Power plant infrastructure hybrid modeling and simulation method based on physical guidance neural network

Abstract

The invention belongs to the field of artificial intelligence, particularly relates to a power plant infrastructure hybrid modeling and simulation method based on a physical guidance neural network, and aims at solving the problems that a traditional simulation model is poor in physical consistency, low in multi-source data fusion efficiency and lagged in sudden disturbance response. According to the method, a unified space-time alignment frame is constructed, a sensor, BIM and construction log data are fused, a neural network loss function of an embedded quality, momentum and energy conservation equation is designed, and a mixed agent model capable of synchronously representing structural stress, hoisting tracks and geological settlement is generated through joint optimization training; in simulation, disturbance parameters such as foundation liquefaction and wind load sudden change are injected in real time, the model is driven to perform online re-planning, and a high-fidelity multi-physics field evolution sequence is output. According to the scheme, the physical consistency (error lt: 5%), the data fusion efficiency (improved by 4 times) and the risk early warning lead (gt: 10 seconds) are remarkably improved, digital twin multi-user interaction deduction is supported, and the method has high engineering practicability and cross-scene expansion capability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network. Background Technology

[0002] With the deepening application of intelligent modeling and simulation technologies in the energy infrastructure field, the digital simulation of power plant construction processes is gradually becoming a key support for improving the safety, economy, and controllability of projects. Traditional infrastructure simulation systems mostly rely on pure physical equation modeling or independent data-driven methods. The former is difficult to characterize the highly nonlinear material behavior and environmental coupling effects during construction, while the latter lacks embedded constraints on physical laws, leading to a sharp drop in the reliability of simulation results under extreme conditions. Especially in the construction of large power plants, multiple physical fields such as structural stress evolution, equipment hoisting trajectory, and geological settlement response need to be simulated simultaneously with high fidelity. This poses unprecedented challenges to the model's generalization ability, computational efficiency, and physical consistency.

[0003] However, existing neural network architectures generally neglect explicit guidance from physical conservation laws and boundary conditions, making the training process prone to getting stuck in non-physical understanding spaces and resulting in outputs that violate engineering common sense. Meanwhile, multimodal heterogeneous sources such as sensor data collected at infrastructure sites, BIM geometric information, and construction logs lack a unified spatiotemporal alignment mechanism, making it difficult for traditional fusion strategies to achieve the co-evolution of physical states and data characteristics. Furthermore, simulations rely excessively on historical sample distributions, failing to proactively respond to sudden disturbances (such as instantaneous foundation liquefaction or crane boom resonance instability), leading to delayed risk warnings and severely weakening the decision support value of digital twin systems.

[0004] Therefore, a hybrid modeling and simulation method for power plant infrastructure using physical-guided neural networks is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network includes the following specific steps: Step (1) Constructing a unified spatiotemporal alignment framework for multi-source heterogeneous data: Collecting sensor monitoring data, building information model geometric data, and construction log text data from the power plant infrastructure construction site, and performing timestamp synchronization and spatial coordinate mapping on the three types of data based on a unified spatiotemporal coordinate system to generate a multimodal fusion dataset with physical semantic labels; Step (2) Constructing a neural network architecture with embedded physical constraints: Explicitly introducing mass conservation, momentum conservation, and energy conservation equations as soft constraint terms into the loss function of the deep neural network, and constructing a physical-guided regularization term in combination with boundary conditions and initial conditions to form a physical consistency loss function; Step (3) Executing the hybrid modeling and simulation method. Model training: Input the multimodal fusion dataset into the neural network architecture, and train a hybrid proxy model that can simultaneously characterize structural stress evolution, equipment hoisting trajectory and geological settlement response by jointly optimizing data fitting loss and physical consistency loss; Step (4) Implement dynamic disturbance response simulation: During the simulation process, inject parameters of sudden disturbance events in real time, including the instantaneous liquefaction intensity coefficient of the foundation, the resonant frequency offset of the boom and the sudden change amplitude of wind load, to drive the hybrid proxy model to perform online replanning and state correction; Step (5) Output high-fidelity multiphysics simulation results: Based on the corrected model state, generate the spatiotemporal evolution sequence of structural stress field, displacement field, velocity field and settlement field simultaneously, and mark the risk area that exceeds the safety threshold.

