A coal bed gas well gas production rate np-lstm combined prediction method

Abstract

The application discloses a coalbed methane well gas production NP-LSTM combined prediction method and belongs to the technical field of oil and gas field development prediction. A two-level model fusion prediction framework is constructed by adopting a meta-learning idea. Firstly, in the base model layer, a NeuralProphet model is used to capture the seasonality and event regularity of gas production and other drainage parameters, and an LSTM model is used to learn deep nonlinear time series dependence between multiple parameters. Then, in the meta-model layer, the prediction results of the two base models are spliced with static geological parameters such as coal seam depth, thickness and gas saturation, and input into a random forest regression model for nonlinear fusion, and finally, high-precision gas production prediction values are output. The advantages of the two types of base models are integrated in a soft fusion manner, and the geological static information is effectively combined, so that the accuracy and robustness of the coalbed methane well gas production prediction are significantly improved, and reliable data support is provided for accurate development and production control of the coalbed methane field.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development engineering technology, specifically to a method for predicting coalbed methane well production using the NP-LSTM joint prediction method based on meta-learning, which integrates NeuralProphet (NP) and LSTM time series prediction models with random forest. Background Technology

[0002] The gas production of coalbed methane wells is a key indicator for evaluating development effectiveness and formulating drainage and production systems. However, the dynamics of gas production are subject to the dual and complex influence of static geological parameters such as coal seam depth, thickness, and gas saturation, as well as dynamic drainage and production parameters such as bottom hole flowing pressure and flow rate. This results in strong nonlinearity, nonstationarity, and time-series correlation, making it extremely challenging to accurately predict its changing trends.

[0003] Currently, methods for predicting coalbed methane well production mainly fall into two categories: traditional methods based on mathematical statistics and physical models, and data-driven methods based on machine learning. Traditional methods (such as the production decline curve method and the material balance method) typically rely on strong theoretical assumptions, have poor adaptability to complex geological conditions, and limited prediction accuracy. With the development of artificial intelligence technology, machine learning models, especially deep learning models such as Long Short-Term Memory (LSTM) networks, have been applied to time series forecasting because they can capture long-term dependencies in data. However, single LSTM models perform poorly when dealing with data that has strong seasonality or periodicity, or is affected by sudden human events (such as equipment adjustments or changes in drainage systems), and are difficult to effectively integrate static geological features.

[0004] In recent years, some models that combine the advantages of traditional time series models (such as ARIMA) and modern neural networks (such as Facebook Prophet and its neural network variant NeuralProphet) have demonstrated strong capabilities in time series forecasting, especially in handling trend, seasonal, and holiday event effects. However, the NeuralProphet model is weaker than LSTM in capturing complex, deep nonlinear coupling relationships between variables.

[0005] Therefore, existing technologies lack a comprehensive prediction model that can simultaneously consider the seasonality, event-based patterns, and deep nonlinear dependencies of time-series data, and effectively integrate static geological parameters. This results in the prediction accuracy of coalbed methane well production failing to meet the needs of refined production management. Summary of the Invention

[0006] Technical Problem: The purpose of this invention is to overcome the shortcomings of the existing technology and provide a more accurate and robust NP-LSTM joint prediction method for coalbed methane well production. This method overcomes the limitations of a single model, comprehensively utilizes the advantages of different types of models, and effectively integrates dynamic drainage data with static geological data, thereby achieving a more comprehensive and accurate dynamic prediction of gas production.

[0007] Technical solution: In order to achieve the above objectives, the present invention provides a joint NP-LSTM prediction method for coalbed methane well production: a meta-learning prediction model with a two-level model fusion prediction framework structure is constructed, including a base model layer and a meta-model layer. The base model layer includes a NeuralProphet (NP) base model and an LSTM base model, and the meta-model layer includes a random forest regression model.

[0008] In the base model layer, the NeuralProphet base model is used to capture the seasonal and event-based patterns in the time series data of coalbed methane well production parameters, and the LSTM base model is used to learn the deep nonlinear time series dependencies between production parameters. In the meta-model layer, the prediction results of the NeuralProphet base model and the LSTM base model are combined with static geological parameters and then input into the random forest regression model for nonlinear fusion, finally outputting a high-precision gas production prediction value.

