Oilfield production monitoring method using carbon quantum dot tracer

Abstract

This invention provides a method for monitoring oilfield productivity using carbon quantum dot tracers, comprising: step S100, obtaining the convection and diffusion characteristics of the oilfield by injecting carbon quantum dot tracers and inert tracers, and establishing an oilfield monitoring feature set; step S200, selecting the main features for oilfield monitoring from the oilfield monitoring feature set using Spearman correlation coefficient and random forest model; and step S300, inputting the main features into a neural network to monitor oilfield productivity, wherein the neural network is a Wiener process optimized particle swarm optimization-backpropagation neural network.

Description

Technical Field

[0001] This invention relates to an oilfield monitoring technology, and more particularly to an oilfield production capacity monitoring method using carbon quantum dot tracers. Background Technology

[0002] Accurate production forecasting is crucial for the efficient and economical development of oil and gas reservoirs. Traditional oilfield production forecasting methods rely on empirical formulas and numerical simulations, which have limited influencing factors and make numerous ideal assumptions when forecasting production. However, many factors influence oilfield production, and their interactions are complex. Continuing to use traditional reservoir production forecasting methods may lead to significant calculation errors and potentially economic losses in oilfield development. Patent CN115977599A utilizes tracers for oilfield monitoring, but uses relatively few features, including well connectivity analysis results, tracer test results, pressure drop test results, and water absorption profile test results. Patent CN103958643A provides a carbon-based fluorescent tracer for reservoir detection, making carbon quantum dot tracer monitoring of oilfields possible. However, the oilfield features obtained through carbon quantum dot tracers are numerous and complex; not all features are suitable for monitoring a specific oilfield. Summary of the Invention

[0003] The purpose of this invention is to provide a method for monitoring oilfield productivity using carbon quantum dot tracers, comprising:

[0004] Step S100: Obtain the convection and diffusion characteristics of the oilfield by injecting carbon quantum dot tracers and inert tracers, and establish an oilfield monitoring feature set;

[0005] Step S200: Select the main features of oilfield monitoring from the oilfield monitoring feature set using the Spearman correlation coefficient and random forest model.

[0006] Step S300: Input the main features into the neural network to monitor the oilfield production capacity. The neural network is a Wiener process optimized particle swarm optimization-backpropagation neural network.

[0007] Furthermore, in step S100, the convection term features include {porosity, effective permeability, oil volume factor, oil saturation, reservoir pressure, clay volume, oil viscosity}; the diffusion term features include {flowback proppant volume, proppant ratio, fracture cross section, fracture spacing}.

[0008] Further, in step S100, the following parameters of the carbon quantum dot tracer and the inert tracer are recorded to obtain the detection feature set of the oilfield: the first breakthrough time and peak concentration time of the carbon quantum dot tracer, the first breakthrough time and peak concentration time of the inert tracer, the total amount of carbon quantum dot tracer injected and the total amount of flowback, the total amount of inert tracer injected and the total amount of flowback, the length of the main flow line between the injection and production wells of the inert tracer, and the flow cross-sectional area of ​​the carbon quantum dot tracer.

[0009] Furthermore, step S200 specifically includes the following processes:

[0010] Step S201: Evaluate the monotonic relationship between each feature and oilfield production capacity using Spearman correlation coefficient;

[0011] Step S202: Determine the importance of each feature to oilfield productivity using a random forest tree model;

[0012] Step S203: Select the main features for oilfield monitoring based on monotonicity and importance.

[0013] Furthermore, in step S203, features with strong positive correlation coefficients to Spearman and the top K features with the highest correlation coefficients in the random forest tree model are selected as the main features for oilfield monitoring.

[0014] Furthermore, in step S300, the Wiener process is used in the neural network to obtain the random perturbations d(f) of characteristic porosity, effective permeability, oil saturation, reservoir pressure, oil viscosity, flowback proppant volume, fracture cross-section, and fracture spacing.

[0015] d(f)=μ f dt+σ f dW f

[0016] Where f represents the feature, μ f For the deterministic drift term of feature f, σ f Let dW be the randomness parameter of feature f. f For Wiener increments.

[0017] Furthermore, in the neural network, the particles are updated using a random perturbation d(f).

[0018] X t+Δt =X t +d(f)

[0019] Among them, X t Let t be the characteristic parameter value at time t, and Δt be the time step.