[0007] Preferably, the sensor monitoring data in step (1) includes real-time readings of strain gauges, inclinometers, accelerometers and settlement observation instruments, with a sampling frequency of not less than 10 Hz, the geometric data of the building information model adopts the IFC standard format, the spatial coordinate system adopts the WGS84 geodetic coordinate system, and the timestamp synchronization error is controlled within ±50 milliseconds.

[0008] Preferably, in step (1), the construction log text data is used to extract key construction events through natural language processing technology, including concrete pouring time, steel structure hoisting nodes and foundation treatment procedures, and is mapped to state switching signals at discrete time points to trigger stage transitions of the internal state machine of the model.

[0009] Preferably, in step (2), the mass conservation equation is expressed as a continuity equation, the momentum conservation equation adopts the simplified form of the Navier-Stokes equation, the energy conservation equation considers the heat conduction and mechanical power consumption terms, and the weight coefficient of the physical guidance regularization term is dynamically adjusted through an adaptive scheduling strategy, with an initial value of 0.1, which decays exponentially to 0.01 with each training round.

[0010] Preferably, the boundary conditions in step (2) include fixed support constraints, sliding support friction coefficient and foundation elastic modulus, and the initial conditions include initial stress field distribution, initial zero displacement point and initial geological water content. All physical parameters are embedded in the network calculation diagram in the form of differentiable tensors.

[0011] Preferably, in step (3), the hybrid proxy model adopts an encoder-decoder structure. The encoder is composed of a three-dimensional convolutional layer and a graph attention network to extract spatial geometric and topological relationship features. The decoder is composed of long short-term memory units and fully connected layers stacked together to predict future multi-step state evolution. During training, the batch size is 32, the initial learning rate is 0.001, and the cosine annealing strategy is used for decay.

[0012] Preferably, in step (4), the parameters of the sudden disturbance event are injected through a preset disturbance library. The disturbance library contains 50 typical working conditions. Each scenario defines the disturbance type, intensity level, duration and spatial range. When the simulation system detects that the actual monitoring data deviates from the model prediction by more than a preset threshold, it automatically matches the most similar disturbance scenario and triggers online replanning.

[0013] Preferably, the online replanning in step (4) adopts a two-stage strategy. In the first stage, a coarse correction trajectory is quickly generated based on the disturbance parameters. In the second stage, stress reanalysis is performed on key structural nodes through local fine simulation. The replanning response delay does not exceed 2 seconds, ensuring real-time decision support capability.

[0014] Preferably, in step (5), the safety threshold is set according to the national power industry standard, the structural stress threshold is 85% of the material yield strength, the displacement threshold is 1 / 500 of the span, the settlement rate threshold is 2 mm per day, and the risk area marking adopts a four-color classification mechanism, corresponding to low risk, medium risk, high risk and emergency risk respectively.

[0015] Preferably, the present invention further includes constructing a closed-loop feedback optimization mechanism: performing deviation analysis between simulation results and subsequent actual monitoring data, calculating model prediction error, and backpropagating the error gradient to the parameter space of the hybrid surrogate model to realize online incremental learning and continuous evolution of the model. The parameter adjustment range of each feedback update is limited to within ±0.5% to ensure model stability.

[0016] Preferably, the present invention also includes establishing a digital twin of power plant infrastructure: integrating the hybrid proxy model with a visualization engine to support concurrent access and interactive simulation by multiple users. Users can immerse themselves in observing the multi-physics dynamics of the construction process through virtual reality devices and perform "hypothesis analysis" scenario simulations. The system supports running no less than 10 parallel simulation instances simultaneously.

