A new well dynamic production prediction method based on static oil and gas field data

Abstract

The present invention proposes a method for predicting the dynamic production of new wells based on static oil and gas field data. The method includes: summarizing and preprocessing the oil and gas well data; training a BP neural network cumulative production prediction model and optimizing the model's hyperparameters; training a BP-LSTM monthly production prediction model and optimizing the model's hyperparameters; and obtaining predicted values for the monthly production data through denormalization. This model utilizes the strong nonlinear fitting capabilities of neural network models and the strong predictive capabilities of LSTM neural networks for time series data to solve the problem of making nonlinear predictions of the dynamic monthly production data of new wells when only static oil and gas field data is known. The model is suitable for rapid estimation of production trends during the development of new wells and is particularly suitable for use by mine engineering and technical personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas production, and more particularly to a method for predicting the dynamic production of new wells based on static oil and gas field data. Background Art

[0002] Many established dynamic and static production prediction methods have been adopted in the actual development of oil and gas fields. In predicting oil and gas well productivity, machine learning methods that have been widely used over the past decade include multivariate regression analysis, RF models, and neural networks. Other machine learning methods, such as support vector regression (SVM) and gradient boosting regression (GBM), have also been explored. Alatrach et al. selected a low-permeability carbonate reservoir in Abu Dhabi, UAE. They analyzed parameters such as well location coordinates and static information (TVD, initial saturation, etc.) into an ANN prediction model. They used oil, gas, and water production, reservoir pressure, and water saturation as output parameters, developed separate models, and conducted predictions. The results demonstrated that the neural network model exhibited good nonlinear fitting. To overcome the limitations of traditional neural networks in oil and gas production prediction, Chakra et al. applied a high-order neural network (HONN) to predict cumulative oil production. The resulting high-order neural network model overcomes the limitations of traditional methods and can represent both linear and nonlinear data. In order to solve the problem that conventional production forecasting methods cannot take into account the asynchrony and lag of time series influencing factors, Gu Jianwei and others proposed to use time series analysis methods to predict production. He combined time series with Kalman filters and established a differential autoregressive integrated moving average model of production in time series based on historical production data. After that, he combined the ARIMA model with the Kalman filter to construct a production forecasting model, and carried out application research of machine learning models on actual oil field data. Finally, he used data fitting and prediction test methods to evaluate the rationality of the algorithm and made predictions on production data. In the prior art, a method and system for rapid prediction of single-well production (202110715729.7) provides a method for predicting production using an LSTM model, specifically including: 1) obtaining the static parameters and dynamic parameters of each oil well respectively, and constructing corresponding LSTM models respectively; 2) performing sensitivity analysis on the network calculation parameters of each LSTM model; 3) establishing a correlation between the degree of daily oil production fluctuation and the optimal network calculation parameters, and forming a relationship chart between the degree of daily oil production fluctuation and the optimal network calculation parameters; 4) determining the optimal LSTM model; 5) obtaining the static parameters and dynamic parameters of the oil well to be predicted, and inputting them into the optimal LSTM model to obtain the production forecast value of the oil well to be predicted. Although this method also uses LSTM, it relies on complete historical oil and gas field exploitation data. Static data and dynamic parameters must be used to realize the production prediction of the oil well, which has limitations for the dynamic production prediction of new wells with only static data.

[0003] While traditional models have achieved some success in production forecasting, they all have their shortcomings, primarily in their scope of application, i.e., varying block characteristics, applicable to different oil and gas field production and development stages, and varying prediction requirements and process complexity. These existing technologies only consider the impact of either static oil and gas data or dynamic oil and gas field data on production capacity forecasting. Given only static oil and gas field data, these existing technologies are unable to predict the dynamic production of new wells. Summary of the Invention

[0004] The purpose of the present invention is to solve the technical problem that the above-mentioned existing technology cannot predict the dynamic production of new wells when only static oil and gas field data are known, and to provide a new well dynamic production prediction method based on static oil and gas field data.