[0009] When applied to new well prediction, the time series data of drainage parameters, including seasonal and event-related information, are input into the base model layer, and then combined with the static geological parameters of the new well into the meta-model layer to finally obtain the final new well gas production prediction sequence.

[0010] The above-mentioned NP-LSTM joint prediction method for coalbed methane well production includes the following steps:

[0011] (a) Data preprocessing: Collect time-series data of static geological parameters and drainage parameters of multiple coalbed methane wells in the target coalbed methane production area; perform cross-well normalization on the static geological parameters to eliminate the influence of dimensions; perform outlier detection and cleaning on the time-series data of drainage parameters, and perform cross-well normalization.

[0012] (b) Dataset Construction: The time-series data of the drainage parameters of each coalbed methane well are arranged in ascending order by timestamp; and divided by time point t. m The full-cycle data of each well is divided into training period data and validation period data as boundaries. A sliding window sample builder with a preset historical window length and a prediction window length is used to perform rolling sampling on the training period data and the validation period data respectively to generate corresponding training sample sets and validation sample sets. Each sample includes dynamic drainage features and static geological features within the historical window, as well as the target gas production label within the prediction window.

[0013] (c) NeuralProphet base model construction and training: Four independent NeuralProphet base models were constructed for the four drainage parameters; custom event terms were configured for the four NeuralProphet base models to integrate human drainage event information; the gas production data in the training period data was used to train the gas production prediction NeuralProphet base model; the bottom hole flowing pressure, liquid column height and stroke frequency in the training period data were used to train the corresponding drainage parameter prediction NeuralProphet base models, and the gas production data was added as an external covariate to all three drainage parameter prediction NeuralProphet base models;

[0014] (d) LSTM base model construction and training: Construct a multivariate, multi-output LSTM base model to simultaneously learn the deep temporal dependencies between the four sorting parameters; use the training sample set obtained in step (b) to supervise the training of the LSTM base model, and use the validation sample set to monitor the training process to prevent overfitting.

[0015] (e) Base model prediction: Add event labels representing human intervention to the time series data of the sampling parameters; use the trained NeuralProphet base model to perform historical fitting on the training period to obtain the predicted values ​​of the training period data. And forward prediction is performed on the verification period to obtain the predicted value of the verification period data. Using the trained LSTM base model, forward propagation is performed on both the training and validation sample sets to obtain the predicted values ​​for the training sample set. and the predicted values ​​of the validation set samples ;

[0016] (f) Meta-model dataset construction: Using the random forest regression model sample builder, for each time point of the training period data, the predicted values ​​of the training period data at the corresponding time point are generated. Predicted values ​​of training sample set The static geological parameters are used to construct the input feature vector of the meta-model; the actual gas production at that time point is used as the target label to generate the meta-model training sample set; for each time point of the validation period data, the predicted value of the validation period data at the corresponding time is also used. Validate the predicted values ​​of the sample set The static geological parameters are combined to form the input feature vector of the meta-model; the actual gas production at that time point is used as the target label to generate the meta-model validation sample set.

[0017] (g) Meta-model fusion and final prediction: Construct a random forest regression model as a meta-model; use the meta-model training sample set to train the random forest regression model, learn the nonlinear mapping relationship from the prediction results of the base model layer and static geological parameters to the final gas production; apply the trained random forest regression model, fuse the prediction results of the base model layer of the new sample with the static geological parameters, and output the final predicted value of coalbed methane well gas production.

[0018] The static geological parameters include at least three or more of the following: coal seam burial depth, coal seam thickness, gas saturation, Langmuir volume, Langmuir pressure, critical desorption pressure, and reservoir permeability; the drainage parameters include at least four or more of the following: gas production, bottom hole flowing pressure, flushing frequency, casing pressure, and dynamic fluid level depth.