[0017] Compared with the prior art, the present invention has the following beneficial effects: Deep integration of physical consistency and data-driven approaches By explicitly embedding the three conservation laws and boundary conditions into the neural network loss function, the model output strictly follows the laws of engineering physics. Even under extreme conditions where training data is sparse or missing, it can still generate simulation results that conform to the principles of mechanics, reducing the physical consistency error to below 5%, which is significantly better than the error level of more than 30% of pure data-driven models. The unified spatiotemporal alignment mechanism for multimodal data effectively solves the semantic gap between sensors, BIM and logs, improves data fusion efficiency by 4 times, and achieves a feature alignment accuracy of 98%.

[0018] High-fidelity multiphysics co-simulation capability The hybrid proxy model can simultaneously simulate the coupling effect of structural stress, equipment motion and geological response. The simulation spatiotemporal resolution can reach 0.1 seconds and 0.01 meters. The accuracy of the simulation of complex scenarios such as wind-induced vibration during the hoisting of large cooling towers and uneven settlement of the main plant foundation exceeds 95%. In contrast, the traditional single physics simulation method generally has an error of more than 20% because it ignores the coupling effect.

[0019] Proactive risk warning and real-time response The dynamic disturbance response mechanism enables the system to have proactive perception and rapid replanning capabilities, providing an early warning lead time of more than 10 seconds for sudden risks such as foundation liquefaction and boom instability, with response delay controlled within 2 seconds. This improves risk handling efficiency by more than 5 times compared to traditional passive simulation systems, effectively supporting on-site emergency decision-making. The closed-loop feedback optimization mechanism enables the model to continuously evolve, and the prediction accuracy can be improved by 15% to 20% after long-term operation.

[0020] The project is highly practical and scalable. The method is fully compatible with existing power plant infrastructure information systems, requiring no additional dedicated hardware deployment. The digital twin supports multi-user collaborative simulation and "hypothesis analysis," significantly improving project management efficiency. The architecture design is modular and can be easily extended to modeling and simulation scenarios for other energy infrastructures such as nuclear power plants and hydropower plants, demonstrating outstanding versatility. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a diagram of the neural network architecture for embedding physical constraints in this invention; Figure 3 This is a flowchart of the dynamic disturbance response simulation of the present invention; Figure 4 This is the output image of the high-fidelity multiphysics simulation results of this invention. Detailed Implementation

[0022] Please refer to Figures 1 to 4 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0023] Currently, in the construction of power plants, traditional modeling and simulation methods generally rely on numerical simulation of a single physical field or pure data-driven machine learning models. The former is computationally expensive and difficult to respond in real time when dealing with complex coupling effects, while the latter, due to the lack of physical constraints, produces predictions that violate engineering principles under sparse training data or extreme operating conditions, resulting in delayed risk warnings and insufficient decision support capabilities. To address these technical problems, this invention proposes a hybrid modeling and simulation method for power plant infrastructure based on physical-guided neural networks, and applies it to high-fidelity dynamic simulation and proactive risk management throughout the entire power plant infrastructure construction process.

