Coal bed methane well small sample production prediction method based on transfer learning combined with LSTM

Abstract

This invention relates to the field of coalbed methane development technology and discloses a method for predicting small-sample coalbed methane well production based on transfer learning combined with LSTM. The method includes the following steps: Step 1: Select coalbed methane wells with sufficient data as the source domain and select small-sample coalbed methane wells to be predicted as the target domain. Collect raw data from both the source and target domains, including parameters related to the drainage process. Step 2: Preprocess the raw data from the source and target domains respectively to obtain preprocessed source and target domain data. A feature selection method combining variance inflation factor analysis, correlation clustering, and bicorrelation verification is used to eliminate multicollinearity and low-correlation features, identifying core influencing factors as model inputs. Simultaneously, the key features selected from the source domain are directly used in the target domain fine-tuning stage, simplifying the hyperparameter tuning process and improving model training efficiency and prediction stability.

Description

Technical Field

[0001] This invention relates to the field of coalbed methane development technology, specifically to a method for predicting small sample production of coalbed methane wells based on transfer learning combined with LSTM. Background Technology

[0002] Coalbed methane (CBM), as a clean and efficient unconventional natural gas resource, holds significant strategic importance for optimizing energy structure, reducing carbon emissions, and ensuring energy security. CBM well production forecasting is a core component of CBM field development planning, dynamic production control, and economic benefit assessment. Accurate production forecasting provides a scientific basis for optimizing development plans and adjusting drainage processes, thereby improving the overall development efficiency and resource recovery rate of CBM fields. As CBM development expands into areas with complex geological conditions, single-well production is influenced by a combination of factors, including geological parameters, engineering measures, and drainage processes, exhibiting significant nonlinear and temporal fluctuations. This places higher demands on the adaptability and accuracy of production forecasting technologies.

[0003] In the field of coalbed methane well production prediction, existing technologies can be mainly divided into two categories: traditional prediction methods and machine learning prediction methods. Traditional prediction methods are represented by numerical simulation and empirical formula methods. Numerical simulation methods establish mathematical models of coalbed methane seepage and combine them with basic data such as reservoir parameters and fluid properties for numerical solutions. This can reflect the physical mechanism of coalbed methane extraction. However, this method relies on a large number of accurate geological and engineering parameters, the modeling process is complex, the calculation cost is high, and it is highly sensitive to parameters. It is difficult to quickly adapt to complex geological conditions and output reliable prediction results. Empirical formula methods fit statistical relationships between production and influencing factors based on field measured data. They are characterized by simple calculation and strong applicability. However, due to the limitations of the fitted samples, their prediction range and accuracy are difficult to meet the actual needs under complex working conditions. Especially in the stage of drastic dynamic changes in production, the prediction error increases significantly.

[0004] With the development of artificial intelligence technology, machine learning methods, with their powerful nonlinear fitting and feature mining capabilities, have been widely used in coalbed methane well production prediction. Among them, recurrent neural networks (RNNs) and their improved models, long short-term memory neural networks (LSTMs), have shown outstanding performance. By introducing core components such as forget gates, input gates, output gates, and cell states, LSTMs effectively solve the gradient vanishing or gradient exploding problems that traditional RNNs encounter when processing long-term time-series data. They can accurately capture long-term dependencies and dynamic changes in production time-series data, making them one of the mainstream models for time-series production prediction.

