Emedia:ISSN1529-7306 Emedia/Online Inc.

Emedia: ISSN 1529-7306 Vol.39_1, 1-17

Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution,

and reproduction in any medium, provided the original work is properly cited.

Coalbed Methane Production Forecasting Based on LSTM

and Transfer Learning Integration

Wenjie Li¹,2, Fengnian Wang¹,2, Chenglong Qiu¹, Chenchen Jia¹, Jinxin Meng3*,

Yuxia Wang3, Bei Zhu3

1.Xinjiang Yaxin Coalbed Methane Resources Technology Research Co., Ltd., Urumqi

830009, Xinjiang Uygur Autonomous Region, China

2.Key Laboratory of Coalbed Methane Exploration and Development in Xinjiang

3. School of Petroleum Engineering, Yangtze University, Wuhan 430100, Hubei, China

*Corresponding Author:Jinxin Meng

Email:i2684501257@163.com

Received 26 March 2026; Accepted 12 July 2026

Abstract Traditional Long Short-term memory (LSTM) method faces a lot of challenges

during coalbed methane production prediction, such as poor generalization on the small

datasets, difficult hyper parameter turning and poor accuracy in prediction. Due to these

challenges this research introduces a new method which is the combination of LSTM and

Transfer Learning inorder to encounter to these issues. This approach combines the Person

Correlation Coefficient together with the Grey relational degree to identify production

factors which is more important by using a “Pre-training Hierarchical Fine-Tuning” transfer

learning framework. In order to maintain the ability of extracting temporal features, the

LSTM layer uses a low learning rate while fully connected layer uses a high learning rate for

faster convergence. When we applied this new strategy to the FSL-34 well, it’s R² value

increased from 0.8 to 0.93 and the Mean Absolute Percentage Error (MAPE) decreased from

0.614% to 0.288%, without forgetting in the situation of small datasets around 100, R²

increased greatly from -2.3 to 0.94 and the MAPE decreased by 37%. These two examples

shows clearly the advantages of the new proposed method over Traditional LSTM models.

While using Maximum Mean Discrepancy (MMD) domain arrangement helps bridge the

gap between separate domains, utilizing the transfer of knowledge from a source domain

reduces the data requirement in the target domain. I general, this strategy provides a very

effective coalbed methane production prediction in cases with scanty data samples.

Keywords Coalbed methane; Long Short-Term Memory network; Transfer learning;

Production prediction; Small-sample prediction

1 Introduction

To predict accurately and correctly production of coalbed methane (CBM) is very

important for maximizing recovery rate and improve development plan[1]. Its very

important to Characterize in a correct way the complex relationships between different

parameters including production parameters, engineering parameters and geological.

Production decline analysis methods and Traditional numerical simulations offering a

more effective way to address reservoir heterogeneity and production uncertainties due to

raised of Machine Learning in recent years[2]. One of the very important and well known

Wenjie Li et al.

application method in the field for it’s ability of identifying long-term dependencies in

time series data is Long Short-Term memory denoted as LSTM in short. Due to it’s special

“Gating Mechanism”, the model proved to very good and flexible in predicting CBM

Production. Apart from it’s greater efficiency and it’s important to the field, LSTM models

also faces some challenges in real time applications. One of the basic Challenge faces

LSTM model lie in it’s complex hyperparameter combinations[3] which needs a number

of experiment inorder to get ideal configurations. Apart from that challenge, LSTM models

incase of small data samples become very sensitive to noise otherwise the performance

decrease it’s ability. These constraints hinder the training efficiency and generalization

performance of LSTM models in real-world applications.[4-5].

The problem of limited data and model customization is solved by using the knowledge

from a source domain to improve performance in a target area. In this study, a new

prediction method is introduced that combines Long Short-Term Memory (LSTM)

networks with transfer learning techniques [6]. Improving accuracy of prediction,

simplifying hyper parameter tuning and reducing errors on dealing with small datasets are

the three main Challenge discussed by this new research by using a

"Pre-training-hierarchical fine-tuning" technique to refine the model building and

integrating multi-dimensional feature selection techniques inorder to identify critical

variables. In a case study of horizontal wells in the Fuchang District’s eight coal seam, the

technique shows very good results in providing a promising technical path for predicting

coal seam gas production. The Potential of transfer learning in improving prediction

models and opens up new possibilities for optimizing gas production forecasts is

highlighted by this study.

