Grounding grid intelligent operation and maintenance system based on digital twinning and computer equipment

Abstract

The invention provides a grounding grid intelligent operation and maintenance system based on digital twinning and computer equipment. A multi-factor coupled corrosion kinetic equation is generated based on a laboratory grounding grid steel sheet accelerated corrosion test, and the corrosion kinetic equation is used for predicting the instantaneous corrosion rate of a galvanized steel sheet and estimating the mass corrosion loss of the steel sheet within specified time; combining the corrosion kinetic equation with the simulation model to establish a steel sheet corrosion data set, and training a physical information neural network to construct a corrosion prediction physical model; training the corrosion prediction physical model by adopting a neural network regression algorithm; constructing a grounding grid digital twin model by taking the corrosion prediction physical model as a base and combining field environment monitoring data; defining a Markov decision process in the digital twin environment of the grounding grid; and based on the feedback and reward function of the digital twin model of the grounding grid, training by adopting a near-end strategy optimization algorithm to obtain an intelligent decision-making body, and continuously optimizing the decision-making of the intelligent decision-making body through an online fine tuning function.

Description

TECHNICAL FIELD

[0001] The application relates to the field of power detection and operation and maintenance, and particularly relates to a grounding grid intelligent operation and maintenance system based on digital twinning and computer equipment. BACKGROUND

[0002] The grounding grid is a key infrastructure for ensuring the safe operation of electrical equipment in a power system. The grounding grid is buried in the soil through a galvanized steel conductor in a mesh structure, forming a low-impedance grounding path to discharge fault current and lightning current. In the long-term operation process, the galvanized steel material of the grounding grid is affected by the coupling of multiple physical fields in the soil environment, including electrochemical corrosion, microbial corrosion and stray current corrosion, which leads to the reduction of the conductor cross section, the increase of the grounding resistance, and even the grounding failure, which threatens the safe and stable operation of the power grid. At present, the corrosion prediction method for the grounding grid in the related technology has the following problems:

[0003] (1) Single corrosion prediction method:

[0004] Based on the physical driving method, such as electrochemical polarization curve analysis and finite element multi-field coupling simulation, the corrosion mechanism can be revealed, but it is difficult to accurately quantify the soil physical and chemical parameters, such as pH value, water content, and pore ratio, and the nonlinear influence of dynamic fluctuations on the corrosion rate, especially in the scenes of seasonal temperature and humidity changes, sudden heavy rain and the like, the prediction error is significant.

[0005] Based on the pure data-driven method, such as the time series prediction model, although the historical detection data can be used to learn the statistical law, there is a lack of physical constraints on the corrosion electrochemical process, which leads to insufficient reliability of extrapolation prediction.

[0006] (2) Limitations of digital twinning application:

[0007] The existing digital twinning model mostly uses simplified two-dimensional geometric modeling and static parameter assignment, and cannot realize high-fidelity restoration of the three-dimensional topological structure of the grounding grid and dynamic visualization of the corrosion process.

[0008] Weak virtual-real interaction capability: Most systems only realize one-way data mapping (physical → virtual), lack real-time decision feedback mechanism based on the simulation results of the twin body, and are difficult to support closed-loop operation and maintenance.

[0009] (3) Insufficient intelligence of operation and maintenance decision:

[0010] The traditional maintenance strategy relies on periodic detection and manual experience decision, and has problems such as response lag, excessive maintenance or insufficient maintenance, etc.

[0011] Although some researches try to introduce reinforcement learning into equipment maintenance, the training environment is mostly based on simplified mathematical models, and the multi-scale deduction capability of the digital twinning system for complex soil corrosion environment is not fully utilized, which limits the generalization of the strategy in the actual scene. SUMMARY

[0012] Therefore, the application provides a grounding grid intelligent operation and maintenance system based on digital twinning and a computer device.

[0013] In a first aspect, the embodiments of the application provide a grounding grid intelligent operation and maintenance system based on digital twinning, comprising: a corrosion evolution law modeling module, an agent model construction module, a grounding grid digital twinning model establishment module, and an intelligent decision-making body construction module.

[0014] The corrosion evolution law modeling module is configured to:

[0015] The mapping relationship between the resistance change rate and the mass loss of the galvanized steel sheet is established based on the grounding grid resistance time series data in the long-term detection report and the Faraday electrolysis law.

[0016] Based on the laboratory grounding grid steel sheet accelerated corrosion test, a multi-factor coupled corrosion kinetics equation is generated, which is used to predict the instantaneous corrosion rate of the galvanized steel sheet based on the soil environment data measured on site, and the mass corrosion loss of the steel sheet is obtained by integrating the specified time.

[0017] The agent model construction module is configured to:

[0018] The corrosion kinetics equation is combined with a finite element electromagnetic-electrochemical joint simulation model to establish a soil environment parameter and steel sheet mass loss dataset, a physical information neural network is trained based on the dataset to construct a corrosion prediction physical model.

[0019] The neural network regression algorithm is used to train the corrosion prediction physical model, wherein the input parameters include the soil physicochemical parameters, the grounding grid topology structure, and the detected grounding grid resistance data, and the output parameter is the probability distribution of the corrosion degree of the galvanized steel sheet within a specified time span, which is determined by the instantaneous corrosion rate of the galvanized steel sheet obtained by prediction, and the corrosion prediction physical model is used to predict the corrosion degree of the galvanized steel sheet in the soil based on the soil environment data simulated by the Monte Carlo method.

[0020] The grounding grid digital twinning model establishment module is configured to:

[0021] The corrosion prediction physical model is combined with the environmental monitoring data on site to construct the grounding grid digital twinning model.

[0022] Since the decision-making agent needs to interact with the environment, it updates through interaction with the environment and feedback given by the built-in reward function of the environment. However, the actual environment cannot meet the requirements of instant updating, evolution and feedback. Therefore, a grounding net digital twin model needs to be established, which needs to meet two basic functions. First, the corrosion evolution of the grounding net needs to be carried out according to the input environmental parameters, grounding net parameters and detection data. Here, the above-mentioned corrosion prediction physical model can be directly used to meet the demand. Second, feedback needs to be given to the maintenance suggestion of the agent according to the running situation of the grounding net, the influence of the maintenance action performed and the cost of the action. A reward function needs to be defined. Finally, the digital twin model needs to have the ability to call the algorithms of other modules and realize visualization as the carrier of various modules.

[0023] The intelligent decision-making module is used for:

[0024] The Markov decision process is defined in the grounding net digital twin environment, and the parameters of the Markov decision process include: state space, action space and reward function.

[0025] Based on the feedback of the grounding net digital twin model and the reward function, the proximal policy optimization algorithm is used to train the LSTM neural network to obtain the intelligent decision-making body, and the decision-making of the intelligent decision-making body is optimized through the online fine-tuning function to output the grounding net life cycle maintenance strategy.