[0020] Compared with the prior art, the present invention has the following advantages: (1) Carbon quantum dot tracers and inert tracers are used to obtain oilfield characteristics. Inert tracers can make up for the problem that the adsorption of carbon quantum dot tracers affects the accuracy of characteristics; (2) Spearman correlation coefficient and random forest are used to comprehensively evaluate all characteristics and select the most critical characteristics of the oilfield to be monitored, which reduces the computation speed of subsequent models; (3) Wiener process is used to optimize PSO. By using the random perturbation of Wiener process, the deadlock of PSO algorithm getting stuck in local optimum is broken, and the global search capability and adaptability of PSO algorithm are improved.

[0021] The present invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0023] Figure 2 Heatmap for Spearman correlation coefficient analysis.

[0024] Figure 3 This is a schematic diagram of the WP-PSO process.

[0025] Figure 4 This diagram illustrates a comparison of the prediction results between the WP-PSO-BP and PSO-BP models. Detailed Implementation

[0026] Combination Figure 1 This paper presents an oilfield monitoring method using carbon quantum dot tracers, primarily for monitoring oilfield production capacity. The method utilizes carbon quantum dot tracers to acquire oilfield characteristics, employs a random forest tree method to screen characteristic factors affecting the production capacity of a specific oilfield, and uses a Wiener-Particle Swarm Optimization algorithm-based neural network to establish a monitoring model. In oilfield development, reservoirs can be classified into homogeneous reservoirs and heterogeneous reservoirs based on their homogeneity. For homogeneous reservoirs, the error in detection using carbon quantum dot tracers originates from adsorption and is less likely to cause significant errors; however, for heterogeneous reservoirs, the carbon quantum dot tracer is also susceptible to interference from the heterogeneous reservoir itself. This embodiment primarily focuses on monitoring heterogeneous reservoir oilfields using carbon quantum dot tracers. The method involved in this embodiment includes the following steps:

[0027] Step S100: Collect oilfield characteristics using carbon quantum dot tracers to construct an oilfield monitoring feature set;

[0028] Step S200: Select the main features for oilfield monitoring from the oilfield monitoring feature set using Spearman correlation coefficient and random forest method;

[0029] Step S300: Input the characteristic factors required by the oilfield into the neural network for production capacity detection. The neural network is an oilfield monitoring model established using a backpropagation neural network with particle swarm optimization. During the model establishment process, the Wiener process is used to optimize the model.

[0030] Carbon quantum dot (CQD) tracers offer advantages such as high sensitivity and long-term monitoring capabilities for oilfield monitoring. While CQDs have relatively low adsorption capacity, they can be affected by adsorption from formation minerals in certain situations, potentially leading to variations in tracer concentration and impacting the accuracy of monitoring results. Therefore, in step S100, CQD tracers and inert tracers (e.g., NaBr solution) are added to the injection wells in the oilfield. Due to differences in molecular size, permeability, adsorption capacity, and diffusion capacity, inert tracers, under ideal conditions, only reflect the fluid movement itself. The breakthrough time of the inert tracer is purely the fluid transport time, and the retention degree of carbon quantum dots can be deduced from the time difference between the two. Especially in heterogeneous reservoir scenarios, using an inert tracer as a control group to calculate relevant oilfield characteristic parameters, compared to using a carbon quantum dot tracer for correction, improves the accuracy by approximately three times. The error source in the former is only measurement noise, while the latter's noise originates from the adsorption error of CQDs and interference from heterogeneous reservoirs.

[0031] The characteristics to be obtained in step S100 include porosity (P), effective permeability (EP), and oil volume factor (VF). O ), oil saturation (S) O ), reservoir pressure (P) R ), clay volume (V) C ), oil viscosity (V) O ), volume of proppant returned (V) P ), proppant ratio (R) P ), crack section (S) F ), crack spacing (D) F ). Among them, oil viscosity (V O The data was not obtained by measuring carbon quantum dot tracers, but rather by PVT measurements. Porosity (P), effective permeability (EP), and oil volume factor (VF) were also measured. O ), oil saturation (S) O ), reservoir pressure (P) R ), clay volume (V) C ), oil viscosity (V) O The convective term represents the process by which the overall fluid motion carries the tracer during migration; the volume of the backflow proppant (V) is a characteristic of the convective term. P ), proppant ratio (R) P ), crack section (S) F ), crack spacing (D) F() is a diffusion term characteristic, describing the spontaneous dispersion of the tracer due to random collisions.