[0019] In step (a), the normalization process uses the maximum-minimum normalization method to map each parameter value to the [0,1] interval; the outlier detection and cleaning uses a statistical method based on interquartile range (IQR) to identify and replace outlier data points.

[0020] In step (b), the time segmentation point Through formula The calculation determines that, among which This is the start time of the sampling process. The end time of the sampling process is denoted as R, which represents the segmentation ratio and ranges from 0.7 to 0.9. The historical window length L of the sliding window sample builder ranges from 20 to 50 time units, and the prediction window length F is 1 time unit. The sliding window sample builder reads t from the time series data of the sampling parameters each time. i to t i+L The drainage parameters within a time period constitute the dynamic characteristics of the sample, while the static geological parameters of the coalbed methane well constitute the static characteristics of the sample. i+L+1 The time point sampling parameters serve as the sample labels, when Read as the first sample When reading the second sample...; when the sliding window sample builder constructs the training sample set... When the sliding window sample builder constructs the validation set... .

[0021] In step (c), the parameters of the NeuralProphet base model are set as follows: the number of lag times (n-lags) is between 20 and 50, n-forecasts is set to 1, yearly-seasonality is set to "True", weekly-seasonality is set to "False", daily-seasonality is set to "False", seasonality-mode is set to "multiplicative", learning-rate is between 0.001 and 0.1, and epoch is between 20 and 100; the configuration of the custom event item is as follows: mode is set to "additive", lower-window is between -3 and 0, upper-window is between 0 and 10, and regularization weight is between 0.5 and 1.0.

[0022] In step (d), the LSTM base model is a stacked network structure, which includes, in sequence: an input layer with dimensions [L, ... F_dynamic], where L is the historical window length and F_dynamic is the number of dynamic sorting features; the first LSTM layer has 64 to 256 neurons; the first Dropout layer has a dropout rate of 0.1 to 0.3; the second LSTM layer has 32 to 128 neurons; the second Dropout layer has a dropout rate of 0.1 to 0.3; a flattening layer; the first fully connected layer has 32 to 128 neurons and uses the ReLU activation function; the output layer has the number of neurons equal to the number of predicted target sorting parameters F_target, and the output dimension is [F_target]; the training configuration of the LSTM base model is: using the Adam or RMSprop optimizer, the loss function is mean squared error or mean absolute error, the epoch (training rounds) is between 50 and 150, and the learning rate is between 0.001 and 0.1; an early stopping mechanism is added, and the patience value is between 10 and 50.

[0023] In step (e), the operational event types marked by the event include: adjustment of the sampling system, equipment maintenance, and unblocking operations; the operational event is defined by an event name string and an event impact time window; the predicted value of the training period data. Covering the NeuralProphet base model for training time data tL+1 to t m Predicted values ​​of sampling parameters within a given time period, and verification data prediction values. Including the NeuralProphet base model for validation data t m+L+1 to t n Predicted values ​​of sampling parameters within a given time period; predicted values ​​of the training sample set. Including LSTM base models on the training sample set t L+1 to t m Predicted values ​​of sampling parameters within a given time period, and predicted values ​​of the validation set samples. Including LSTM base models on the test sample set t m+L+1 to t n Predicted values ​​of sampling parameters within a given time period.

[0024] In step (f), the random forest regression model sample builder constructs samples according to time points: for training period data, each reading time t j NeuralProphet base model training data prediction values Predicted values ​​of training sample set and LSTM base model The normalized static geological parameters of the well, along with all the normalized parameters of the well, are used as features of the meta-model training samples, with time t as the metric. j The actual gas production is used as the label for the training samples of the meta-model, where t j The range of values ​​is from t L+1 to t m For data during the verification period, the reading time is t each time. k NeuralProphet base model validation period data prediction values Validate the predicted values ​​of the LSTM base model on the sample set. The normalized static geological parameters of the coalbed methane well, along with all other parameters, are used as features of the meta-model validation sample, with time t as the metric. k The actual gas production is used as the label for the meta-model validation sample, where t k The range of values ​​is from t m+L+1 to t n .

