production provided by the application; Figure 5 is a coalbed methane well production dynamic parameter scatter plot provided by the application; Figure 6 is a comparison diagram of real daily gas production and CNN-LSTM-Attention combined XGBoost hybrid model predicted daily gas production results provided by the application. DETAILED DESCRIPTION

[0029] In order to have a clearer understanding of the technical features, objectives and effects of the application, the specific embodiments of the application will be described in detail with reference to the accompanying drawings.

[0030] Figure 1 A schematic diagram of a coalbed methane well production prediction method based on a hybrid model of the present embodiment is shown. In the present embodiment, the coalbed methane well production prediction method based on a hybrid model comprises the following steps: S1: pre-processing and dividing the historical dynamic production parameters to obtain the training set and the test set; In an exemplary embodiment, the historical dynamic production parameters include oil pressure data, casing pressure data, orifice plate data, upflow pressure data, upflow temperature data, daily water production data, and daily gas production data. In an exemplary embodiment, step S1 specifically comprises: S11: processing the abnormal values and missing values in the historical dynamic production parameters to obtain the historical dynamic production data; S12: standardizing the historical dynamic production data to obtain the standardized historical data; S13: dividing the standardized historical data to obtain the training set and the test set; S2: constructing a preliminary prediction model, training the preliminary prediction model using the training set, and obtaining the trained preliminary prediction model; In an exemplary embodiment, the preliminary prediction model comprises an input layer, a CNN layer, an LSTM layer, an attention layer and a fully connected layer connected in sequence; In an exemplary embodiment, the CNN layer comprises a convolution layer and a pooling layer connected in sequence; the convolution layer is used for feature extraction, and a weighted sum is calculated by sliding a convolution kernel over an input function; the pooling layer is used for feature dimension reduction, and the size of the data space is reduced by aggregating elements within a pooling window; In an exemplary embodiment, the update formula of the LSTM layer is:

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] wherein, , respectively represent the forget gate, the input gate and the output gate; t represents a time slice; z represents an activation function; and respectively represent a weight and a bias; represents the hidden layer state of a neuron at an instant ; is the input vector of the instant ; represents the output of the candidate cell state; represents the output of the candidate cell state; represents the cell state before updating; represents the cell state after updating; represents the weight matrix of the cell state after updating; tanh is the hyperbolic tangent function; is the hidden layer state.

[0037] In an exemplary embodiment, the specific expression of the attention layer is as follows:

[0038]

[0039]

[0040]

[0041] wherein, denotes the unnormalized attention score, measuring the importance of the current time step t ; V denotes the weight vector, mapping the attention space to a scalar score, finally generating the attention score (scalar); denotes the activation function, introducing a nonlinear function; the hidden layer state of the current time step, i.e., the output of the CNN-LSTM; is the hidden layer state of the historical time step, i.e., the hidden layer state of the time step i ; is the weight matrix, transforming the current hidden layer state to map to the attention space; is the weight matrix, transforming the historical hidden layer state to map to the same space; b denotes the bias variable, increasing the flexibility of the model; exp denotes the exponential function, converting the score to a positive number, which is convenient for Softmax normalization; denotes the normalized attention weight, indicating the contribution proportion of ; F denotes the context vector, focusing on the information of all historical time steps, and the weight is dynamically allocated by the attention mechanism; denotes the updated hidden layer state; f denotes the fusion function.

[0042] S3: Use the trained preliminary prediction model to predict the test set to obtain the preliminary prediction result; S4: According to the preliminary prediction result, use the XGBoost algorithm to obtain the daily gas production prediction result; In an exemplary embodiment, the objective function of the XGBoost algorithm is:

[0043] wherein, denotes the objective function; i represents the serial number of the i th sample; is the total number of samples in the data set; denotes the loss function, and respectively represent the actual true value and the model prediction value of the th sample, measuring the difference between the two; denotes the serial number of the tree in XGboost, i.e., the kA tree; K The total number of decision trees; The model of the k The model of the decision tree, including tree structure and parameters (leaf weights); The regularization term of the k The regularization term of the tree, which controls the model complexity.

[0044] In an exemplary embodiment, the coalbed methane well production prediction method based on a hybrid model further comprises: according to the daily gas production prediction result, using the determination coefficient and the mean square error to evaluate the prediction performance.

[0045] In some embodiments, the coalbed methane well production prediction method based on a hybrid model described above can also be implemented in the following way.

