Method for rapidly judging retrograde condensation of complex medium condensate gas reservoir based on POA-LSTM-TCN

Abstract

The invention discloses a POA-LSTM-TCN-based quick judgment method for reverse condensation of a complex medium condensate gas reservoir, and relates to the technical field of oil and gas field development engineering. The method comprises the following steps: acquiring parameters, including reservoir physical property parameters, fluid parameters and production data, of a target block when reverse condensation does not occur, and preprocessing the parameters to obtain a historical data set; a POA-LSTM-TCN combination model is established; based on actual production data, the oil-gas ratio in the production process is calculated through the trained model, a calculation prediction result is judged with the real oil-gas ratio, and if the calculation prediction result is compared with the real oil-gas ratio, the deviation of the calculation prediction result at a certain time point or a certain time period is larger than a first threshold value, the oil-gas ratio is determined to be the real oil-gas ratio. If yes, it shows that the reservoir has the retrograde condensation phenomenon at the time point or the time period. According to the method, abnormal fluctuation and prediction of the gas-oil ratio are combined, the reverse condensation phenomenon can be judged more accurately and rapidly in the hole-seam-fault complex medium gas reservoir, and higher early warning precision is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas field development engineering, and particularly to a rapid judgment method for retrograde condensation in complex medium condensate gas reservoirs based on POA-LSTM-TCN. Background Art

[0002] The pore-fracture-fault complex medium gas reservoir is a complex reservoir structure composed of pores, fractures and faults, with high heterogeneity and fluid flow complexity. During the seepage process of the gas reservoir, the cross-flow behavior between different media plays a crucial role. According to the multiple-medium theory, the matrix part in the gas reservoir has a relatively low permeability and mainly serves as the storage space for gas, while the main flow path of gas is through the fracture system for cross-flow. Due to the large difference in the flow capacity of multiple media, during the development of the gas reservoir, the pressure difference inside different media is also very large, and it is entirely possible to have a situation where the fracture pressure is extremely low while the matrix pressure is very high.

[0003] During the development of condensate gas reservoirs, when the gas pressure drops below the dew point pressure of the condensate oil, the condensate oil will precipitate from the gas phase and preferentially precipitate in the areas with lower pressure. Traditional gas reservoir engineering methods usually judge the occurrence of retrograde condensation phenomenon by the average reservoir pressure: assuming that the pressure changes in each medium in the gas reservoir are uniform, so the average reservoir pressure can be calculated and compared with the dew point pressure of the condensate oil to judge whether the retrograde condensation phenomenon has occurred. When the average reservoir pressure is lower than the dew point pressure, it is considered that the retrograde condensation phenomenon has occurred and the condensate oil precipitates from the gas phase. This method is simple and easy to implement in theory, so it is widely used in the prediction of retrograde condensation in gas reservoir engineering.

[0004] However, in the pore-fracture-fault complex medium gas reservoir, there are significant pressure differences between the matrix, fractures and faults. Specifically, the pressure in the fracture system and fractures is usually relatively low and may have dropped below the dew point pressure, resulting in the preferential precipitation of condensate oil in these areas. Due to the relatively low permeability of the matrix part, the gas flow is restricted, and the pressure may not drop below the dew point, so the precipitation of condensate oil is delayed and the retrograde condensation phenomenon does not occur in the gas reservoir in the matrix. This phenomenon leads to the fact that the average reservoir pressure calculated by the traditional method cannot accurately reflect the true state of each part of the reservoir. Especially when retrograde condensation has occurred in the fractures but not in the matrix, using the average pressure to judge the retrograde condensation phenomenon obviously has a large deviation.

[0005] Due to the fact that the retrograde condensation phenomenon usually does not have an obvious dynamic response in the initial stage, traditional pressure monitoring methods cannot detect the differences between fractures and matrices in a timely manner, making it difficult to quickly judge the occurrence of retrograde condensation. This not only affects the production dynamics of the reservoir but also poses great challenges to the development and management of gas reservoirs. Although numerical simulation methods can be used to make judgments to a certain extent, those skilled in the art know that numerical simulation methods need to be based on geological models and experimental tests of fluid PVT parameters. The cost of these two basic studies is over one million yuan each. It not only has the problem of low accuracy but also requires a long simulation time and high cost. Summary of the Invention

[0006] To solve at least one of the above problems, the present invention proposes a rapid method for judging retrograde condensation in a complex medium condensate gas reservoir based on POA-LSTM-TCN.

