Method for predicting chemical flooding heavy oil yield based on STL-TCN model

Abstract

The invention discloses a chemical flooding heavy oil yield prediction method based on an STL-TCN model, and relates to the field of chemical flooding heavy oil yield prediction, and the method comprises the following steps: obtaining a sample data set; carrying out denoising and normalization preprocessing on the data; dividing into a test set and a verification set; performing oil production prediction by using the test set; and verifying with a true value, and updating model parameters to obtain a final optimized neural network model. The method not only can consider the influence of related factors on the oil production, but also can consider the correlation of data in time, is high in prediction precision, and can be used as a new method applied to the prediction of the oil production of the thickened oil.

Description

Technical Field

[0001] The invention relates to the field of chemical flooding heavy oil production prediction, in particular to a method for chemical flooding heavy oil production prediction based on an STL-TCN model. Background Art

[0002] In recent years, with the advancement of artificial intelligence algorithms, artificial intelligence prediction algorithms have gradually been applied to oil production prediction. Currently, the artificial intelligence algorithms used in heavy oil production prediction mainly include support vector machine regression and back propagation neural network, but these two methods cannot consider the temporal correlation of data and are more sensitive to the selection of independent variables.

[0003] To this end, we invented a method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model. This method can not only consider the impact of relevant factors on oil production, but also consider the temporal correlation of data. It has high prediction accuracy and strong adaptability, and can be used as a new method for predicting the production of heavy oil by chemical flooding. Summary of the invention

[0004] The object of the present invention is to provide a method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model, comprising the following steps:

[0006] Step 1: Obtain actual dynamic production data from the oil field block;

[0007] Step 2: Use Pearson correlation to test the correlation of influencing factors;

[0008] Step 3: Remove outliers from the initial sample data set in step 1 and standardize and normalize the data set;

[0009] Step 4: The data set is divided into training set and test set;

[0010] Step 5: Process the data using the STL time series decomposition method;

[0011] Step 6: Build the STL-TCN time series convolutional neural network model;

[0012] Step 7: Train the model and evaluate its prediction accuracy.

[0013] Preferably, the actual dynamic production data in step 1 is the output data and dynamic production data within any time period.

[0014] Preferably, in step 2, Pearson correlation analysis is used to screen out factors that relatively have a greater impact on heavy oil production.

[0015] Preferably, in step 3, the production data is standardized and normalized, and the standardization scales the target variable x to the range of [0,1], and saves the scaling parameters x_min and x_max. The standardization method is:

[0016]

[0017] where x * is the standardized data value, x is the data of the data set, μ is the mean of the data set, and σ is the standard deviation of the data set;

[0018] Normalize the feature variable x, also scale it to the range [0,1], and save the scaling parameters x_min and x_max,

[0019] The normalization method is:

[0020]

[0021] where x * is the standardized data value, x is the data of the data set, and x min is the minimum value of the sample in the data set, x max is the maximum value of samples in the data set.

[0022] Preferably, the first 80% of the data set processed in step 3 is used as a training set, and the last 20% is used as a test set.

[0023] Preferably, in step 6, STL is used to decompose the time series data to obtain seasonality, trend and residual components;

[0024] STL is a nonparametric method for smoothing based on locally weighted regression:

[0025] Y t =T t +S t +R t

[0026] where Y t is the original time series value at time point t; T t It is the trend component at time point t, reflecting the changing trend of the time series; S t is the seasonal component at time point t, reflecting periodic and repetitive fluctuations; R t is the residual component at time point t, reflecting random fluctuations or noise that cannot be explained by trend or seasonality.

[0027] S t =LOESS(Yt ,seasonal window)

[0028] The seasonal window is the length of the local smoothing window, which is usually equal to the period length of the time series.

[0029] T t =LOESS(Y t ′,trend window)

[0030] where Y t ′ is the deseasonalized data; trend window is the window length used to extract long-term trends, which is usually longer than the seasonal window;

[0031] R t =Y t -T t -S t

[0032] The seasonal and trend components are retained and input into the TCN model as exogenous variables so that the model can utilize the information of these components and the original sequence at the same time, which can better reflect the seasonality and trend of oil production and has better robustness.