[0005] However, existing LSTM-based coalbed methane well production prediction technologies still suffer from several technical shortcomings that urgently need to be addressed: First, the models are highly dependent on training data. The training effect of LSTM models is highly dependent on sufficient high-quality time-series data. However, in actual coalbed methane development, newly commissioned wells and wells with complex geological conditions often suffer from insufficient data accumulation. In small sample scenarios, the models struggle to fully learn the correlation between production and influencing factors, easily leading to overfitting, low prediction accuracy, and large trend fitting deviations, which cannot meet actual production needs. Second, the rationality of feature selection is insufficient. Coalbed methane production is influenced by numerous factors, covering multiple dimensions such as geology, engineering, and drainage. Existing technologies often directly use the original parameters as model inputs without fully considering the multicollinearity problem among features, resulting in increased model complexity and decreased generalization ability. At the same time, some methods only select features through single correlation analysis, making it difficult to comprehensively identify core factors strongly correlated with production, further affecting prediction accuracy. Third, the model has poor transfer adaptability. The geological conditions and drainage processes of different coalbed methane wells vary, resulting in different distribution characteristics of production data. Existing LSTM models are mostly trained independently for single well conditions, and cannot reuse the effective information contained in the historical data of wells already in production. In the prediction of new wells or wells with scarce data, a lot of hyperparameter debugging and model training are required again, which not only increases the model deployment cost but also reduces the prediction efficiency. Fourth, the data preprocessing is not targeted enough. The raw data collected in the field of coalbed methane wells is easily affected by factors such as equipment failure, power outages, and environmental interference, and contains outliers, missing values, and random noise. Existing preprocessing methods mostly use simple methods such as deleting outliers and filling missing values ​​with the mean, without fully considering the temporal integrity of production data. This can easily destroy the temporal correlation characteristics of the data, resulting in poor quality of model input data and thus affecting the reliability of prediction results.

[0006] The aforementioned technical problems make it difficult for existing coalbed methane well production prediction technologies to meet the actual needs of efficient coalbed methane field development in terms of adaptability to small sample scenarios, prediction accuracy, and model deployment efficiency. Therefore, developing a coalbed methane well production prediction method that can reduce data dependence, improve small sample prediction accuracy, and simplify model training process has become an urgent technical challenge to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for predicting small-sample production of coalbed methane wells based on transfer learning combined with LSTM, thus solving the problems mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting small-sample production of coalbed methane wells based on transfer learning combined with LSTM, comprising the following steps: Step 1: Select coalbed methane wells with sufficient data as the source domain and select small sample coalbed methane wells to be predicted as the target domain. Collect the raw data of the source domain and the target domain. The raw data includes parameters related to the drainage process. Step 2: Preprocess the raw data of the source domain and the target domain respectively to obtain preprocessed source domain data and target domain data; Step 3: Perform feature filtering on the preprocessed source domain data and target domain data to determine the key features affecting coalbed methane well production; Step 4: Construct a basic prediction model based on LSTM, which includes a two-layer LSTM layer and a fully connected layer; Step 5: Use the preprocessed source domain data to pre-train the basic prediction model to obtain the pre-trained model; Step 6: Employ a hierarchical learning rate fine-tuning strategy to transfer the parameters of the pre-trained model to the target domain, and fine-tune the model using the pre-processed target domain data to obtain the final yield prediction model. Step 7: Input the small sample data of the target domain into the final production prediction model and output the coalbed methane well production prediction results.

[0009] Preferably, the relevant parameters of the drainage process in step 1 include at least one of the following: bottom hole pressure, dynamic fluid level, casing pressure, stroke, number of strokes, cumulative water production, fracturing fluid flowback rate, daily water production, torque, current, rotational speed, system pressure, and pump efficiency.

[0010] Preferably, the data preprocessing in step 2 includes outlier handling, missing value imputation, and data smoothing, as follows: Outlier handling: outliers in the original data are identified using the Z-score normalization method and set as missing values; Missing value imputation: missing values ​​are imputed using linear interpolation; Data smoothing: random noise in the original data is eliminated using Kalman filtering to improve data reliability.

[0011] Preferably, the feature selection in step 3 includes multicollinearity removal and bicorrelation verification, as follows: Multicollinearity removal: Variance expansion factor analysis is used to determine whether multicollinearity exists in the data, and highly correlated features are grouped using correlation clustering, with one representative feature retained in each group; Bicorrelation verification: The relationship between the selected features and coalbed methane production is analyzed by combining grey relational analysis and Pearson coefficient, and features with strong correlation to production are retained as key features.