[0007] The technical solution of the present invention is as follows: A rapid method for judging retrograde condensation in a complex medium condensate gas reservoir based on POA-LSTM-TCN, comprising the following steps:

[0008] S1. Obtain the parameters of the target block when retrograde condensation has not occurred, including reservoir physical property parameters, fluid parameters, and production data, and preprocess them to obtain a historical data set;

[0009] S2. Establish a POA-LSTM-TCN combined model: Establish an LSTM model and a TCN model respectively, and use POA to optimize the hyperparameters of the LSTM model and the TCN model respectively; with the oil-gas ratio as the output, use the historical data set to train the POA-LSTM-TCN combined model. During the training process, set weights for the prediction results of the LSTM model and the TCN model respectively, and use the mean square error as the loss function to obtain the trained model;

[0010] S3. Based on the actual production data, calculate the oil-gas ratio during the production process through the trained model, and judge the calculated prediction result and the true oil-gas ratio: If the deviation of the calculated prediction result from the true oil-gas ratio at a certain time point or a certain time period is greater than the first threshold, it indicates that retrograde condensation has occurred in the reservoir at this time point or this time period.

[0011] Advantageous Effects:

[0012] The present invention judges the occurrence of retrograde condensation by monitoring the abnormal change of the gas-oil ratio: By accurately capturing the local abnormal fluctuation of the gas-oil ratio, the possible occurrence of retrograde condensation in the gas reservoir can be detected earlier and a warning can be given. This method can overcome the shortcoming that traditional methods cannot effectively reflect pressure unevenness.

[0013] Meanwhile, the POA-LSTM-TCN combined machine learning model of the present invention can more accurately reveal the differences between fractures and matrixes in complex medium gas reservoirs compared with traditional machine learning models, and has higher accuracy and adaptability.

[0014] In summary, by combining the abnormal fluctuations of the gas-oil ratio and the calculation prediction results, the present invention provides a unique and innovative solution that can more accurately and quickly judge the retrograde condensation phenomenon in the pore-fracture-fault complex medium gas reservoir and achieve higher early warning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a comparison chart of the verification results of the POA-LSTM-TCN combined model and other models;

[0016] Figure 2 It is a pressure change chart during the numerical simulation process;

[0017] Figure 3 It is a change chart of the calculated prediction results and the simulated true oil-gas ratio over time obtained by using the method of this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0018] The following will clearly and completely describe the specific embodiments of the present invention in conjunction with the examples and the drawings. Obviously, the described examples are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0019] A method for quickly judging retrograde condensation in a complex medium condensate gas reservoir based on POA-LSTM-TCN includes the following steps:

[0020] S1. Obtain the parameters of the target block when retrograde condensation has not occurred, including reservoir physical property parameters, fluid parameters, and production data, and preprocess them to obtain a historical data set;

[0021] In this step, the selected parameters include date, well number, formation, oil production method, daily liquid production, daily oil production, daily water production, daily gas production, gas-oil ratio, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, and temperature after the choke. Delete the irrelevant columns in the data set: well number, formation, oil production method, daily liquid production, and daily oil production, and calculate the cumulative gas production. Therefore, the final parameters are daily water production, daily gas production, gas-oil ratio, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, temperature after the choke, and cumulative gas production.

[0022] The above-mentioned parameters can all be obtained from the production record data of the well site, so the acquisition cost is relatively low.

[0023] After obtaining the above parameters, since the number of parameters is relatively large and the data scales of each parameter data are different, the calculation efficiency will be greatly reduced in the subsequent simulation calculation process. Therefore, in this embodiment, it is also necessary to preprocess the data.

[0024] During preprocessing, first perform a correlation analysis on the parameters. In the prior art, there are many methods for correlation analysis, such as the common Pearson correlation coefficient method, Spearman correlation coefficient method, etc., all of which can be applied to this embodiment. However, considering the characteristics of each data, in this embodiment, the Pearson correlation coefficient method is preferably used to perform a correlation analysis on each parameter. After obtaining the analysis results, select the parameters with a correlation greater than the second threshold. For the second threshold, it can be set according to the actual situation, such as 0.5.

[0025] After selecting the corresponding parameters, since the dimensions and numerical ranges of each parameter vary significantly, it is also necessary to perform normalization processing on them to facilitate subsequent calculations. The normalization processing method is a conventional method in the art, and its specific operation will not be elaborated here.