[0025] In step (g), the hyperparameters of the random forest regression model are optimized by grid search or random search, wherein the number of trees n_estimators ranges from 50 to 500, the maximum depth max_depth ranges from 5 to 20, and the minimum number of leaf node samples min_samples_leaf ranges from 1 to 5.

[0026] Beneficial Effects: By adopting the above technical solution, this invention breaks through the limitation of traditional time series models that rely solely on historical dynamic data, ensuring that the prediction results are always constrained by the reservoir geological background, thus greatly improving the scientific rigor and practicality of the model. First, by constructing a hybrid base model combining NeuralProphet and LSTM, complementary advantages are achieved. The NeuralProphet base model excels at capturing explicit trends, seasonality, and human-influenced patterns in gas production data; while the LSTM base model is adept at learning and memorizing complex, deep nonlinear temporal dependencies between production and drainage parameters. The final random forest meta-model can nonlinearly fuse prediction results from these two different perspectives and incorporate static geological constraints, resulting in final predictions that far surpass any single model in accuracy and reliability. Second, the hybrid framework possesses stronger explanatory power and adaptability to complex gas production dynamics in the real world. It can handle periodic fluctuations caused by seasonal changes and effectively respond to the impact of sudden events such as equipment adjustments and changes in work schedules. Through the event terms of NeuralProphet, it can also characterize the complex mechanisms of dynamic coupling between multiple variables such as bottom hole flowing pressure and liquid column height, and through LSTM... This comprehensive capture capability enables the model to exhibit stronger stability and robustness when facing gas wells with different geological conditions and production histories. Thirdly, it innovatively uses static geological parameters as one of the core input features of the meta-model; the main advantages of this invention compared to existing technologies are:

[0027] ① Overcome the problem that a single LSTM model is insufficient in capturing explicit seasonal and event-based patterns in the data.

[0028] ② It compensates for the weakness of the single NeuralProphet model in learning and expressing complex dynamic nonlinear relationships among multiple variables.

[0029] ③ Establish an effective mechanism to deeply integrate the prediction results from different base models with invariant static geological parameters in order to improve the accuracy and generalization ability of the final prediction.

[0030] ④ The prediction model can adaptively respond to the impact of human-induced gas production control measures (events) on gas production dynamics. Attached Figure Description

[0031] Figure 1 This is a flowchart of the NeuralProphet-LSTM joint prediction method for coalbed methane well production of the present invention.

[0032] Figure 2 This is a diagram of the joint prediction architecture of the NeuralProphet-LSTM hybrid model for coalbed methane well production in this invention. Detailed Implementation

[0033] An embodiment of the present invention will be further described below with reference to the accompanying drawings:

[0034] The present invention provides a joint NP-LSTM prediction method for coalbed methane well production: a meta-learning prediction model with a two-level model fusion prediction framework structure is constructed, including a base model layer and a meta-model layer. The base model layer includes a NeuralProphet (NP) base model and an LSTM base model, and the meta-model layer includes a random forest regression model.

[0035] In the base model layer, the NeuralProphet base model is used to capture the seasonal and event-based patterns in the time series data of coalbed methane well production parameters, and the LSTM base model is used to learn the deep nonlinear time series dependencies between production parameters. In the meta-model layer, the prediction results of the NeuralProphet base model and the LSTM base model are combined with static geological parameters and then input into the random forest regression model for nonlinear fusion, finally outputting a high-precision gas production prediction value.

[0036] When applied to new well prediction, the time series data of new well production parameters, including seasonal and event-related information, are input into the base model layer, and then combined with the static geological parameters of the new well into the meta-model layer to finally obtain the final new well gas production prediction sequence.

[0037] The specific steps of the NP-LSTM joint prediction method for coalbed methane well production are as follows:

[0038] (a) Data preprocessing: Collect time-series data of static geological parameters and drainage parameters of multiple coalbed methane wells in the target coalbed methane production area; perform cross-well normalization on the static geological parameters to eliminate the influence of dimensions; perform outlier detection and cleaning on the time-series data of the drainage parameters, and perform cross-well normalization; the static geological parameters include at least three or more of the following: coal seam depth, coal seam thickness, gas saturation, Langmuir volume, Langmuir pressure, critical desorption pressure, and reservoir permeability; the drainage parameters include at least four or more of the following: gas production, bottom hole flowing pressure, flushing frequency, casing pressure, and dynamic fluid level depth; the normalization process uses the maximum-minimum normalization method to map each parameter value to the [0,1] interval; the outlier detection and cleaning uses a statistical method based on interquartile range (IQR) to identify and replace outlier data points.