[0026] According to the above system of the embodiment of the application, the following additional technical features can also be provided:

[0027] In the above technical solution, optionally, the multiple factors include: soil pH value, soil porosity ratio, water content, soil resistivity and sensitivity weight coefficient of galvanized steel sheet to corrosion rate.

[0028] In any of the above technical solutions, optionally, the proxy model construction module is further used for:

[0029] After training the corrosion prediction physical model using the neural network regression algorithm, the corrosion degree of the galvanized steel sheet at different positions of the grounding net is calculated through the mapping relationship and the grounding net resistance data in the field monitoring report;

[0030] The field measured soil environment data and the grounding net resistance data in the field monitoring report are input into the corrosion prediction physical model, and the predicted corrosion degree of the galvanized steel sheet is output;

[0031] The corrosion degree and the predicted corrosion degree are compared to optimize the corrosion prediction physical model.

[0032] In any of the above technical solutions, optionally, the corrosion evolution rule modeling module includes:

[0033] The data acquisition unit is configured to acquire, through an ion selective electrode sensor, a TDR time domain reflection moisture sensor and a four-probe resistance meter, soil pH, water content, soil porosity ratio and soil resistivity of soil where the grounding grid is located in real time, and a sampling interval is less than or equal to 60 minutes.

[0034] The corrosion kinetics modeling unit is configured to establish a corrosion kinetics equation based on Faraday's law of electrolysis and an Arrhenius equation.

[0035] The model verification unit is configured to simulate different environmental combinations through a laboratory test box to adjust a soil multi-factor coupling effect function in the corrosion kinetics equation.

[0036] In any of the above technical solutions, the corrosion evolution modeling module further comprises:

[0037] The data processing unit is configured to remove noise data by using an isolated forest-based anomaly detection algorithm and fill in missing values by using a spatiotemporal Kriging interpolation method.

[0038] In any of the above technical solutions, the proxy model construction module comprises:

[0039] The machine learning proxy model unit is configured to determine input parameters and output parameters, preliminarily construct a corrosion prediction physical model by using a neural network regression algorithm based on a physical information neural network, and perform transfer learning because the machine learning proxy model is established on the basis of small steel sheet samples in a laboratory and there is a certain gap between the actual grounding grid.

[0040] The transfer learning optimization unit is configured to use weights of the preliminarily constructed corrosion prediction physical model as initial values, perform self-adaptive training by using grounding grid data in a field monitoring report, make a mean square error of the corrosion prediction physical model less than or equal to 5%, and obtain the corrosion prediction physical model.

[0041] In any of the above technical solutions, the proxy model construction module further comprises:

[0042] The dynamic calibration unit is configured to calculate corrosion degrees of galvanized steel sheets at different positions of the grounding grid, output the corrosion degrees based on the corrosion prediction physical model, and perform online update on parameters of the corrosion prediction physical model based on a Kalman filter model when a deviation between the corrosion degrees and the predicted corrosion degrees is greater than or equal to 15%.

[0043] In any of the above technical solutions, the intelligent decision-making module comprises:

[0044] The state space construction unit is configured to define a state space, and a state vector S of the state space is t =[H1,…,H n ,Ccost ,R risk E env ], where H i C represents the health index of the i-th galvanized steel sheet node, n represents the number of galvanized steel sheet nodes, and C cost R represents the cumulative maintenance cost. risk E represents the probability of failure risk. env Representing a set of environmental parameters, the formula for calculating the health index is: H i =1-d i / d th d i Indicates the degree of corrosion, d th This represents the failure threshold, and the degree of corrosion is estimated using a corrosion prediction physical model.

[0045] Action space design unit: Defines maintenance actions as a multi-dimensional discrete space, including three types of maintenance strategies and their combinations: maintenance scope, maintenance method, and maintenance timing;

[0046] Reward function generation unit: Calculates the reward function, which is:

[0047] r t =w1(1-R) risk )+w2(1-C cost / C max )-w3I fail

[0048] Where r1 represents the reward function, w1, w2, and w3 represent the adaptive weights, and R... risk C represents the probability of failure risk. cost C represents the cumulative maintenance cost. max Indicates the cost of the largest maintenance operation, I fail Represents the fault indication function;

[0049] Hierarchical decision-making unit: A multi-agent collaborative decision-making architecture is adopted, with the upper-level agent planning the allocation of the annual maintenance budget and the lower-level agent executing the specific maintenance operations corresponding to the budget allocation.

[0050] Optionally, in any of the above technical solutions, the system further includes:

[0051] A grounding grid digital twin environment construction module for forming the grounding grid digital twin environment required for intelligent decision-making;

[0052] The grounding grid digital twin environment construction module is specifically used for:

[0053] Develop a three-dimensional visualization platform based on Unreal Engine; wherein, the parameterized three-dimensional model is generated by analyzing the grounding grid CAD drawing, and the corrosion depth is rendered by thermal map gradient; integrate Python computing kernel, real-time call corrosion prediction model and reinforcement learning inference engine;

[0054] Establish a bidirectional data channel; wherein, the bidirectional data channel includes uplink and downlink, the uplink maps sensor data to the grounding grid digital twin model through the MQTT protocol, and drives the dynamic update of the corrosion state, the downlink encodes the decision action of the intelligent decision body into PLC control instructions, and issues them to the field maintenance equipment through the gRPC interface;

[0055] Deploy a lightweight TensorRT inference engine in the Python environment embedded in Unreal Engine, quantize and compress the corrosion prediction agent model, and ensure that the time consumption of single forward propagation is ≤33ms;

[0056] The rendering thread and the physical computing thread exchange data through shared memory and set priority strategy to ensure that the numerical calculation delay fluctuation is ≤±5% when the visualization refresh rate is ≥60FPS.

[0057] In the second aspect, the embodiments of the present application provide a computer device, which comprises the digital-twin-based grounding grid intelligent operation and maintenance system of the first aspect.

[0058] The digital-twin-based grounding grid intelligent operation and maintenance system and the computer device of the embodiments of the present application relate to a grounding grid corrosion state prediction and full life cycle operation and maintenance decision scheme combining multi-physical field simulation and deep reinforcement learning, realize three-dimensional visualization real-time simulation of the corrosion process of the grounding grid by constructing a digital twin environment, and dynamically optimize the maintenance strategy combined with the deep reinforcement learning algorithm, thereby improving the corrosion prediction accuracy and reliability, and solving the problems of lagging corrosion evaluation of the traditional grounding grid, dependence of the maintenance decision on experience, and insufficient economy.