[0032] Specifically, carbon quantum dot tracers (CODs) and inert tracers at a concentration ratio of 1:1 are added to the injection wells of the oilfield. Fluid samples are collected periodically (<0.5h) from the production wells, and the first breakthrough time t of the CQDs and inert tracers are recorded respectively. i,CQDs t i,inert Record the peak concentration time t of CQDs and inert tracers. p,CQDs t p,inert The backflow CQDs were quantitatively analyzed using a high-precision fluorescence spectrophotometer, and the backflow inert tracer was analyzed using a chromatograph. The following characteristics of the oilfield were calculated:

[0033] (1) Porosity (P)

[0034] Porosity refers to the ratio of pore volume to total volume in a rock. The greater the porosity, the more oil can usually be stored in the reservoir, thus providing a material basis for higher production capacity.

[0035]

[0036] Where, ρ in,CQDs Injecting flux into carbon quantum dot tracers, L inert It is the length of the mainstream flow line between injection and production wells for inert tracers. A represents the cross-sectional area of ​​the carbon quantum dot tracer, i.e., the effective cross-sectional area of ​​the carbon quantum dot tracer flowing between injection and production wells; since the breakthrough time of the inert tracer is purely the fluid transport time, t is used. i,inert To correct the breakthrough time of carbon quantum dot tracers, using L inert To determine the length of the main flow line between injection and production wells.

[0037] (2) Effective penetration rate (EP)

[0038] Effective permeability refers to the actual permeability at which a certain phase fluid can flow in a reservoir. The higher the effective permeability, the stronger the fluid's flow capacity in the reservoir. Effective permeability is not related to the adsorption of carbon quantum dot tracers, but is related to the actual flow rate of the oil. The effective permeability EP can be obtained by the following formula.

[0039]

[0040] Where μ is the fluid viscosity, This represents the average pressure gradient between injection and production wells.

[0041] (3) Oil volume factor (VF) O )

[0042] The oil volume factor is a measure of the ratio of the volume of formation oil under standard surface conditions to the volume under formation conditions. It reflects the volume change of formation oil under different pressure and temperature conditions and has a significant impact on production capacity prediction.

[0043]

[0044] Among them, A k Q is the well-controlled cross-sectional area. O To the oil production of production wells, Oil / water viscosity (test data) These are the inversion values ​​for carbon quantum dots (obtained through laboratory calibration).

[0045] (4) Oil saturation (S) O )

[0046] Oil saturation refers to the proportion of oil volume in a reservoir to the effective pore volume of the reservoir. It is an important parameter for measuring the oil reserves and flow capacity of a reservoir.

[0047]

[0048] Among them, R water To calibrate the resistivity under pure water conditions in the laboratory, γ is the core calibration index, α is the core coefficient, and R... t This represents the true resistivity of the formation.

[0049] (5) Reservoir pressure (P) R )

[0050] Reservoir pressure refers to the pressure exerted on fluids within a reservoir, reflecting the reservoir's storage state and flow capacity.

[0051]

[0052] Where, ρ in,inert , where h is the inert tracer injection flow rate, and h is the effective vertical thickness of the fluid flowing in the reservoir.

[0053] (6) Clay volume (V) C )

[0054] The type, content, occurrence, and physical properties of clay minerals have a significant impact on the physical properties of reservoirs. The higher the clay mineral content, the lower the porosity and permeability of the sandstone, and the worse the reservoir performance.

[0055]

[0056] Among them, M in,CQDs M out,CQDs These represent the total amount injected into CQDs and the total amount returned, respectively, k CQDsLet ρ be the adsorption coefficient of CQDs on the clay surface. clay V is the density of clay. out,CQDs For the backflow volume of CQDs, C out,CQDs The average concentration of CQDs in the effluent.

[0057] (7) Return proppant volume (V) P )

[0058] During flowback, the backflow of proppant can reduce the conductivity of artificial fractures. If the flowback rate is too high, a large amount of proppant will flow back into the wellbore with the flowback fluid, and may even be carried to the surface. This will not only reduce the conductivity of the fractures, but may also lead to the accumulation of sand at the bottom of the well, burying the gas layer and affecting the production of oil and gas.