[0039] (b) Dataset Construction: Arrange the time-series data of the drainage parameters of each well in ascending order by timestamp; divide the data by time point t. mTo define the boundaries, the full-cycle data of each well is divided into training and validation periods. A sliding window sample builder with preset historical and prediction window lengths is used to perform rolling sampling on the training and validation period data, generating corresponding training and validation sample sets. Each sample mainly consists of dynamic drainage features and static geological features within the historical window, and a target gas production label within the prediction window. The time division points... Through calculation The calculation determines that, among which This is the start time of the sampling process. The end time of the sampling process is defined as R, which is the segmentation ratio, ranging from 0.7 to 0.9; preferably R = 0.8. The historical window length L of the sliding window sample builder ranges from 20 to 50 time units, and the prediction window length F is 1 time unit; preferably L = 30 and F = 1. The sliding window sample builder reads t from the time series data of the sampling parameters each time. i to t i+L The drainage parameters within a time period constitute the dynamic characteristics of the sample, while the geological parameters of the coalbed methane well constitute the static characteristics of the sample. i+L+1 The time point sampling parameters serve as the sample labels, when Read as the first sample When reading the second sample...; when the sliding window sample builder constructs the training sample set... When the sliding window sample builder constructs the validation set... .

[0040] (c) NeuralProphet Base Model Construction and Training: Multiple independent NeuralProphet base models were constructed for the four drainage parameters: gas production, bottom hole flowing pressure, liquid column height, and stroke frequency. Custom event terms were configured for each NeuralProphet base model to incorporate human-generated drainage event information. A gas production prediction NeuralProphet base model was trained using gas production data from the training period data. Corresponding drainage parameter prediction NeuralProphet base models were trained using the bottom hole flowing pressure, liquid column height, and stroke frequency from the training period data, with gas production data added as an external covariate to all three drainage parameter prediction NeuralProphet base models. The parameters of the NeuralProphet base models were set as follows: lag time (n-lags) was between 20 and 50, n-forecasts (prediction time) was set to 1, and year... The ly-seasonality is set to "True", weekly-seasonality to "False", daily-seasonality to "False", seasonality-mode to "multiplicative", learning-rate to between 0.001 and 0.1, and epoch to between 20 and 100. The custom event item is configured as follows: mode to "additive", lower-window to -3 to 0, upper-window to 0 to 10, and regularization weight to between 0.5 and 1.0.

[0041] (d) LSTM Base Model Construction and Training: A multivariate, multi-output LSTM base model is constructed to simultaneously learn the deep temporal dependencies between the four sorting parameters; the LSTM base model is trained under supervision using the training sample set obtained in step (b), and the training process is monitored using a validation sample set to prevent overfitting; the LSTM base model is a stacked network structure, which includes, in sequence: an input layer with dimensions [L, F_dynamic], where L is the length of the historical window and F_dynamic is the number of dynamic sorting features; a first LSTM layer with 64 to 256 neurons; a first Dropout layer with a dropout rate between 0.1 and 0.3; a second LSTM layer with 32 to 128 neurons; a second Dropout layer with a dropout rate between 0.1 and 0.3; a flattening layer; a first fully connected layer with 32 to 128 neurons, using the ReLU activation function; and an output layer with the number of neurons equal to the number of predicted target sorting parameters F_target, and an output dimension of [F_target].

[0042] The training configuration of the LSTM base model is as follows: use Adam or RMSprop optimizer, loss function is mean squared error or mean absolute error, epoch (training rounds) is between 50 and 150, learning-rate is between 0.001 and 0.1; add early stopping mechanism, patience value is between 10 and 50.