2 Principles and Construction of LSTM Transfer Learning

2.1 Fundamentals of LSTM Neural Networks

An improved version of traditional RNNs, pointing the problem of vanishing or exploding

gradients by utilizing a "Gate mechanism" is known as Long Short-term Memory (LSTM)

Networks [7]. This mechanism contain the Forget gate, output gate, input gate and Cell

state, which make the network to gain important information over long range. LSTM

networks are very excellent at identifying long-term dependencies in data.

Forget Gate: It Uses the sigmoid function, which is expressed in equation number (1)

below to determine whether to keep or discard historical data.:

ft=σwf󰇟ht-1,xt󰇠+bf (1)

Where: xt is the input information at time step t, σ is the sigmoid activation function,

ht-1 is the hidden layer state at time step t-1, wf , bf are the weights and bias of the forget

gate.

Input gate: Filters new information and updates candidate states. The Formulas are as

follows:

it=σ󰇛wi󰇟ht-1,xt󰇠+bi󰇜 (2)

'= tanh󰇛wC󰇟ht-1,xt󰇠+bC󰇜 (3)

Where:Ct

' represents the candidate state after tanh transformation; information in may be

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

updated to the current memory time step.wi andbi denote the input gate's weight and bias,

respectively.wC andbC denote the candidate state's weight and bias, respectively.

Cell state update: Updated based on the decision from the forget gate and input gate, as

follows:

Ct=ftCt-1+itCt

' (4)

Where:Ct andCt-1 represent the memory cell states at time t and t-1, respectively.

Output Gate: Uses the following formula to generate the hidden layer state at the current

time step:

ot=σ󰇛wo󰇟ht-1,xt󰇠+bo󰇜 (5)

ht=ottanh Ct (6)

Fig. 1 LSTM Network Unit Structure at Time t

2.2 Transfer Learning Fusion Strategy

Pre-training and fine-tuning are the primary uses of transfer learning, an effective tool in

machine learning. This study focuses on the "pre-training - hierarchical fine-tuning"

framework in the context of coalbed methane wells[1,13]. The three-step approach, which

shows how transfer learning can optimize methane well analysis, consists of data

pre-processing, model pre-training, and fine-tuning.

For building a basic or fundamental model during the first training phase, the source

domain of FSL-30 dataset is used. The aim was to gain a basic and fundamental

understanding of extracting significant temporal patterns associated with the different

generation of coalbed methane using effective “LSTM + Fully connected layer” approach.

Parameters of the model are shifted from the source domain towards the targeted ones, this

is happening during Fine-Tuning phase of the FSL- 32 wells and FSL-34 wells. To insure

peak performance, a Hierarchical learning rate technique is used. The LSTM Layer was

given a lower learning rate of 1e-5 inorder to extract time features efficiently while the

Wenjie Li et al.

fully connected layer was given a greater learning rate of 1e-4 so as to quickly adjust to the

data distribution of the target domain. Apart from that, Techniques for feature engineering

optimization are also used. In order to prevent future information leakage, training samples

are enlarged using a sliding window technique, lag features are added to capture temporal

autocorrelation, and data is divided depending on time order.

INPUT LSTM

LR:0.002 FC

LR:0.001 OUTPUT

source

domain

INPUT LSTM

LR:5e-5 FC

LR:1e-4 OUTPUT

target

domain

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

Fig. 2 Schematic Diagram of Transfer Learning Principle

2.3 MMD for Assessing Domain Similarity Distribution

In transfer learning, the similarity between the source domain and the target domain is the

core prerequisite for selecting transfer strategies and ensuring transfer effectiveness,

especially in energy engineering scenarios such as coalbed methane/oil daily gas

production prediction - these scenarios have strong domain dependencies, and data

distribution is significantly affected by geological and engineering conditions. Similarity

directly determines the feasibility and efficiency of "transfer". The Maximum Mean