[0059] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, the specific embodiments of the present application can be implemented according to the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0060] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0061] Figure 1 The structure block diagram of the digital-twin-based grounding grid intelligent operation and maintenance system of the embodiments of the present application is shown;

[0062] Figure 2 A schematic diagram of data collection of the ground net intelligent operation and maintenance system according to an embodiment of the present application is shown.

[0063] Figure 3 A schematic diagram of a process of constructing a corrosion prediction proxy model according to an embodiment of the present application is shown.

[0064] Figure 4 A schematic diagram of various types of constructing a corrosion prediction proxy model according to an embodiment of the present application is shown.

[0065] Figure 5 A schematic diagram of constructing a ground net digital twin model according to an embodiment of the present application is shown.

[0066] Figure 6 A schematic diagram of a process of training an intelligent decision body according to an embodiment of the present application is shown.

[0067] Figure 7 A schematic diagram of training an intelligent decision body using a PPO algorithm according to an embodiment of the present application is shown.

[0068] Figure 8 A comparison diagram between an intelligent decision body and a conventional strategy according to an embodiment of the present application is shown.

[0069] Figure 9 A comparison diagram between an intelligent decision body and a conventional strategy according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described clearly below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0071] The terms “first”, “second”, and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by “first”, “second”, etc. are generally of a kind and do not limit the number of objects, for example, the first object can be one or more. In addition, “and / or” in the specification and claims indicates at least one of the connected objects, and the character “ / ” generally indicates that the objects before and after are in an “or” relationship.

[0072] The application provides an intelligent operation and maintenance scheme for a grounding grid based on digital twinning and deep reinforcement learning. By constructing a multi-modal data fusion corrosion prediction model and a virtual-real interactive decision training environment, precise prediction of corrosion states and dynamic optimization of operation and maintenance strategies are achieved throughout the life cycle. The following will describe the digital twinning-based intelligent operation and maintenance system for a grounding grid and the computer device provided by the embodiments of the application in detail in conjunction with the drawings and specific embodiments and their application scenarios. In the case of no conflict, the embodiments and features in the embodiments described below can be combined with each other.

[0073] The application provides an intelligent operation and maintenance system for a grounding grid based on digital twinning, as shown in the figure. The intelligent operation and maintenance system 10 for the grounding grid comprises a corrosion evolution law modeling module 11, an agent model construction module 12, a grounding grid digital twinning model establishment module 13, and an intelligent decision body construction module 14. Figure 1

[0074] The corrosion evolution law modeling module 11 is configured to:

[0075] The mapping relationship between the resistance change rate and the mass loss of the galvanized steel sheet is established by combining the Faraday electrolysis law with the grounding grid resistance time series data in the long-term detection report.

[0076] Based on the laboratory grounding grid steel sheet accelerated corrosion test, a multi-factor coupled corrosion kinetics equation is generated. The corrosion kinetics equation is used to predict the instantaneous corrosion rate of the galvanized steel sheet based on the soil environment data measured on site, and the mass corrosion loss of the steel sheet is obtained by integrating the corrosion rate within a specified time. The multi-factors include the soil pH value, the soil porosity ratio, the water content, the soil resistivity, and the sensitivity weight coefficient of the galvanized steel sheet itself resistance to the corrosion rate.

[0077] The agent model construction module 12 is configured to:

[0078] The corrosion kinetics equation is combined with the finite element electromagnetic-electrochemical joint simulation model to establish a soil environment parameter and steel sheet mass loss dataset. A physical information neural network is trained based on the dataset to construct a corrosion prediction physical model.

[0079] The neural network regression algorithm is used to train the corrosion prediction physical model. The input parameters include the soil physicochemical parameters, the grounding grid topology structure, and the detected grounding grid resistance data. The output parameter is the probability distribution of the corrosion degree of the galvanized steel sheet within a specified time span. The probability distribution of the corrosion degree of the galvanized steel sheet within the specified time span is determined by the predicted instantaneous corrosion rate of the galvanized steel sheet. The corrosion prediction physical model is used to predict the corrosion degree of the galvanized steel sheet in the soil by Monte Carlo simulation of the soil environment data.

[0080] ​The corrosion degree of the galvanized steel sheet at different positions of the grounding grid is calculated through the mapping relationship and the grounding grid resistance data in the field monitoring report.

[0081] The field measured soil environment data and the grounding grid resistance data in the field monitoring report are input into the corrosion prediction physical model, and the predicted corrosion degree is output.

[0082] The corrosion degree and the predicted corrosion degree are compared, and the corrosion prediction physical model is optimized.

[0083] The grounding grid digital twin model establishment module 13 is configured to:

[0084] The corrosion prediction physical model is used to construct the grounding grid digital twin model in combination with the field environment monitoring data.

[0085] The intelligent decision-making module 14 is configured to:

[0086] A Markov decision process is defined in the grounding grid digital twin environment, and parameters of the Markov decision process include a state space, an action space, and a reward function.

[0087] Based on the feedback of the grounding grid digital twin model and the reward function, an LSTM neural network is trained using a proximal policy optimization algorithm to obtain an intelligent decision-making body, and the decision-making of the intelligent decision-making body is optimized through an online fine-tuning function to output a grounding grid life cycle maintenance strategy.

[0088] The grounding grid intelligent operation and maintenance system 10 further includes a grounding grid digital twin environment construction module 15, which is specifically configured to:

[0089] A three-dimensional visualization platform is developed based on the Unreal Engine;

[0090] A bidirectional data channel is established, which includes an uplink and a downlink. The uplink maps sensor data to the grounding grid digital twin model through the MQTT protocol and drives dynamic updates of the corrosion state, and the downlink encodes decision-making actions of the intelligent decision-making body into PLC control instructions and issues them to field maintenance equipment through a gRPC interface.

[0091] A lightweight TensorRT inference engine is deployed in a Python environment embedded in the Unreal Engine to quantitatively compress the corrosion prediction proxy model.

[0092] The rendering thread and the physical calculation thread exchange data through shared memory and set priority strategies.

[0093] In one embodiment, as Figure 1As shown, the corrosion evolution modeling module 11 includes: a data acquisition unit 111, a corrosion kinetics modeling unit 112, a model verification unit 113, and a data processing unit 114. The proxy model construction module 12 includes: a machine learning proxy model unit 121, a transfer learning optimization unit 122, and a dynamic calibration unit 123. The intelligent decision module 14 includes: a state space construction unit 141, an action space design unit 142, a reward function generation unit 143, and a hierarchical decision unit 144.

[0094] The working method of the grounding grid intelligent operation and maintenance system 10 is as follows:

[0095] Step 1: Multi-dimensional data acquisition

[0096] (1) Data source configuration:

[0097] Figure 2 The schematic diagram of the data acquisition of the grounding grid intelligent operation and maintenance system of the embodiment of the application is shown, mainly including soil environment monitoring and grounding grid resistance detection.

