Non-diagenetic system hydrate slurry formation volume fraction prediction method based on LSTM (Long Short Term Memory)
Through the improved LSTM model, learning the generation rules of non-diagenetic hydrate slurry is solved, and the problem that the prior art is difficult to accurately predict the generation volume and behavior of hydrate slurry is achieved, achieving more accurate prediction and higher production efficiency.
Patent Information
- Application Number
- CN202411797577.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-03-25
AI Technical Summary
The prior art is difficult to accurately predict the generation volume and behavior of non-diagenetic hydrate slurries in complex environments, resulting in problems such as pipeline blockage, equipment wear and energy loss.
Using the improved LSTM model, the generation volume fraction of hydrate slurry is monitored and predicted in real time by learning the slurry hydrate generation rules. The model consists of a first layer LSTM network composed of 64 memory units, a Dropout layer, a second layer LSTM network composed of 32 memory units, and a full connection layer cascade, which can more effectively process multidimensional dynamic data.
The improved LSTM model can more accurately adapt to hydrate slurry behavior prediction in complex environments, improve prediction accuracy and generalization capabilities, help improve production system efficiency and reduce energy losses.
Smart Images
Figure CN119694451A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of non-diagenetic system hydrate slurry, and in particular to a method for predicting the generated volume fraction of non-diagenetic system hydrate slurry based on LSTM. Background Art
[0002] In recent years, non-diagenetic hydrate deposits have gradually become an important target for natural gas development due to their wide distribution and huge reserves. The generation and distribution of hydrate slurries in non-diagenetic systems are affected by many factors such as temperature, pressure, particle content, flow rate and salinity. Especially in the special environment of deep ocean water or low temperature and high pressure, the dynamic behavior of slurries poses new challenges to the efficiency and safety of natural gas extraction. The uncertainty of slurry flow characteristics may lead to problems such as pipeline blockage, equipment wear and energy loss.
[0003] Since the formation and decomposition process of slurry hydrate is highly nonlinear and time-varying, traditional thermodynamic models and experimental methods are difficult to accurately predict the behavior in actual production, so new methods are urgently needed to predict the formation volume of hydrate. Summary of the invention
[0004] In view of the shortcomings of the existing methods, the present invention is based on an improved LSTM model to learn the laws of hydrate formation in slurry, providing the possibility for real-time monitoring and prediction. Compared with traditional methods, the improved LSTM model can more effectively process multi-dimensional dynamic data, has strong generalization ability, and can more accurately adapt to the prediction of hydrate slurry behavior in complex environments.
[0005] The technical solution adopted by the present invention is: the method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM comprises the following steps:
[0006] Step 1, collecting hydrate slurry parameters;
[0007] As a preferred embodiment of the present invention, the hydrate slurry parameters include: time, temperature, pressure, gas consumption, salt concentration, sand concentration, and hydrate volume fraction.
[0008] As a preferred embodiment of the present invention, the gas consumption calculation formula is:
[0009] (1)
[0010] Where R is the gas constant, n g (t) represents the gas consumption at time t, P(t) represents the pressure value at time t, V g (t) represents the volume of carbon dioxide gas at time t, z(t) represents the compression factor at time t, and T(t) represents the temperature value at time t.
[0011] Step 2: cleaning and pre-treating the hydrate slurry parameters;
[0012] As a preferred implementation of the present invention, the preprocessing includes: regularization processing.
[0013] Step 3: Construct an improved LSTM model, take hydrate slurry parameters as input and hydrate volume fraction as output, and train the improved LSTM model;
[0014] As a preferred embodiment of the present invention, the improved LSTM model is cascaded by a first-layer LSTM network consisting of 64 memory units, a first-layer Dropout layer, a second-layer LSTM network consisting of 32 memory units, a second Dropout layer and a 32-dimensional fully connected layer.
[0015] Step 4: Compare the predicted and observed values of the volume fractions of the training set and the test set, using the performance evaluation index as the inspection standard;
[0016] As a preferred implementation of the present invention, the performance evaluation indicators include: MAE, RMSE, MBE, and R2.
[0017] As a preferred implementation of the present invention, cross-validation is performed on the improved LSTM model.
[0018] Beneficial effects of the present invention:
[0019] 1. Use historical data to train the improved LSTM network to capture the law of hydrate formation in slurry, making it possible for real-time monitoring and prediction;
[0020] 2. Compared with traditional methods, the cascade of dual LSTM networks can more effectively process multi-dimensional dynamic data, has strong generalization ability, and can more accurately adapt to the prediction of hydrate slurry behavior in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is the improved LSTM network structure diagram of the present invention;
[0022] Figure 2 It is a comparison chart of the predicted slurry generation volume fraction and the measured value of the present invention. DETAILED DESCRIPTION
[0023] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.