[0043] (e) Base model prediction: Add event labels representing human intervention to the time series data of the sampling parameters; use the trained NeuralProphet base model to perform historical fitting on the training period to obtain the predicted values ​​of the training period data. And forward prediction is performed on the verification period to obtain the predicted value of the verification period data. Using the trained LSTM base model, forward propagation is performed on both the training and validation sample sets to obtain the predicted values ​​for the training sample set. and the predicted values ​​of the validation set samples The types of operational events marked by the event markers include, but are not limited to: adjustments to the drainage system, equipment maintenance, and unblocking operations; each operational event is defined by an event name string and an event impact time window; the predicted values ​​of the training period data... Covering the NeuralProphet base model for training time data t L+1 to t m The predicted values ​​of the sampling parameters within the specified time period, and the predicted values ​​of the verification data. Including the NeuralProphet base model for validation data t m+L+1 to tn Predicted values ​​of sampling parameters within a given time period; predicted values ​​of the training sample set. Including LSTM base models on the training sample set t L+1 to t m Predicted values ​​of sampling parameters within a given time period, and predicted values ​​of the validation set samples. Including LSTM base models on the test sample set t m+L+1 to t n Predicted values ​​of sampling parameters within a given time period.

[0044] (f) Meta-model dataset construction: Using the random forest regression model sample builder, for each time point of the training period data, the predicted values ​​of the training period data at the corresponding time point are generated. The predicted values ​​of the training sample set The static geological parameters are used to construct the input feature vector of the meta-model; the actual gas production at that time point is used as the target label to generate the meta-model training sample set; for each time point of the validation period data, the predicted value of the validation period data at the corresponding time point is... The predicted values ​​of the verification sample set The static geological parameters are used to construct the input feature vector of the meta-model through feature concatenation; the actual gas production at that time point is used as the target label to generate the meta-model validation sample set; the random forest regression model sample builder constructs samples according to time points: for training period data, each reading time t j NeuralProphet base model training data prediction values Predicted values ​​of training sample set and LSTM base model The normalized static geological parameters of the well, along with all the normalized parameters of the well, are used as features of the meta-model training samples, with time t as the metric. j The actual gas production is used as the label for the training samples of the meta-model, where t j The range of values ​​is from t L+1 to t m For data during the verification period, the reading time is t each time. k NeuralProphet base model validation period data prediction values Validate the predicted values ​​of the LSTM base model on the sample set. The normalized static geological parameters of the well, along with all the normalized parameters of the well, are used as features of the meta-model validation sample, with time t as the metric. k The actual gas production is used as the label for the meta-model validation sample, where t k The range of values ​​is from t m+L+1 to t n .

[0045] (g) Meta-model fusion and final prediction: Construct a random forest regression model as a meta-model; use the training sample set of the meta-model to train the meta-model and learn the nonlinear mapping relationship from the prediction results of the base model and geological parameters to the final gas production; apply the trained meta-model to fuse the prediction results of the base model and geological parameters of the new samples, and output the final predicted value of coalbed methane well gas production; the hyperparameters of the random forest regression model are optimized by grid search or random search, wherein the value of the number of trees n_estimators ranges from 50 to 500, the value of the maximum depth max_depth ranges from 5 to 20, and the value of the minimum number of leaf node samples min_samples_leaf ranges from 1 to 5.

Claims

1. A joint NP-LSTM prediction method for coalbed methane well production, characterized in that: A meta-learning prediction model with a two-level model fusion prediction framework is constructed, which includes a base model layer and a meta-model layer. The base model layer includes NeuralProphet (NP) base model and LSTM base model, and the meta-model layer includes a random forest regression model. In the base model layer, the NeuralProphet base model is used to capture the seasonal and event-based patterns in the time series data of gas production parameters of coalbed methane wells, and the LSTM base model is used to learn the deep nonlinear time series dependencies between production parameters. At the meta-model layer, the prediction results of the NeuralProphet base model and the LSTM base model are combined with static geological parameters and then input into the random forest regression model for nonlinear fusion, finally outputting a high-precision gas production prediction value. When applied to new well prediction, the time series data of drainage parameters, including seasonal and event-related information, are input into the base model layer, and then combined with the static geological parameters of the new well into the meta-model layer to finally obtain the final new well gas production prediction sequence.