Discrepancy (MMD) can be used to compare the distribution differences between two

datasets using a Reproducible Kernel Hilbert Space (RKHS), which can help determine the

similarity between source and target domain data. Its square form is theoretically defined

as:

MMD2(P,Q)=Ex,x'∼P(k(x,x'))+Ey,y'∼Q(k(y,y'))-2Ex∼P,y∼Q(k(x,y))

(7)

Ex,x'∼P(k(x,x')): The expected value of the kernel function for two independent samples x

and x 'in distribution P, reflecting the correlation between samples within P;Ey,y'∼Q(k(y,y')):

reflects the correlation of samples within the distribution Q; Ex∼P,y∼Q(k(x,y)): The

expected value of the kernel function between sample x of distribution P and sample y of

distribution Q, reflecting the correlation between samples of different distributions

However, in practical applications, it is not possible to directly obtain the true expected

distribution, and empirical estimation is required through a limited sample set. Assuming

sample set X={xi}i=1

n (from distribution p), (Y={yj}j=1

m) (from j=1 to m, from

distribution q), the unbiased empirical estimation formula for MMD is:



MMD2=1

n(n-1) k

i≠j(xi,xj)+ 1

m(m-1) k

i≠j(yi,yj)-2

nm k

i,j(xi,yj)

(8)

n: Source domain sample size, m: Target domain sample size;k

i≠j (xi,xj): The sum of the

kernel function values of all non repeating sample pairs in the source

domain; 1

m(m-1) k

i≠j (yi,yj) : The sum of the kernel function values of all non repeating

sample pairs in the target domain;k

i,j(xi,yj) : The sum of the kernel function values of all

cross sample pairs in the source and target domains (i traverses the source domain, j

traverses the target domain).

2.4 Principles of MMD Domain Alignment

One of a widely used technique in transfer learning and domain adaptation is called MMD

domain alignment. The method involves mapping data from the source/target domain to a

Reproducing Kernel Hilbert Space (RKHS) and then minimizing the distance between

their distributions to ensure convergence, aiming to achieve efficient distribution

arrangement.

In feature mapping, the feature extractors include LSTM hidden layers that seek to extract

features of both two ways which is source domain sample and target domain samples.

Wenjie Li et al.

These features are then transformed to the RKHS with the help of kernel functions, such as

Gaussian or linear kernels. The method has the advantage of efficiently capture a nonlinear

correlations between characteristics without having to do more complicated computations

in higher dimensions.

The value of MMD is calculated by an independent empirical formula. This involves the

determination of the mean cross-correlation of the source domain and target domains and

the mean autocorrelation in each of them. Due to analytical reasons, a measure of the

difference between the distributions can be generated by comparing these values.

MMD calculation: Its an objective empirical formula that measures the difference in

distribution by combining source domain/target domain’s autocorrelation and

cross-correlation between the two domains. In practical model training, the MMD loss and

task loss are kept to provide a total loss function that strikes a balance between task

performance and domain alignment effectiveness and the formula is as follows:

L=Ltask+λ∼MMD2(XS,XT) (9)

To obtain a balanced relationship between task performance and domain alignment is

very important in Machine Learning Models. The core task loss, Ltask is determined by

the Mean Absolute Error (MAE). To control the model focus much on domain alignment at

the expense of task performance, a balancing hyper parameter, λ is used to adjust the

weight of the MMD loss. XS represent the source domain and XT denote the features of

the target domains.

3 Experimental Design and Results Analysis

3.1 Experimental Data and Preprocessing for the Study Area

An investigation based on the use of MMD approach was subjected on three horizontal

wells in Fukang District 8. The target domain of fine-tuning was Wells FSL-32 and

FSL-34 and the source domain of pre-training was well FSL-30. To ensure accuracy and

reliability, data were collected on a daily basis and subjected to a thorough planning

process.

Malfunction of the equipment and power cuts were also a problem that was occasioned

lacking or atypical values in the process of coalbed methane data collection. To make sure

that the results are properly analysed and interpreted, Z-score normalisation was used to

identify exceptions that would be treated as missing values.