[0098] At the key nodes of the grounding grid, such as branch junctions, edge extension sections, and high current density areas, monitoring terminals are deployed, and each terminal is integrated to include:

[0099] Four-probe resistance meter (measurement accuracy ±0.1 mΩ): record the grounding resistance value at 60-minute intervals;

[0100] Soil parameter sensor group: including pH electrode (range 0-14, accuracy ±0.2), TDR time domain reflection moisture sensor (range 0-100%, resolution 0.5%), temperature and humidity composite probe (temperature range -20℃-80℃, accuracy ±0.5℃; humidity range 0-100% RH, accuracy ±3% RH).

[0101] Thus, the resistivity, pH value, water content, porosity ratio, and temperature parameters of the soil where the grounding grid is located are collected in real time, and the sampling frequency is not less than 1 time / 60 minutes.

[0102] The corrosion evolution database across the time scale (year / month / day) can be constructed by combining the grounding resistance time series data in the historical detection report and the laboratory grounding grid steel sheet accelerated corrosion test data.

[0103] (2) Data transmission and storage:

[0104] LoRa wireless networking technology is adopted to converge data to the edge gateway, and after AES-256 encryption, the data is uploaded to the cloud platform; the database is stored in time-space dimensions, the timestamp is aligned to the millisecond level, and the spatial coordinates are accurate to 0.0001°.

[0105] Step 2: Multi-source heterogeneous data fusion and preprocessing

[0106] (1) Adopting improved Isolation Forest algorithm to detect outliers:

[0107] Set the number of subsampling times to 100 and the depth of sub-tree to 15 layers; mark the records with resistance mutation variables exceeding ±3σ (σ is the standard deviation of historical data) as outliers; combine with expert experience library (such as lightning event log) for manual review to eliminate misjudgment samples.

[0108] (2) Adopting spatio-temporal Kriging interpolation method to fill in missing values:

[0109] Construct a semi-variogram function model:

[0110] γ(h,t) = N ugget +(S ill -N ugget )·(1-exp(-3h / a h -3t / a t ))

[0111] Wherein, γ(h,t) represents the spatio-temporal semi-variogram value, which quantifies the strength of data variability between two points (or observation values) separated by spatial distance h and time interval t. The larger γ(h,t) is, the greater the difference (or uncertainty) under the given spatio-temporal distance.

[0112] h represents the spatial distance, which refers to the Euclidean distance between two spatial positions (such as sensor positions, geographic coordinate points), with the unit of meter (m). t represents the time interval, which refers to the time difference between two observation time points, with the unit of hour.

[0113] a h is the spatial range, which represents the range of spatial correlation, when the spatial distance h reaches a h , the value of exp(-3h / a h ) is about exp(-3)≈0.05, which makes the (1-exp(-3h / a h )) term close to 0.95. At this time, γ(h,t) is very close to its maximum value (S ill ). Therefore, a h defines the maximum distance at which data has significant spatial correlation. Beyond this distance, spatial correlation becomes weak and the semi-variogram value tends to be stable (reaches the base value). a h = 50m, which means that within the spatial range of this study, data points have strong spatial correlation within 50 meters.

[0114] a t is the time range, which represents the range of time correlation, similar to a h , when the time interval t reaches at The value of the term exp(-3t / a t ) is approximately exp(-3)≈0.05, which makes the term (1-exp(-3t / a t )) close to 0.95. At this time, γ(h,t) is very close to its maximum value (S ill ). Therefore, a t defines the maximum time interval for which the data has significant temporal correlation. Beyond this interval, the temporal correlation becomes weak and the semi-variogram values tend to be stationary (reach a baseline value). a t =24hour, which means that the data points have strong temporal correlation within 24 hours (1 day), indicating that the data may have daily periodicity (such as day / night, temperature variation, human activity cycle, etc.).

[0115] S ill is the baseline value, representing the stationary value that the semi-variogram eventually reaches. N ugget is the initial value, representing the value of the semi-variogram when the spatio-temporal distance is zero (h=0,t=0), N ugget =0.15.

[0116] (3) Feature engineering and dimensionality reduction:

[0117] Perform standardization (Z-score normalization) on the original data to eliminate dimension differences.

[0118] Step 3: Corrosion feature quantification

[0119] (1) Mass loss calculation:

[0120] Based on Faraday's law of electrolysis, derive the mapping relationship between resistance change rate and steel mass loss:

[0121]

[0122] where Δm is the steel mass loss, M is the molar mass of the steel, n is the number of charge transfers, F is the Faraday constant, A(t) is the real-time effective corrosion area, i corr (t) is the resistance change rate.

[0123] Combine the resistance-cross-sectional area relationship R=ρ·L / (A0-ΔA), where A0 is the initial cross-sectional area and ΔA is the corrosion loss area, to iteratively solve the steel mass loss Δm.

[0124] (2) Extraction of environmental impact factors:

[0125] Calculate the dynamic influence coefficient of soil parameters:

[0126] pH influence factor: f pH= 1 - 0.15|pH - 7.2|, valid when pH ∈ [5.0, 9.5]; the effect of pH on corrosion rate is that the more acidic, the faster the corrosion, and weak alkaline also accelerates corrosion, but when the alkalinity rises to a certain extent, the corrosion rate will decrease instead.

[0127] Moisture content impact factor: f θ = 0.8θ 0.3 , θ is the volume moisture content; the effect of moisture content on steel sheet corrosion is a parabolic curve that first increases and then decreases, with a maximum point, i.e. the maximum moisture content.

[0128] Temperature correction factor: f T = exp(E a / R·(1 / 298 - 1 / (T + 273))), E a = 45 kJ / mol, T is the temperature.

[0129] Soil resistivity: f ρsoil = 0.423ρ soil 0.213 The greater the soil resistivity, the faster the steel sheet corrosion rate.

[0130] (3) To provide the parameters required by the physical model for corrosion prediction, data needs to be extracted from three dimensions, including: physical characteristics: calculate the mass loss rate of steel sheet based on Faraday's law, grounding grid resistance value, and correct the temperature impact factor combined with Arrhenius equation; environmental characteristics: calculate the sliding average of soil physicochemical parameters, seasonal fluctuation coefficient and mutation detection index; structural characteristics: analyze the topological connection relationship of grounding grid.

[0131] Step 4: Accelerated corrosion test of laboratory grounding grid steel sheet

[0132] (1) Test piece preparation and test device:

[0133] Test piece processing: material: Q235 hot-dip galvanized steel sheet (thickness 5 mm, galvanized layer 80 μm), cut into 100 mm × 20 mm standard test sample; pretreatment: sequentially ultrasonic cleaning with acetone for 15 min, deionized water washing, weighing to 0.1 mg accuracy.