[0026] After normalization processing, it is also necessary to perform an ADF test on these parameter data. The purpose of the test is to determine whether each parameter data is stationary time series data. If it is not stationary sequence data, it is also necessary to perform stationary processing. These operations are also conventional means in the art, so their specific operations will not be elaborated here. In particular, for the parameter cumulative gas production, since it shows an upward trend itself, it belongs to non-stationary characteristics. However, since it conforms to the actual situation, it will not have a negative impact on the subsequent overall analysis and is regarded as an acceptable characteristic and does not require stationary processing.

[0027] At the same time, for the above historical data, it is required that it is the historical data when no retrograde condensation occurs. Specifically, it refers to the historical data when the average formation pressure and the minimum bottom-hole flowing pressure of a single well are both much higher than the dew point pressure and the pressure in the fracture is higher than the dew point pressure. This enables the subsequent prediction process to be carried out under the condition of no retrograde condensation.

[0028] S2. Establish a POA-LSTM-TCN combined model: respectively establish an LSTM model and a TCN model, and use POA to optimize the hyperparameters of the LSTM model and the TCN model respectively; use the oil-gas ratio as the output, and train the POA-LSTM-TCN combined model with the historical data set. During the training process, set weights for the prediction results of the LSTM model and the TCN model respectively, and use the mean square error as the loss function to obtain the trained model;

[0029] For the LSTM model and the TCN model, each has its own advantages: the LSTM is good at capturing sequential information with long-term dependencies, while the TCN is outstanding in multi-scale feature learning and computational efficiency. After combining the two, their respective advantages can be fully utilized: it can not only accurately capture the long-term trend of the gas-oil ratio data, but also efficiently process the short-term fluctuations in the data. This combined model can better adapt to the diversity and complexity of the gas-oil ratio data in complex medium gas reservoirs, improving the prediction accuracy and response speed.

[0030] For the LSTM model and the TCN model, it is usually determined according to their loss functions. In this embodiment, the mean square error is used as their loss function.

[0031] POA, that is, the Pelican Optimization Algorithm, its main operation process is as follows:

[0032] When the pelican determines the prey's position, it moves towards it. The prey's position is randomly generated in the search space. If the fitness of the new position is better than the current best solution, the best solution is updated and it moves in this direction; otherwise, the distance is increased.

[0033]

[0034] Among them, P1 is the first stage; X i,j is the state of the i-th pelican in the j-th dimension; is the new state of the i-th pelican in the j-th dimension at P1; rand is a random number between 0 and 1; p j is the position of the prey in the j-th dimension; I is a random number equal to 1 or 2; F p is its objective function value; F i is the objective function value of the i-th pelican.

[0035] If the value of the objective function is improved at this position, the new position of the pelican is accepted. In this type of update, it is called an effective update to prevent moving to a non-optimal area.

[0036]

[0037] Among them, is the new state of the i-th pelican; is the objective function value of the first stage.

[0038] The pelican spreads its wings on the water surface to attract and capture the prey. This strategy results in more prey being captured in the attack area, and the effective update is used again to accept or reject the new pelican position.

[0039]

[0040] Among them, is the new state of the i-th pelican in the second stage in the j-th dimension; R is an integer of 0 or 2; is X i,j the neighborhood radius of; t is the current iteration number; T is the maximum iteration number.

[0041] After the updates in the first and second stages, the algorithm updates the current best solution according to the new state of the population and the objective function value, and enters the next iteration until the maximum iteration number is reached. The finally output best solution is the quasi-optimal solution of the problem.

[0042] Use POA to optimize the hyperparameters of the LSTM model and the TCN model. The optimized models can be named the optimal LSTM model and the optimal TCN model. Subsequently, using the dataset in S1, with the oil-gas ratio as the output and the remaining parameters as the input, train the optimal LSTM model and the optimal TCN model. During the training process, set the weights of 0 to 1 and the loop step size for the optimal LSTM model and the optimal TCN model respectively. Using the mean square error as the loss function, calculate the loss function value between the output result calculated by the following formula and the true value. When the loss function value is the smallest, the weight at this time is the optimal weight, and then a model with the optimal weight combination is obtained;

[0043]

[0044] In the formula, pred represents the predicted value of the weight combination; w1 represents the weight of LSTM; w2 represents the weight of TCN; yhat_lstm represents the predicted value of LSTM; yhat_tcn represents the predicted value of TCN; 1e-10 represents the extremely small constant 1×10 -10 .

[0045] For the above step size, it can be set according to the actual situation, such as the conventional 0.01, 0.02, etc., or the step size can be adjusted according to the actual situation.

[0046] After the above operations are completed, the entire model is trained and can be used for subsequent predictions.