Assuming missing values. The deletion of data might be distroyed since coalbed methane

is time-series data, which can result in the time continuity of data being interrupted. Linear

interpolation is applied in the case of imputation. The linear interpolation equation is:

y=y0+(y1-y0)⋅ x-x0

x1-x0

(10)

Where x0, y0 and x1, y1 represent coordinates of two known points.

Kalman filtering is employed to eliminate anomalies caused by random noise deviations,

enhancing data reliability.

3.2 Evaluation Metrics and Model Parameters

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

MAPE, RMSE, MAE and R² are the examples of the indicators used to evaluate the

performance of the model. A higher accuracy model is reflected in R² value closer to 1.

These indicators help in comprehensive assessment of the model's effectiveness.

The calculation formulas for the four metrics are as follows:

MAE=1



i=1 yi-yi



| (11)

RMSE=





(

i=1 yi-y



i)2 (12)

MAPE=100%



yi-yi





i=1 (13)

R2=1-(

i=1 yi-yi

)2

(

i=1 yi-y)2 (14)

Where: yi represent the predicted coalbed methane production value, and yi

 is the

representation of the actual value.

Setting parameters is an important part of time series analysis. To create sample data, time

intervals must be segmented. A sliding window technique is utilized to convert time series

data into inputs for the machine learning model[7]. This approach includes adjusting a

fixed-size window to obtain repeated training samples. In order to capture autocorrelation

and trend information inside the time series, lagged aspects of the target variable are also

included. Adam, an adaptive optimisation technique, is used as the optimiser for model

training. The initial learning rate (Lr) is set at 0.002, the batch size (batch_size) is 32, the

sliding step size (input_size) is 5, and the lag feature step size is 1 through search and

debugging. To improve the modelling process, a two-layer LSTM with 128 hidden layer

units is used.

3.3 Analysis of Key Factors Affecting Coalbed Methane Production

Coalbed methane production is a complex process influenced by different factors, including

geological parameters, engineering measures parameters, and production process

parameters[15]. While technical measurements are the outcome of drilling and fracturing

operations, geological parameters are established by inherent features of the coal seam[16].

These variables have little effect on the forecast of production dynamics and are fixed for

individual wells. The success of coalbed methane generation is largely dependent on the

characteristics of the production process after drilling and fracturing. The production process

is strongly influenced by variables such bottom-hole pressure, dynamic fluid level, casing

pressure, stroke, and stroke rate. These metrics show dynamic trends over time and are

gathered as time-series data. To comprehend and maximise coalbed methane production, it is

essential to analyse these tendencies. To increase the effectiveness and efficiency of

industrial processes, researchers concentrate on finding patterns in these time-series data. In

conclusion, optimising coalbed methane production requires the interaction of geological,

technical, and production process characteristics. By examining, analyzing and

understanding these factors, operators can greatly improve the productivity of coalbed

methane wells.

Variance Inflation Factor (VIF) analysis was used in the study to address multi-collinearity.

Strong multicollinearity was indicated by a VIF larger than 10, which suggested eliminating

one feature. Moderate multicollinearity was indicated by VIF scores between 5 and 10,

Wenjie Li et al.

necessitating additional research. No multicollinearity issues were identified by VIF values

less than 5. Table 1 contains detailed results.

Table 1 Variance Inflation Factor (VIF) Analysis

Feature

VIF

Collinearity Level

Cumulative Water Production

999.00

Severe Multicollinearity

Fracturing fluid return rate

999.00

Severe multicollinearity

Daily water production

24.36

Severe multicollinearity

Wellhead Pressure

23.07

Severe multicollinearity

Actual liquid level

17.40

Severe multicollinearity

Torque %/Current A

5.07

Moderate multicollinearity

RPM/Stroke

5.02

Moderate multicollinearity

System Pressure

4.60

No multicollinearity

System Pressure

3.57

No multicollinearity

Pump Efficiency

2.82

No multicollinearity

We inserted strongly associated features using correlation clustering, retain only one

representative feature in a group. This improved our model's interpretability and reduced

problems of the multicollinearity. Because of this our analysis was more effective and

efficient and produced better insights without inflating AIGC rates.

