A method and device for predicting the productivity of a tight gas well based on artificial intelligence

Abstract

The application provides a dense gas well productivity prediction method and device based on artificial intelligence, and belongs to the technical field of oil and gas productivity intelligent prediction, and comprises the following steps: obtaining production data of a first historical time period of a target offshore dense gas well, preprocessing the production data of the first historical time period, and constructing a historical production data set; constructing a training model based on multi-feature extraction and time sequence capture, training the training model by using the historical production data set, and generating a fusion type gas well productivity prediction model; inputting target production data of the target offshore dense gas well into the fusion type gas well productivity prediction model, and outputting predicted productivity. The application can balance calculation efficiency and prediction accuracy, realize efficient productivity prediction with deep coupling of feature extraction and time sequence capture, and provide reliable technical support for rapid decision-making and economic optimization of offshore dense gas field development.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent prediction of oil and gas production capacity, and in particular to a method and apparatus for predicting the production capacity of tight gas wells based on artificial intelligence. Background Technology

[0002] Offshore tight gas fields with low permeability possess abundant reserves and promising development prospects. However, these fields face significant challenges due to their high development difficulty and low single-well productivity. For reservoir development to be feasible, the economic benefits of the reserves must exceed the development costs. Currently, long horizontal open-hole wells, with their large unloading area and low flow resistance, are widely used in offshore tight gas field development. Well production is a core indicator reflecting reservoir storage and permeability. Therefore, accurate production capacity assessment is crucial for evaluating development value and ensuring the economic feasibility of development plans. Production capacity forecasting is an important means of measuring the production capacity of typical wells, and formulating subsequent production enhancement measures based on forecast results is of great significance for achieving efficient development of offshore tight gas fields.

[0003] Commonly used production capacity prediction methods include decline curve analysis and numerical simulation. For example, CN110084431A and CN120719957A disclose methods for production capacity prediction using decline curve analysis and numerical simulation, respectively. However, decline curve analysis has limitations in accuracy and interpretability, making it difficult to accurately describe the production characteristics of offshore gas wells. While numerical simulation can achieve high-precision simulations by increasing grid complexity, it is complex to operate and computationally time-consuming. Furthermore, it suffers from the difficulty in obtaining actual effective data. These issues hinder current production capacity prediction for offshore tight gas fields.

[0004] Therefore, improving the accuracy of productivity evaluation of offshore tight gas wells and reducing computational load are urgent technical problems to be solved. Summary of the Invention

[0005] To address the limitations of existing declining curve analysis methods in terms of accuracy and the complexity and time-consuming computation of numerical simulation methods, this invention provides an artificial intelligence-based method and apparatus for predicting the production capacity of tight gas wells. This method achieves deep coupling prediction of offshore tight gas well production capacity through "multi-feature extraction" and "time-series dynamic capture," thereby improving the accuracy of gas well production capacity evaluation and reducing the computational load of production capacity prediction.

[0006] This invention provides a method for predicting the production capacity of tight gas wells based on artificial intelligence, comprising:

[0007] The production data of the target offshore tight gas well for the first historical period is obtained. The production data of the first historical period is preprocessed to construct a historical production dataset. The production data of the first historical period includes production date, well opening time, tubing pressure, and daily gas production. The preprocessing of the production data of the first historical period includes normalization processing of the production data of the first historical period. A training model based on multi-feature extraction and temporal capture is constructed, and the training model is trained using the historical production dataset to generate a fusion gas well production capacity prediction model. The training model based on multi-feature extraction and temporal capture includes a first CNN layer, a second CNN layer, an LSTM layer, and a fully connected layer. The first CNN layer and the second CNN layer both include convolutional layers and activation layers for extracting local and global multi-features from the historical production data. The LSTM layer is used for temporal feature capture, and the fully connected layer is used to map the temporal features extracted by the LSTM layer to specific daily gas production prediction values. The target production data of the target offshore tight gas well is input into the fusion gas well production capacity prediction model, and the predicted production capacity is output. The target production data includes the production date, well opening time, tubing pressure, and daily gas production of the second historical time period. The target production data is preprocessed and then input into the fusion gas well production capacity prediction model, and the daily gas production is output.

[0008] Optionally, the step of constructing a training model based on multi-feature extraction and time-series capture, and training the training model using the historical production dataset to generate a fusion gas well production capacity prediction model includes dividing the historical production dataset into a training set and a test set, training the training model based on multi-feature extraction and time-series capture using the training set, and evaluating the performance of the trained training model based on multi-feature extraction and time-series capture using the test set.

[0009] Optionally, a sliding window method is used to divide the training and test sets, with the window size set to [value missing]. The sliding step size is set to ， Before use Production data forecast for the day +1 day to + Daily gas production.

[0010] Optionally, in the hyperparameters of the first CNN layer and the second CNN layer, the kernel size is 3, the stride is 1, the padding is 1, and the activation function is the ReLU function.

[0011] Optionally, in the hyperparameters of the LSTM layer, the size of the hidden layer is 1, and the number of LSTM layers is 4.

[0012] Optionally, the torch deep learning module can be used to build a training model based on multi-feature extraction and temporal capture.

[0013] Optionally, the Adam optimizer can be used to optimize the model parameters, and the model's representational ability can be enhanced by dynamically adjusting the learning rate.

