A neural additive model-based robust prediction method and system for oilfield production

Abstract

The application discloses a kind of oilfield production robust prediction method and system based on neural additive model, it is related to petroleum well production prediction technical field, including: based on neural additive model SMART, by obtaining the input data consisting of multidimensional time series data, model training is carried out, and prediction model is constructed;Based on prediction model, using sparse learning strategy, mode-based measurement method and non-convex optimization algorithm, model optimization is carried out, and the production of oilfield is predicted according to the optimized prediction model.The application combines neural network and additive model, introduces mode-based measurement, sparse learning and non-convex optimization algorithm, effectively improves the accuracy, robustness, interpretability and efficiency of oilfield production prediction method, so that it is more suitable for application in actual scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil well production prediction, in particular to an oilfield production robust prediction method and system based on neural additive model. BACKGROUND

[0002] In the prior art, oilfield production prediction methods are mainly divided into linear model methods and deep learning model methods. Linear models (such as AR, ARIMA) have good interpretability, but perform poorly in handling high-dimensional data and complex nonlinear relationships. Deep learning models (such as RNN, LSTM, Transformer) can capture complex nonlinear relationships, but lack interpretability. Neural additive models (NAMs) combine the expressive power of neural networks and the interpretability of additive models, but lack robustness in handling noise and outliers in oilfield data.

[0003] Oil well production prediction is of great significance for intelligent monitoring of oil wells and improving oil recovery. Linear models cannot handle high-dimensional data and complex nonlinear relationships. Deep learning models lack interpretability and are difficult to explain the contribution of each feature to the prediction result. Neural additive models (NAMs) perform poorly in handling noise and outliers, lacking robustness.

[0004] Oil well production prediction is of great significance for intelligent monitoring of oil wells and improving oil recovery. Oilfield data has the characteristics of high dimensionality, complex nonlinear relationship, noise and outlier interference, and dynamic time series. In order to overcome the shortcomings of existing methods in oil well production prediction, the present application proposes an oilfield production robust prediction method based on neural additive model (SMART), which has the ability to handle high-dimensional data and complex nonlinear relationships, and is robust to data with noise and outliers. SUMMARY

[0005] In order to solve the above problems, the purpose of the present application is to provide an oilfield production robust prediction technology based on neural additive model to overcome the shortcomings of existing methods in oil well production prediction.

[0006] In order to achieve the above technical purpose, the present application provides an oilfield production robust prediction method based on neural additive model, comprising the following steps:

[0007] Based on the neural additive model SMART, by acquiring input data composed of multi-dimensional time series data, model training is performed to construct a prediction model;

[0008] Based on the prediction model, a sparse learning strategy, a mode-based measurement method and a non-convex optimization algorithm ProxSAGA are used for model optimization, and the oilfield production is predicted according to the optimized prediction model.

[0009] Preferably, in the process of obtaining input data, based on the production characteristics and development history data of the water drive oilfield, the characteristics for oilfield production prediction are screened, and the correlation degree between the input variables and the output variables is measured by the Pearson correlation coefficient, and the input data is obtained according to the measurement result, wherein the input data includes daily oil production, daily water production, cumulative oil production, water cut, flow pressure, cumulative liquid production and permeability.

[0010] Preferably, in the process of constructing the neural additive model SMART, the output values of all basis functions are summed and a bias term is added to construct the neural additive model SMART.

[0011] Preferably, in the process of model optimization, a mode-based measurement method is used to evaluate the difference between the model predicted value and the actual value, and the influence of outliers is reduced by minimizing the deviation between the predicted value and the conditional mode.

[0012] Preferably, in the process of performing mode measurement, based on the mode regression problem, the conditional mode function is obtained, and the mode regression risk is introduced, and the kernel density estimation is used to realize the mode measurement.

[0013] Preferably, in the process of model optimization, by jointly sparsely regularizing the weights of the neural network corresponding to each basis function, the relevant features are selected, and the influence of redundant features is reduced.

[0014] Preferably, in the process of model optimization, the non-convex optimization algorithm ProxSAGA is used, which introduces a double-mini-batch gradient estimation mechanism and combines with a proximal operator to solve the non-convex optimization problem and improve the optimization efficiency and convergence speed of the model.