In order to calculate the correlation coefficient matrix, The relation between each pair of

original features is assessed. We can identify these correlations by setting a correlation

threshold of 0.68, which denotes highly connected characteristics and groupings. In order to

create the final feature set, one feature is selected as a representative from each category. This

approach guarantees precision and effectiveness in choosing the most pertinent and varied

aspects.

Highly associated features were found in correlation analysis in table 2, and Figure 4 shows

how cluster analysis divided them into three different categories. These results highlight the

connections between variables and contribute to a better comprehension of the data patterns

found in the research.

Table 2 Pairs of Highly Correlated Features

Feature 1

Feature 2

Correlation

Coefficient

Bottom Hole Pressure

Torque %/Current A

-0.855279

Bottom Hole Pressure

Actual liquid level

0.961906

Bottom Hole Pressure

Cumulative Water

Production

-0.883171

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

Feature 1

Feature 2

Correlation

Coefficient

Bottom Hole pressure

Fracturing fluid recovery

rate

-0.885690

Casing pressure

System Pressure

0.749062

Torque %/Current A

Actual Liquid Level

-0.832141

Torque %/Current A

Cumulative Water

Production

0.760135

Torque %/Current A

Fracturing Fluid Return

Rate

0.762208

Actual liquid level

Cumulative Water

Production

-0.867059

Cumulative Water

Production

Fracturing fluid return rate

0.999957

Actual liquid level

Fracturing fluid return rate

-0.869312

Daily Water Production

Rotational Speed/Stroke

Rate

0.701321

Pump Efficiency

Cumulative Water

Production

0.683532

Fig. 3 Correlation-Based Clustering Results

By using a dual correlation verification approach, We examined the complex relationships

between daily water production, cumulative water production, casing pressure and daily

gas production. Pearson's correlation coefficient for linear relationships and Grey

correlation for nonlinear relationships were both used in the investigation. The findings

provided very clear insights into the complex relationships between the variables by

Wenjie Li et al.

revealing clear nonlinear correlation patterns.

Table 3 Nonlinear Correlation Matrix

Feature

Grey

Correlation

Degree

Pearson

Correlation

Coefficient

Nonlinearity

Strength

Relationship

Type

Well

Pressure

0.728

0.145

0.583

Moderate

Nonlinear

Relationship

Daily

Water

Yield

0.560

0.120

0.440

Weakly

nonlinear

relationship

Cumulati

ve water

productio

0.647

0.859

-0.212

Strong linear

relationship

Table 3 shows the moderate nonlinearity relationship for well pressure which implies not

to be much important. Daily water production also not suitable for retention due to its low

linear correlation and weak nonlinear association. On the other hand, future model

projections will use cumulative water production as an input characteristic because it

shows a significant linear relationship.

3.4 Model Performance Validation

This study sought to assess the effectiveness of a long short-term memory network (LSTM)

transfer learning fusion model in predicting the performance of coalbed methane wells in

the Fukucang area [3, 12]. Using data from three horizontal wells which are FSL-30,

FSL-32, and FSL-34 the study compared the capability of the conventional sample

prediction model and few sample prediction. Research aim to illustrate the benefits of the

LSTM transfer learning fusion model over conventional LSTM models by carrying out

experiments, evaluating goodness of fit using R² values, and analyzing a variety of error

indicators, including mean absolute error, root mean square error, and mean absolute

percentage error[17,18,19]. The results indicated reduced in a data dependence, increased

in prediction accuracy, and simplifying hyper parameters which highlighting the promise

of this model in predicting coalbed methane well performance.

3.4.1 Conventional Sample Prediction Results

Before moving the learnt parameters to the FSL-32 well and FSL-34 wells for additional

fine-tuning, the conventional method of sample-based prediction need a pre-training the

model on data from the FSL-30 well. This procedure enables the model to capture

temporal information from the original source well while adapting to the distinct data

patterns of the target wells. This approach enables the model to effectively and correctly

generate accurate prediction based on several reliable datasets.