[0134] (2) Test system setup:

[0135] 1) Electrolytic cell configuration:

[0136] Working electrode: galvanized steel test piece (exposed area 10 cm 2 ); Reference electrode: saturated calomel electrode (SCE);

[0137] Auxiliary electrode: platinum sheet electrode; electrolyte: simulated soil leaching solution (NaCl 0.1 mol / L + Na2SO4 0.05 mol / L + CaCO3 0.02 mol / L).

[0138] 2) Environmental control module:

[0139] Thermostatic bath: temperature control range 5-50℃, accuracy ±0.5℃; gas injection system: adjust dissolved oxygen concentration (0-8 mg / L); pH automatic titrator: maintain solution pH value within the set value ±0.1 range.

[0140] (3) Multi-factor coupling corrosion test

[0141] 1) Single variable control test:

[0142] Fixed basic conditions (T = 25℃, pH = 6.5, DO = 5 mg / L), sequentially change: pH gradient: 6.5, 7.5, 8.5; humidity gradient: 10%, 20%, 30%; temperature gradient: 10℃, 25℃, 40℃.

[0143] Each group of tests lasted 72 hours, and samples were taken every 8 hours to measure: electrochemical test: potentiodynamic polarization scanning was performed using Gamry 3000 electrochemical workstation (scanning rate 0.5 mV / s); weight loss measurement: after the test, the corrosion products were removed (500 mL HCl + 3.5 g of methylene tetramine solution was soaked for 10 min), dried and weighed to calculate the corrosion rate.

[0144] 2) Multi-factor orthogonal test:

[0145] L9(3 3 ) orthogonal table was used to design the test combination to investigate the effects of pH (6.5, 7.5, 8.5), humidity gradient (10%, 20%, 30%), temperature (10℃, 25℃, 40℃), and three factors;

[0146] Through orthogonal analysis, the main effect factors were determined, and the electrochemical corrosion equation of the steel sheet was established to understand the relationship between the mass loss of the steel sheet and the external environment.

[0147] Step 5: Model construction

[0148] (1) Physical calculation path design:

[0149] Based on the accelerated corrosion test data of the steel sheet, the corrosion kinetics equation was established as:

[0150] Δm = ∫K·i corr ·exp(-E a / RT)·f(pH, θ, ρ soil )dt

[0151] In the experiment, the mass loss of the steel sheet is recorded as Am, where K is a proportional constant, i corr is the corrosion current density, E a is the activation energy, R is the gas constant, T is the soil temperature, f(pH, θ, p soil ) is the soil multi-factor coupling effect function, which is determined by the soil pH, moisture content θ and soil resistivity p soil .

[0152] The soil multi-factor coupling effect function needs to determine the coefficient through regression. After obtaining f(pH, θ, p soil ) through regression analysis, the mass loss of the steel sheet per unit time, i.e. the theoretical prediction of the corrosion rate v phys , can be finally determined through the geometric parameters of the steel sheet, which is the instantaneous corrosion rate.

[0153] (2) Data-driven path design:

[0154] Figure 3 The flowchart for building the corrosion prediction proxy model according to the embodiments of the present application is shown.

[0155] 1) Build a physics-informed neural network (PINN):

[0156] Input layer: soil physicochemical parameters, grounding grid topology features and corrosion rate v phys calculated by the corrosion kinetics equation; hidden layer: adopt a gated recurrent unit (GRU) to capture the time sequence dependence, and a parallel residual convolution module to extract spatial features; output layer: predict the corrosion degree distribution d pred in the future At time.

[0157] 2) Design a compound loss function:

[0158] L total = a·MSE(d pred , d real ) + b·|v phys -v pred |

[0159] Where a, b are adaptive weight coefficients, balancing data fitting and physical consistency constraints.

[0160] 3) Align field-laboratory data:

[0161] Convert the laboratory accelerated test results to the equivalent years in the actual soil environment:

[0162] t field = t lab ·v field / v lab

[0163] where v field is the field corrosion rate, v lab is the corrosion rate under laboratory accelerated conditions.

[0164] 4) Fine-tuning of the corrosion prediction physical model:

[0165] Due to the differences between experimental conditions and field environmental conditions, and more influencing variables in the field, it is necessary to use transfer learning to fine-tune the neural network of the corrosion prediction physical model, and use the Levenberg-Marquardt algorithm to optimize the parameters of the last layer of the neural network, so that the determination coefficient R 2 ≥ 0.92 between the predicted value and the field measured value.

[0166] The finally established corrosion prediction physical model will be implanted in the digital twin model to control the corrosion evolution process of the grounding grid. The corrosion depth spatial and temporal distribution map is established as the base of the digital twin model.

[0167] Figure 4 The various types of schematic diagrams for constructing the corrosion prediction agent model (i.e., the grounding grid corrosion degradation model) of the embodiments of the present application are shown.

[0168] Step 6: Construction of the grounding grid digital twin model

[0169] Figure 5 The schematic diagram for constructing the grounding grid digital twin model of the embodiments of the present application is shown. Since the decision-making agent needs to interact with the environment, it is updated through interaction with the environment and feedback given by the reward function built in the environment. However, the actual environment cannot meet the requirements of instant updating, evolution and feedback. Therefore, it is necessary to establish a grounding grid digital twin model, which needs to meet 2 basic functions. First, it needs to evolve the corrosion of the grounding grid according to the input environmental parameters, grounding grid parameters and detection data. Here, the corrosion prediction physical model obtained above can be directly used to meet the demand. Second, it needs to calculate the cost corresponding to the maintenance action according to the running situation of the grounding grid and the executed maintenance action, and give feedback to the agent. The reward function needs to be defined. Finally, the digital twin model needs to have the ability to call the algorithms of other modules and realize visualization as the carrier of various modules.

[0170] Step 7: Construction of the digital twin environment

[0171] Deep reinforcement learning requires agents to be trained in the built environment, constantly interacting with the built environment, and updating the agents iteratively through the feedback provided by the environment. Since not all agents can be directly trained in the actual environment, a digital twin environment similar to the actual environment needs to be built. The embodiment is based on experimental and detection data to build a high-fidelity digital twin environment, realize corrosion dynamic simulation, maintenance action interaction and reinforcement learning environment support. The specific implementation steps are as follows:

[0172] (1) Develop a three-dimensional visualization platform based on Unreal Engine:

[0173] 1) CAD3D modeling:

[0174] Import the grounding grid design drawing (in DWG format), parse the geometric topology structure through the Datasmith plug-in of Unreal Engine, parameterize the conductor nodes, and bind the material properties (initial thickness, galvanized layer thickness, resistivity).

[0175] 2) Generate LOD (Level of Detail) model:

[0176] The first-level model retains the complete topology (accuracy 0.1 mm), and the second-level model is simplified for fast rendering.