[0015] Preferably, in the process of model optimization, the non-convex optimization algorithm ProxSAGA includes the following steps:

[0016] Initialize the model parameters, calculate the initial gradient estimate, determine the step size μ, the regularization coefficient λ and the mini-batch size b;

[0017] Randomly select two mini-batch sample index sets for each iteration, each set containing b samples;

[0018] Subtract the corresponding historical gradient from the gradient calculated by the mini-batch sample, and add the global gradient estimate to calculate the modified gradient direction;

[0019] First, gradient descent is performed, and then the proximal operator is used to process the regularization term to obtain new parameters;

[0020] The historical parameters of the mini-batch sample are replaced by the current parameters, and the historical parameters of the remaining samples remain unchanged;

[0021] scaling the global gradient estimate by a ratio;

[0022] Finally, the parameters after iteration are returned as the model result.

[0023] The application discloses an oilfield production robust prediction system based on a neural additive model, comprising:

[0024] A model construction module is configured to perform model training and construct a prediction model by acquiring input data composed of multidimensional time series data based on a neural additive model SMART.

[0025] A production prediction module is configured to perform model optimization by adopting a sparse learning strategy, a mode-based measurement method and a non-convex optimization algorithm ProxSAGA based on the prediction model, and predict the oilfield production according to the optimized prediction model.

[0026] The application discloses the following technical effects:

[0027] The application adopts the mode-based measurement method to effectively handle the influence of non-Gaussian noise and abnormal values.

[0028] The application has the ability of explainability and processing high-dimensional data and complex nonlinear relationships, and is robust to data with noise and abnormal values.

[0029] The application combines neural networks and additive models, introduces mode-based measurement, sparse learning and non-convex optimization algorithms, effectively improves the accuracy, robustness, explainability and efficiency of the oilfield production prediction method, and makes it more suitable for practical scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 is a prediction method flowchart according to the application;

[0032] Figure 2 is a SMART model structure diagram according to the application;

[0033] Figure 3 is a structure diagram of a submodel in the SMART model according to the application;

[0034] Figure 4 is a non-convex optimization algorithm ProxSAGA flowchart according to the application;

[0035] Figure 5 This is the feature correlation quantitative analysis described in this invention;

[0036] Figure 6 This is the 2020 oil well production curve described in this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0038] like Figures 1-6 As shown, this invention provides a robust oilfield production prediction method based on a neural additive model, such as... Figure 1 As shown, it includes the following steps:

[0039] Step 1: Data Preprocessing

[0040] Based on the production characteristics of water-driven oilfields and historical development data, 15 characteristics were qualitatively selected: effective thickness, permeability, well spacing, daily oil production, daily water production, water cut, cumulative oil production, cumulative fluid production, cumulative water production, recovery rate, static pressure, flowing pressure, pump depth, stroke, and stroke frequency. The Pearson correlation coefficient was used to measure the correlation between two variables X and Y, with values ​​ranging from -1 to 1. The larger the absolute value of the correlation coefficient, the stronger the correlation. Absolute values ​​between 0.8 and 1.0 indicate extremely strong correlation; absolute values ​​between 0.6 and 0.8 indicate strong correlation. Based on the qualitative analysis results, the Pearson correlation coefficient was used for quantitative analysis. Features with an absolute correlation value greater than 0.6 were used as model inputs. The model input features were determined to be daily oil production, daily water production, cumulative oil production, water cut, flowing pressure, cumulative fluid production, and permeability. Figure 5 As shown.

[0041] Data from January to December 2020 was selected as the validation set, data from January to December 2021 was selected as the test set, and all other training samples were used as the training set. To ensure the stability of the model training and yield prediction results, a normalization method was adopted to map the data to the [0,1] interval. The input data is multidimensional time series data, containing K=7 variables and N time steps.

[0042] The data is divided into input-output pairs using a sliding window approach, with the input time series length L = 6, so the input is the characteristic value of K variables at L time steps, and the output is the target variable value at the L+1 time step.

[0043] The multivariate time series dataset is divided into input X and output Y, where input X contains the "input" of all input-output pairs, and output Y contains the "output" of all input-output pairs.

[0044] Step 2: Model construction:

[0045] For each target variable, a prediction model based on a neural additive model is constructed, as shown in Figure 2 The model is composed of multiple basis functions, each of which is constructed by a neural network, each neural network corresponding to an input variable and an output variable, and the output of each neural network corresponds to the contribution of the characteristic value of the input variable at the current time step to the output.

[0046] That is, the output values of all basis functions are summed up and a bias term is added, and finally the predicted value of the target variable at the current time step is obtained (as shown in Figure 6 The SMART expression of the model is as follows:

[0047]

[0048] where b is the bias term, x is the input, y is the predicted value of the output, P = L x K is the number of input characteristic values, and f is the basis function.