[0047] At the same time, during the entire training process, similar to the conventional operation, divide the historical dataset in S1 into a training set and a validation set according to a ratio of 7:3. The training set is used for the entire training process, and the validation set is used to verify the accuracy of the trained model.

[0048] S3. Based on the actual production data, calculate and predict the oil-gas ratio during the production process through the trained model, and judge the calculated and predicted result with the true oil-gas ratio: If the deviation between the calculated and predicted result and the true oil-gas ratio is greater than the first threshold at a certain time point or a certain time period, it means that retrograde condensation occurs in the reservoir at this time point or this time period.

[0049] For actual production data, it is basically the same as the parameters in S1. For example, parameters such as daily water production, daily gas production, oil-gas ratio, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, temperature after choke, and cumulative gas production can be obtained relatively easily after production. Among them, the oil-gas ratio, as the true oil-gas ratio, does not participate in subsequent calculation predictions but is only used for comparison.

[0050] When calculating the oil-gas ratio during the production process using the trained model, the input parameters are the same as the optimized parameters in S1. At the same time, when performing calculation predictions, normalization processing is also required. The processed data is input into the trained model for prediction, and it will output a calculation prediction result, which is the predicted oil-gas ratio.

[0051] At the same time, the reason for making the above judgment in this embodiment is as follows: The true oil-gas ratio can reflect the true condensate situation in the gas reservoir. And the data in S1 are all data when no retrograde condensation occurs. After training the model in S2 with these data, it is more adaptable to the working conditions when no retrograde condensation occurs.

[0052] When retrograde condensation occurs in the reservoir, there will be large fluctuations between the actual result and the predicted value. Through these fluctuations, it can be shown that retrograde condensation has occurred in the gas reservoir. For the first threshold, in the embodiments of the present invention, it is set as the deviation of the oil-gas ratio prediction value and the simulated true oil-gas ratio of ±2σ. Of course, those skilled in the art can also set different first thresholds according to the specific situation of this block.

[0053] To further illustrate the method of the embodiments of the present invention, specific examples are given below.

[0054] Target block characteristics:

[0055] 1. Geological characteristics: Bozige gas reservoir is located in the Kelasu structural belt of the Kuqa depression in the northern Tarim Basin. The target reservoirs are the Bashijiqike Formation and the Baxigai Formation, with a burial depth exceeding 6000m. The development and evolution of the stratigraphic structure in the Kelasu structural belt are mainly controlled by regional thrust faults, forming a large number of faults of different scales, with the fracture scale ranging from micrometers to hundreds of meters. Natural fractures are well developed in the gas reservoir. Research shows that the main dynamic source of the structural fractures in the reservoir is the lateral tectonic "compaction" effect. The main strike is relatively consistent with the direction of the current maximum horizontal principal stress. The development and distribution of fractures are jointly affected by lithology, layer thickness, overlying gypsum salt layer, structural style, and fault distance. The reservoir of the block is a low-porosity and ultra-low-permeability tight gas reservoir. The reservoir porosity is mainly concentrated in 3.5% - 10%, with an average value of 6.1%; the reservoir permeability is mainly concentrated in 0.055mD - 1mD, with an average value of 0.178mD; the permeability and storage characteristics of the reservoir are controlled by the degree of fracture development.

[0056] 2. Fluid characteristics: This block belongs to a high-temperature and ultra-high-pressure gas reservoir. The original formation pressure of the gas reservoir at a depth of 6128.89 m (altitude -4101.89 m) is 115.76 MPa, the formation temperature is 128.65 °C, and the pressure coefficient is 1.93, belonging to a high-temperature and ultra-high-pressure system. The temperature gradient is 1.67 °C / 100 m, and the pressure gradient is 0.38 MPa / 100 m. Its natural gas has a high methane content, is rich in light hydrocarbons, and has a low non-hydrocarbon gas content. The relative density of natural gas is 0.64, the highest methane content is 87.34%, the second highest ethane content is 6.93%, the content of propane and above is 3.46%, the nitrogen content is 1.96%, the carbon dioxide content is 0.31%, and there is no sulfur content. Characteristics of condensate oil: Light condensate oil, with an average density of 0.198 g / cm 3 , no sulfur content, low viscosity, and the average dynamic viscosity is 1.26 mPa·s; high wax content, ranging from 11.4% to 20.6%, low gum and asphaltene content, with average values of 0.31% and 0.43% respectively. The original formation pressure in the TB block is higher than the dew point pressure. At 127.9 °C, the dew point pressure is 53.41 MPa.