[0059] A standard curve of "CQDs fluorescence intensity - proppant concentration" was established in the laboratory. This curve was obtained by measuring the fluorescence intensity of a known volume of labeled proppant sample in a simulated reflux environment; the total reflux proppant volume was calculated.

[0060] V P =∑C·Q·Δt

[0061] Where C is the proppant concentration obtained from the "CQDs fluorescence intensity-proppant concentration" standard curve during sampling, Q is the backflow rate at the sampling time, and Δt is the sampling time interval.

[0062] (8) Propionate ratio (R) P )

[0063] The proppant ratio refers to the ratio of proppant dosage to fracture volume. As the proppant ratio increases, the fracture conductivity increases, as more proppant can more effectively keep the fracture open, reduce fracture closure, and thus improve fluid flow capacity.

[0064] Crack volume V was obtained using an inert tracer. f

[0065]

[0066] Among them, M in,inert M out,inert These represent the total amount of inert tracer injected and the total amount of backflow, respectively. C in,inert The initial concentration of the inert tracer in the crack;

[0067] Calculate the volume V of the retained proppant. retained,prop

[0068] V retained,prop =V in,prop -V P

[0069] Among them, V in,prop This represents the initial injection volume of the proppant.

[0070] Calculate the proppant ratio

[0071] R P =V retained,prop / V f .

[0072] (9) Crack section (S) F )

[0073] As the main channel for fluid flow, the cross-sectional area of ​​a crack directly affects the fluid's flow velocity and flow rate. A larger crack cross-section can reduce fluid flow resistance, thereby increasing production capacity.

[0074] Through the crack top depth h top and bottom depth h bottom Calculate the crack height h f

[0075] h f =h top -h bottom

[0076] Calculate the crack section S F

[0077] S F =V f / h f .

[0078] (10) Crack spacing (D) F )

[0079] Fracture spacing is the distance between adjacent fracturing clusters within a horizontal well section. Excessive spacing can cause unfractured areas to become production blind spots, while insufficient spacing can suppress fracture width, leading to sand blockage or uneven proppant distribution. Therefore, it is necessary to calculate the maximum and minimum values ​​of the spacing, i.e., D. F ∈[D Fmin D Fmax ],in

[0080]

[0081] Where v is Poisson's ratio, E is Young's modulus, t is the estimated production time, and P is... e P w These are the boundary pressure and the bottom hole flowing pressure, respectively. t This is the overall compression coefficient.

[0082] In step S200, different feature factors have different impacts on the monitoring of production capacity in different oilfields. When constructing an oilfield monitoring model, analyzing the correlation between specific oilfield production capacity and features is crucial. Too many features may increase model complexity and potentially lead to overfitting; while features with the least impact may reduce model accuracy. By evaluating and selecting features that have the main influence, weakly correlated features can be eliminated while retaining strongly correlated features, thus improving the accuracy of model predictions. Based on this, this embodiment first uses the Spearman correlation coefficient to evaluate the monotonic relationship between each feature and the oilfield monitoring objective, obtaining a preliminary order of feature influence; then, a random forest tree model is used to train the data to evaluate the comprehensive impact of features on oilfield monitoring. The Spearman correlation coefficient can address the impact of feature correlation on evaluation in the random forest model, and the random forest model can address the inability of the Spearman correlation coefficient to capture complex non-monotonic relationships. Based on this, the specific steps of step S200 are as follows:

[0083] Step S210: Use Spearman correlation coefficient for multiple samples to assess the monotonic relationship between each feature and the oilfield monitoring objective;

[0084] Step S220: Set output as the dependent variable to train the random forest regression model and calculate feature importance;

[0085] Step S230: Evaluate the order of influence of features.

[0086] In step S210, features are calculated for each sample across multiple samples, and the Spearman correlation coefficient r between each feature and output is calculated. The samples are then sorted by absolute value to determine the strength of the association between features and output. The Spearman correlation coefficient r is calculated using the following formula.

[0087]

[0088] Where M is the total number of samples, x m Let y be the feature value in the m-th sample. m Let be the value of the production capacity in the m-th sample. The larger the absolute value of the Spearman correlation coefficient, |r|, the stronger the monotonic correlation between the feature and oilfield production. Figure 2 The image shows a heatmap of Spearman correlation coefficient analysis calculated from data obtained during monitoring of an oil field in this embodiment.