[0177] 3) Corrosion state visualization mapping:

[0178] Color coding: dynamically adjust the material diffuse color based on corrosion depth: healthy state corresponds to RGB(0, 255, 0), which means corrosion depth ≤0.2mm; moderate corrosion corresponds to RGB(255, 165, 0), which means 0.2mm<depth≤0.5mm; severe corrosion corresponds to RGB(255, 0, 0), which means depth>0.5mm.

[0179] 4) Corrosion simulation:

[0180] Dynamically adjust the corrosion level through the shader, and increase the color depth by 1% for every 1% increase in corrosion level.

[0181] (2) Random corrosion evolution simulation:

[0182] 1) Multi-factor coupled random model:

[0183] Based on the laboratory orthogonal test data, construct the corrosion rate probability density function:

[0184] v stoch = v det ·(1+0.3·N(0, 1))

[0185] Where, v detFor deterministic corrosion rate, N(0, 1) is a standard normal distribution random variable.

[0186] 2) Introduce spatial correlation:

[0187] Gaussian random field is used to simulate the spatial heterogeneity of soil parameters, and the covariance function is:

[0188] C(r) = σ 2 · exp(-r 2 / (2l 2 ))

[0189] Where r is the distance between nodes, l is the correlation length of galvanized steel sheet, and σ = 0.15 is the variance.

[0190] 3) Real-time simulation engine:

[0191] Create a corrosion simulation blueprint in Unreal Engine: call Python script to calculate the corrosion increment of each node every frame; generate corrosion products (such as rust particle effects) through Niagara particle system; update the corrosion state every 60 seconds, and the time acceleration ratio can be adjusted (1x-1000x).

[0192] (3) Maintenance action interaction mechanism:

[0193] 1) Virtual maintenance operation design

[0194] The action instruction set is defined as shown in Table 1:

[0195] Table 1

[0196]

[0197] 2) Virtual-real interaction protocol development

[0198] ① Instruction transmission interface:

[0199] Use gRPC+Protobuf to build a two-way communication protocol, and encode virtual maintenance actions as Protobuf messages. Fields include: action type, target location (latitude and longitude), parameter list; physical execution results (such as actual repair effects) are returned to the digital twin through MQTT.

[0200] Perform communication delay compensation, deploy NTP time synchronization server to ensure that the virtual-real time axis deviation is ≤50ms.

[0201] ② Effect feedback mechanism:

[0202] After the maintenance operation is executed, trigger the dynamic response calculation of the digital twin: call the finite element analysis kernel to recalculate the electric field distribution; update the environmental parameters in the corrosion model (such as the corrosion inhibitor concentration diffusion model).

[0203] Generate maintenance effect report in Unreal Engine: Show resistance contrast before and after repair, remaining life prediction in holographic projection form; generate interactive decision board through Unreal Motion Graphics (UMG).

[0204] 3) Deep reinforcement learning environment integration:

[0205] ① Reinforcement learning interface encapsulation

[0206] State space encapsulation: Extract state vector from digital twin (updated every 60 seconds).

[0207] Data format conversion: Use Unreal Engine's JsonUtilities module to serialize state to JSON.

[0208] ② Training environment acceleration optimization

[0209] Distributed simulation cluster:

[0210] Deploy multiple Unreal Engine digital twin environments (run independent digital twin copies for each instance); implement parallel policy evaluation through Ray framework, training throughput improved by 8-12 times.

[0211] Asynchronous training architecture:

[0212] Learner: Run PPO algorithm to update policy network; Worker: Each worker runs a digital twin environment instance synchronously; Replay Buffer: Store state-action-reward tuples, support priority sampling.

[0213] ③ Real-time reward function calculation

[0214] Reward visualization: Display the contribution of each region to the total reward in the form of a dynamic heat map in the digital twin interface.

[0215] Step 8: Training of full life cycle maintenance intelligent decision body based on PPO algorithm

[0216] This embodiment details the use of proximal policy optimization (PPO) algorithm to train intelligent agents in a digital twin environment to generate optimal grounding net life cycle maintenance strategies. Figure 6 The flowchart of training intelligent decision body of the embodiment of the application is shown, Figure 7 The schematic diagram of training intelligent decision body of the embodiment of the application using PPO algorithm is shown.

[0217] The specific implementation steps are as follows:

[0218] (1) Markov decision process:

[0219] 1) State space definition:

[0220] S t = [H1,..., H n , C cost , R risk , E env ]

[0221] H i = 1 - d i / d th , d i is the corrosion degree, d th is the failure threshold, C cost is the cumulative maintenance cost, R risk is the failure risk probability, E env is the environmental parameter set.

[0222] 2) Action space design:

[0223] Maintenance range: local coating repair (1-3 nodes), section replacement (4-10 nodes), global corrosion inhibitor injection; Maintenance timing: immediate execution, delayed to the next detection cycle, trigger additional detection.

[0224] In one embodiment, the discrete-continuous mixed space design is shown in Table 2:

[0225] Table 2

[0226]

[0227] 3) Reward function optimization:

[0228] r t = w1·(1-R risk ) + w2·(1-C cost / C max ) - w3I fail

[0229] Wherein, r t represents the reward function, w1, w2, w3 respectively represent the adaptive weight, w1, w2 dynamically adjust with the seasonal variation of power grid load, R risk represents the failure risk probability, C cost represents the cumulative maintenance cost, C max represents the maximum maintenance action cost, I fail represents the failure indication function.

[0230] The calculation of I fail is the case where the resistance of the steel sheet in the grounding grid exceeds the threshold value, and the determination criteria are as follows:

[0231]

[0232] wherein: d th =0.5Ω is the corrosion failure threshold, N is the number of steel sheets or measurement points in the grounding grid.

[0233] R risk is the grounding grid failure risk value, which is estimated by the corrosion rate as follows:

[0234]

[0235] v max is the maximum corrosion rate of the grounding grid, v phys is the current measured corrosion rate.

[0236] 4) Hierarchical decision-making mechanism:

[0237] Macro-strategy layer: uses the Proximal Policy Optimization (PPO) algorithm to plan the annual maintenance budget allocation; Micro-execution layer: reinforcement learning agents generate specific maintenance plans for corroded steel sheets.

[0238] (2) Intelligent decision-making body training, PPO algorithm implementation:

[0239] 1) Network architecture design:

[0240] Utilize the Actor-Critic dual network structure.

[0241] ① Actor network:

[0242] Input layer: state vector (dimension N+5); Hidden layer: 3 fully connected layers (512→256→128), ReLU activation; Output layer: discrete action: Softmax output maintenance type probability distribution; Continuous action: Tanh output range [-1, 1], linearly mapped to the actual parameter interval.