[0014] The present invention provides an artificial intelligence-based tight gas well production capacity prediction device, including a historical production dataset construction module, a fusion gas well production capacity prediction model generation module, and a production capacity prediction module; The historical production dataset construction module is used to construct a historical production dataset by preprocessing the production data of the target offshore tight gas well for the first historical period. The production data for the first historical period includes the production date, well opening time, tubing pressure, and daily gas production. The preprocessing of the production data for the first historical period includes normalization of the production data for the first historical period. The fusion-based gas well production capacity prediction model generation module is used to construct a training model based on multi-feature extraction and time-series capture, and to train the training model using the historical production dataset to generate a fusion-based gas well production capacity prediction model. The training model based on multi-feature extraction and time-series capture includes a first CNN layer, a second CNN layer, an LSTM layer, and a fully connected layer. The first CNN layer and the second CNN layer both include convolutional layers and activation layers, which are used to extract local and global multi-features from the historical production data. The LSTM layer is used for time-series feature capture, and the fully connected layer is used to map the time-series features extracted by the LSTM layer to specific daily gas production prediction values. The production capacity prediction module is used to input the target production data of the target offshore tight gas well into the fusion gas well production capacity prediction model and output the predicted production capacity. The target production data includes the production date, well opening time, tubing pressure, and daily gas production of the second historical time period. The target production data is preprocessed and then input into the fusion gas well production capacity prediction model to output the daily gas production.

[0015] This invention provides an artificial intelligence-based method and apparatus for predicting the production capacity of tight gas wells. First, it acquires production data for a first historical period from the target offshore tight gas well. After preprocessing the production data for this first historical period, a historical production dataset is constructed. The production data for this first historical period includes production date, well start-up time, tubing pressure, and daily gas production. The preprocessing of the production data for this first historical period includes normalization. Then, a training model based on multi-feature extraction and time-series capture is constructed, and the training model is trained using the historical production dataset to generate a fusion-based gas well production capacity prediction model. The training model based on multi-feature extraction and time-series capture includes a first C... The model consists of an NN layer, a second CNN layer, an LSTM layer, and a fully connected layer. Both the first and second CNN layers include convolutional and activation layers for extracting local and global multi-features from the historical production data. The LSTM layer is used for temporal feature capture, and the fully connected layer maps the temporal features extracted by the LSTM layer to specific daily gas production prediction values. Finally, the target production data of the target offshore tight gas well is input into the fusion-based gas well production capacity prediction model, which outputs the predicted production capacity. The target production data includes the production date, well opening time, tubing pressure, and daily gas production for a second historical time period. Preprocessing the target production data before inputting it into the fusion-based gas well production capacity prediction model outputs the daily gas production, achieving the following beneficial effects: 1. This invention constructs an integrated prediction model that combines multi-feature collaborative extraction and precise capture of time-series dependencies through a fusion architecture of CNN and LSTM. On the one hand, it leverages the powerful local feature extraction capabilities of CNN to deeply explore the nonlinear correlations between input features such as tubing pressure, well opening time, and daily gas production, fully releasing the information value of multi-source production data. On the other hand, it relies on the processing advantages of LSTM for long-term time-series data to accurately extract the dynamic patterns of offshore gas well production changes over time, thereby breaking through the technical limitations of single intelligent models and significantly improving the accuracy and reliability of production capacity prediction. 2. Through its optimized model architecture, this invention can significantly shorten the production capacity forecasting cycle while ensuring that the forecasting accuracy meets engineering requirements. This effectively meets the "quick and efficient" decision-making needs of offshore gas field development sites and avoids problems such as untimely optimization of development plans and waste of resources caused by forecast lag. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the production capacity prediction method for tight gas wells based on artificial intelligence, according to an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the artificial intelligence-based tight gas well production prediction method based on the training model of CNN-LSTM, according to an embodiment of the present invention.

[0018] Figure 3 This is a graph showing the historical production data changes of a certain offshore tight gas well.

[0019] Figure 4 This is a schematic diagram illustrating the data partitioning using the sliding window method in the tight gas well production prediction method based on artificial intelligence according to an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram of the convolution operation of the CNN layer in the tight gas well production prediction method based on artificial intelligence according to an embodiment of the present invention.

[0021] Figure 6 This is a graph showing the variation of model training error in the tight gas well production prediction method based on artificial intelligence according to an embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram comparing the predicted results with the actual values ​​of a certain offshore tight gas well.

[0023] Figure 8 This is a scatter plot of the predicted and actual values ​​for a certain offshore tight gas well.

[0024] Figure 9 This is a schematic diagram of the structure of an artificial intelligence-based tight gas well production prediction device according to an embodiment of the present invention.

[0025] Figure labeling: 10, Historical production dataset construction module; 20, Integrated gas well production capacity prediction model generation module; 30, Production capacity prediction module. Detailed Implementation

[0026] The following describes specific embodiments and appendices. Figure 1 -Appendix Figure 9 The invention is described in detail so that those skilled in the art can more fully understand its purpose, features and effects.

[0027] Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In the event of any discrepancy between the definitions of terms in this invention and their commonly understood meaning by one of ordinary skill in the art to which this invention pertains, the definitions provided in this invention shall prevail.