[0089] In step S220, a new dataset of features is randomly selected with replacement each time, and each decision tree in the random forest model is trained independently. The prediction results of all decision trees are averaged to obtain the maximum displacement and the final displacement regression prediction result of the final oilfield production. Specifically...

[0090] Step S221: Determine the number of decision trees, the feature selection method, and the growth method of the decision trees;

[0091] Step S222, Construct the root node: Place all training data in the root node, select an optimal feature, and use the optimal feature as the segmentation criterion for the current node;

[0092] Step S223, Data Segmentation: The training dataset is segmented into subsets based on the selected features, with each subset corresponding to a value of the feature; if these subsets can be correctly classified, leaf nodes are constructed and these subsets are assigned to the corresponding leaf nodes; if there are still subsets that cannot be correctly classified, the optimal features are reselected for these subsets, and they are further segmented and corresponding nodes are constructed.

[0093] Step S224, recursively construct subtrees: repeatedly split the data subset on each leaf node to construct subtrees until the maximum tree body is satisfied;

[0094] Step S225, Generate leaf nodes: Generate leaf nodes and assign them classification or regression results.

[0095] The optimal feature acquisition method in step S222 is as follows:

[0096] Step S2221, define dataset D m ={D m,1 D m,2 ,...,D m,Nm From the feature set {P, EP, VF} O ,S O ,P R V C V O V P ,R P ,S F D F Random sampling in};

[0097] Step S2222, obtain the nth feature D m,n Information entropy H(D) m,n )

[0098]

[0099] in, J is the number of samples in the j-th category for the n-th feature. m,n The total number of categories for the samples with the nth feature;

[0100] Step S2223, obtain the nth feature D m,n The conditional entropy H(D) m |D m,n )

[0101]

[0102] Where, N m For dataset D m The total number of features;

[0103] Step S2224, obtain the nth feature D m,n For dataset D m Degree of influence G(D) m D m ,n)

[0104] G(D m D m,n )=H(D m,n )-H(D m |D m,n )

[0105] Step S2225, select max(G(D) m D m The eigenvalues ​​corresponding to n) are taken as the optimal eigenvalues.

[0106] In step S230, the Spearman correlation coefficient and the feature influence order obtained from the random forest algorithm are compared and comprehensively evaluated. During the monitoring of a certain oil field, from... Figure 2 Spearman correlation coefficient analysis yielded porosity (P), effective permeability (EP), and oil saturation (S). O ), crack section (S) F ), reservoir pressure (P) R ), volume of proppant returned (V) P ), crack spacing (D) F The ratio of proppant to target capacity is strongly positively correlated with the target capacity. P The volume of clay (V) is positively correlated with the target production capacity. C ), oil volume factor (VF) O ), oil viscosity (V) O The porosity (P) and crack spacing (D) show a weak negative correlation with the target production capacity. Table 1 shows that the porosity (P) and crack spacing (D) are... F Effective permeability (EP), oil saturation (S) O ), volume of proppant returned (V) P ), reservoir pressure (P) R () is the core feature of model prediction.

[0107] Table 1. Correlation obtained from Random Forest

[0108] feature P EP <![CDATA[VF O ]]> <![CDATA[S O ]]> <![CDATA[P R ]]> <![CDATA[V C ]]> <![CDATA[V O ]]> <![CDATA[V P ]]> <![CDATA[R P ]]> <![CDATA[S F ]]> <![CDATA[D F ]]> Correlation 0.92 0.87 -0.2 0.83 0.81 -0.2 -0.2 0.82 0.45 0.66 0.91

[0109] according to Figure 2 As shown in Table 1, a comprehensive analysis yielded Table 2, in which the importance of the features was ranked as follows: porosity (P) > effective permeability (EP) > oil saturation (S). O Reservoir pressure (P) R Volume of proppant returned (V) P Crack section (S) F Crack spacing (D) F ).