[0033] Preferably, in step 6, the number of model layers, optimizer type and activation function type in the encoder and decoder are set;

[0034] The encoder compresses the exogenous variables X and oil production Y into an internal representation that can be processed by the machine. It accepts the input of features and oil production, extracts features by dilated convolution and batch normalization, adds residual connections to prevent gradient disappearance or explosion, and outputs the input features and oil production summary through the activation function. The main parts of the encoder are:

[0035]

[0036] Where L t is the feature representation output by the encoder, which contains a summary of past features; d is the dilation factor; N is the size of the convolution kernel; k is the index of the convolution kernel position; K(k) is the value of the convolution kernel at the kth position;

[0037] The decoder accepts the potential output from the encoder and the features of the future moment, and passes through the dense layer and batch normalization and activation function ReLU, the exogenous variables and the potential output L of the encoder. t The residual function is used to combine and finally a fully connected layer is used to output the oil production prediction results at the future moment. The main parts of the decoder are:

[0038] θ t+1 =R(X t+1)+L t

[0039] Where R(X t ) Based on the exogenous variable X t The residual function, L t is the output of the encoder, θ t is the output of the decoder, and then θ is transformed into t Mapped to the final predicted value;

[0040] The decoder calculates the probability distribution based on the data characteristics and outputs the probability distribution with confidence intervals of 0.1-0.9 and 0.001-0.999 respectively;

[0041] The prediction framework is:

[0042]

[0043] where y 1:t is the historical observation value, y t+ω is the future target value, X t+ω is an exogenous variable in the future, and Ω is the time step of the forecast.

[0044] Preferably, in step 7, the processed training set is input into the STL-TCN model to obtain the trained STL-TCN model; the test set is input into the trained STL-TCN model to obtain the chemical flooding heavy oil production prediction data;

[0045] Use MSE, RMSE, MAE, and MAPE to evaluate the model and its prediction accuracy. The closer to 0, the better the model effect.

[0046]

[0047] Where n is the number of data points, y i is the true value, is the predicted value.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The use of TCN models for oil production forecasting has significant innovations and advantages in dealing with long-distance dependencies, parallel computing, model transparency, and flexibility. These features enable TCN to not only improve prediction accuracy in practical applications, but also reduce training time and computing costs, and provide more interpretable results, thereby providing strong support for production decisions in the oil industry.

[0050] (2) The STL time series decomposition method can help reveal the structural components in time series data. By decomposing the data, long-term trends and cyclical fluctuations can be more clearly understood, which helps to better understand the basic pattern of chemical flooding heavy oil production. Before modeling, decomposing the data through STL can help better prepare the data and help build a more accurate prediction model. Understanding the trend and seasonal components of the data can help choose a suitable model. STL can help extract these components from the time series, so that this information can be added to the model, improving the accuracy of the prediction and enhancing the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method for prediction of the present invention;

[0052] Figure 2 is the Pearson correlation coefficient heat map;

[0053] Figure 3 It is the STL decomposition result diagram of the cumulative oil production of the production wells;

[0054] Figure 4 It is a graph of predicted values ​​and true values ​​in the training set;

[0055] Figure 5 It is a forecast chart of the cumulative oil production of production wells. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The present invention provides a method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model, comprising the following steps:

[0058] Step 1: Obtain actual dynamic production data from the oil field block;

[0059] The data for the whole year of 2020 from January 1, 2020 to December 31, 2020 were selected from the daily production data and dynamic production data of the chemical flooding heavy oil field block from August 1, 2019 to the present. Including the time of injection wells and production wells, oil production, bottom hole pressure, oil production rate, water content, cumulative liquid production, porosity, permeability, saturation of chemical flooding heavy oil field blocks, and daily injection pressure and daily injection volume of production wells, the actual dynamic production data obtained are used as the initial sample data set of the neural network.

[0060] Step 2: Use Pearson correlation to test the correlation of influencing factors;

[0061] Pearson correlation analysis was used to process the time, oil production, bottom hole pressure, oil production rate, water content, cumulative liquid production, daily injection pressure and daily injection volume of injection wells and production wells. The results are as follows Figure 3 As shown in the figure, the bottom hole pressure of the production well (Kpa), the daily injection volume of the injection well (m3), the water content of the production well (%), and the cumulative liquid production of the production well (m3) are positively correlated with the oil production. The correlation coefficients of these parameters are close to 0.9 or even higher. The injection pressure of the injection well (Kpa) has almost no effect on the oil production, and the oil production rate of the production well (m 3 / d) has a correlation coefficient of -0.54. Although there is a certain negative correlation, the impact is relatively small. When the velocity increases, the cumulative oil production may decrease.

[0062] Step 3: Remove outliers from the initial sample data set in step 1 and standardize and normalize the data set;

[0063] The production data is standardized and normalized. Standardization scales the target variable x to the range [0,1].

[0064] Step 4: The data set is divided into training set and test set;

[0065] The first 80% of the processed data set in step 3 is used as the training set, and the last 20% is used as the test set.