[0012] Preferably, in the variance inflation factor analysis, a variance inflation factor greater than 10 is determined to be severe multicollinearity, a variance inflation factor between 5 and 10 is determined to be moderate multicollinearity, and a variance inflation factor less than 5 is determined to be non-multicollinearity; in the correlation clustering method, a correlation threshold is set, and highly correlated feature groups are identified based on this threshold.

[0013] Preferably, in step 4, when constructing the basic prediction model, a sliding window is introduced to convert time series data into model input, and multiple training samples are created by moving a fixed-size window; at the same time, the lagged features of the target variable are introduced to capture the autocorrelation and trend information in the time series.

[0014] Preferably, in step 5, the pre-training process uses an adaptive optimization algorithm as the optimizer, sets hyperparameters such as initial learning rate, batch size, sliding step size, lag feature step size, and number of LSTM hidden layer units, and uses mean squared error as the loss function; during pre-training, the training set and test set are divided according to a preset ratio to avoid interference from future information leakage on model training.

[0015] Preferably, the layered learning rate fine-tuning strategy in step 6 is as follows: the LSTM layer is set with a smaller learning rate to protect the temporal feature extraction capability obtained through pre-training; the fully connected layer is set with a larger learning rate to accelerate parameter update speed and enable the model to quickly adapt to the target domain data distribution.

[0016] Preferably, after outputting the production forecast results in step 7, the forecast results are evaluated using mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination. The closer the coefficient of determination is to 1 and the smaller the other indicators are, the higher the accuracy of the forecast model.

[0017] Preferably, the small sample data of the target domain is time-series data with no more than 100 samples, and the temporal integrity of the data is preserved.

[0018] This invention provides a method for predicting small-sample production of coalbed methane wells based on transfer learning combined with LSTM. It has the following beneficial effects: 1. This invention employs a transfer learning strategy of "source domain pre-training - target domain hierarchical fine-tuning" and combines the efficient extraction capability of a two-layer LSTM network for temporal features. It effectively reuses the production correlation patterns contained in the source domain coalbed methane wells with sufficient data, significantly reduces the dependence on small sample data in the target domain, and achieves accurate prediction of coalbed methane well production in small sample scenarios. This solves the problems of low prediction accuracy and poor trend fitting of traditional LSTM models under small sample data.

[0019] 2. This invention employs a feature selection method that combines variance inflation factor analysis, correlation clustering, and bicorrelation verification to eliminate multicollinearity and low-correlation features, identify core influencing factors as model inputs, and directly reuse key features selected from the source domain during the target domain fine-tuning stage. This simplifies the hyperparameter tuning process and improves model training efficiency and prediction stability.

[0020] 3. This invention employs a layered learning rate fine-tuning strategy, setting a smaller learning rate for the LSTM layer to protect the temporal feature extraction capability obtained through pre-training, and setting a larger learning rate for the fully connected layer to quickly adapt to the target domain data distribution. Combined with data preprocessing procedures such as Z-score standardization, linear interpolation, and Kalman filtering, this invention achieves a significant reduction in prediction error and optimization of model adaptability, making the prediction results more consistent with actual production conditions. Attached Figure Description

[0021] Figure 1 This is a diagram of the LSTM network unit structure at a given moment in this invention. Figure 2 This is a diagram of the transfer learning neural network structure of the present invention; Figure 3 This is a schematic diagram of the correlation clustering results of the present invention; Figure 4 This is a schematic diagram of the pre-training visualization results of the present invention; Figure 5 This is a schematic diagram comparing the predicted and actual daily gas production values ​​of different models in this invention. Figure 6 This is a schematic diagram of the small sample prediction results of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 - Appendix Figure 6 This invention provides a method for predicting small sample production of coalbed methane wells based on transfer learning combined with LSTM.