[0028] Currently, the prediction of production capacity of offshore tight gas horizontal wells faces three major challenges: first, it is difficult to achieve a balance between speed and accuracy in prediction; second, the prediction methods are complex and costly, making it difficult to meet the demand for cost reduction.

[0029] The rapid development of artificial intelligence technology has provided a new technical path for solving complex problems in the oil and gas field, showing great application potential, especially in the areas of production capacity prediction and parameter optimization. This invention provides an artificial intelligence-based method and apparatus for predicting the production capacity of tight gas wells, constructing a prediction system that requires no complex manual intervention and has high computational efficiency. This achieves efficient and accurate production capacity prediction, which is of great significance for cost reduction, efficiency improvement, and efficient development of offshore tight gas fields.

[0030] Example 1 As a specific embodiment of the present invention, this embodiment provides a method for predicting the production capacity of tight gas wells based on artificial intelligence, referring to... Figure 1 , Figure 2 The specific steps are as follows: S100: Obtain the production data of the target offshore tight gas well for the first historical period, and construct a historical production dataset after preprocessing the production data of the first historical period. S200. Construct a training model based on multi-feature extraction and time-series capture, and train the training model using the historical production dataset to generate a fusion gas well production capacity prediction model, wherein the training model based on multi-feature extraction and time-series capture is constructed using a CNN-LSTM architecture. S300. Input the target production data of the target offshore tight gas well into the fusion gas well production capacity prediction model and output the predicted production capacity.

[0031] The artificial intelligence-based tight gas well production prediction method provided in this embodiment combines the multi-feature extraction advantages of CNN (Convolutional Neural Network) with the temporal dependency capture capability of LSTM (Long Short-Term Memory) to construct a fusion-type temporal prediction model. This becomes an important solution to break through the technical bottleneck of production prediction for offshore tight gas horizontal wells, and provides technical support for achieving higher accuracy and higher efficiency in gas well production prediction.

[0032] As a core variant of RNNs (Recurrent Neural Networks), LSTM possesses the unique ability to capture long-term dependencies and nonlinear characteristics in time-series data, exhibiting significant advantages such as high prediction accuracy, superior computational efficiency, and strong adaptability to dynamic operating conditions. However, the production capacity of offshore tight gas wells is the result of the coupling effects of multiple factors. Its changes are dynamically influenced by multi-dimensional well control parameters such as tubing size, wellhead pressure, well operation status, and reservoir pressure, with complex nonlinear relationships existing between these parameters. While LSTM models perform excellently in processing single-channel time-series data, they have significant limitations in uncovering the interaction relationships between different feature dimensions and extracting multi-parameter coupling features. This makes it difficult to fully utilize the information value of multi-source production data, and there is still room for improvement in prediction accuracy.

[0033] With its powerful local feature extraction capabilities, CNN can effectively capture the spatial correlation features between multi-dimensional data, which can make up for the shortcomings of LSTM model in mining multi-feature interaction relationships. Combining CNN and LSTM to build a CNN-LSTM-based fusion gas well production capacity prediction model can make offshore gas well production capacity prediction more accurate.

[0034] The artificial intelligence-based tight gas well production prediction method in this embodiment solves the problems of low accuracy of traditional decline curve analysis method and low efficiency of numerical simulation method. While ensuring that the prediction accuracy meets the engineering requirements, it significantly shortens the prediction cycle and provides timely support for rapid on-site decision-making.

[0035] Furthermore, in S100 of this embodiment, the target offshore tight gas well refers to a gas well that is capable of normal gas production and whose production capacity is to be predicted, and the first historical time period refers to a period of normal gas production in the past, the length of which is set according to the production capacity prediction demand.

[0036] Specifically, the production data acquired for the first historical time period of the target offshore tight gas well includes the production date, well start-up time, tubing pressure, and daily gas production. Well start-up time refers to the daily operating time of the gas well, such as 5 hours, 10 hours, or 24 hours. The production date can be expressed in days, such as day 1, day 10, or day 100. The production date, well start-up time, tubing pressure, and daily gas production are all dynamic parameters and are the core parameters of gas well production. Well start-up time, tubing pressure, and daily gas production are all time-series data, and the relationship between the three is non-linear.

[0037] In this embodiment, the production data focuses on four core dynamic parameters: production date, well opening time, tubing pressure, and daily gas production. This accurately covers the key influencing factors of offshore tight gas well productivity, avoids redundant data interference, and provides a high-quality data foundation for subsequent model training.

[0038] In one example, production data for a certain offshore tight gas well, including well start time, tubing pressure, and daily gas production, are obtained over the past two years. Figure 3 As shown, production dynamics exhibit significant nonlinear characteristics, making prediction difficult using traditional forecasting methods.

[0039] Further, in S100, the preprocessing of the production data for the first historical time period includes normalizing the production data for the first historical time period. Preferably, in this embodiment, the Min-Max Normalization method is used to normalize the production data for the first historical time period, linearly mapping the production data for the first historical time period to [0, 1].

[0040] The minimum-maximum normalization formula is as follows:

[0041] In the formula, For the normalized value of the parameter, The original value of the parameter. This is the minimum value of the original parameter. This is the maximum value of the original value of the parameter.