[0110] Table 2 Comprehensive Analysis Table

[0111]

[0112] In step S300, a WP-PSO-BP network model is established to monitor the oilfield. The BP neural network is a multi-layer feedforward network that utilizes backpropagation of errors for nonlinear fitting, enabling it to handle high-dimensional data and complex nonlinear relationships. Particle Swarm Optimization (PSO) is a swarm intelligence-based search algorithm suitable for optimizing nonlinear functions in multidimensional space. This algorithm operates by randomly generating a swarm of particles in space. These particles are evaluated based on fitness values ​​determined by the objective function and move continuously in space at a certain speed. Global optimization is achieved through multiple iterations. In each iteration, the individual optimal value and the swarm optimal value are obtained by comparing the fitness values ​​of each particle. The particle's velocity and position are updated based on the positions of the individual and swarm optimal values. However, PSO suffers from issues such as excessively fast convergence, poor global search capability, strong weight dependence, and tendency for iterative population information to become uniform. The Wiener process (WP) is a continuous-time stochastic process. Because the motion of carbon quantum dot tracers in a fluid is subject to random collisions with liquid molecules—collisions that are independent, continuous, and random—its diffusion process can be approximated as a Wiener process. These characteristics are obtained through relevant parameters of the carbon quantum dot tracer. In practical applications, random fluctuations exist in the Wiener process. The Wiener process (WP) can be improved by using random perturbations to break the stalemate of local optima in the Particle Swarm Optimization (PSO) algorithm, thereby enhancing the global search capability and adaptability of the PSO algorithm. Therefore, this embodiment constructs a WP-PSO-BP network to monitor the oil field, using WP to update the random parameters in the PSO.

[0113] Specifically, a PSO particle is defined by two core vectors: the position vector. and velocity vector Where the position vector This indicates that the vector to be optimized is directly encoded, specifically the velocity vector. This indicates the direction and step size for controlling the position update. In this embodiment, the parameters to be optimized in the position vector include {porosity, effective permeability, oil volume factor, oil saturation, reservoir pressure, clay volume, oil viscosity, proppant volume, proppant ratio, fracture cross-section, fracture spacing}, i.e., {P, EP, VF}. O ,S O ,P R V C V O V P ,R P ,S F D F In this embodiment, WP significantly optimizes the performance of PSO (Particle Swarm Optimization) in reservoir parameter inversion by introducing stochastic parameters. These stochastic parameters are mainly generated by the combined effects of diverse sedimentary environments, diagenetic influences, tectonic activity, biological activity, fluid flow, and random geological processes. The stochastic parameters are shown below:

[0114] θ P Porosity exhibits random fluctuations around its mean in its spatial distribution;

[0115] θ EP The intensity of random permeability is the intensity of random perturbation, which represents the random perturbation of the spatial variability of permeability or porosity. The sources of randomness include formation stress sensitivity (fracturing / closure), particle migration and blockage, and dynamic changes in the fracture network.

[0116] The intensity of random fluctuations in oil saturation indicates that the fingering phenomenon at the micropore scale leads to non-uniform displacement. The sources of randomness include the non-uniform displacement front, the random effect of capillary forces, and the evolution of water channeling.

[0117] The intensity of random fluctuations in reservoir pressure represents the random changes in pressure over time or space. Sources of randomness include disturbances from fluid extraction / injection, changes in formation connectivity, and intrusion of edge and bottom water.

[0118] The proppant volume random fluctuation coefficient represents the uncertainty of proppant distribution. The randomness comes from the fact that the proppant reflux rate is affected by the random fluctuation of the production regime (flow rate / pressure).

[0119] The random fluctuation coefficient of oil viscosity indicates the sudden change in viscosity caused by temperature fluctuations and other factors. The sources of randomness include non-uniform changes in the formation temperature field and asphaltene precipitation.

[0120] The random fluctuation coefficient of the fracture cross section represents the ratio of the standard deviation to the mean of the fracture cross section size. The sources of randomness include proppant embedding / fracture, geostress redirection, and fluctuations in fracturing fluid flowback rate.

[0121] The random fluctuation coefficient of fracture spacing represents the ratio of the standard deviation to the mean of fracture spacing. The sources of randomness include proppant embedding / fracture, geostress redirection, and fluctuations in fracturing fluid flowback rate.