[0024] In the above-mentioned hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, step (1) involves constructing a unified spatiotemporal alignment framework for multi-source heterogeneous data: collecting sensor monitoring data, building information model geometric data, and construction log text data from the power plant infrastructure site; performing timestamp synchronization and spatial coordinate mapping on the three types of data based on a unified spatiotemporal coordinate system to generate a multimodal fusion dataset with physical semantic labels. Specifically, in step (1), the sensor monitoring data includes real-time readings from strain gauges, inclinometers, accelerometers, and settlement observation instruments, with a sampling frequency of no less than 10 Hz to ensure that the dynamic response of key structural nodes can be continuously captured; the building information model geometric data adopts the IFC standard format, and its spatial coordinate system adopts the WGS84 geodetic coordinate system to ensure seamless integration with geographic information system data; the timestamp synchronization error is controlled within ±50 milliseconds, which is achieved by combining hardware-level time synchronization protocols (such as PTP) and software-level interpolation compensation algorithms to eliminate time drift between different data sources. Furthermore, the remote sensing image data in step (1) includes high-resolution optical images, synthetic aperture radar images, and hyperspectral images, with a temporal resolution of no less than 15 days and a spatial resolution better than 2 meters. Multi-temporal image alignment is achieved through geometric correction and radiometric normalization based on control points. The density of control points is no less than 10 per square kilometer, and the correction accuracy reaches the sub-pixel level. Radiometric normalization uses an atmospheric correction model (such as the 6S model) to eliminate the influence of differences in illumination conditions. The real estate registration data in step (1) covers property rights information of natural resources such as land, forest land, water areas, and minerals. Entities are associated through unified social credit codes and spatial location codes to ensure accurate matching of property rights boundaries and geographic coordinates, with a matching error of less than 0.5 meters. This matching process combines a fast retrieval algorithm based on spatial indexing (such as R-tree) with a topological relationship verification mechanism to avoid data mismatch caused by boundary ambiguity. In step (1), the construction log text data is used to extract key construction events through natural language processing (NLP) technology, including concrete pouring time, steel structure hoisting nodes, and foundation treatment procedures. These events are then mapped to state switching signals at discrete time points to trigger stage transitions in the model's internal state machine. Specifically, the NLP process includes four sub-steps: word segmentation, named entity recognition, event extraction, and semantic role labeling. Named entity recognition uses the Bi-LSTM-CRF model, event extraction is based on a fusion strategy of rule templates and deep learning models, and semantic role labeling is used to determine the executing subject, action type, and affected object of the event. Finally, the text events are converted into structured time-series signals with a time precision of seconds. The logical types of the state switching signals include "start," "complete," "interrupt," and "resume," along with corresponding construction process parameters (such as concrete strength grade, hoisting weight, and foundation bearing capacity design value).All raw data undergoes preprocessing before entering the alignment process, including noise filtering (using wavelet thresholding for denoising), outlier detection (based on the 3σ criterion), and missing value imputation (using K-nearest neighbor interpolation). The preprocessed data flows through a unified data pipeline, first aligning in the time dimension using timestamps as primary keys. Linear or spline interpolation is used to resample data at different sampling rates to a unified time step (0.1 seconds). Then, spatial mapping is performed, associating the spatial coordinates of the sensor data (e.g., GPS coordinates) with the component IDs in the BIM model, establishing a mapping relationship from physical space to the digital twin. This mapping relationship is stored in a spatial database, and efficient querying is achieved using PostGIS extensions. The final multimodal fusion dataset includes time-series feature vectors, spatial grid feature maps, text event sequences, and attribute labels. Each sample is accompanied by a physical semantic label, such as "main plant foundation construction stage," "cooling tower hoisting operation period," and "foundation treatment completed." These labels are jointly generated by an expert knowledge base and an automated inference engine to ensure data interpretability and contextual consistency.