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

(1) Source Domain Pre-training Results (FSL-30 Well)

Daily samples from the FSL-30 well dataset are subjected to a number of preprocessing

procedures, such as data smoothing, missing value imputation and outlier elimination. The

80% of samples dataset is divided into a training set and a 20% of samples test set at an 8

by 2 ratio in accordance with time-series data analysis standards. This division makes sure

that the model training process is not interfered by forward information leakage. The

FSL-30 well dataset is now ready for efficient analysis and model building.

A fully connected layer structure and a two-layer LSTM were used in the pre-training

stage. Adam optimizer with a batch size of 32, 0.002 initial learning rate, a sliding window

stride of 5 for the input size, an LSTM hidden layer size of 128, a lag feature stride of 1

and the mean squared error (MSE) as the loss function was chosen after hyper parameters

were adjusted using grid search.

The pre-training results are shown in Figure 4, which shows an R² value of 0.94 on the test

set and a Model Absolute Percentage Error (MAPE) of 3.4feature remarkable point is that

the residuals cluster close to zero and exhibit no discernible skew and also the projected

values closely match the actual values over time. This implies that the time relationship

between coalbed methane production and cumulative water production was successfully

captured by the pre-trained model. It can therefore be used as a reliable benchmark for

additional improvement in the intended field. As a result, this pre-trained model has the

potential to improve accuracy and efficacy in future applications.

Fig. 4 Pre-training Visualization Results

(2) Target Domain Fine-Tuning Results (FSL-32, FSL-34)

A hierarchical learning rate technique used to maximize the performance of the model

during the Fine-tuning process in specific domains. The model was able to maintain the

Wenjie Li et al.

temporal feature extraction skills learnt during pre-training while avoiding the core

parameters from being overloaded by the small data samples in the target domain by

employing a lower learning rate of 1×10-5 in the long-term and short-term memory layers.

The fully connected layer, on the other hand, was given a higher learning rate of 1×10-4 in

order to speed up change in parameters and allow the model to quickly adjust to the local

data distribution of the FSL-32 and FSL-34 wells. The most important thing is that, the

“Cumulative water production" feature from the source domain was directly used, while

the input features of the target domain remained unchanged. Additionally, a double-layer

learning rate adjustment made the model transfer procedure more simple while

maintaining the same structure of the LSTM network. Both well FSL-32 and FSL-34 well

saw remarkable performance gains with the improved model. For the FSL-32 well, R² rose

from 0.70 to 0.90 equivalent to 77% increase, while the test of set MAPE dropped from

10.40% to 6.1%. The model's predicted accuracy was further improved by the notable

decreases in the MAE and RMSE values. More Significant improvements were also noted

for the FSL-34 well test set. R² improved from 0.80 to 0.93 equivalent to 86% percent of

increased value, while the MAPE dropped from 0.614% to 0.288%. Additionally, the MAE

and RMSE values significantly dropped, confirming the improved model’s efficiency

precisely forecasting the FSL-34 production of the well.

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

Fig. 5 Comparison of Daily Gas Production Predictions and Actual Values from

Different Models for FSL-32 and FSL-34

Table 4 Performance Comparison of Different Models for Predicting Daily Gas

Production

Well

No.

Model

MAE

RMSE

MAPE

R²

FSL-32

Model

in this paper

167.05

250.25

6.1%

0.9

LSTM

280.73

432.17

10.40%

0.7

FSL-34

Model

in this paper

17.07

24.45

0.288%

0.9

LSTM

36.54

42.19

0.614%

0.8

The traditional LSTM prediction curve shown in Figure 5, shows the yield's overall trend

but is inaccurate when it comes to numerical fitting. However, the integrated model's

prediction had little inaccuracy and closely matched the actual numbers, especially during

periods of stable production. This highlights how well the hierarchical fine-tuning

approach adjusts to the particular domain data.

3.4.2 Small-Sample Prediction

Wenjie Li et al.

We reduced the FSL-34 dataset to about 100 data points while keeping the gas production

data from the first 100 days in order to assess the prediction performance of the suggested

model. We were able to evaluate the accuracy of the model using this lower sample

prediction dataset. We used the MMD domain alignment technique described in Section

2.4 to handle the possible decrease in distribution similarity brought on by data deletion.