[0005] The present invention is achieved through the following technical solution: a new well dynamic production prediction method based on static oil and gas field data, comprising the following steps:

[0006] Step S1: Summarization and preprocessing of oil and gas well static data;

[0007] Step S2: training the BP (back propagation) neural network cumulative yield prediction model and optimizing the model hyperparameters;

[0008] Step S3: training the BP-LSTM (long short-term memory network) monthly output forecasting model and optimizing the model hyperparameters;

[0009] Step S4: Obtain the predicted value of the new well's dynamic monthly production data through inverse normalization processing.

[0010] In a preferred embodiment of the present invention, in step S1, the static data of the oil and gas wells includes static oil and gas field data and cumulative production data of each well, and the oil and gas well data are dimensionlessly processed, and the calculation formula is:

[0011]

[0012] Where x is the sample data, dimensionless, μ is the mean of the sample data, dimensionless, σ is the standard deviation of the sample data, dimensionless, and x′ is the dimensionless data.

[0013] In a preferred embodiment of the present invention, in step S2, the training method of the BP neural network cumulative yield prediction model is: selecting relu as the intermediate activation function in the hidden layer; selecting sigmoid as the activation function in the output layer, and the sigmoid can map the output vector to between 0 and 1 for easy calculation and processing.

[0014] In a preferred embodiment of the present invention, in step S2, the activation function in the hidden layer is:

[0015]

[0016] Where x is the data calculated by the hidden layer of the neural network and is dimensionless.

[0017] In a preferred embodiment of the present invention, in step S2, the activation function in the output layer is:

[0018]

[0019] Where x is the data calculated by the output layer of the neural network and is dimensionless.

[0020] In a preferred embodiment of the present invention, in step S2, the training process of the BP neural network cumulative yield prediction model includes two steps: forward propagation and back propagation.

[0021] In a preferred embodiment of the present invention, in step S2, during the forward propagation process, for a neural network having L layers, the general calculation formula at the j-th neuron of the L-th layer is:

[0022]

[0023] Where: n l Indicates the number of neurons involved in the calculation of the L-th layer, It means that the output value of the L-1 layer, i.e. the previous layer, is used as the input value of the L layer. represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer, represents the bias coefficient of the jth neuron in the Lth layer, and f represents the activation function.

[0024] In a preferred embodiment of the present invention, in step S2, during the iterative process of back propagation, the stochastic gradient descent method is used to derive the loss function to update the weight and bias parameters. The weight update formula is as follows:

[0025]

[0026]

[0027] Among them: α is the learning rate, that is, the step factor when the gradient descends in a certain direction, represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer, represents the bias coefficient of the jth neuron in the Lth layer, J(L) represents the loss function of the Lth layer, It means to find the partial derivative of a term.

[0028] In a preferred embodiment of the present invention, in step S3, the input layer and hidden layer of the LSTM neural network are both LSTM layers, the LSTM layer is composed of LSTM storage units, the output layer is a fully connected layer of the neural network, and multiple LSTM layers constitute a DLSTM model; the BP-LSTM monthly output forecasting model is established using the keras package in Python.

[0029] Compared with the prior art, the present invention has the following advantages: (1) It provides a new well dynamic production prediction method based on static oil and gas field data. Compared with the traditional dynamic and static production prediction method, the use of machine learning methods to predict oil and gas field production dynamics has its inherent advantages. It does not require formula derivation and geological modeling. The data is directly derived from actual production data, which effectively avoids human interference and improves the efficiency of production capacity prediction. (2) The present invention is based on a neural network method, in which the neural network model has strong nonlinear fitting capabilities, which solves the problem of making nonlinear predictions of new well monthly production dynamic data when only static oil and gas field data are known. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of a new well dynamic production prediction method based on static oil and gas field data of the present invention;

[0031] Figure 2 This is a statistical chart of the time T that the oil well has been put into production according to an embodiment of the present invention;

[0032] Figure 3 This is a structural diagram of a BP neural network cumulative yield prediction model according to an embodiment of the present invention;