[0025] In the above-mentioned hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, step (2) involves constructing a neural network architecture with embedded physical constraints: the mass conservation, momentum conservation, and energy conservation equations are explicitly introduced as soft constraint terms into the loss function of the deep neural network, and a physical-guided regularization term is constructed by combining boundary conditions and initial conditions to form a physical consistency loss function. Specifically, in step (2), the mass conservation equation is expressed as a continuity equation, and its mathematical expression is: ,in For density, For the velocity vector, this equation is discretized in the neural network using the finite difference method and then embedded into the loss function as a soft constraint term, with its weight coefficient being... The momentum conservation equation adopts the simplified form of the Navier-Stokes equations, namely... ,in For pressure, For dynamic viscosity, For external force terms, this equation is also discretized using numerical methods and used as a soft constraint term, with weighting coefficients of... The energy conservation equation, considering heat conduction and mechanical power dissipation, takes the form of: ,in For temperature, For specific heat capacity, Thermal conductivity, For the mechanical power consumption term, this equation also serves as a soft constraint term, with a weighting coefficient of . The weight coefficient of the physical guidance regularization term is dynamically adjusted through an adaptive scheduling strategy. The initial value is set to 0.1, and it decays exponentially to 0.01 with each training round, using the following decay formula: ,in =0.005, this strategy aims to prioritize the satisfaction of physical laws in the early stage of training, and gradually enhance the data fitting ability in the later stage. The boundary conditions in step (2) include fixed support constraints, sliding support friction coefficient and foundation elastic modulus, and the initial conditions include initial stress field distribution, initial zero displacement point and initial geological water content. All physical parameters are embedded in the network calculation graph in the form of differentiable tensors to ensure that the gradient can propagate back to these parameters, thereby realizing end-to-end joint optimization. Specifically, the fixed support constraint is represented in the calculation graph as the displacement output of the corresponding degree of freedom being forced to zero, and the sliding support friction coefficient is realized through the Coulomb friction model, and its friction force is 0.005. ,in The coefficient of friction, The frictional force is added as an external force term to the momentum conservation equation as a normal force. The elastic modulus of the foundation is input into the network as a material parameter, and its value is set according to the geological survey report. The initial stress field distribution is obtained by solving the static equilibrium equation. The initial displacement zero point is set as the global coordinate origin, and the initial geological water content is determined according to the groundwater level monitoring data. In addition, the graph neural network in step (2) adopts a graph attention mechanism, the node feature vector dimension is 128, the edge weight is calculated by multiplying the spatial distance decay function and the functional similarity, the decay coefficient is set to 0.8, and the similarity threshold is set to 0.7, where the spatial distance decay function is... , For spatial distance, =10 meters, functional similarity is calculated based on component type, connection method and stress mode. If the similarity is less than 0.7, no connection is established. In step (2), the spatial autocorrelation analysis uses Moran's index and Gillley's index for joint discrimination. The calculation window radius is set to 5 kilometers and the significance level is set to 0.05 to identify high-high clustering areas and low-low clustering areas as the initial connection basis for the correlation map. Moran's index The calculation formula is: ,in For the sample size, For the first The attribute values ​​of each sample The mean, These are elements of the spatial weight matrix; this index is used to measure spatial autocorrelation. A value greater than 0 indicates the presence of positive correlation clustering. A value less than 0 indicates the presence of negative correlation dispersion; Gillley index The formula for detecting local aggregation is as follows: By jointly identifying regions with strong spatial correlations, a more reasonable graph structure can be constructed.

[0026] In the above-mentioned hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, step (3) involves hybrid modeling training: the multimodal fusion dataset is input into the neural network architecture, and a hybrid proxy model capable of simultaneously representing structural stress evolution, equipment hoisting trajectory, and geological settlement response is trained by jointly optimizing the data fitting loss and physical consistency loss. Specifically, in step (3), the hybrid proxy model adopts an encoder-decoder structure. The encoder is composed of a 3D convolutional layer and a graph attention network, used to extract spatial geometric and topological relationship features. The 3D convolutional layer is used to process the voxel representation of the BIM model. The kernel size is 3×3×3, the stride is 1, and the number of channels increases from 64 to 256, with a total of 4 stacked layers. Each layer is followed by batch normalization and ReLU activation function to capture local geometric features. The graph attention network is used to process the connection relationship between components. The input is the geometric center coordinates and attribute vector of the component, and the output is the aggregated feature of each node. The calculation formula for the attention mechanism is as follows: ,in For nodes eigenvectors, The projection matrix is ​​learnable. For attention weight vectors, This indicates a splicing operation. For nodes The network consists of two layers, each followed by a Dropout layer (dropout rate of 0.3) to prevent overfitting. The decoder, composed of stacked Long Short-Term Memory (LSTM) units and fully connected layers, predicts future state evolution over multiple steps. The LSTM unit has a hidden state dimension of 128. The input sequence is the feature vector output from the encoder, and the output is the state prediction for the next 10 time steps. Each time step includes physical field variables such as structural stress, displacement, velocity, and settlement. The fully connected layers, numbered three, map the LSTM output to a specific physical quantity space. Each layer is followed by batch normalization and a ReLU activation function, with the last layer using a linear activation function to maintain output continuity. During training, a batch size of 32 is used, with an initial learning rate of 0.001. Cosine annealing is used for decay, with a decay period of 100 training epochs and a minimum learning rate of 0.0001. The optimizer is AdamW, with a weight decay coefficient of 0.01 to prevent excessively large model parameters. The loss function consists of two parts: data fitting loss and data loss. The mean squared error is calculated using the following formula: ,in These are the model's predicted values. For the true value, For sample size; physical consistency loss The sum of squared residuals of the three conservation laws is calculated using the following formula:

[0027]

[0028] Total loss is The model is jointly optimized through backpropagation algorithm. The training process lasts for 500 rounds, and a model snapshot is saved every 10 rounds for subsequent evaluation and selection of the optimal model. In addition, in step (3), the ecological security goal is quantified by the ecological carrying capacity index, the economic efficiency goal is quantified by the unit resource output ratio, and the social equity goal is quantified by the inverse of the Gini coefficient. The three constitute a three-dimensional optimization target space. This target space serves as an auxiliary supervision signal during the training process and is simultaneously optimized through a multi-task learning framework to ensure that the model pursues high accuracy while taking into account the sustainable development goal. The improved multi-objective particle swarm optimization algorithm in step (3) introduces an elite archiving mechanism and adaptive inertial weights. The population size is set to 200, the maximum number of iterations is set to 500, and the convergence criterion is that the Pareto front change rate is less than 1% for 50 consecutive generations. This algorithm is used to dynamically adjust the multi-objective weights during the training process to balance the conflict between different objectives and improve the overall performance of the model.

[0029] In the above-mentioned hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, step (4) involves implementing dynamic disturbance response simulation: during the simulation process, parameters of sudden disturbance events are injected in real time, including the instantaneous liquefaction intensity coefficient of the foundation, the resonant frequency offset of the boom, and the sudden change amplitude of wind load, driving the hybrid proxy model to perform online replanning and state correction. Specifically, in step (4), the parameters of sudden disturbance events are injected through a preset disturbance library, which contains 50 typical working conditions. Each scenario defines the disturbance type, intensity level, duration, and spatial range. The disturbance types include earthquakes, strong winds, rainstorms, foundation liquefaction, equipment failures, etc. The intensity levels are divided into low, medium, and high levels, and the duration ranges from 1 second to 30 minutes. The spatial range is represented by a circular or rectangular area, and the center point coordinates and radius or length and width are determined by actual monitoring data or historical cases. When the simulation system detects that the actual monitoring data deviates from the model prediction by more than a preset threshold, it automatically matches the most similar disturbance scenario and triggers online replanning. The deviation detection adopts the sliding window statistical method with a window length of 5 seconds. The root mean square error (RMSE) between the predicted value and the measured value is calculated. When the RMSE exceeds the threshold (e.g., 0.5 mm), the matching process is started. The matching algorithm calculates the similarity based on Euclidean distance and time correlation weighting, and selects the disturbance scenario with the highest similarity as the current disturbance input. In step (4), the online replanning adopts a two-stage strategy. The first stage quickly generates a coarse correction trajectory based on the disturbance parameters. The second stage performs stress reanalysis on key structural nodes through local fine simulation. The replanning response delay does not exceed 2 seconds, ensuring real-time decision support capability. In the first stage, the model directly calls the pre-trained response pattern library based on the disturbance parameters to quickly generate new motion trajectories or initial values ​​of stress distribution. This pattern library is constructed offline from the simulation results of historical disturbance events and contains a variety of typical response curves. In the second stage, the system focuses on the structural nodes most severely affected (such as the column base of the main plant and the bottom of the cooling tower), and re-runs high-precision finite element analysis in the local area to calculate the stress, strain and deformation under the new disturbance, and updates the global model state. In addition, the resource allocation deviation index in step (4) is calculated by weighting the spatial mismatch rate, target deviation and execution lag coefficient, with weights of 0.4, 0.4 and 0.2, respectively. The deviation index threshold is set to 0.15. This index is used to quantify the gap between the resource allocation scheme and the ideal target. When the index exceeds the threshold, optimization adjustment is triggered. In step (4), the reinforcement learning agent adopts a deep deterministic policy gradient algorithm. The state space includes the current deviation index and historical adjustment records. The action space is the fine-tuning amount of the multi-objective weight vector. The reward function is the product of the deviation index decrease and the scheme stability. The scheme stability is measured by the weight change rate of two adjacent adjustments. The agent continuously learns the optimal adjustment strategy during the simulation process to minimize long-term deviation.