[0042] The production data of the first historical time period is preprocessed by normalization to construct a historical production dataset, which is used for model training in S200.

[0043] By employing min-max normalization to map the data to the [0,1] interval, model bias caused by differences in parameter scales is effectively eliminated, improving model generalization ability and ensuring that multiple parameters can collaboratively participate in feature extraction. A structured historical production dataset is formed through a standardized preprocessing workflow, ensuring data consistency and usability while providing standardized data format support for model training in S200, reducing data adaptation costs for model training.

[0044] Furthermore, the construction of a training model based on multi-feature extraction and time-series capture, and the training of the training model using the historical production dataset to generate a fusion gas well production capacity prediction model, includes dividing the historical production dataset into a training set and a test set, training the training model based on multi-feature extraction and time-series capture using the training set, and evaluating the performance of the trained training model based on multi-feature extraction and time-series capture using the test set.

[0045] Optionally, the training set and test set can be split in a 7:3 ratio, meaning that 70% of the data in the historical production dataset is used for model training, and 30% is used for model performance evaluation and does not participate in the training process. In other feasible embodiments, the ratio of the training set and test set can also be 6:4 or 8:2.

[0046] Preferred, such as Figure 4 As shown, a sliding window method is used to partition the training and test sets. The window size is fixed, but the window size and stride can be set as needed. The window size is set to... The sliding step size is set to For example, in this embodiment, the window size is set to 6 and the sliding step is set to 1, that is, the daily gas production of the next day (the 7th day) is predicted using the data of the previous 6 days.

[0047] Specifically, in this embodiment, the formula for predicting the daily gas production on the 7th day is as follows:

[0048] in, This is the predicted daily gas production for day 7. For the parameters of CNN-LSTM, These figures represent the well opening time, tubing pressure, and daily gas production for the first six days.

[0049] In other feasible embodiments, the daily gas production on the 31st day can be predicted using data from the previous 30 days, or the sliding step size can be set to 2 to predict the daily gas production on the 31st day. +1 and the +2 days of daily gas production.

[0050] The sliding window method supports custom window size and step size, which can be adapted to different forecasting cycle requirements, maximize the use of historical production data, and improve the model's ability to capture the trend of capacity changes.

[0051] Furthermore, combined Figure 2 As shown in Table 1, the training model based on multi-feature extraction and temporal capture, constructed using a CNN-LSTM architecture in this embodiment, includes two CNN layers, an LSTM layer, and one fully connected layer. Each CNN layer includes a convolutional layer and an activation layer. The model training process closely matches the production characteristics of offshore tight gas wells, which involve multi-parameter coupling and large temporal fluctuations. The core logic is feature extraction, temporal modeling, and result output, enabling effective model training and ensuring the usability of the resulting fusion-based gas well production capacity prediction model.

[0052] Table 1. Parameters of the training model based on CNN-LSTM

[0053] The input and output dimensions of a CNN layer can be represented as (B, L, C), and the input and output dimensions of an LSTM layer can be represented as (L, B, C), where B represents the batch number, L represents the sequence length, and C represents the number of sequence channels. In this embodiment, the production date, well start time, tubing pressure, and daily gas production of the first 6 days are used as input features, and the daily gas production of the 7th day is used as the output label. Therefore, the batch number of the first CNN layer is 1, which is suitable for the scenario of offshore single-well production capacity prediction. Each batch focuses on the continuous production data of a single well, avoiding cross-interference of data from multiple wells. The sequence length is 6, which is equal to the sliding window size. The number of input channels is 4, corresponding to the four core influencing parameters of offshore tight gas fields: production date, well start time, tubing pressure, and daily gas production. The input dimension is (1, 6, 4), and after processing by convolutional and activation layers, the output dimension is (1, 6, 16). After processing by the first CNN layer, the number of channels increases from 4 to 16, meaning that each time step (day) corresponds to 16 local coupled features, realizing the first transformation from low-dimensional original parameters to high-dimensional local features, and providing a foundation for subsequent deep feature extraction.

[0054] In the first CNN layer, the hyperparameters include a kernel size of 3, a stride of 1, padding of 1, and the ReLU activation function. The convolution operation of the CNN layer in this embodiment is as follows: Figure 5 As shown, the convolutional kernels perform multiplication along the sequence direction and then sum them. Convolutional layers with local weight sharing can achieve preliminary extraction of sequence features, while the ReLU activation function introduces non-linear expressive power, thereby enhancing the model's representational ability.

[0055]

[0056] The hyperparameter settings of the first CNN layer ensure that the sequence length remains unchanged after convolution, still 6 days. The convolution kernel slides along the time series dimension (L=6), covering the four parameter data for 3 days each time, and calculates local feature values ​​by multiplying and summing element-wise. For example, convolution operations are performed on the "well start time (e.g., 20h / 22h / 24h) and tubing pressure (e.g., 18MPa / 16MPa / 15MPa)" data for days 1-3 to capture the local correlation feature of prolonged well start time accompanied by decreased tubing pressure. This feature is a key signal of energy release and production capacity changes in offshore tight gas wells.

[0057] By activating the convolution results using the ReLU function, invalid negative features are filtered out, such as the abnormal local relationship where well opening time is shortened but tubing pressure rises, while retaining the nonlinear information of valid features, thus enhancing the model's adaptability to complex production conditions of offshore gas wells.