2. The NP-LSTM joint prediction method for coalbed methane well production according to claim 1, characterized in that, The joint forecasting method includes the following steps: (a) Data preprocessing: Collect time-series data of static geological parameters and drainage parameters of multiple coalbed methane wells in the target coalbed methane production area; perform cross-well normalization on the static geological parameters to eliminate the influence of dimensions; perform outlier detection and cleaning on the time-series data of drainage parameters, and perform cross-well normalization. (b) Dataset Construction: The time-series data of the drainage parameters of each coalbed methane well are arranged in ascending order by timestamp; and divided by time point t. m The full-cycle data of each well is divided into training period data and validation period data as boundaries. A sliding window sample builder with a preset historical window length and a prediction window length is used to perform rolling sampling on the training period data and the validation period data respectively to generate corresponding training sample sets and validation sample sets. Each sample includes dynamic drainage features and static geological features within the historical window, as well as the target gas production label within the prediction window. (c) NeuralProphet base model construction and training: Four independent NeuralProphet base models were constructed for the four drainage parameters; custom event terms were configured for the four NeuralProphet base models to integrate human drainage event information; the gas production data in the training period data was used to train the gas production prediction NeuralProphet base model; the bottom hole flowing pressure, liquid column height and stroke frequency in the training period data were used to train the corresponding drainage parameter prediction NeuralProphet base models, and the gas production data was added as an external covariate to all three drainage parameter prediction NeuralProphet base models; (d) LSTM base model construction and training: Construct a multivariate, multi-output LSTM base model to simultaneously learn the deep temporal dependencies between the four sorting parameters; use the training sample set obtained in step (b) to supervise the training of the LSTM base model, and use the validation sample set to monitor the training process to prevent overfitting. (e) Base model prediction: Add event labels representing human intervention to the time series data of the sampling parameters; use the trained NeuralProphet base model to perform historical fitting on the training period to obtain the predicted values ​​of the training period data. And forward prediction is performed on the verification period to obtain the predicted value of the verification period data. Using the trained LSTM base model, forward propagation is performed on both the training and validation sample sets to obtain the predicted values ​​for the training sample set. and the predicted values ​​of the validation set samples ; (f) Meta-model dataset construction: Using the random forest regression model sample builder, for each time point of the training period data, the predicted values ​​of the training period data at the corresponding time point are generated. Predicted values ​​of training sample set The static geological parameters are used to construct the input feature vector of the meta-model; the actual gas production at that time point is used as the target label to generate the meta-model training sample set; for each time point of the validation period data, the predicted value of the validation period data at the corresponding time is also used. Validate the predicted values ​​of the sample set The static geological parameters are combined to form the input feature vector of the meta-model; the actual gas production at that time point is used as the target label to generate the meta-model validation sample set. (g) Meta-model fusion and final prediction: Construct a random forest regression model as a meta-model; use the meta-model training sample set to train the random forest regression model, learn the nonlinear mapping relationship from the prediction results of the base model layer and static geological parameters to the final gas production; apply the trained random forest regression model, fuse the prediction results of the base model layer of the new sample with the static geological parameters, and output the final predicted value of coalbed methane well gas production.

3. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: The static geological parameters include at least three or more of the following: coal seam burial depth, coal seam thickness, gas saturation, Langmuir volume, Langmuir pressure, critical desorption pressure, and reservoir permeability; the drainage parameters include at least four or more of the following: gas production, bottom hole flowing pressure, flushing frequency, casing pressure, and dynamic fluid level depth.

4. The NeuralProphet-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (a), the normalization process uses the maximum-minimum normalization method to map each parameter value to the [0,1] interval; the outlier detection and cleaning uses a statistical method based on interquartile range (IQR) to identify and replace outlier data points.

5. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (b), the time segmentation point Through formula Calculation determined, where This is the start time of the sampling process. The end time of the sampling is R, which is the segmentation ratio and ranges from 0.7 to 0.