[0049] Step 3: Median-based metric:

[0050] A median-based metric method is used to evaluate the difference between the model prediction value and the actual value, which reduces the influence of outliers by minimizing the deviation between the prediction value and the conditional median. Traditional mean-based loss functions are easily affected by outliers, leading to deviation of model prediction results from true values. The median, as the value with the highest frequency in the dataset, can more accurately reflect the central tendency of the data, thereby reducing the influence of outliers.

[0051] 1. Definition of median regression problem

[0052] Assume that the generation process of the sample can be represented as:

[0053] Y t = f * (x t )+ò t

[0054] where the true generating function f* Reflecting the deterministic law between oilfield production and production parameters, Y is the target variable, x is the input feature, t is the time index variable, and ∈ is the noise term. Assuming that for any given x, the mode of the conditional noise ∈ is zero, as follows:

[0055]

[0056] where X is the input random vector, x is the specific value of the input vector, argmax is the peak point of the density function, p ò|X is the conditional probability density function of noise.

[0057] 2. Definition of the conditional mode function:

[0058] The conditional mode function is defined as:

[0059]

[0060] where p Y|X is the conditional density function of Y given X.

[0061] 3. Introduction of the mode regression risk:

[0062] In order to measure the fitting degree of the conditional density of the model f and the true mode regression function f * (x), the invention introduces the mode regression risk R(f), defined as:

[0063]

[0064] where, is the input feature space p Y|X is the conditional density value of the model f given X=x, ρ X is the distribution of the input feature X.

[0065] 4. Kernel density estimation to realize mode measurement:

[0066] The realization of mode measurement relies on the kernel density estimation (KDE) technology. The invention defines a kernel function K, which satisfies the following conditions:

[0067]

[0068] where φ(u) is a symmetric non-negative function, satisfying φ(u)=φ(-u) and φ(u)<φ(0); σ>0 is the bandwidth parameter, controlling the smoothing degree of the kernel function. Common kernel functions include Gaussian kernel, Epanechnikov kernel and triangular kernel, u1 and u2 represent two points in the input space.

[0069] For any measurable function f, the present invention can define an empirical kernel density estimate as:

[0070]

[0071] where, is the data set, N is the total length of the time series; L is the length of the sliding window; y i is the true value at the i-th time step; f(x i ) is the predicted value of the model at the i-th time step. Through kernel density estimation, the present invention can efficiently estimate the unknown probability density function, especially in the case of high-dimensional input features.

[0072] 5. Mode-based metric:

[0073] During the model training process, the present invention uses the mode metric to replace the traditional metric method, as follows:

[0074]

[0075] where: φ is the kernel function, σ is the bandwidth parameter, y i is the true value at the i-th time step, f(x i ) is the predicted value of the model at the i-th time step.

[0076] By maximizing the predicted value of the model will be as close as possible to the conditional mode, thereby reducing the impact of outliers on the model.

[0077] Step five: sparse regularization:

[0078] In high-dimensional time series prediction, feature selection is an important challenge. In order to improve the efficiency and interpretability of the model, the present invention introduces sparse regularization, which selects relevant features and reduces the impact of redundant features by jointly sparsifying the weights of the neural network corresponding to each basis function.

[0079] The sparse regularization term Ω(f) is defined as:

[0080]

[0081] where: P is the number of time series variables, L is the length of the sliding window, W j is the parameter of the j-th feature corresponding subnetwork.

[0082] In the SMART proposed by the present invention, the optimization problem can be represented as:

[0083]

[0084] where: λ is the regularization parameter, controlling the strength of sparse regularization.

[0085] Step six: Non-convex optimization algorithm ProxSAGA:

[0086] In the SMART model, the objective function is non-convex and non-smooth, and the traditional gradient descent method is difficult to optimize directly. In order to solve this problem, the invention adopts the non-convex optimization algorithm ProxSAGA, as shown in Figure 4 Introducing a double mini-batch gradient estimation mechanism, combined with a proximal operator, solving non-convex optimization problems, and improving the optimization efficiency and convergence speed of the model. The following is a detailed description of the non-convex optimization algorithm ProxSAGA, combined with the formula.