[0066] Step 5: Process the data using the STL time series decomposition method;

[0067] Use STL (seasonal-trend decomposition with local weighted regression smoothing) to decompose the time series data to obtain seasonal, trend and residual components. The results are as follows: Figure 3 As shown in the figure, the overall trend of cumulative oil production over time is shown. From 2022, the growth rate tends to be stable. Although the cumulative oil production has increased overall in seasonality, there are still some cyclical fluctuations, especially in the early stage. The residual fluctuated greatly in early 2020 and then tended to be stable. The smaller the fluctuation, the better the model can fit the data and the smaller the error. The seasonal and trend components are retained and input into the TCN model as exogenous variables.

[0068] Step 6: Build the STL-TCN time series convolutional neural network model;

[0069] Set the number of model layers, optimizer type, and activation function type in the encoder and decoder.

[0070] Step 7: Train the model and evaluate its prediction accuracy.

[0071] Input the processed training set into the STL-TCN model to obtain the trained STL-TCN model. The fitting results are as follows: Figure 4 As shown; the test set is input into the trained STL-TCN model to obtain the chemical flooding heavy oil production prediction data. The results are as follows Figure 5 As shown; the model is evaluated by MSE, RMSE, MAE, and MAPE, and the results are shown in Table 1.

[0072] In this embodiment, the actual dynamic production data in step 1 is the output data and dynamic production data within any time period.

[0073] In this embodiment, Pearson correlation analysis is used in step 2 to screen out factors that relatively have a greater impact on heavy oil production.

[0074] In this embodiment, in step 3, the production data is standardized and normalized. The standardization scales the target variable x to the range of [0, 1] and saves the scaling parameters x_min and x_max. The standardization method is:

[0075]

[0076] where x * is the standardized data value, x is the data of the data set, μ is the mean of the data set, and σ is the standard deviation of the data set;

[0077] Normalize the feature variable x, also scale it to the range [0,1], and save the scaling parameters x_min and x_max,

[0078] The normalization method is:

[0079]

[0080] where x * is the standardized data value, x is the data of the data set, and x min is the minimum value of the sample in the data set, x max is the maximum value of samples in the data set.

[0081] Preferably, the first 80% of the data set processed in step 3 is used as a training set, and the last 20% is used as a test set.

[0082] Preferably, in step 6, STL is used to decompose the time series data to obtain seasonality, trend and residual components;

[0083] STL is a nonparametric method for smoothing based on locally weighted regression:

[0084] Y t =T t +St +R t

[0085] where Y t is the original time series value at time point t; T t It is the trend component at time point t, reflecting the changing trend of the time series; S t is the seasonal component at time point t, reflecting periodic and repetitive fluctuations; R t is the residual component at time point t, reflecting random fluctuations or noise that cannot be explained by trend or seasonality.

[0086] S t =LOESS(Y t ,seasonal window)

[0087] The seasonal window is the length of the local smoothing window, which is usually equal to the period length of the time series.

[0088] Tt=LOESS(Yt′,trend window)

[0089] where Y t ′ is the deseasonalized data; trend window is the window length used to extract long-term trends, which is usually longer than the seasonal window;

[0090] R t =Y t -T t -S t

[0091] The seasonal and trend components are retained and input into the TCN model as exogenous variables so that the model can utilize the information of these components and the original sequence at the same time, which can better reflect the seasonality and trend of oil production and has better robustness.

[0092] In this embodiment, the number of model layers, optimizer type and activation function type in the encoder and decoder are set in step 6;

[0093] The encoder compresses the exogenous variables X and oil production Y into an internal representation that can be processed by the machine. It accepts the input of features and oil production, extracts features by dilated convolution and batch normalization, adds residual connections to prevent gradient disappearance or explosion, and outputs the input features and oil production summary through the activation function. The main parts of the encoder are:

[0094]

[0095] Where L tis the feature representation output by the encoder, which contains a summary of past features; d is the dilation factor; N is the size of the convolution kernel; k is the index of the convolution kernel position; K(k) is the value of the convolution kernel at the kth position;

[0096] The decoder accepts the potential output from the encoder and the features of the future moment, and passes through the dense layer and batch normalization and activation function ReLU, the exogenous variables and the potential output L of the encoder. t The residual function is used to combine and finally a fully connected layer is used to output the oil production prediction results at the future moment. The main parts of the decoder are:

[0097] θ t+1 =R(X t+1 )+L t

[0098] Where R(X t ) Based on the exogenous variable X t The residual function, L t is the output of the encoder, θ t is the output of the decoder, and then θ is transformed into t Mapped to the final predicted value;

[0099] The decoder calculates the probability distribution based on the data characteristics and outputs the probability distribution with confidence intervals of 0.1-0.9 and 0.001-0.999 respectively;

[0100] The prediction framework is:

[0101]

[0102] where y 1:t is the historical observation value, y t+ω is the future target value, X t+ω is an exogenous variable in the future, and Ω is the time step of the forecast.