[0033] Figure 4 A comparison chart of the actual monthly production value of a certain well and the monthly production prediction value of the BP cumulative production prediction model in an embodiment of the present invention;

[0034] Figure 5 Schematic diagram of state updates between multiple LSTM layers at time k in an embodiment of the present invention;

[0035] Figure 6 This is a calculation flow chart of a single storage unit and a single time step in the LSTM layer of an embodiment of the present invention;

[0036] Figure 7 Comparison chart of the prediction results and actual values of the BP model and BP-LSTM model in the embodiment of the present invention DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0038] Figure 1 The flowchart of a new well dynamic production prediction method based on static oil and gas field data according to the present invention is shown.

[0039] Taking the M oilfield block as an example, there are 148 oil wells in the area, and production data records are kept from December 2015 to April 2020. The start time of production for each well is different, but the data records are all ended in April 2020. The specific implementation includes the following steps:

[0040] Step S1: Summarization and preprocessing of oil and gas well static data;

[0041] Statistics and analysis were conducted based on the time T (months) each well had been put into production. The data were dimensionless processed to facilitate subsequent modeling applications. Figure 2 It can be seen that the production time T of most oil wells is concentrated between half a year and three years, accounting for about 84.46%, indicating that this is a relatively newly developed block.

[0042] Block M contains 14 types of static oil and gas field data: average porosity, average permeability, average oil saturation (So), thickness percentage of Class III reservoirs, number of designed sections, number of designed clusters, cluster spacing, total sand volume, sand addition per cluster, sand addition intensity, fluid intensity, total fluid volume, horizontal section length, and fracturing fluid volume. Data were normalized using the normal distribution method, ultimately achieving a mean of zero and a variance of 1.

[0043] Step S2: training the BP neural network cumulative yield prediction model and optimizing the model hyperparameters;

[0044] The model's hyperparameters were optimized using a grid search method, and the random seed number was set to a fixed value and single-threaded mode was used to ensure the model's repeatability. The weight initialization method was all standard normal distribution initialization. During the BP neural network model training process, ReLU was selected as the intermediate activation function in the hidden layer. The calculation formula involved is:

[0045]

[0046] Sigmoid is selected as the activation function in the output layer. It can map the output vector to between 0 and 1 for easy calculation and processing. The calculation formula is:

[0047]

[0048] The training process of each BP neural network includes two steps: forward propagation and back propagation. During the forward propagation process, for a neural network with L layers, the general calculation formula at the jth neuron in the Lth layer is:

[0049]

[0050] Where: n l Indicates the number of neurons involved in the calculation of the L-th layer, It means that the output value of the L-1 layer, i.e. the previous layer, is used as the input value of the L layer. Represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer.

[0051] For a training set containing m data, the mean square error is used as its loss function. In order to prevent overfitting, the L2 norm is added. In summary, the loss function of the Lth layer is defined as:

[0052]

[0053] Among them: m represents the number of samples, n represents the dimension of the output value, represents the actual value of the kth dimension of the jth neuron in the Lth layer, represents the predicted value of the kth dimension of the jth neuron in the Lth layer, θ represents the matrix of weights or bias parameters, and λ is the adjustable parameter of the L2 norm.

[0054] In the iterative process of back propagation, the stochastic gradient descent method is used to derive the loss function to update the weight and bias parameters. The update formula is as follows:

[0055]

[0056]

[0057] Among them: α is the learning rate, that is, the step factor when the gradient descends in a certain direction, represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer, represents the bias coefficient of the jth neuron in the Lth layer, J (L) represents the loss function of the Lth layer, It means to find the partial derivative of a term.

[0058] The structure of the BP neural network cumulative output prediction model established in the embodiment is as follows: Figure 3As shown in Figure 2, the model input layer receives normalized data for 14 static factors, plus bias, for a total of 15 dimensions. The number of hidden layers and neurons in each layer is determined using a grid search method. The output layer outputs denormalized data for each well's cumulative production over the previous x months, ..., and the previous x + 5 months, totaling 6 dimensions. Wells with incomplete static factor data are deleted before data preprocessing. In this case, due to the small number of wells with production time of 36 months or more, five BP neural network cumulative production prediction models are established, with values of x = 1, 7, 13, 19, and 25 selected based on the output layer.