[0057] 3. Production data: The Bozai gas reservoir records the daily well production data from December 2, 2018 to February 14, 2023. The data includes daily water production, daily gas production, gas-oil ratio, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, and temperature after the choke, and calculates and adds the cumulative gas production column. Due to the excessive original data, this test example gives some statistical data in Table 1.

[0058] Table 1 Partial statistical data table

[0059]

[0060] The trained model is obtained by training using the methods of S1 and S2 in this embodiment, and then its accuracy is verified using the validation set. The final results are as Figure 1 shown in Table 2; at the same time, in order to illustrate the superiority of the model in the embodiment of the present invention, the POA-TCN model and the POA-LSTM model are used for comparison.

[0061] Table 2 Comparison table of the effects of the model in the embodiment of the present invention and other models

[0062] Model POA-LSTM POA-TCN POA-LSTM-TCN MAE 1.7643 3.7970 1.1114 RMSE 2.5026 4.6920 1.7318 MAPE 3.7435％ 7.7403％ 2.4386％

[0063] As can be seen from Table 1 and Figure 1 it can be seen that the model in the embodiment of the present invention has high accuracy and small error.

[0064] Due to time constraints, there are currently no parameters for retrograde condensation in this block. Therefore, we use some data from this block for numerical simulation to simulate future production results. The specific operations are as follows:

[0065] 1. Based on the geological model of the target block, coarsen the grid and properties of the template block and assign its static data attributes. In this step, based on the existing geological model, to improve the efficiency of numerical simulation, the grid and properties of the target block are reasonably coarsened on the premise of ensuring calculation accuracy. The porosity is coarsened using arithmetic mean, and the permeability is coarsened using geometric mean. Subsequently, smoothing processing is carried out to eliminate extreme values and reduce the differences between adjacent grids, improving the calculation speed and stability. The established geological model is imported into tNavigator in Rescue format and assigned static data attributes, including structural information, matrix porosity and permeability, and fracture porosity and permeability, etc., to establish a basic grid model with physical properties.

[0066] 2. Establish a hybrid model combining a dual-porosity model and a discrete fracture model: Merge the micro-fracture and pore properties and use them as the matrix part, use the medium-small scale fracture properties as the continuous fracture medium part, and treat the large-scale fractures as discrete fracture parts. In this step, a hybrid model combining a dual-porosity model and a discrete fracture model is adopted. The property body after merging the micro-fracture and pore properties is used as the matrix part in the dual-porosity model; the medium-small scale fracture property body is used as the continuous fracture medium part in the dual-porosity model; the large-scale fractures are treated as a discrete fracture model. For the established large-scale fracture system, the connected fracture slices are selected and characterized in tNavigator in the form of discrete fracture slices.

[0067] 3. Set different relative permeability curves and capillary pressure curves according to different parts of the hybrid model. In this step, according to the different relative permeability curve characteristics of gas-water phase seepage in the matrix-micro fracture, medium-small scale fracture, and large fracture-fault regions, different relative permeability curve partitions are set. Moreover, the capillary pressure values and influence degrees in the matrix-micro fracture, medium-small scale fracture, and large fracture-fault regions are different, so different capillary pressure curves need to be set separately. In addition, under different horizons, different positions, different temperature and pressure, and different fluid conditions, the capillary pressure curves may all have obvious differences and may all require different partitions.

[0068] 4. Take the oil and gas in the gas reservoir for PVT experiments, and input them into the PVT module of the hybrid model for fitting until the fluid properties of the hybrid model match the gas reservoir characteristics of the target block. In this step, take the oil and gas in the gas reservoir, and measure the PVT parameters under gas reservoir conditions through experiments, including key data such as dew point pressure, relative volume, liquid content, etc. Input the experimental data into the PVT module of tNavigator for various fitting operations, including constant composition expansion experiments (CCE), bubble point and dew point pressure fitting, etc., to ensure that the fluid properties in the model are highly consistent with the real reservoir characteristics, thereby improving the accuracy of simulation prediction.

[0069] 5. Initialize the hybrid model and conduct reserve calculation. In the gas reservoir initialization stage, according to geological and fluid characteristics, use the traditional gravity vertical model to set basic parameters such as initial gas saturation, irreducible water saturation, and gas-water interface position to construct the gas reservoir model in the initial state. At the same time, carry out reserve calculation in combination with the model structure and physical properties, and compare and verify with the calculation results of the conventional volume method to ensure the accuracy and rationality of the model initialization settings.