[0243] ② Critic network:

[0244] Structure same as Actor network, output is state value V(s) scalar.

[0245] 2) Training process:

[0246] Run K=8 Worker threads in parallel in the digital twin environment, for each thread:

[0247] Reset the environment to a random initial state (different corrosion distribution, soil parameters); Execute the current policy π θ Interact for T=2048 steps, record the trajectory τ=<s t ,a t ,r t ,s t+1 >.

[0248] The agent contains two networks, Actor network and Critic network, whose outputs are different.

[0249] Calculate advantage estimate A t :

[0250]

[0251] Where λ = 0.99, γ = 0.95 are discount factors, mainly considering the inflation of cost over time. V(s t ) represents the state value at the current time, V(s t+1 ) represents the state value at the next time. r t is the reward function, A t is the relative value estimate of the current maintenance action. During training, the smaller the deviation of the value estimate, the better, that is, the smaller the error δ t .

[0252] Policy optimization, the Actor network outputs the probability of adopting each maintenance action. During the training process, the probability of the action that brings more value to the system needs to be increased, and the action that has a negative impact on the system needs to be reduced. Therefore, advantage estimate A t is needed during training to judge the pros and cons of the current action.

[0253] Loss function calculation, policy loss (including clipping mechanism):

[0254] L clip (θ) = E t [ρ t (θ)A t , clip(ρ t (θ), 1-∈, 1+∈)A t ]

[0255] Where ρ t = π θ / π θ,old , π θ is the action probability output by the Actor neural network updated last time, π θ,old is the action probability output by the Actor network when interacting with the digital twin model, ∈ = 0.2 is the clipping size. In summary, the loss functions of the two neural networks are as follows:

[0256] Critic network value function loss:

[0257]

[0258] Actor network entropy regularization term:

[0259] L ENT (θ)=E t [-∑π θ (a|s t )logπ θ (a|s t )]

[0260] Total loss:

[0261] L total =L clip -c1L VF +c2L ENT

[0262] Where c1=0.5, c2=0.01, the role of c2 is to ensure that the probability of the actor outputting certain actions does not become 0, so that it has a very small probability value that may be selected to try, ensuring the exploratory nature of the neural network during the training process.

[0263] 3) Parameter update:

[0264] M=4 rounds of small batch updates are performed using the Adam optimizer (learning rate 3×10 -4 ) with a batch size of 64.

[0265] After completing 10 iterations, test the performance of the strategy in the verification environment, and evaluate the indicators: average reward, number of nodes exceeding corrosion, and unit cost maintenance efficiency. If the evaluation results improve by <1% for 3 consecutive times, trigger the early stop mechanism. Save the Pareto optimal strategy to the strategy library for online deployment and calling.

[0266] Figure 8 and Figure 9 shows a comparison chart of the performance of the agent and the conventional strategy according to the embodiments of the application.

[0267] (3) Online deployment and adaptive optimization:

[0268] 1) Digital twin-physical system synchronization:

[0269] Shadow mode operation: after the agent generates the maintenance strategy, it first simulates the execution of 100 Monte Carlo inferences in the digital twin; calculate the expected return variance, if σ 2 <5% then approve the execution, otherwise trigger manual review; the execution results are written to the SCADA system through the OPC UA protocol to drive the field maintenance equipment.

[0270] 2) Continuous learning mechanism

[0271] Incremental training trigger conditions: environment parameter drift detection (KL divergence > 0.1); strategy performance degradation (average reward decrease by more than 15%).

[0272] Data backfeeding: Add on-site maintenance records (action-effect pairs) to the experience pool and start a round of fine-tuning training every 24 hours.

[0273] This application discloses an intelligent operation and maintenance system for grounding grids based on digital twins. Addressing the problems of traditional grounding grid corrosion assessment methods, such as lag, reliance on manual experience in operation and maintenance decisions, and difficulty in achieving full life-cycle economic optimization, this application establishes a dynamic simulation model of grounding grid corrosion status by constructing a multi-dimensional data fusion digital twin. Specifically, a spatiotemporal distribution database is established based on corrosion data from historical inspection reports. Combining this with corrosion rate characteristic parameters of galvanized steel sheets obtained from accelerated corrosion tests of grounding grid steel sheets in the laboratory, a corrosion-stress electrochemical coupling model is constructed using finite element analysis. A decision network architecture with dual time scales is constructed using deep reinforcement learning algorithms, quantifying grounding grid health status, environmental parameters, and maintenance costs into the state space of a Markov decision process. A multi-objective reward function that balances equipment failure risk and economic losses is designed. The PPO algorithm is used for strategy optimization, ultimately forming a dynamically updatable maintenance decision model. Through iterative optimization via virtual-real interaction, this system achieves a corrosion trend prediction accuracy of over 80% and a 20%-30% reduction in full life-cycle maintenance costs, effectively solving the problem of proactive protection against hidden defects in power infrastructure.

[0274] The specific technical advantages of this application include:

[0275] (1) Improved prediction accuracy:

[0276] Experimental research on the electrochemical mechanism of steel sheets and statistical data show that the corrosion degree prediction error is ≤7% (more than 20% lower than the traditional single model).

[0277] The digital twin environment provides high-fidelity training data, reducing the generalization error of reinforcement learning strategies in real-world scenarios by 28%.

[0278] (2) Optimization of decision-making economy:

[0279] The total life cycle maintenance cost is reduced by 20%-30%, avoiding additional losses caused by over-maintenance and sudden failures;

[0280] A Pareto optimal tradeoff between safety and economy is achieved through a dynamic weighted reward function.

[0281] (3) Enhanced engineering applicability:

[0282] Supports access via mainstream industrial protocols (Modbus, OPC UA) and is compatible with existing power monitoring systems;

[0283] The 3D visualization interface reduces the decision-making and cognitive load of maintenance personnel, and shortens the average fault response time by 60%.

[0284] The embodiment of the present application also provides a computer device comprising each step of the ground grid intelligent operation and maintenance system based on digital twinning and capable of achieving the same technical effects. To avoid repetition, no further description is given here.

[0285] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article, or device that includes a list of elements not only includes those elements, but also includes other elements not explicitly listed, or inherent to such a process, method, article, or device. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or device that includes the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, but can also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted, or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0286] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not restrictive. Those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope protected by the claims.