[0046] In this embodiment, the specific steps of the coalbed methane well production prediction method based on a hybrid model are as follows: Step one: the time series samples obtained by standardizing the historical dynamic production parameters in the development process of a certain coalbed methane well are divided into training set and test set; Step two: the training set is input into the CNN layer for feature extraction, and the bidirectional LSTM layer is used to capture the time series dependence relationship, and the CNN-LSTM model is trained; Step three: the feature sequence output by the CNN-LSTM is weighted and processed based on the attention mechanism, and then input into the fully connected layer for feature extraction and output of the predicted value; Step four: the CNN-LSTM-Attention model prediction result is taken as input, and the XGBoost algorithm is used to further improve the prediction accuracy to obtain the daily gas production prediction result.

[0047] Through the collected dynamic production parameters of the coalbed methane well, including oil pressure (MPa) and casing pressure (MPa), orifice (mm), upflow pressure (MPa), upflow temperature (℃), daily water production (m ), and daily gas production (10 4 m 3 ), the original data set is generated.

[0048] After the historical dynamic production parameters are processed for abnormal values and missing values, they are divided into training set and test set. The training set is input into the CNN-LSTM, and the program self-optimization method is used to determine the optimal hyperparameters of the model. According to the optimal hyperparameters, the performance of the model is optimized.

[0049] The first layer of the CNN-LSTM-Attention model uses a CNN model to extract local space-time features of the input sequence through a convolution kernel function; the second layer uses an LSTM model to model the long-term time dependence of the feature sequence; and the third layer introduces an attention mechanism to assign weights to each time step of the LSTM layer, dynamically capture the differentiated contribution of different time steps in the time series to the prediction result, and weightedly sum the output to obtain the final result.

[0050] In the CNN model, Relu is used as the activation function to avoid gradient disappearance of sequence data and continue space-time feature extraction. The LSTM gating mechanism uses Sigmoid to generate a 0-1 gating signal to control information flow. The LSTM memory unit uses the TANH function for gradient stabilization to smooth the output prediction result. Softmax is used as the activation function of the attention mechanism to ensure that the sum of the time step weights is 1, reflecting attention allocation. Finally, the CNN-LSTM-Attention model is trained and optimized to determine the final dropout threshold. Finally, the XGboost algorithm uses the output results of the CNN-LSTM-Attention model as new input features, uses a tree structure to capture nonlinear feature interactions, and regresses to predict the daily gas production of the coalbed methane well.

[0051] The mean square error (MSE) and the determination coefficient (R 2 ) are used to evaluate and verify the performance of the hybrid model, and finally a high-precision prediction of the coalbed methane well production CNN-LSTM-Attention combined with the XGboost hybrid model is developed.

[0052] In some embodiments, the above-mentioned coalbed methane well production prediction method based on a hybrid model can also be implemented in the following manner.

[0053] As shown in Figure 2 , the coalbed methane well production prediction method based on a hybrid model of the present embodiment includes the following steps: Step 1: Obtain the original data set by collecting and integrating the historical production dynamic parameters of the coalbed methane well, including oil pressure (MPa) and casing pressure (MPa), orifice (mm), upflow pressure (MPa), upflow temperature (℃), daily water production (m 3 ), and daily gas production (10 4 m 3 ), and process the abnormal values and missing values therein, the processing method being as follows:

[0054] wherein is the missing value and abnormal value to be supplemented, , are the actual values before and after , respectively.

[0055] Step 2: Standardize all historical dynamic production data, formula as follows:

[0056] wherein is the parameter to be normalized, represents the minimum value of the parameter, represents the maximum value of the parameter.

[0057] Step 3: Establish a combination of convolutional neural network and long short-term memory network based on attention mechanism, CNN-LSTM-Attention model. As shown in Figure 3 is the schematic diagram of CNN-LSTM-Attention and XGBoost hybrid model; First, the historical dynamic production parameters of coalbed methane wells after standardization are divided into training set and test set, including 299 days of oil pressure (MPa) and casing pressure (MPa), orifice plate (mm), upflow pressure (MPa), upflow temperature (℃), daily water production (m 3 ), daily gas production (10 4 m 3 ).

[0058] Second, neural network parameter setting and optimization, using program self-optimization method to determine the optimal hyperparameters of the model, by inputting the training set into the CNN-LSTM model, the CNN layer is through 128 convolution kernel functions, and the convolution window size is 3. When the number of neurons in the bidirectional LSTM layer is 256, the model performance is optimal.

[0059] Third, the attention mechanism is used to weight and sum the output of each time step of LSTM, and Softmax is used as the activation function of neural network, and the dropout threshold of the model is determined as 0.3.