[0024] like Figure 1As shown, the LSTM-based method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system includes the following steps:
[0025] Step 1, collecting hydrate slurry parameters;
[0026] Collect data such as temperature and pressure of hydrate slurry changing with time under different conditions, as well as the corresponding hydrate volume fraction;
[0027] The hydrate slurry parameters obtained by visualizing the high-pressure reactor equipment include: time, temperature, pressure, gas consumption, salt concentration, sand concentration, and hydrate volume fraction; the hydrate slurry parameters also include other parameters, and these 7 parameters are the optimal parameters in the improved LSTM network; for details of the test equipment, please refer to "Kinetic characterization of gas mixture separation by hydrate method with 1,3-dioxolane and multi-walled carbon nanotubes complex synergistic system" by Lv Xiaofang et al.;
[0028] The hydrate slurry testing device is a visual high-pressure reactor. The entire set of high-pressure reactor equipment includes air intake, data acquisition, cooling, hydrate reaction, detection system, etc.
[0029] The reactor parameters are as follows:
[0030] Design pressure: 0~16Mpa; design temperature: -20~20℃; the controllable temperature range of the cooling system is -20~80℃; experimental materials: deionized water, silt sand, NaCl, carbon dioxide.
[0031] Table 1 shows the measurement range of flow characteristic parameters
[0032] Basic parameters unit scope time s 1~20129 temperature ℃ 1~6.94 pressure MPa 2.5~4 Salt concentration % 25~35 Sand concentration % 10~30 Gas consumption Mol 0~1.17 Hydrate volume fraction % 0~15.15
[0033] The experiment first checked the airtightness of the device, ensured the cleanliness of the reactor, and confirmed that the data transmission was normal; then the vacuum was evacuated for 5 minutes, the prepared sand-salt solution was placed in the reactor, data acquisition was started, and the changes in temperature and pressure data were recorded. At the same time, the water bath temperature was set above the phase equilibrium temperature; then carbon dioxide gas was slowly introduced into the reactor at a rate of 0.1 MPa / min until the pressure reached the preset value, and then the temperature and pressure stabilized period was entered; magnetic stirring was started at a rate of 600 rpm, and the water bath temperature was gradually reduced at a rate of 0.1 K / min; the sudden drop in pressure and the rapid rise in temperature indicated that hydrates began to form. During this period, the data acquisition system collected pressure, temperature and other data; after the hydrate formation was completed, the magnetic stirring was turned off, and finally the water bath was turned off, and the experiment ended.
[0034] The gas consumption calculation formula is:
[0035] (1)
[0036] Where R is the gas constant, n g (t) represents the gas consumption at time t, P(t) represents the pressure value at time t, V g (t) represents the volume of carbon dioxide gas at time t, z(t) represents the compression factor at time t, and T(t) represents the temperature value at time t.
[0037] Step 2: Cleaning and preprocessing the hydrate slurry parameter data;
[0038] Based on the regularization method, the data of static characteristics (such as sand concentration, salt concentration) and dynamic characteristics (temperature, pressure, gas consumption) as well as the corresponding generated hydrate volume fraction data are processed, and the sliding window technology is used to construct the time series input;
[0039] The calculation formula for completing the regularization process is:
[0040] (2)
[0041] Among them, x raw represents the original variable, x norm is the regularized variable, represents the mean of x, represents the standard deviation of x.
[0042] Step 3: Build an improved LSTM model, input hydrate slurry parameters into the model, and output the hydrate formation volume fraction;
[0043] The LSTM model is constructed by adding a gating device (input gate f t , output gate i t , and the forget gate Ot ) to handle the large time span of time series prediction; the three gating devices in the unit control the information flow between units, so that the LSTM model has excellent memory ability when processing sequence data, can capture long-term dependencies, and can effectively solve problems such as gradient disappearance and gradient explosion;
[0044] In each cycle (also called time step), the calculation process of the LSTM network involves three main input vectors and two output vectors, namely:
[0045] c t-1 : represents the state vector stored in the memory cell in the previous cycle, usually called the "cell state" of the previous time step;
[0046] h t-1 : represents the output vector of the previous cycle, usually the "hidden state" of the previous time step;
[0047] x t : The input vector of the current time step, which contains the external information of this time step.
[0048] The output of each loop includes:
[0049] c t : The state of the memory cell after this cycle, usually called the "cell state" of the current time step;
[0050] h t : The output vector of this cycle, usually used as the "hidden state" of this time step, contains the processing results of the current input and the previous input.