9. The historical window length L of the sliding window sample builder ranges from 20 to 50 time units, and the prediction window length F is 1 time unit. The sliding window sample builder reads t from the time-series data of the sampling parameters each time. i to t i+L The drainage parameters within a time period constitute the dynamic characteristics of the sample, while the static geological parameters of the coalbed methane well constitute the static characteristics of the sample. i+L+1 The time point sampling parameters serve as the sample labels, when Read as the first sample When reading the second sample...; when the sliding window sample builder constructs the training sample set... When the sliding window sample builder constructs the validation set... .

6. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (c), the parameters of the NeuralProphet base model are set as follows: the number of lag times (n-lags) is between 20 and 50, n-forecasts is set to 1, yearly-seasonality is set to "True", weekly-seasonality is set to "False", daily-seasonality is set to "False", seasonality-mode is set to "multiplicative", learning-rate is between 0.001 and 0.1, and epoch is between 20 and 100; the configuration of the custom event item is as follows: mode is set to "additive", lower-window is between -3 and 0, upper-window is between 0 and 10, and regularization weight is between 0.5 and 1.

0.

7. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (d), the LSTM base model is a stacked network structure, which includes, in sequence: an input layer with dimensions [L, ... F_dynamic], where L is the historical window length and F_dynamic is the number of dynamic sorting features; the first LSTM layer has 64 to 256 neurons; the first Dropout layer has a dropout rate of 0.1 to 0.3; the second LSTM layer has 32 to 128 neurons; the second Dropout layer has a dropout rate of 0.1 to 0.3; a flattening layer; the first fully connected layer has 32 to 128 neurons and uses the ReLU activation function; the output layer has the number of neurons equal to the number of predicted target sorting parameters F_target, and the output dimension is [F_target]; the training configuration of the LSTM base model is: using the Adam or RMSprop optimizer, the loss function is mean squared error or mean absolute error, the epoch (training rounds) is between 50 and 150, and the learning rate is between 0.001 and 0.1; an early stopping mechanism is added, and the patience value is between 10 and 50.

8. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (e), the operational event types marked by the event include: adjustment of the sampling system, equipment maintenance, and unblocking operations; the operational event is defined by an event name string and an event impact time window; the predicted value of the training period data. Covering the NeuralProphet base model for training time data t L+1 to t m Predicted values ​​of sampling parameters within a given time period, and verification data prediction values. Including the NeuralProphet base model for validation data t m+L+1 to t n Predicted values ​​of sampling parameters within a given time period; predicted values ​​of the training sample set. Including LSTM base models on the training sample set t L+1 to t m Predicted values ​​of sampling parameters within a given time period, and predicted values ​​of the validation set samples. Including LSTM base models on the test sample set t m+L+1 to t n Predicted values ​​of sampling parameters within a given time period.

9. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (f), the random forest regression model sample builder constructs samples according to time points: for training period data, each reading time t j NeuralProphet base model training data prediction values Predicted values ​​of training sample set and LSTM base model The normalized static geological parameters of the well, along with all the normalized parameters of the well, are used as features of the meta-model training samples, with time t as the metric. j The actual gas production is used as the label for the training samples of the meta-model, where t j The range of values ​​is from t L+1 to t m For data during the verification period, the reading time is t each time. k NeuralProphet base model validation period data prediction values Validate the predicted values ​​of the LSTM base model on the sample set. The normalized static geological parameters of the coalbed methane well, along with all other parameters, are used as features of the meta-model validation sample, with time t as the metric. k The actual gas production is used as the label for the meta-model validation sample, where t k The range of values ​​is from t m+L+1 to t n .

10. The NP-LSTM joint prediction method for coalbed methane well production according to claim 2, characterized in that: In step (g), the hyperparameters of the random forest regression model are optimized by grid search or random search, wherein the number of trees n_estimators ranges from 50 to 500, the maximum depth max_depth ranges from 5 to 20, and the minimum number of leaf node samples min_samples_leaf ranges from 1 to 5.

Images