[0058] The output of the first CNN layer serves as the input to the second CNN layer. After processing by convolutional and activation layers, the output size is (1, 6, 32). After processing by the second CNN layer, the number of channels changes from 16 to 32.

[0059] In the second CNN layer, the hyperparameters are: kernel size 3, stride 1, padding 1, and ReLU activation function.

[0060] The convolution kernel slides along the time series dimension, convolving the (1,6,16) feature matrix output by the first CNN layer. For example, it operates on 16 local features from days 2 to 4 to capture the global correlation between well start time, tubing pressure, and daily gas production over three days. For instance, when the well start time is stable at 24 hours and the tubing pressure is maintained at 18 MPa, the daily gas production remains at 100×10⁻⁶. 4 m 3 The characteristic pattern corresponds to the stable production state of offshore tight gas wells.

[0061] The ReLU function is activated again to enhance the effective global features; the final output feature matrix size is (1, 6, 32), and the number of feature channels is increased from 16 to 32, realizing the second transformation from local features to global correlation features, and comprehensively covering the complex interaction relationship of offshore gas well production parameters, such as the chain reaction feature of excessively long well opening time leading to fluid accumulation, which in turn causes a sudden drop in tubing pressure and a decline in daily gas production.

[0062] Through processing by the first and second CNN layers, the interaction relationships between multiple features from different channels are extracted.

[0063] After processing by the second CNN layer, the feature sequence is input into the LSTM layer for temporal information modeling. The LSTM accurately captures temporal dependencies through three gating mechanisms: forget gate, input gate, and output gate. Based on the high-dimensional features extracted by the CNN, the LSTM layer models the long-term dependence of production capacity on time, which is suitable for the characteristics of large temporal fluctuations in the production capacity of offshore tight gas wells and the significant influence of historical production status.

[0064] The output of the second CNN layer is used as the input of the LSTM layer. The (B,L,C) format (1,6,32) output of the second CNN layer is converted into the (L,B,C) format (6,1,32) adapted by the LSTM layer. That is, the sequence length is 6, the batch number is 1, and the number of channels is 32, which conforms to the logic of LSTM processing step by step in time sequence.

[0065] After processing, the output size is (6,1,64), which means the sequence length is 6, the batch number is 1, and the number of channels is 64.

[0066] Among the hyperparameters of the LSTM layer, the hidden size is 1, and the number of LSTM layers is 4.

[0067] The forget gate calculates weights (between 0 and 1) using the sigmoid function to determine whether to retain historical features. For example, for low production characteristics caused by short well opening time on the first day (e.g., 2 hours), the forget gate assigns a low weight (e.g., 0.2) to reduce its interference with subsequent production predictions for days 4-6, thus adapting to the actual pattern that the current production capacity of offshore gas wells is mainly affected by the production status of the past 3-5 days. The input gate filters valid features for the current time step, such as the features of tubing pressure of 18MPa on day 3 and well opening time of 24h, and updates the cell status. The output gate combines the cell state with the current features to output the hidden state at the current time step. Finally, after processing with 4 layers of LSTM, the output temporal feature matrix has a size of (6,1,64), realizing the transformation from global correlation features to temporal dynamic features, capturing long-term temporal patterns such as the daily gas production starting to decline steadily after the oil pipeline pressure has been declining for 3 days.

[0068] The output of the LSTM layer is passed to the fully connected layer, and finally the prediction result is output through the fully connected layer.

[0069] The fully connected layer maps the time-series features extracted by the LSTM layer to specific daily gas production prediction values, completing the final transformation from features to results. The processing procedure is as follows.

[0070] Feature flattening: The (6,1,64) three-dimensional time-series feature matrix output by the LSTM layer is flattened into a one-dimensional vector (length = 6×64=384), eliminating the time-series dimension and integrating all time-series dynamic features, such as the pressure change trend within 6 days and the impact of well opening time adjustment.

[0071] Linear Mapping and Predicted Output: The weight matrix of the fully connected layer linearly maps the 384-dimensional feature vector to a 1-dimensional output value, i.e., the first dimensional feature vector. The predicted daily gas production for +1 day (e.g., day 7) is output in a size of (1,1). During this process, the fully connected layer learns the mapping relationship between the time-series characteristics and production capacity of offshore gas wells, transforming abstract features into specific engineering prediction values, such as 55×10⁻⁶. 4 m 3 .

[0072] The entire data processing process is closely aligned with the offshore tight gas well production prediction scenario: by using two CNN layers to progressively extract the local and global correlations between well opening time, tubing pressure, and daily gas production, the problem of traditional models being unable to quantify multi-parameter nonlinear relationships is solved; the 4-layer LSTM gating mechanism accurately captures the long-term temporal dependence of production capacity, adapting to the characteristics of intermittent production and large fluctuations in production capacity of offshore gas wells; from data normalization to the final prediction output, every step considers the engineering attributes of offshore gas well production data.

[0073] The CNN-LSTM fusion architecture achieves dual core capabilities of multi-feature extraction and temporal sequence capture. CNN mines the nonlinear correlation between parameters, while LSTM captures the dynamic laws of production capacity over time, thus collaboratively solving the problem of predicting the production capacity of offshore tight gas wells by combining multiple factors and temporal fluctuations.