[0060] The convolution layer of convolutional neural network is used for feature extraction, and the weighted sum is calculated by sliding the convolution kernel on the input function.

[0061] Let the input tensor be , where T (time step) is 10 and d (input dimension) is 6. For the first convolution kernel (total of 128): convolution kernel weight: , the value of the first time step of the output feature map is .

[0062] The final output is the concatenation of all convolution kernel results:

[0063] wherein denotes a one-dimensional convolution operation.

[0064] Pooling layers are used for feature dimension reduction, reducing the data space size by aggregating elements within a statistical pooling window.

[0065] Assume the pooling window size is , and the step size is :

[0066] The bidirectional LSTM layer includes a forget gate , an input gate , an output gate , , and a candidate memory at the current time. The "gate" can control the update and replacement of incoming information, and effectively solve the long-term dependence problem of information. The input gate controls the incoming information, the forget gate controls the incoming information, the forget gate controls the retention and discard of information, and the output gate controls how much information can be output as the current time according to the current state. The update formula of the LSTM network at time is:

[0067] The input gate is:

[0068] The candidate memory is:

[0069] The memory update is:

[0070] The output gate is:

[0071] The hidden state is:

[0072] The attention mechanism is introduced to assign the probability weight of the LSTM network hidden layer, so that the model is easier to process the long time sequence dependence relationship, improve the interpretability of the neural network and the model performance, can highlight the important information on the coalbed methane production, and enhance the accuracy of the coalbed methane production prediction model.

[0073] The specific expression is as follows:

[0074] In the formula, e represents an unnormalized attention score, which measures the importance of to the current time t ; Vdenotes the weight vector, which maps the attention space to the scalar score, and finally generates the attention score (Scalar). denotes the activation function, which introduces a nonlinear function. The hidden layer state at the current time, which is the output of the CNN-LSTM. is the hidden layer state at the historical time, and at the time step i is the hidden layer state at the time step is the weight matrix, which transforms the current hidden layer state into the attention space. is the weight matrix, which transforms the historical hidden layer state into the same space. is the bias variable, which increases the flexibility of the model. b

[0075]

[0076] In the formula exp denotes the exponential function, which converts the score to a positive number, which is convenient for Softmax normalization. denotes the normalized attention weight, which represents the contribution ratio of .

[0077]

[0078] In the formula, F represents the context vector, which focuses on all historical time information, and the weight is dynamically allocated by the attention mechanism.

[0079] The results are integrated by the fusion function f , which represents the updated hidden layer state:

[0080] Fourth, build a CNN-LSTM-Attention model. The first layer uses a CNN model to extract local spatiotemporal features of the input sequence through a convolution kernel; the second layer uses an LSTM model to establish long-term temporal dependencies of the feature sequence; the third layer introduces an attention mechanism to calculate the weighted sum of the outputs of each time step of LSTM, thereby outputting the prediction result. As shown in Figure 4 is the model structure diagram of the CNN-LSTM-Attention combined XGBoost hybrid model for predicting coalbed methane production; Step 4, use the output of the CNN-LSTM-Attention model as new input using the XGboost algorithm, and use the tree structure to capture nonlinear feature interactions to regress and predict the daily gas production of the coalbed methane well.​​​​

[0081] The XGBoost algorithm is a high-level form of ensemble learning technique that uses gradient boosting to optimize decision tree models to provide accurate estimates of target variables based on input features. The XGBoost algorithm improves the model sequentially through an objective function, which is a combination of a loss function and a regularization term. The objective function can be written as:

[0082] wherein, represents the loss function, represents the regularization term.

[0083] As Figure 5 shown is a scatter plot of coalbed methane well production dynamic parameters; Step 5, the coefficient of determination (R2) and mean squared error (MSE) are used to evaluate the prediction performance of the hybrid model.

[0084]

[0085]

[0086] wherein, represents the actual observation value of the i-th sample, is the model prediction value of the i-th sample, represents the mean of all observations. The correlation coefficient between the actual value and the predicted value of the coalbed methane well daily gas production is 0.945, and the mean squared error is 0.015. As shown is a comparison chart of real daily gas production and CNN-LSTM-Attention combined with XGBoost hybrid model prediction of daily gas production.

[0087] Figure 6 In summary, the hybrid model of convolutional neural network based on attention mechanism and long short-term memory network combined with XGBoost has higher accuracy and better generalization ability. By combining the advantages of the two neural networks, the shortcomings of a single model in the prediction process are compensated. The attention mechanism can focus on important information affecting coalbed methane production, thereby improving the accuracy of production prediction. The model established uses the stochastic gradient descent method, which makes it less likely to fall into local minimum, and is conducive to finding the overall optimal solution, achieving high-precision output prediction of coalbed methane production.