[0051] The formula for the corresponding unit update process is as follows:
[0052] (3)
[0053] (4)
[0054] (5)
[0055] (6)
[0056] (7)
[0057] (8)
[0058] in, is the output of the previous unit; Represents the input of the current unit; , , , , , , , It is a trainable parameter in the model and is continuously updated during the training process; , It is the activation function of the model, responsible for improving the nonlinear expression ability of the model.
[0059] The improved LSTM model structure of the present invention includes:
[0060] The first layer of LSTM network has a 7-dimensional input feature and a 64-dimensional output feature. The batch dimension is set to batch_first=True.
[0061] The first Dropout layer has a dropout rate of 0.3 to reduce overfitting;
[0062] The second layer of LSTM network has an input feature of 64 dimensions, an output feature of 32 dimensions, and the batch dimension is set to batch_first=True;
[0063] The second Dropout layer, with a dropout rate of 0.3, is used to further reduce overfitting;
[0064] The fully connected layer has an input feature of 32 dimensions and an output feature of 1, which outputs the final prediction value of the model.
[0065] That is, the hydrate slurry parameters at the current moment are used to predict the hydrate volume fraction at the next moment.
[0066] The LSTM model is improved by using historical data training, and the generalization ability of the model is improved through cross-validation.
[0067] Step 4: Compare the predicted and observed values of the volume fraction of hydrate slurry in the non-diagenetic system in the training set and the test set, and use MAE, RMSE, MBE, and R2 as test criteria to verify the rationality of the simulation results;
[0068] Using MAE, RMSE, MBE, R 2 As a test standard, the formula is:
[0069] (9)
[0070] (10)
[0071] (11)
[0072] (12)
[0073] Where n represents the length of the time series in the training set and the test set.
[0074] Table 2 shows the MAE, RMSE, MBE, and R of the training and test sets. 2 Indicator comparison
[0075] MAE RMSE MBE <![CDATA[R 2 ]]> Train 0.0675 0.1074 0.112 0.9956 Test 0.0682 0.1034 0.1142 0.9931
[0076] Figure 2 The good fitting effect between the experimental prediction value and the actual experimental value is demonstrated, indicating that it is completely feasible to use the improved LSTM model to predict the volume fraction of hydrate slurry in non-diagenetic system; this result not only verifies the effectiveness of the improved LSTM model in this field, but also provides strong technical support for the optimization of natural gas storage and transportation system; by accurately predicting the volume fraction of hydrate slurry in non-diagenetic system, the improved LSTM model can provide important decision-making basis for the optimization of production process and help improve the efficiency of production system.
[0077] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.
Claims
1. A method for predicting the volume fraction of hydrate slurry generated in non-diagenetic systems based on LSTM, characterized in that: The following steps are involved: Step 1, collecting hydrate slurry parameters; Step 2: cleaning and pre-treating the hydrate slurry parameters; Step 3: Construct an improved LSTM model, take hydrate slurry parameters as input and hydrate volume fraction as output, and train the improved LSTM model.
2. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 1, characterized in that: The improved LSTM model is cascaded by a first-layer LSTM network consisting of 64 memory units, a first-layer Dropout layer, a second-layer LSTM network consisting of 32 memory units, a second Dropout layer, and a 32-dimensional fully connected layer.
3. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 1 is characterized in that: Performance evaluation indicators include: MAE, RMSE, MBE, and R2.
4. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 1, characterized in that: Cross-validation was performed on the improved LSTM model.
5. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 1, characterized in that: Hydrate slurry parameters include: time, temperature, pressure, gas consumption, salt concentration, sand concentration, and hydrate volume fraction.
6. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 5, characterized in that: The gas consumption calculation formula is: （1） Where R is the gas constant, n g (t) represents the gas consumption at time t, P(t) represents the pressure value at time t, V g (t) represents the volume of carbon dioxide gas at time t, z(t) represents the compression factor at time t, and T(t) represents the temperature value at time t.
7. The method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM according to claim 1, characterized in that: Preprocessing includes: Regularization processing.
8. The LSTM-based non-diagenetic system hydrate slurry generation volume fraction prediction system is characterized by: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the method for predicting the volume fraction of hydrate slurry generated in a non-diagenetic system based on LSTM as described in any one of claims 1 to 7.
9. A computer readable medium storing a computer program code, characterized in that: When the computer program code is executed by a processor, the method for predicting the generated volume fraction of hydrate slurry in a non-diagenetic system based on LSTM is implemented as described in any one of claims 1 to 7.