[0103] In this embodiment, in step 7, the processed training set is input into the STL-TCN model to obtain the trained STL-TCN model; the test set is input into the trained STL-TCN model to obtain the chemical flooding heavy oil production prediction data;

[0104] Use MSE, RMSE, MAE, and MAPE to evaluate the model and its prediction accuracy. The closer to 0, the better the model effect.

[0105]

[0106] Where n is the number of data points, y i is the true value, is the predicted value.

[0107]

[0108]

[0109] Table 1 Model prediction accuracy

[0110] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model, characterized in that: The following steps are involved: Step 1: Obtain actual dynamic production data from the oil field block; Step 2: Use Pearson correlation to test the correlation of influencing factors; Step 3: Remove outliers from the initial sample data set in step 1 and standardize and normalize the data set; Step 4: The data set is divided into training set and test set; Step 5: Process the data using the STL time series decomposition method; Step 6: Build the STL-TCN time series convolutional neural network model; Step 7: Train the model and evaluate its prediction accuracy.

2. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: The actual dynamic production data in step 1 are the output data and dynamic production data within any time period.

3. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: In step 2, Pearson correlation analysis is used to screen out the factors that relatively have a greater impact on heavy oil production.

4. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: In step 3, the production data is standardized and normalized. Standardization scales the target variable x to the range [0,1] and saves the scaling parameters x_min and x_max. The standardization method is: where x * is the standardized data value, x is the data of the data set, μ is the mean of the data set, and σ is the standard deviation of the data set; Normalize the feature variable x, also scale it to the range [0,1], and save the scaling parameters x_min and x_max, The normalization method is: where x * is the standardized data value, x is the data of the data set, and x min is the minimum value of the sample in the data set, x max is the maximum value of samples in the data set.

5. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: The first 80% of the processed data set in step 3 is used as the training set, and the last 20% is used as the test set.

6. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: In step 6, STL is used to decompose the time series data to obtain seasonal, trend and residual components; STL is a nonparametric method for smoothing based on locally weighted regression: Y t =T t +S t +R t where Y t is the original time series value at time point t; T t It is the trend component at time point t, reflecting the changing trend of the time series; S t is the seasonal component at time point t, reflecting periodic and repetitive fluctuations; R t is the residual component at time point t, reflecting random fluctuations or noise that cannot be explained by trend or seasonality. S t =LOESS(Y t ,seasonal window) The seasonal window is the length of the local smoothing window, which is usually equal to the period length of the time series. T t =LOESS(Y t ′,trend window) where Y t ′ is the deseasonalized data; trend window is the window length used to extract long-term trends, which is usually longer than the seasonal window; R t =Y t -T t -S t The seasonal and trend components are retained and input into the TCN model as exogenous variables so that the model can utilize the information of these components and the original sequence at the same time, which can better reflect the seasonality and trend of oil production and has better robustness.

7. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: In step 6, set the number of model layers, optimizer type, and activation function type in the encoder and decoder; The encoder compresses the exogenous variables X and oil production Y into an internal representation that can be processed by the machine. It accepts the input of features and oil production, extracts features by dilated convolution and batch normalization, adds residual connections to prevent gradient disappearance or explosion, and outputs the input features and oil production summary through the activation function. The main parts of the encoder are: Where L t is the feature representation output by the encoder, which contains a summary of past features; d is the dilation factor; N is the size of the convolution kernel; k is the index of the convolution kernel position; K(k) is the value of the convolution kernel at the kth position; The decoder accepts the potential output from the encoder and the features of the future moment, and passes through the dense layer and batch normalization and activation function ReLU, the exogenous variables and the potential output L of the encoder. t The residual function is used to combine and finally a fully connected layer is used to output the oil production prediction results at the future moment. The main parts of the decoder are: i t+1 =R(X t+1 )+L t Where R(X t ) Based on the exogenous variable X t The residual function, L t is the output of the encoder, θ t is the output of the decoder, and then θ is transformed into t Mapped to the final predicted value; The decoder calculates the probability distribution based on the data characteristics and outputs the probability distribution with confidence intervals of 0.1-0.9 and 0.001-0.999 respectively; The prediction framework is: where y 1:t is the historical observation value, y t+ω is the future target value, X t+ω is an exogenous variable in the future, and Ω is the time step of the forecast.

8. The method for predicting the production of heavy oil by chemical flooding based on the STL-TCN model according to claim 1, characterized in that: In step 7, the processed training set is input into the STL-TCN model to obtain the trained STL-TCN model; the test set is input into the trained STL-TCN model to obtain the chemical flooding heavy oil production prediction data; Use MSE, RMSE, MAE, and MAPE to evaluate the model and its prediction accuracy. The closer to 0, the better the model effect. Where n is the number of data points, y i is the true value, is the predicted value.

Images