[0088] The present embodiment provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the above-mentioned coalbed methane well production prediction method based on a hybrid model.

[0089] The present embodiment provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the above-mentioned coalbed methane well production prediction method based on a hybrid model. ​​

[0090] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, and these all belong to the protection of the present application.

Claims

1. A method for predicting production of coalbed methane wells based on a mixture model, characterized in that, The method comprises the following steps: S1: preprocessing and dividing historical dynamic production parameters to obtain a training set and a test set; S2: constructing a preliminary prediction model, training the preliminary prediction model using the training set to obtain a trained preliminary prediction model; S3: predicting the test set using the trained preliminary prediction model to obtain a preliminary prediction result; S4: obtaining a daily gas production prediction result using an XGBoost algorithm according to the preliminary prediction result.

2. The hybrid model based coal bed gas well production prediction method of claim 1, wherein, The historical dynamic production parameters include oil pressure data, casing pressure data, orifice plate data, upflow pressure data, upflow temperature data, daily water production data, and daily gas production data.

3. The hybrid model based coal bed gas well production prediction method of claim 1 or 2, wherein, Step S1 specifically comprises: S11: processing abnormal values and missing values in the historical dynamic production parameters to obtain historical dynamic production data; S12: standardizing the historical dynamic production data to obtain standardized historical data; S13: dividing the standardized historical data to obtain a training set and a test set.

4. The method of claim 1 or 2, wherein, The preliminary prediction model comprises an input layer, a CNN layer, an LSTM layer, an attention layer, and a fully connected layer connected in sequence.

5. The hybrid model based coal bed gas well production prediction method of claim 4, wherein, The CNN layer comprises a convolution layer and a pooling layer connected in sequence; the convolution layer is used for feature extraction, and a weighted sum is calculated by sliding a convolution kernel on an input function; the pooling layer is used for feature dimension reduction, and the size of the data space is reduced by aggregating and statistically pooling elements in a window.

6. The hybrid model based coal bed gas well production prediction method of claim 4, wherein, The update formula of the LSTM layer is as follows: ， ， ， ， ， ， wherein, , , respectively represent a forget gate, an input gate, and an output gate; represents a time slice; represents an activation function; and respectively represent a weight and a bias; represents a hidden layer state at an instant ; is an input vector at an instant; represents an output of a candidate cell state; represents a cell state before update; represents a cell state after update; represents a weight matrix of the cell state after update; is a hyperbolic tangent function; is a hidden layer state.

7. The method of claim 4, wherein, The specific expression of the attention layer is as follows: ， ， ， ， in, Represented as an unnormalized attention score, measuring For the current moment The importance of; The weight vector represents the attention space mapped to scalar scores, ultimately generating attention scores. , is a scalar; This represents the activation function, introducing a non-linear function; This represents the current hidden layer state, which is the output of the CNN-LSTM. For the hidden state of a historical moment, at time step The hidden layer state; The weight matrix transforms the current hidden layer state. ,Will Mapped to attention space; The weight matrix is ​​used to transform the historical hidden layer states. ,Will Mapped to Same space; Represented as bias variables, this increases the flexibility of the model; This represents an exponential function, converting fractions to positive numbers for easier... Normalization; Represented as normalized attention weights, indicating The contribution ratio; F represents the context vector, which focuses on information from all historical moments, and the weights are dynamically allocated by the attention mechanism; This is represented as a fusion function.

8. The hybrid model based coal bed gas well production prediction method of claim 1 or 2, wherein, The objective function of the XGBoost algorithm is as follows: ， where, is the objective function; i represents the sequence number of the th sample; is the total number of samples in the dataset; denotes the loss function, and denote the actual true value and the model predicted value of the th sample, respectively, measuring the difference between them; denotes the sequence number of the tree in XGboost, i.e., the th tree; represents the total number of decision trees; denotes the model of the th decision tree, containing the tree structure and parameters, i.e., leaf weights; denotes the regularization term for the k th tree, controlling the model complexity.

9. The method of claim 1 or 2, wherein, The coalbed methane well production prediction method based on a hybrid model further comprises: evaluating the prediction performance using a determination coefficient and a mean square error according to the daily gas production prediction result.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the coalbed methane well production prediction method based on a hybrid model according to any one of claims 1-9.

Images