[0074] Preferably, the torch deep learning module is used to construct a training model based on multi-feature extraction and temporal capture; the Adam optimizer is used to optimize the model parameters, and the learning rate is dynamically adjusted to enhance the model's representation ability.

[0075] In one example, the following is a partial code snippet showing how to build a training model based on multi-feature extraction and temporal capture using the torch deep learning module: import torch import torch.nn as nn class CNN_LSTM_Predictor(nn.Module): def __init__(self): super(CNN_LSTM_Predictor, self).__init__() # First CNN layer: Extracting local multi-parameter coupled features (4 input channels → 16 output channels) self.cnn1 = nn.Conv1d(in_channels=4, out_channels=16, kernel_size=3, stride=1, padding=1) self.relu = nn.ReLU() # Activation function enhances nonlinearity, adapting to complex correlations of gas well parameters # Second CNN layer: Deepen global feature fusion (16 channels → 32 channels) self.cnn2 = nn.Conv1d(in_channels=16, out_channels=32, kernel_size=3, stride=1, padding=1) # LSTM layer: Capturing temporal dynamics (input 32 feature dimensions, 4 stacked layers to enhance long dependency capture) self.lstm = nn.LSTM(input_size=32, hidden_size=1, num_layers=4,batch_first=False) # Fully connected layer: Mapped to predicted daily gas production (flattened features → 1D output) self.fc = nn.Linear(6 1, 1)# 6 represents the sequence length, and 1 represents the size of the LSTM hidden layer. def forward(self, x): # Input x dimension: (B, L, C) = (1, 6, 4), adapted for single-well single-batch time series data x = x.permute(0, 2, 1) # Convert to CNN-adapted format: (B, C, L) = (1, 4, 6) # First CNN layer processing: Convolution → Activation x = self.relu(self.cnn1(x)) # Output: (1, 16, 6) # Second CNN layer processing: Convolution → Activation x = self.relu(self.cnn2(x)) # Output: (1, 32, 6) # Convert to LSTM adapted format: (L, B, C) = (6, 1, 32) x = x.permute(2, 0, 1) # LSTM layer processing: Output (seq_len, batch, num_directions) hidden_size) = (6,1, 1) lstm_out, _ = self.lstm(x) # Flattened sequence characteristics: (1, 6) 1) = (1, 6) x = lstm_out.view(1, -1) # Fully connected layer output prediction: (1, 1) (corresponding to daily gas production for 1 day) output = self.fc(x) return output The Torch deep learning module can dynamically compute graphs to adapt to data characteristics. Offshore tight gas well production data may have dynamic adjustment requirements such as missing data and outlier correction. Torch's dynamic graphs can adapt to data format changes in real time without recompiling the network, making it more flexible than TensorFlow static graphs. Layer parameters are precisely controllable, with hyperparameters directly defined via nn.Conv1d (1D convolution to adapt to time-series data) and nn.LSTM, such as a 4-layer LSTM with a hidden layer size of 1, adapting to the core requirements of multi-feature extraction and time-series capture. GPU acceleration improves training efficiency. Offshore gas well historical production data is large (e.g., 4 parameters per day over two years). Torch supports CUDA acceleration, which can shorten the model training cycle by more than 50%, adapting to the "quick and efficient" on-site decision-making needs. It has strong modularity and reusability; model classes can be directly transferred to multiple offshore tight gas wells, requiring only adjustments to the input data without reconstructing the network structure, reducing the cost of large-scale applications.

[0076] The Adam optimizer combines momentum gradient descent with an adaptive learning rate, which is adapted to multiple parameter scales. Even after normalization, well opening time and tubing pressure still exhibit implicit scale differences. Adam calculates the first moment (momentum) and second moment (gradient variance) of each parameter and assigns adaptive learning rates to different parameters. For example, a smaller learning rate is assigned to the weights related to tubing pressure, while a larger learning rate is assigned to the weights related to well opening time, avoiding slow convergence of some parameters due to a single learning rate. The momentum term accelerates convergence. The model contains a hybrid architecture of CNN and LSTM with a large number of parameters (the weights of the CNN layer alone reach 3×4×16+3×16×32=1728). Adam's momentum term (default β1=0.9) can accumulate historical gradient directions, accelerating the convergence of cross-layer parameters (such as the connection weights between CNN features and LSTM temporal features) and reducing training oscillations. To prevent gradient vanishing: Although the LSTM layer alleviates gradient vanishing through a gating mechanism, there is still a risk of gradient decay with 4 stacked layers. Adam corrects the gradient through the second moment to avoid the ineffective training of deep parameters (such as the weights of the 4th layer LSTM) due to excessively small gradients.