The prediction result can be seen in Figure 6 below, demonstrating the effectiveness of the

model in analyzing gas production trends.

Fig. 6 Small-Sample Prediction Results

Table 6 Performance Comparison of Different Models for Predicting Daily Gas

Production

Model

MAE

RMSE

MAPE

R²

Model in this

paper

32.17

54.67

3.03%

0.9

LSTM

390.73

425.30

40.93%

-2.3

The actual production curve and the predicted curve from the LSTM model differ

significantly, as shown in the graph in Figure 6. However, there are very few differences

between the prediction curve of an alternate model and the actual production trend. By

showing a more realistic depiction of the true production curve, this alternative model

performs better than the LSTM model, providing a more accurate reflection of reality.

The suggested model has a reduced prediction error and a higher R² value when compared

to the LSTM model in Table 6. The suggested model significantly improves from -2.3 to

0.94 with the same expected days. In comparison to the LSTM model, the mean relative

error is lowered by about 37%, and the MAE and RMSE errors also show significant

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

reductions. These results indicate the wonderful performance of the proposed model in

obtaining accuracy.

Predictive accuracy has significantly improved when LSTM is mixed into transfer

learning models. The application of the MMD domain alignment technique helps to reduce

inter-domain variations, which facilitates the transfer learning process. This is particularly

helpful for datasets that exhibit notable distribution differences. Through fine-tuning the

target domain using this model, adjustments to the learning rates of the LSTM layer and

the fully connected layer are all that is required[9, 10, 11]. This successfully resolves the

problem of excessive parameter adjustment that is frequently linked to LSTM models. A

viable approach to maximizing prediction accuracy across many datasets is provided by

the LSTM-integrated transfer learning model.

4. Conclusion

1. The Pearson correlation coefficient and grey correlation analysis were used to obtain the

required foundation data for the model prediction (cumulative water production).

2. The difficulty of managing several hyper parameters in LSTM models is addressed by

varying learning rates across layers. Reducing domain differences, reducing the need for

target domain data, and enabling accurate predictions with small sample numbers are all

made possible by using source domain knowledge transfer and MMD domain alignment.

This method improves LSTM models' efficacy and efficiency across a range of prediction

tasks..

3. Significant gains in performance metrics show that the LSTM transfer learning fusion

model performs better than the basic LSTM model. For instance, the average relative error

of FSL-32 dropped from 10.4% to 6.1%, and the R² value increased from 0.7 to 0.9.

Similarly, the error rate of FSL-34 decreased from 0.614% to 0.288%, while the R² value

climbed from 0.8 to 0.93. Furthermore, the R² value for small sample prediction saw a

great increase from -2.3 to 0.94. Overall, the model achieved an impressive reduction of

approximately 37% in the average relative error.

Acknowledgements

Project Number: 2024B01017-3; Project Name: Evaluation of Deep Coalbed Methane

Resources in Key Basins of Xinjiang and Development of Favorable Area Selection and

Development Technology - Subproject Three: Research on the Recoverability Evaluation

and Engineering Technology System of Deep Coalbed Methane

References

[1] Li Yuan, Guo Dali, Kang Yunwei. Dynamic Prediction of Coalbed Methane Production

Integrating Attention Mechanisms [J]. Science, Technology and Engineering, 2023,

23(02): 550-557.

[2] Han Yong. Adaptability Analysis and Optimization Design of Fracturing Technology in

Coalbed Methane Fracturing and Production [J]. Geological Equipment, 2025, 26(04):

41-45.

Wenjie Li et al.

[3] Yu Yong, Dai Wujie, Hu Jun. Application of Intelligent Optimization Machine Learning

Methods Based on Cross-Validation in Nozzle Profile Optimization [J]. Journal of

Beijing Institute of Technology, 2025, 45(08).

[4] Shang Tingting, Fu Qiming, Cui Zhiming, et al. Research on Building Energy

Consumption Prediction Method Based on Sliding Window Transfer LSTM [J].