[0122] Where, θ * Let σ be the random parameter corresponding to the convection term characteristic. * These are the random parameters corresponding to the diffusion term characteristics. Oil volume factor (VF) O ), clay volume (V) C ), proppant ratio (R) P There is no corresponding random parameter, oil volume factor (VF) O The randomness mainly stems from random fluctuations in pressure rather than independent random processes; the clay volume (V) C ) and proppant ratio (R P In PSO, it is often treated as a constant.

[0123] A stochastic perturbation model is constructed for the above stochastic parameters using the Wiener process:

[0124] d(P)=μ P dt+θ P dW P

[0125] d(EP)=μ EP dt+θ EP dW EP

[0126]

[0127] Where μ is a deterministic drift term, determined by the variation trend of the acquired features {porosity, effective permeability, oil volume factor, oil saturation, reservoir pressure, clay volume, oil viscosity, proppant volume, proppant ratio, fracture cross section, fracture spacing}; dW is the Wiener increment, which is random noise following a normal distribution N(0,dt).

[0128] In each iteration of PSO, a Wiener process perturbation is added to each particle to simulate parameter randomness. In this embodiment, the fitness function of the PSO optimization algorithm is based on the MSE loss function, with the addition of a random perturbation to the Wiener path, i.e.

[0129]

[0130] Where N is the number of Wiener paths and T is the monitoring period. It is the i-th Wiener path.

[0131] The particle update process using the Wiener process is as follows:

[0132] X t+Δt =X t +d(f)

[0133] Among them, X t Let be the characteristic parameter value at time t, Δt be the time step, and d(f) represent the random perturbation of the corresponding feature. dt is the continuous-time differential symbol, representing an infinitesimal time increment; in practice, dt is replaced by a finite-step Δt for discretized random updates.

[0134] The optimized PSO algorithm process after Wiener process is as follows: Figure 3 As shown, the WP-PSO-BP network is trained using a training set and tested for its realism and applicability using a test set. Figure 4 It can be seen that the WP-PSO-BP model involved in this embodiment is closer to the actual production capacity than the BP neural network model that only uses PSO.

Claims

1. A method for monitoring oilfield productivity using carbon quantum dot tracers, characterized in that, include: Step S100: Obtain the convection and diffusion characteristics of the oilfield by injecting carbon quantum dot tracers and inert tracers, and establish an oilfield monitoring feature set; Step S200: Select the main features of oilfield monitoring from the oilfield monitoring feature set using the Spearman correlation coefficient and random forest tree model. Step S300: Input the main features into the neural network to monitor the oilfield production capacity. The neural network is a Wiener process optimized particle swarm optimization-backpropagation neural network. In step S100, the convection term features include {porosity, effective permeability, oil volume factor, oil saturation, reservoir pressure, clay volume, oil viscosity}; the diffusion term features include {flowback proppant volume, proppant ratio, fracture cross section, fracture spacing}. In step S100, the following parameters of carbon quantum dot tracer and inert tracer are recorded to obtain the detection feature set of the oilfield: the first breakthrough time and peak concentration time of carbon quantum dot tracer, the first breakthrough time and peak concentration time of inert tracer, the total amount of carbon quantum dot tracer injected and the total amount of flowback, the total amount of inert tracer injected and the total amount of flowback, the length of the main flow line between injection and production wells of inert tracer, and the flow cross-sectional area of ​​carbon quantum dot tracer.

2. The method according to claim 1, characterized in that, Step S200 specifically includes the following process: Step S201: Evaluate the monotonic relationship between each feature and oilfield production capacity using Spearman correlation coefficient; Step S202: Determine the importance of each feature to oilfield productivity using a random forest tree model; Step S203: Select the main features for oilfield monitoring based on monotonicity and importance.

3. The method according to claim 2, characterized in that, In step S203, features with strong positive correlation coefficients to Spearman and the top K features with the highest correlation coefficients in the random forest tree model are selected as the main features for oilfield monitoring.

4. The method according to claim 2, characterized in that, In step S300, the Wiener process is used in the neural network to obtain the random perturbations d(f) of characteristic porosity, effective permeability, oil saturation, reservoir pressure, oil viscosity, flowback proppant volume, fracture cross-section, and fracture spacing. ， Where f represents a feature. For the deterministic drift term of feature f, Let f be the randomness parameter of feature f. For Wiener increments.

5. The method according to claim 4, characterized in that, In neural networks, particles are updated using random perturbations d(f). ， in, The characteristic parameter value at time t, For time step.

Images