[0024] Long Short-Term Memory (LSTM) neural networks are a special type of recurrent neural network (RNN) specifically designed to solve the vanishing or exploding gradient problems of traditional RNNs. The innovation of LSTM lies in the introduction of a "gate" mechanism to control the flow of information. It includes key components such as cell states, forget gates, input gates, and output gates.

[0025] The forget gate determines which information needs to be forgotten, and the update formula at time t is as follows:

[0026] In the formula: The input information at time t, It is the sigmoid activation function. The hidden layer state at time t-1 , The weights and biases of the forget gate.

[0027] The input gate determines which new information needs to be stored, and the update formula at time t is as follows:

[0028]

[0029] In the formula: The information is the candidate state after the tanh transformation; the information in it may be updated to the current memory time. , The weights and biases of the input gate, , The weights and biases of the candidate states.

[0030] The cell state is updated based on the decisions of the forget gate and the input gate. The update formula at time t is as follows:

[0031] In the formula: and The states of memory cells at time t and t-1 are respectively.

[0032] The output gate determines the output at the current time step, and the update formula at time t is as follows:

[0033]

[0034] Transfer learning is a machine learning technique mainly divided into two types: pre-training and fine-tuning, and feature transfer. This study uses the first type, namely transfer: transferring knowledge learned on one task (source domain) to another related task (target domain); learning: continuing to learn on the target domain to adapt the model to the new data distribution. The transfer strategy is hierarchical learning rate fine-tuning. LSTM layers use a smaller learning rate to better extract temporal features, while fully connected layers use a larger learning rate to map the prediction results.

[0035] 1. Experimental data and preprocessing in the study area This study uses three horizontal wells in the Fukang Eighth District as research objects. Well FSL-30 was used as the source domain pre-training dataset, and wells FSL-32 and FSL-34 were used as the target domain fine-tuning dataset. Data sampling was performed once daily. Data preprocessing workflow: Outlier Handling: During coalbed methane data collection, many missing or outlier values ​​may occur due to equipment malfunctions or power outages. Z-score standardization is used to identify outliers and treat them as missing values.

[0036] Missing values ​​were imputed. Since coalbed methane data is time-series data, arbitrary deletion might disrupt its temporal sequence. Therefore, linear interpolation was used to impute missing values. The formula for linear interpolation is:

[0037] In the formula, ( , )and( , ( ) are two known points.

[0038] Data smoothing. Kalman filtering is used to eliminate anomalies caused by random noise deviations, improving data reliability.

[0039] 2. Evaluation Indicators and Model Parameters Evaluation metrics. Mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) are used. 2 The overall performance of the model is evaluated, where R... 2 The closer the value is to 1 and the smaller the other indicators are, the higher the model accuracy.

[0040] The formulas for calculating the four indicators are as follows:

[0041]

[0042]

[0043]

[0044] In the formula: This is the predicted value for coalbed methane production. This is the actual value.

[0045] Model parameters. In dealing with time series problems, it is necessary to divide time segments to construct sample data [7]. A sliding window is introduced to convert time series data into machine learning model input, and multiple training samples are created by moving a window of a fixed size. At the same time, the lag features of the target variable are introduced to capture the autocorrelation and trend information in the time series. During model training, the optimizer uses the adaptive optimization algorithm (Adam). After search and debugging, the initial learning rate (Lr) is set to 0.002, the batch size (batch_size) is set to 32, the sliding step size (input_size) is set to 5, and the lag feature step size is set to 1. A two-layer LSTM is used, and the number of hidden layer units is set to 128.