Journal of Suzhou University of Science and Technology (Natural Science Edition),

2025, 42(03): 58-69.

[5] Huang Qianguang, You Yuwen, Li Jintao. Fault Identification of Variable Air Volume

Air Conditioning Systems Based on Transfer Learning Using MCNN-BiLSTM [J].

Journal of Tianjin University of Urban Construction, 2025, 31(04): 285-294.

[6] Zhang Xianmin, Wang Xuebo, Deng Ze, et al. Review of Research Progress on

Enhanced Recovery Technologies for Coalbed Methane [J/OL]. Drilling and

Production Technology, 2025,(05):169-177.

[7] Yu Dechao, Han Qi, Zhou Junhua, et al. Research on Dynamic Capacity Expansion

Risk Prediction Method for High-Voltage Transmission Lines Based on Time-Series

Data Correlation Analysis [J]. Intelligent Buildings and Smart Cities, 2025, (09):

26-28.

[8] Kang Lifang, Wang Zhenguo, Zhai Jiayu, et al. Exploration Practice and Insights of

Deep Coalbed Methane in Shenfu Area, Eastern Margin of Ordos Basin [J].

Petrochemical Technology, 2025, 32(06): 285-287.

[9] Cang Qing, He Hao, Deng Fan, et al. Research on Changes in Hydraulic Structures

Along the Yangtze River Based on Data Analysis and Grey Correlation Degree [J].

Water Resources Science and Technology & Economy, 2025, 31(09): 10-15.

[10] Chen Zhe, Zhou Jinhui, Jin Donghui, et al. Load Forecasting Method for Integrated

Energy Systems Based on TimeVAE and Transfer Learning [J]. Zhejiang Electric

Power, 2025, 44(09): 58-69.

[11] Yu Hongchi, Cui Mimi. Progress in Multi-Source Sensor Transfer Learning for Heavy

Metal Tracing Monitoring in Rivers [J]. Experiment and Analysis, 2025, 3(03): 71-75.

[12] Zhao Mengxue, Shi Xiaohu, Bian Jing, et al. A Transfer Learning-Based Method for

Predicting Drilling Overflow[J]. Journal of Inner Mongolia University (Natural

Science Edition), 2025, 40(05):42-50.

[13] Zhang E, Liu Y, Lin G, et al. A transfer learning method of collaborating random walk

and adaptive instance normalization for inscription image denoising [J]. Engineering

Applications of Artificial Intelligence, 2025, 162(PB): 112458-112458.

[14] Klaib FA, Saif A, Alhosanie NT, et al. Automated assessment and detection of third

molar and inferior alveolar nerve relations using UNet and transfer learning models[J].

Scientific Reports, 2025, 15(1):34529-34529.

[15] Prabhas S K, Basem A, Lakshmi L, et al. A Deep learning framework for brain tumor

detection using CNNs and transfer learning on MRI scans[J]. Systems and Soft

Computing, 2025, 7200389-200389.

[16] Li Juhua, Chen Chen, Xiao Jialin, et al. Prediction of Production for Multi-stage

Fractured Shale Gas Wells Based on Random Forest Algorithm [J]. Journal of Yangtze

University (Natural Science Edition), 2020, 39(6): 168-174.

[17] Ji Lei, Li Juhua, Xiao Jialin, et al. Application of Random Forest Algorithm in the

Multi-stage Fracturing Well Reformation of Shale Gas Field [J]. Petroleum Geology &

Oilfield Development in Daqing, 2020, 17(4): 34-38

[18] Li Jiaojie, Qin Shunli, Wang Jie, et al. Application of Random Forest Algorithm in

Jimsuasal Shale Oil Reservoir [J]. Journal of Yangtze University (Natural Science

Edition), 2023, 20(2): 69-76

Coalbed Methane Production Forecasting Based on LSTM and Transfer Learning Integration

[19]Liang Chenggang, Li Jihua*, Chen Yiwai, et al. Evaluation of Reservoir Production

Capacity of Shale Oil Reservoirs Based on Naive Bayes Algorithm [J]. Journal of

Shenzhen University (Science & Technology), 2023, 40(1): 66-73