[0077] In one example, the following is a portion of the code that uses the Adam optimizer to optimize model parameters: # Initialize the model, loss function (MSE adapted to regression task), and Adam optimizer model = CNN_LSTM_Predictor() criterion = nn.MSELoss() # Mean squared error adaptation for continuous daily gas production prediction optimizer = torch.optim.Adam(model.parameters(), lr=0.001) # Initial learning rate 0.001 # Training Loop for epoch in range(500): # This application example trains for 500 epochs. optimizer.zero_grad() # Clears gradients to zero to prevent accumulated bias. # Input data: (1, 6, 4) (Production data for the first 6 days) inputs = torch.tensor(history_data, dtype=torch.float32) # Tag: (1, 1) (Actual daily gas production on day 7) labels = torch.tensor(target_gas, dtype=torch.float32) # Forward propagation prediction outputs = model(inputs) # Calculate the loss (MSE of predicted value vs. actual value) loss = criterion(outputs, labels) # Backpropagation to find gradient loss.backward() # Update Adam optimizer parameters optimizer.step() As a supplement to the Adam optimizer, the learning rate is adjusted in stages to solve the problem of convergence stagnation in the later stages of model training.

[0078] Using Torch's built-in ReduceLROnPlateau scheduler, the learning rate is dynamically adjusted based on the test set loss. Example code is below: # Dynamic Learning Rate Scheduler: If the loss does not decrease after 5 rounds, the learning rate is halved. scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,mode='min', patience=5, factor=0.5) # Add a scheduler to the training loop for epoch in range(500): # (Forward propagation, backward propagation, and parameter update steps are the same as above) # Calculate the validation loss using the test set after each training round. val_loss = calculate_val_loss(model, val_data, val_labels) # Adjust the learning rate based on validation loss scheduler.step(val_loss) By dynamically adjusting the learning rate, local optima can be avoided. In the training of offshore gas well production prediction models, it is easy to get stuck in local optima due to an excessively high learning rate (such as the predicted value consistently deviating from the actual value by 5%). Dynamically lowering the learning rate can help the model escape local optima in the later stages of training.

[0079] Furthermore, by dynamically adjusting the learning rate, training speed and accuracy can be balanced. In the early stages of training, a larger learning rate (0.001) is used to quickly approximate the optimal solution, while in the later stages, a smaller learning rate (such as 0.0001) is used to finely adjust the parameters, enabling the model to converge within 500 rounds, which is 30% more efficient than training with a fixed learning rate. The dynamic learning rate also enhances the model's generalization ability, reducing the model's overfitting to training data noise (such as abnormal production data caused by reservoir heterogeneity), allowing the model to maintain high accuracy in new well prediction.

[0080] In one example, the training loss changes as follows: Figure 6 As shown, after 500 training epochs, the error steadily decreased, and the training error was very close to the test error, indicating extremely low overfitting and good model training performance. Both the training and test sets exhibited small prediction errors, with mean squared errors (MSE) as low as 0.00054 and 0.00067, respectively.

[0081] Reference Figure 7 This is a comparison between the predicted and actual production capacity of a certain offshore tight gas well. It can be seen that the predicted daily and cumulative production curves match the actual values ​​very well, indicating that the model has strong generalization ability. Furthermore, referring to... Figure 8 The scatter plot of predicted and observed values ​​shows that the coefficients of determination (R²) for both the training and test sets reach a high level of 0.89. Although some biases appear in the training set (mainly due to reservoir heterogeneity and noise in the yield data), most points fall near the y=x line, confirming the effectiveness of the CNN-LSTM framework in this embodiment.

[0082] Finally, after training, a fusion-type gas well production capacity prediction model is obtained to predict the future production capacity of offshore tight gas wells. For example, the daily gas production of the next day is predicted based on the historical production data of the past 6 days. The production date, well opening time, tubing pressure and daily gas production of the past 6 days are preprocessed and input into the fusion-type gas well production capacity prediction model, and the predicted value of the daily gas production of the next day is output.

[0083] The target production data and historical production data use the same parameter types and preprocessing standards to ensure that the data format input to the model is consistent, reduce data adaptation errors, and improve the reliability of prediction results.

[0084] Existing AI prediction methods, such as single LSTM models, can only process single-channel time-series data. For example, they can only extract relevant features before and after production data, and are insufficient in capturing the interrelationships between different data features, failing to effectively mine multi-parameter interaction relationships. Gas well production is affected by multiple factors, and processing only the single-channel time-series data of daily gas production can easily lead to large prediction biases. This embodiment of the AI-based tight gas well production prediction method incorporates well opening time and tubing pressure in addition to the daily gas production to be predicted, forming a multi-channel production prediction. A fusion gas well production prediction model is established using a CNN-LSTM architecture. Through feature extraction and time-series representation, offshore tight gas well production prediction is performed to extract local nonlinear features and time dependencies, thereby achieving efficient and accurate production prediction. This solves the problems of low accuracy and poor efficiency of traditional prediction methods and the insufficient multi-feature processing capabilities of single intelligent models. The fusion gas well production prediction model directly outputs the daily gas production prediction results, which can be directly used to evaluate the development value of gas wells and formulate production enhancement measures, providing core technical support for cost reduction, efficiency improvement, and development of offshore tight gas fields.

[0085] Example 2 As another specific embodiment of the present invention, this embodiment provides a tight gas well production capacity prediction device based on artificial intelligence, referring to... Figure 9 It includes a historical production dataset construction module 10, a fusion-type gas well production capacity prediction model generation module 20, and a production capacity prediction module 30. in, The historical production dataset construction module 10 is used to construct a historical production dataset after preprocessing the production data of the first historical period of the target offshore tight gas well. The fusion-type gas well production capacity prediction model generation module 20 is used to construct a training model based on multi-feature extraction and time-series capture, and to train the training model using the historical production dataset to generate the fusion-type gas well production capacity prediction model. The production capacity prediction module 30 is used to input the target production data of the target offshore tight gas well into the fusion gas well production capacity prediction model and output the predicted production capacity.