Claims

1. A grounding grid intelligent operation and maintenance system based on digital twin, characterized in that, include: Corrosion evolution law modeling module, proxy model construction module, grounding grid digital twin model establishment module, and intelligent decision-making body construction module; The corrosion evolution law modeling module is used for: By using the time-series data of grounding grid resistance in long-term test reports and combining it with Faraday's law of electrolysis, a mapping relationship between the rate of resistance change and the mass loss of galvanized steel sheets was established. Based on accelerated corrosion tests of steel sheets in laboratory grounding grids, a multi-factor coupled corrosion kinetic equation is generated. This corrosion kinetic equation is used to predict the instantaneous corrosion rate of galvanized steel sheets based on field-measured soil environmental data. The corrosion loss of steel sheet mass is obtained by integrating over a specified time period. The proxy model construction module is used for: By combining the corrosion kinetic equation with the finite element electromagnetic-electrochemical joint simulation model, a dataset of soil environmental parameters and steel sheet mass loss is established. Based on the dataset, a physical information neural network is trained to construct a physical model for corrosion prediction. The corrosion prediction physical model is trained using a neural network regression algorithm. The input parameters include soil physicochemical parameters, grounding grid topology, and detected grounding grid resistance data. The output parameter is the probability distribution of the corrosion degree of galvanized steel sheets within a specified time span. The probability distribution of the corrosion degree of galvanized steel sheets within a specified time span is determined by the predicted instantaneous corrosion rate of the galvanized steel sheets. The corrosion prediction physical model is used to predict the corrosion degree of galvanized steel sheets in soil using soil environmental data simulated by Monte Carlo simulation. The grounding grid digital twin model building module is used for: Based on the aforementioned corrosion prediction physical model and combined with on-site environmental monitoring data, a digital twin model of the grounding grid is constructed. The intelligent decision-making module is used for: A Markov decision process is defined in the digital twin environment of the grounding grid. The parameters of the Markov decision process include: state space, action space, and reward function. Based on the feedback from the digital twin model of the grounding grid and the reward function, a near-end policy optimization algorithm is used to train an LSTM neural network to obtain an intelligent decision-making body. The decision-making body is then optimized through online fine-tuning to output a full life-cycle maintenance strategy for the grounding grid.

2. The system according to claim 1, characterized in that, Multiple factors include: The sensitivity weighting coefficients of soil pH, soil porosity, moisture content, soil resistivity, and the resistance of galvanized steel sheet to corrosion rate.

3. The system according to claim 1, characterized in that, The proxy model construction module is also used for: After training the corrosion prediction physical model using a neural network regression algorithm, the degree of corrosion of galvanized steel sheets at different locations of the grounding grid is calculated using the mapping relationship and the grounding grid resistance data in the field monitoring report. The soil environmental data measured on site and the grounding grid resistance data in the on-site monitoring report are input into the corrosion prediction physical model, and the predicted corrosion degree of the galvanized steel sheet is output. By comparing the stated degree of corrosion with the predicted degree of corrosion, the physical model for corrosion prediction is optimized.

4. The system according to claim 1, characterized in that, The corrosion evolution modeling module includes: The data acquisition unit is used to collect soil pH, moisture content, porosity and resistivity of the soil where the grounding grid is located in real time through ion-selective electrode sensor, TDR time-domain reflectometry moisture sensor and four-probe resistor, with a sampling interval of ≤60 minutes. The corrosion kinetics modeling unit is used to establish corrosion kinetic equations based on Faraday's law of electrolysis and the Arrhenius equation. The model validation unit is used to simulate different environmental combinations in a laboratory test chamber and adjust the soil multi-factor coupling effect function in the corrosion kinetic equation.

5. The system according to claim 4, characterized in that, The corrosion evolution modeling module also includes: The data processing unit is used to process the data collected by the data acquisition unit, using an anomaly detection algorithm based on isolated forest to remove noisy data, and using spatiotemporal kriging interpolation to fill in missing values.

6. The system according to claim 3, characterized in that, The proxy model construction module includes: A machine learning agent model unit is used to determine the input parameters and the output parameters. Based on the physical information neural network, a neural network regression algorithm is used to train the model and initially construct a physical model for corrosion prediction. The transfer learning optimization unit uses the weights of the initially constructed corrosion prediction physical model as initial values ​​and performs adaptive training using grounding grid data from the field monitoring report to ensure that the mean square error of the corrosion prediction physical model is ≤5%, thus obtaining the corrosion prediction physical model.

7. The system according to claim 6, characterized in that, The proxy model construction module also includes: The dynamic calibration unit is used to calculate the corrosion degree of galvanized steel sheets at different locations of the grounding grid, and output the predicted corrosion degree based on the corrosion prediction physical model. When the deviation between the corrosion degree and the predicted corrosion degree is ≥15%, the parameters of the corrosion prediction physical model are updated online based on the Kalman filter model.

8. The system according to claim 1, characterized in that, The intelligent decision-making module includes: State space construction unit, used to define the state space, wherein the state vector S of the state space is... t =[H1,…,H n C cost ,R risk E env ], where H i C represents the health index of the i-th galvanized steel sheet node, n represents the number of galvanized steel sheet nodes, and C cost R represents the cumulative maintenance cost. risk E represents the probability of failure risk. env Representing a set of environmental parameters, the formula for calculating the health index is: H i =1-d i / d th d i Indicates the degree of corrosion, d th The failure threshold is indicated, and the degree of corrosion is estimated using the corrosion prediction physical model. The action space design unit is used to define maintenance actions as a multi-dimensional discrete space, including three types of maintenance strategies and their combinations: maintenance scope, maintenance method, and maintenance timing. The reward function generation unit is used to calculate the reward function, which is: r t =w1(1-R risk )+w2(1-C cost / C max )-w3I fail Where, r t Let R represent the reward function, w1, w2, and w3 represent the adaptive weights, and R be the value of the reward function. risk C represents the probability of failure risk. cost C represents the cumulative maintenance cost. max Indicates the cost of the largest maintenance operation, I fail Represents the fault indication function; The hierarchical decision-making unit is used to adopt a multi-agent collaborative decision-making architecture. The upper-level agent plans the allocation of the annual maintenance budget, and the lower-level agent executes the specific maintenance operations corresponding to the budget allocation.

9. The system according to claim 1, characterized in that, Also includes: A grounding grid digital twin environment construction module for forming the grounding grid digital twin environment required by the intelligent decision-making body; The grounding grid digital twin environment construction module specifically includes: Develop a 3D visualization platform based on Unreal Engine; A two-way data channel is established; wherein, the two-way data channel includes an uplink and a downlink. The uplink maps sensor data to the digital twin model of the grounding grid through the MQTT protocol and drives the dynamic update of corrosion status. The downlink encodes the decision actions of the intelligent decision-making body into PLC control instructions and sends them to the field maintenance equipment through the gRPC interface. Deploy a lightweight TensorRT inference engine in the Python environment embedded in Unreal Engine to quantize and compress the erosion prediction agent model; The rendering thread and the physics calculation thread exchange data through shared memory and set priority strategies.

10. A computer device, characterized in that, Including the grounding grid intelligent operation and maintenance system based on digital twins as described in any one of claims 1 to 9.

Images