[0046] 3. Analysis of the main controlling factors of coalbed methane The factors influencing coalbed methane (CBM) production are complex and involve numerous parameters, primarily including geological parameters, engineering parameters, and drainage process parameters. Geological parameters are determined by the inherent properties of the coal seam. Engineering parameters are generated by drilling and fracturing processes; both are fixed for a single well and have no significant impact on the dynamic numerical prediction of single-well production. Drainage process parameters are mainly generated during drainage control. After drilling and fracturing, drainage parameters such as bottomhole pressure, dynamic fluid level, casing pressure, stroke, and stroke frequency become the main factors affecting CBM well production. Most drainage process parameters are time-series data with consistent time intervals and continuous sampling over time, exhibiting dynamic trends over time. Uncovering these patterns is the main research focus for time-series data.

[0047] Multicollinearity removal. Variance inflation factor (VIF) analysis was used to determine the presence of multicollinearity. VIF > 10: Severe multicollinearity, it is recommended to delete one feature. VIF 5-10: Moderate multicollinearity, attention is needed. VIF < 5: No multicollinearity problem. Specific results are shown in Table 1.

[0048] Table 1. Analysis of Variance Inflation Factor (VIF) ; Correlation clustering is employed to group highly correlated features, retaining only one representative feature in each group. This effectively improves model interpretability and addresses the multicollinearity problem. A correlation coefficient matrix is ​​calculated from the original feature pairs. Highly correlated groups are identified based on a correlation threshold, and representative features within each group are selected to ultimately obtain the feature set. The correlation threshold is the core parameter used to define highly correlated features and complete the grouping.

[0049] The correlation analysis results show that the calculation of highly correlated features is shown in Table 2. With a correlation threshold set at 0.68, correlation clustering analysis was performed, as follows: Figure 3 As shown, a total of 3 highly correlated feature groups were identified.

[0050] Table 3-5-2 Highly Correlated Features ; Table 3. Nonlinear Correlation Status Table ; Table 3 shows that the characteristic importance of casing pressure is relatively low, indicating a moderately nonlinear relationship; the nonlinear strength of daily water production is weak (0.440), and the linear correlation is low (0.120). Therefore, neither is retained. The feature with a strong linear relationship, cumulative water production, is retained and used as an input feature in subsequent model predictions.

[0051] 4. Model Performance Validation To systematically verify the predictive performance of the LSTM-transfer learning fusion model, this study conducted experiments from two dimensions: conventional sample prediction and small sample prediction. Three horizontal coalbed methane wells—FSL-30 (source domain), FSL-32, and FSL-34 (target domain)—in the Fukang No. 8 area were used as research objects. By comparing the model with the traditional LSTM model, the results were analyzed based on error metrics (MAE, RMSE, MAPE) and goodness-of-fit (R²). 2 A dual perspective is used to quantitatively evaluate the advantages of the model in terms of improved accuracy, reduced data dependence, and simplified hyperparameters.

[0052] The conventional sample prediction uses "source domain pre-training - target domain fine-tuning" as the core process. First, the model is pre-trained based on the data-rich FSL-30 well. Then, the pre-trained parameters are transferred to FSL-32 and FSL-34 wells for hierarchical fine-tuning. This ensures that the model adapts to the target domain data distribution while retaining the ability to extract time-series features from the source domain.

[0053] (1) Source domain pre-training results (FSL-30 well) The FSL-30 well dataset is sampled once a day. After outlier removal (Z-score standardization), missing value imputation (linear interpolation), and data smoothing (Kalman filtering) preprocessing, it strictly follows the characteristics of time-series data and divides the training set (first 80% of samples) and test set (last 20% of samples) in an 8:2 ratio to avoid interference with model training from future information leakage.

[0054] The pre-training stage adopts a basic structure of "two-layer LSTM + fully connected layer". The hyperparameters are determined after grid search optimization: the initial learning rate of Adam optimizer is 0.002, the batch size is 32, the sliding window step size is 5, the lag feature step size is 1, the number of hidden units in LSTM is 128, and the loss function is mean squared error (MSE).