[0059] The data set was divided into a training set and a test set ratio of 8:2. The model was established and predicted based on the keras package in Python. The parameter information of the five BP neural network cumulative yield prediction models is shown in Table 1.

[0060] Table 1 Parameter information statistics of BP neural network cumulative production prediction model

[0061]

[0062] A typical well in Block M was selected. The cumulative production forecast for the typical well from the first 1 to 30 months was obtained. The monthly production change of the typical well in the first 30 months was obtained by difference calculation. Figure 4 .

[0063] Step S3: Train the BP-LSTM monthly output forecasting model and optimize the model hyperparameters;

[0064] The BP-LSTM monthly output forecasting model construction method in the embodiment is as follows: the input layer and hidden layer of the LSTM neural network are both LSTM layers, which are composed of LSTM storage units, and the output layer is a fully connected layer of the neural network. Figure 5 The figure below is a schematic diagram of the state update of multiple LSTM layers at time k. This is a horizontal division of the LSTM neural network model (different layers at the same time), where multiple LSTM layers constitute the DLSTM model.

[0065] At time k, input data x k and the hidden state of the LSTM layer 1 storage unit at the previous time step At the same time, it is input into the storage unit of LSTM layer 1, and the hidden state of the current storage unit of LSTM layer 1 is calculated and output. On the one hand, The storage unit of LSTM layer 1 is input to the next time step. On the other hand, the hidden state and the hidden state of the LSTM layer 2 storage unit at the previous time step At the same time, it is input into the storage unit of LSTM layer 2, and the hidden state of the current storage unit of LSTM layer 2 is calculated and output. and On the one hand, it will be input into the storage unit of LSTM layer 2 at the next time step, and on the other hand, it will be combined with the hidden state of LSTM layer 3 storage unit at the previous time step. At the same time, it is input into the storage unit of LSTM layer 3. And so on until the data is input into the storage unit of the last LSTM layer.

[0066] The most critical thing in a single LSTM layer is the state of the storage unit, which runs along the entire LSTM structure and is regulated by three gate structures. This belongs to the vertical division of the LSTM neural network model (the same layer at different times), such as Figure 6 Shown in each A gate structure processing is performed at each location, and the adjustment process of the state at time k is completed by the following three steps:

[0067] Step 1: The forget gate is based on the current input x k and the output h at the previous moment k-1 The formula for deciding which part of the memory needs to be forgotten, the input value z and the forget gate F are as follows:

[0068] z=tanh(w z [h k-1 ,x k ])

[0069] F=sigmoid(w F [h k-1 ,x k ])

[0070] Step 2: Input gate according to x k and h k-1 Decide what information to add to state c k-1 Generate a new state c k , which can be divided into two steps. First, the input gate decides which values to update, and then creates a new state c k , input gate I and new state c k The formula is as follows:

[0071] I=sigmoid(w I [h k-1 ,x k ])

[0072] c k =Fc k-1 +Iz

[0073] Step 3: Get the new state c in LSTMk After that, the output gate generates the output h at the current moment k , output gate O and current output h k The formula is:

[0074] O=sigmoid(w O [h k-1 ,x k ])

[0075] h k =Otanh(c k )

[0076] Among them, w F 、w I 、w O 、w z are all weight matrices.

[0077] A BP-LSTM monthly production prediction model was established using the monthly production data obtained from the BP neural network cumulative production prediction model. The prediction idea is a one-step sliding window prediction, that is, the data from the previous t moments are used to predict the data at the t+1 moment (t=4 in this embodiment). The dynamic monthly production data is first converted into the format of supervised data, and then processed using the maximum and minimum normalization method (the formula is as follows). After being divided into a training set: test set ratio of 7:3, it is input into the model, and finally the predicted value of the monthly production data is obtained after denormalization.

[0078] Step S4: Obtain the predicted value of the new well's dynamic monthly production data through inverse normalization processing.