[0087] 1. Input training set Initialize parameters (The historical gradient estimate of all samples is initialized as the initial weight), step size μ>0, regularization coefficient λ, batch size b, initial global gradient estimate

[0088] 2. Main loop step (s=0,1,...,S-1):

[0089] Step 1: Random batch sampling, select two independent batches I s , J s ;

[0090] Step 2: Calculate random gradient estimate

[0091] Step 3: Proximal update parameters Where the proximal operator is defined as If Otherwise Where M is the number of network layers of the subnetwork;

[0092] Step 4: Update historical parameters For j∈J s , For Only update the batch J s Corresponding historical parameters The rest remain unchanged, where Indicates the historical parameters of the network corresponding to the (s+1)th sample and the jth feature;

[0093] Step 5: Update global gradient estimate

[0094]

[0095] 3. Algorithm output parameters of the Sth iteration:

[0096] The present application introduces conditional mode regression, reduces the influence of outliers on the model by minimizing the deviation between the predicted value and the conditional mode; significantly improves the robustness of the model in the non-Gaussian noise environment, and avoids the interference of outliers on the prediction result.

[0097] The present application jointly sparsely regularizes the overall weight of the neural network corresponding to each basis function, selects relevant features, and reduces the influence of redundant features; improves the feature selection ability and interpretability of the model, and reduces the interference of irrelevant features on the prediction result.

[0098] The present application adopts the non-convex optimization algorithm ProxSAGA, solves the non-convex optimization problem by introducing a double-mini-batch gradient estimation mechanism and combining with the proximal operator, and significantly improves the calculation efficiency and optimization performance of the model, which is suitable for large-scale high-dimensional data sets;

[0099] The present application adopts the kernel density estimation technology to realize the mode measurement, and efficiently estimates the conditional density function; provides reliable theoretical support for mode regression in non-Gaussian noise environment.

[0100] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system) and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of the flows and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a machine that implements the function specified in the flowchart and / or block diagram. Figure 1 The function of one flow or multiple flows and / or blocks Figure 1 The device of the function of one block or multiple blocks.

[0101] In the description of the present application, it should be understood that the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0102] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for robust prediction of oilfield production based on a neural additive model, characterized in that, The method comprises the following steps: Based on the neural additive model SMART, the input data composed of multi-dimensional time series data is obtained, the model is trained, and the prediction model is constructed; Based on the prediction model, the sparse learning strategy, the mode-based measurement method and the non-convex optimization algorithm ProxSAGA are used for model optimization, and the oilfield production is predicted according to the optimized prediction model; In the process of obtaining input data, based on the production characteristics and development history data of water drive oilfield, the characteristics used for oilfield production prediction are screened, and the correlation between input variables and output variables is measured by Pearson correlation coefficient, and the input data is obtained according to the measurement result, wherein the input data includes daily oil production, daily water production, cumulative oil production, water cut, flow pressure, cumulative liquid production and permeability; In the process of constructing the neural additive model SMART, the output values of all basis functions are summed and a bias term is added to construct the neural additive model SMART; When the model is optimized, the mode-based measurement method is used to evaluate the difference between the model prediction value and the actual value, which is used to reduce the influence of abnormal values by minimizing the deviation between the prediction value and the conditional mode; When the mode measurement is performed, the conditional mode function is obtained based on the mode regression problem, and the mode regression risk is introduced, and the kernel density estimation is used to realize the mode measurement; When the model is optimized, the joint sparse regularization is performed on the weights of the neural network corresponding to each basis function, the relevant features are selected, and the influence of redundant features is reduced; When the model is optimized, the non-convex optimization algorithm ProxSAGA is used, the double-mini-batch gradient estimation mechanism is introduced, and the proximal operator is combined to solve the non-convex optimization problem, and the optimization efficiency and convergence speed of the model are improved; When the model is optimized, the non-convex optimization algorithm ProxSAGA comprises the following steps: Initialize the model parameters, calculate the initial gradient estimation, determine the step size μ, the regularization coefficient λ and the mini-batch size b; Randomly select two mini-batch sample index sets each containing b samples in each iteration; The gradient calculated by the mini-batch sample is subtracted from the corresponding historical gradient, and then the global gradient estimation is added to calculate the modified gradient direction; First, gradient descent is performed, and then the proximal operator is used to process the regularization term to obtain new parameters; The historical parameters of the mini-batch sample are updated to the current parameters, and the historical parameters of the remaining samples remain unchanged; Adjust the global gradient estimation in proportion; Finally, the parameters after iteration are returned as the model result.

2. A system for implementing the method of robust prediction of oilfield production based on a neural additive model according to claim 1, characterized in that, It comprises: The model construction module is used for obtaining the input data composed of multi-dimensional time series data based on the neural additive model SMART, training the model, and constructing the prediction model; The production prediction module is used for obtaining the input data composed of multi-dimensional time series data based on the neural additive model SMART, training the model, and constructing the prediction model.

Images