[0055] Visualization of pre-training results, such as Figure 4As shown, the model's performance on the test set is as follows: Mean Absolute Percentage Error (MAPE) 3.43%, and Coefficient of Determination (R²) 2.5%. 2 The predicted value is 0.94, and the time-series trend of the predicted value is highly consistent with the actual value. The residual distribution is concentrated near zero and has no obvious skewness. This indicates that the pre-trained model has fully grasped the time-series correlation between coalbed methane production and the main controlling factor (cumulative water production), and can serve as a reliable basic model for fine-tuning the target domain.

[0056] In the target domain fine-tuning stage, a layered learning rate strategy was adopted. A small learning rate (1e-5) was set for the LSTM layers to protect the temporal feature extraction capabilities obtained from pre-training and prevent core parameters from being "covered" by small samples in the target domain. A larger learning rate (1e-4) was set for the fully connected layers to accelerate parameter updates and enable them to quickly adapt to the local data distribution of wells FSL-32 and FSL-34. During fine-tuning, the target domain input features were not re-selected; the "cumulative water production" feature determined by the source domain was directly used. The LSTM network structure (number of layers, number of hidden units) remained consistent with the pre-trained model. Model transfer was achieved through only two layers of learning rate adjustment, significantly simplifying the hyperparameter tuning process.

[0057] The model prediction results for wells FSL-32 and FSL-34 after fine-tuning are shown in Table 3. The MAPE of the FSL-32 well test set decreased from 10.40% to 6.1% compared to the traditional LSTM, and R... 2 The MAE and RMSE decreased from 280.73 and 432.17 to 167.05 and 250.25, respectively, as the MAE increased from 0.70 to 0.90. For the FSL-34 well test set, the MAPE decreased from 0.614% with traditional LSTM to 0.288%, and R... 2 The MAE and RMSE decreased from 36.54 and 42.19 to 17.07 and 24.45, respectively, as the MAE increased from 0.80 to 0.93.

[0058] Table 4 Comparison of the predictive effects of different models on daily gas production ; While traditional LSTM prediction curves can capture the overall trend of production, they have obvious defects in numerical fitting accuracy. The fusion model's predicted values ​​almost overlap with the actual values, and it can maintain extremely low errors even in the stable production phase, fully verifying the adaptability of the hierarchical fine-tuning strategy to the target domain data.

[0059] A comparison of the effects of different models on predicting daily gas production for FSL-34 was conducted, such as... Figure 5 The above shows the prediction results of LSTM fusion transfer learning.

[0060] Here, the FSL-34 dataset is reduced to approximately 100 samples to preserve the temporal sequence. Only the gas production data from the first 100 days are retained. This dataset is used as a small-sample prediction dataset to explore whether the model in this paper can achieve effective prediction results.

[0061] Table 5 Comparison of the predictive effects of different models on daily gas production ; Depend on Figure 6 It can be seen that the prediction curve of the LSTM model is completely unrelated to the actual production curve, while the prediction curve of the model in this paper is basically consistent with the trend of the actual production curve and has a smaller error. Overall, it is more consistent with the prediction curve of the LSTM model and more in line with reality.

[0062] As shown in Table 5, under the same number of prediction days, the prediction error of the model in this paper is lower than that of the LSTM model, and R0 2 The accuracy improved from -2.63 to 0.86. The mean relative error decreased by approximately 74%. Both the MAE error and RMSE error were significantly reduced compared to the LSTM model.