[0070] 6. Conduct simulation production with the hybrid model, and correct its production based on the historical production data of the target block until they match. Based on the historical data such as gas production, oil production, formation pressure, and flowing pressure of actual gas wells, input them as constraint conditions into the numerical simulator, and conduct history matching by adjusting the model parameters. The "constant gas production" strategy is adopted in the simulation process to gradually correct the consistency between the model output and the actual production curve, accurately restore the productivity and pressure response of single wells and the overall block, and provide a reliable basis for subsequent model prediction.

[0071] 7. Conduct simulation production based on the corrected model, and use the simulation production data as the simulation real results. In this step, the simulation real results include not only the simulated real oil-gas ratio, but also parameters such as daily water production, daily gas production, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, temperature after choke, and cumulative gas production in the future production process.

[0072] Carrying out simulation production through the above steps is a conventional means in this field, and those skilled in the art can adjust its simulation process according to the actual situation.

[0073] The situation after simulation production using the above numerical simulation method is as Figure 2 、 Figure 3 shown, where Figure 2 is the pressure change diagram in the numerical simulation process; Figure 3 is the diagram of the change of the oil-gas ratio of the calculation prediction result obtained by the method of this embodiment and the simulated real oil-gas ratio over time. From Figure 2It can be seen that in June 2031, the pressure in the seam was lower than the dew point pressure, and retrograde condensation occurred in the fracture. However, if judged by conventional methods, such as the average pressure, retrograde condensation still did not occur in the reservoir at this time, indicating that conventional methods are not applicable to complex media. From Figure 3 It can be seen that retrograde condensation occurred in June 2031, which is consistent with Figure 2 the results, indicating that the method of the present invention can effectively judge the retrograde condensation phenomenon in complex media reservoirs.

[0074] The above are only the preferred embodiments of the present invention and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the disclosed technical content without departing from the technical solution of the present invention. However, as long as it does not depart from the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A rapid judgment method for retrograde condensation in complex media condensate gas reservoirs based on POA-LSTM-TCN, characterized in that, It includes the following steps: S1. Obtain the parameters of the target block when no retrograde condensation occurs, including reservoir physical property parameters, fluid parameters, and production data, and preprocess them to obtain a historical data set; S2. Establish a POA-LSTM-TCN combined model: Establish an LSTM model and a TCN model respectively, and use POA to optimize the hyperparameters of the LSTM model and the TCN model respectively; Use the oil-gas ratio as the output, and use the historical data set to train the POA-LSTM-TCN combined model. During the training process, set weights for the prediction results of the LSTM model and the TCN model respectively, and use the mean square error as the loss function to obtain the trained model; S3. Based on the actual production data, calculate and predict the oil-gas ratio during the production process through the trained model, and judge the calculated and predicted result with the true oil-gas ratio: If the deviation of the calculated and predicted result from the true oil-gas ratio at a certain time point or a certain time period is greater than the first threshold, it indicates that retrograde condensation has occurred in the reservoir at this time point or during this time period.

2. The method according to claim 1, characterized in that In S1, the parameters include daily water production, daily gas production, gas-oil ratio, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, temperature after the choke, and cumulative gas production.

3. The method according to claim 1, characterized in that, In S1, the preprocessing method includes: performing a correlation analysis on the parameters, selecting the parameters with a correlation greater than the second threshold, and then performing normalization processing and ADF test on them.

4. The method according to claim 1, characterized in that, In S2, when training the POA-LSTM-TCN combined model, first determine the optimal hyperparameters of the LSTM model and the TCN model based on the loss function and POA to obtain the optimal LSTM model and the optimal TCN model; Subsequently, set weights of 0 to 1 for the optimal LSTM model and the optimal TCN model respectively and set the loop step size, and use the mean square error as the loss function to calculate the loss function value between the output result calculated by the following formula and the true value. When the loss function value is the smallest, the weight at this time is the optimal weight, and then a model with an optimal weight combination is obtained; In the formula, pred represents the predicted value combined with weights; w1 represents the weight of LSTM; w2 represents the weight of TCN; yhat_lstm represents the predicted value of LSTM; yhat_tcn represents the predicted value of TCN; 1e-10 represents the extremely small constant 1×10 -10 .

5. The method according to claim 1, wherein In S3, the actual production data includes daily water production, daily gas production, water-gas ratio, oil pressure, casing pressure, back pressure, wellhead temperature, temperature after the choke, and cumulative gas production.

6. The method according to claim 1, characterized in that, In S3, the first threshold is 2 times the deviation.

Images