[0086] The AI-based offshore tight gas well production prediction device provided in this embodiment combines the advantages of CNN's multi-feature extraction with LSTM's temporal dependency capture capability to construct a fusion-type temporal prediction model. This becomes an important solution to break through the technical bottleneck of offshore tight gas horizontal well production prediction, and provides technical support for achieving higher accuracy and higher efficiency in gas well production prediction.

[0087] Existing conventional production capacity forecasting methods often fail to balance computational speed and accuracy, making it difficult to meet the "quick and efficient" requirements of field operations. The production capacity forecasting method and apparatus for tight gas wells based on artificial intelligence provided by this invention, through an optimized model architecture, can significantly shorten the production capacity forecasting cycle while ensuring forecast accuracy meets engineering requirements. This effectively meets the "quick and efficient" decision-making needs of offshore gas field development, avoiding problems such as untimely optimization of development plans and resource waste caused by forecast lag.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting the production capacity of tight gas wells based on artificial intelligence, characterized in that, The method includes: The production data of the target offshore tight gas well for the first historical period is obtained. The production data of the first historical period is preprocessed to construct a historical production dataset. The production data of the first historical period includes production date, well opening time, tubing pressure, and daily gas production. The preprocessing of the production data of the first historical period includes normalization processing of the production data of the first historical period. A training model based on multi-feature extraction and temporal capture is constructed, and the training model is trained using the historical production dataset to generate a fusion gas well production capacity prediction model. The training model based on multi-feature extraction and temporal capture includes a first CNN layer, a second CNN layer, an LSTM layer, and a fully connected layer. The first CNN layer and the second CNN layer both include convolutional layers and activation layers for extracting local and global multi-features from the historical production data. The LSTM layer is used for temporal feature capture, and the fully connected layer is used to map the temporal features extracted by the LSTM layer to specific daily gas production prediction values. The target production data of the target offshore tight gas well is input into the fusion gas well production capacity prediction model, and the predicted production capacity is output. The target production data includes the production date, well opening time, tubing pressure, and daily gas production of the second historical time period. The target production data is preprocessed and then input into the fusion gas well production capacity prediction model, and the daily gas production is output.

2. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 1, characterized in that, The process of constructing a training model based on multi-feature extraction and time-series capture, and training the training model using the historical production dataset to generate a fusion-type gas well production capacity prediction model, includes dividing the historical production dataset into a training set and a test set, training the training model based on multi-feature extraction and time-series capture using the training set, and evaluating the performance of the trained training model based on multi-feature extraction and time-series capture using the test set.

3. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 2, characterized in that, The training and test sets are divided using a sliding window method, with the window size set to [value missing]. The sliding step size is set to ， Before use Production data forecast for the day +1 day to + Daily gas production.

4. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 1, characterized in that, The production data for the first historical time period is normalized using minimum-maximum normalization.

5. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to any one of claims 1-4, characterized in that, In the hyperparameters of the first and second CNN layers, the kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU.

6. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 5, characterized in that, In the hyperparameters of the LSTM layer, the size of the hidden layer is 1, and the number of LSTM layers is 4.

7. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 6, characterized in that, The Torch deep learning module was used to build a training model based on multi-feature extraction and temporal capture.

8. The method for predicting the production capacity of tight gas wells based on artificial intelligence according to claim 6, characterized in that, The Adam optimizer is used to optimize the model parameters, and the model's representation ability is enhanced by dynamically adjusting the learning rate.

9. A production capacity prediction device for tight gas wells based on artificial intelligence, characterized in that, The device includes a historical production dataset construction module (10), a fusion gas well production capacity prediction model generation module (20), and a production capacity prediction module (30). The historical production dataset construction module (10) is used to construct a historical production dataset after preprocessing the production data of the target offshore tight gas well in the first historical period. The production data of the first historical period includes production date, well opening time, tubing pressure, and daily gas production. The preprocessing of the production data of the first historical period includes normalization of the production data of the first historical period. The fusion-type gas well production capacity prediction model generation module (20) is used to construct a training model based on multi-feature extraction and time-series capture, and to train the training model using the historical production dataset to generate a fusion-type gas well production capacity prediction model. The training model based on multi-feature extraction and time-series capture includes a first CNN layer, a second CNN layer, an LSTM layer, and a fully connected layer. The first CNN layer and the second CNN layer both include convolutional layers and activation layers, which are used to extract local and global multi-features from the historical production data. The LSTM layer is used for time-series feature capture, and the fully connected layer is used to map the time-series features extracted by the LSTM layer to specific daily gas production prediction values. The production capacity prediction module (30) is used to input the target production data of the target offshore tight gas well into the fusion gas well production capacity prediction model and output the predicted production capacity. The target production data includes the production date, well opening time, tubing pressure, and daily gas production of the second historical time period. The target production data is preprocessed and then input into the fusion gas well production capacity prediction model to output the daily gas production.

Images