[0063] In summary, LSTM-integrated transfer learning models significantly improve model accuracy. The model is pre-trained in the source domain and then fine-tuned in the target domain to reduce inter-domain differences, allowing the use of source domain knowledge to reduce the amount of data needed in the target domain, thus facilitating small-sample prediction [9,10,11]. In the target domain fine-tuning process of the model presented in this paper, only the learning rates of the LSTM layers and fully connected layers need to be adjusted, solving the problem of LSTM models requiring the tuning of numerous hyperparameters.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting small-sample production of coalbed methane wells based on transfer learning combined with LSTM, characterized in that, Includes the following steps: Step 1: Select coalbed methane wells with sufficient data as the source domain and select small sample coalbed methane wells to be predicted as the target domain. Collect the raw data of the source domain and the target domain. The raw data includes parameters related to the drainage process. Step 2: Preprocess the raw data of the source domain and the target domain respectively to obtain preprocessed source domain data and target domain data; Step 3: Perform feature filtering on the preprocessed source domain data and target domain data to determine the key features affecting coalbed methane well production; Step 4: Construct a basic prediction model based on LSTM, which includes a two-layer LSTM layer and a fully connected layer; Step 5: Use the preprocessed source domain data to pre-train the basic prediction model to obtain the pre-trained model; Step 6: Employ a hierarchical learning rate fine-tuning strategy to transfer the parameters of the pre-trained model to the target domain, and fine-tune the model using the pre-processed target domain data to obtain the final yield prediction model. Step 7: Input the small sample data of the target domain into the final production prediction model and output the coalbed methane well production prediction results.

2. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, The relevant parameters of the drainage process mentioned in step 1 include at least one of the following: bottom hole pressure, dynamic fluid level, casing pressure, stroke, number of strokes, cumulative water production, fracturing fluid flowback rate, daily water production, torque, current, rotational speed, system pressure, and pump efficiency.

3. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, The data preprocessing described in step 2 includes outlier handling, missing value imputation, and data smoothing, as follows: Outlier handling: outliers in the original data are identified using the Z-score normalization method and set as missing values; Missing value imputation: missing values ​​are imputed using linear interpolation; Data smoothing: random noise in the original data is eliminated using Kalman filtering to improve data reliability.

4. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, The feature selection in step 3 includes multicollinearity removal and bicorrelation verification, as follows: Multicollinearity removal: Variance expansion factor analysis is used to determine whether multicollinearity exists in the data, and highly correlated features are grouped using correlation clustering, with one representative feature retained in each group; Bicorrelation verification: The relationship between the selected features and coalbed methane production is analyzed by combining grey relational analysis and Pearson coefficient, and features with strong correlation to production are retained as key features.

5. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 4, characterized in that, In the variance inflation factor analysis, a variance inflation factor greater than 10 is considered severe multicollinearity, a variance inflation factor between 5 and 10 is considered moderate multicollinearity, and a variance inflation factor less than 5 is considered non-multicollinearity. In the correlation clustering method, a correlation threshold is set, and highly correlated feature groups are identified based on this threshold.

6. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, In step 4, when building the basic prediction model, a sliding window is introduced to convert time series data into model input, and multiple training samples are created by moving a fixed-size window; at the same time, the lagged features of the target variable are introduced to capture the autocorrelation and trend information in the time series.

7. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, In step 5, the pre-training process uses an adaptive optimization algorithm as the optimizer, setting hyperparameters such as initial learning rate, batch size, sliding step size, lag feature step size, and number of LSTM hidden layer units, and using mean squared error as the loss function. During pre-training, the training set and test set are divided according to a preset ratio to avoid interference from future information leakage on model training.

8. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, The layered learning rate fine-tuning strategy described in step 6 is as follows: the LSTM layer is set with a smaller learning rate to protect the temporal feature extraction capability obtained through pre-training; the fully connected layer is set with a larger learning rate to accelerate parameter update speed and enable the model to quickly adapt to the target domain data distribution.

9. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to claim 1, characterized in that, After outputting the production forecast results in step 7, the forecast results are evaluated using mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination. The closer the coefficient of determination is to 1 and the smaller the other indicators are, the higher the accuracy of the forecast model.

10. The method for predicting small-sample coalbed methane well production based on transfer learning and LSTM according to any one of claims 1-9, characterized in that, The small sample data in the target domain is time-series data with no more than 100 samples, and the temporal integrity of the data is preserved.

Images