[0079] The denormalization formula is:

[0080]

[0081] The model was built using the keras package in Python. The random seed number was set to a fixed value and single-threaded to ensure model reproducibility. The Adam method was selected for weight parameter optimization. The activation function for the LSTM layer was the tanh function, and the output layer function was the sigmoid function. A grid search method was used to optimize hyperparameters. The number of LSTM layers was set to 2, the number of neurons per layer was 6, and the dropout ratio was 0.4. Figure 7The results show that the BP-LSTM monthly production forecasting model can roughly predict the changing trend of monthly production, and the relative error of the model test set is smaller than the relative error of the prediction results of the BP cumulative production forecasting model, which shows that the prediction performance of the BP-LSTM model is better than that of the BP model. This is because the BP-LSTM method not only considers the relationship between static data and dynamic data, but also the relationship between dynamic data. Therefore, the present invention can make nonlinear predictions of the dynamic monthly production data of new wells when only static oil and gas field data is known. It is suitable for rapid estimation of production trends during the development of new wells and is particularly suitable for use by mine engineering and technical personnel.

[0082] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technology and method of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A new well dynamic production prediction method based on static oil and gas field data, characterized in that: The following steps are involved: Step S1: Summarization and preprocessing of oil and gas well static data; Step S2: training the BP neural network cumulative yield prediction model and optimizing the model hyperparameters; Step S3: Train the BP-LSTM monthly output forecasting model and optimize the model hyperparameters; Step S4: Obtain the predicted value of the new well's dynamic monthly production data through denormalization processing; In step S2, the training method of the BP neural network cumulative yield prediction model is as follows: Relu is selected as the intermediate activation function in the hidden layer; Sigmoid is selected as the activation function in the output layer. Sigmoid can map the output vector to between 0 and 1 for easy calculation and processing; In step S2, the training process of the BP neural network cumulative yield prediction model includes two steps: forward propagation and back propagation; In step S2, during the forward propagation process, for a neural network having L layers, the general calculation formula at the j-th neuron in the L-th layer is: Where: n l Indicates the number of neurons involved in the calculation of the L-th layer, It means that the output value of the L-1 layer, i.e. the previous layer, is used as the input value of the L layer. represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer, represents the bias coefficient of the jth neuron in the Lth layer, and f represents the activation function; In step S2, during the iterative process of back propagation, the stochastic gradient descent method is used to derive the loss function to update the weight and bias parameters. The weight update formula is as follows: Among them: α is the learning rate, that is, the step factor when the gradient descends in a certain direction, represents the weight value between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer, represents the bias coefficient of the jth neuron in the Lth layer, J(L) represents the loss function of the Lth layer, It means to find the partial derivative of a term.

2. The method for predicting dynamic production of new wells based on static oil and gas field data according to claim 1, characterized in that: In step S1, the static data of oil and gas wells includes static oil and gas field data and cumulative production data of each well, and the oil and gas well data are dimensionlessly processed, and the calculation formula is: Where x is the sample data, dimensionless, μ is the mean of the sample data, dimensionless, σ is the standard deviation of the sample data, dimensionless, and x′ is the dimensionless data.

3. The method for predicting dynamic production of new wells based on static oil and gas field data according to claim 2, characterized in that: In step S2, the activation function in the hidden layer is: Where x is the data calculated by the hidden layer of the neural network and is dimensionless.

4. The method for predicting dynamic production of new wells based on static oil and gas field data according to claim 3, characterized in that: In step S2, the activation function in the output layer is: Where x is the data calculated by the output layer of the neural network and is dimensionless.

5. The method for predicting dynamic production of new wells based on static oil and gas field data according to claim 4, characterized in that: In step S3, the input layer and hidden layer of the LSTM neural network are both LSTM layers, the LSTM layer is composed of LSTM storage units, the output layer is a fully connected layer of the neural network, and multiple LSTM layers constitute a DLSTM model; the BP-LSTM monthly output forecasting model is established using the keras package in Python.

Images