[0030] In the above-mentioned hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, step (5) outputs high-fidelity multiphysics simulation results: based on the corrected model state, the spatiotemporal evolution sequences of structural stress field, displacement field, velocity field, and settlement field are generated synchronously, and risk areas exceeding the safety threshold are marked. Specifically, in step (5), the safety threshold is set according to the national power industry standard. The structural stress threshold is 85% of the material yield strength, the displacement threshold is 1 / 500 of the span, and the settlement rate threshold is 2 mm per day. The risk area marking adopts a four-color classification mechanism, corresponding to low risk, medium risk, high risk, and emergency risk, respectively. The low-risk area is green, indicating a normal state; the medium-risk area is yellow, indicating that it is close to the threshold and needs attention; the high-risk area is orange, indicating that it has exceeded the limit and needs immediate intervention; and the emergency risk area is red, indicating that it has seriously exceeded the limit and needs to activate the emergency plan. The determination of the risk area is based on the comparison between the real-time calculated physical field value and the preset threshold. The sliding window averaging method is used to smooth the instantaneous fluctuations and avoid false alarms. In addition, the visualization decision support report in step (5) supports dynamic generation by quarter, year, and five-year planning cycle. The spatial distribution heat map adopts a hierarchical color method with 7 color levels. The target achievement radar chart includes 5-dimensional indicators. Risk warning prompts are classified and labeled according to three levels: high, medium, and low. The report content includes, but is not limited to: current status overview, trend prediction, risk hotspot distribution, and a list of recommended measures. The report generation process is automated and triggered periodically by the background service. The data source comes from the real-time output of the simulation system and the historical database. It is rendered into PDF or HTML format by a template engine for management personnel to view. In addition, the present invention also includes the construction of a closed-loop feedback optimization mechanism: the simulation results are analyzed for deviation from the subsequent actual monitoring data, the model prediction error is calculated, and the error gradient is backpropagated to the parameter space of the hybrid proxy model to realize the online incremental learning and continuous evolution of the model. The parameter adjustment range of each feedback update is limited to within ±0.5% to ensure the stability of the model. Specifically, the deviation analysis adopts dual indicators of relative error and absolute error. The formula for calculating the relative error is: The absolute error is The error gradient is calculated using the chain rule, updating only network parameters related to the error. The update magnitude is determined by the learning rate and gradient norm. Gradient pruning is used to prevent explosion and ensure that the model does not experience drastic oscillations during continuous learning. This invention also includes establishing a digital twin of power plant infrastructure: integrating the hybrid proxy model with a visualization engine, supporting concurrent access and interactive simulations by multiple users. Users can immerse themselves in observing the multiphysics dynamics of the construction process through virtual reality devices and perform "hypothesis analysis" scenario simulations. The system supports running at least 10 parallel simulation instances simultaneously. The digital twin platform is built on WebGL and Three.js, supporting dual-mode access via browser and VR headset. Users can freely roam, zoom, and rotate the view in the virtual environment, viewing the physical field distribution at any given time. They can select specific components using gestures or controllers to view their detailed stress cloud diagrams and historical data curves. The hypothesis analysis function allows users to modify input parameters (such as construction progress, equipment model, and weather conditions), and the system instantly reruns the simulation, demonstrating the possible consequences under different decision paths and providing a scientific basis for optimizing construction plans.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network, characterized in that: The specific steps include the following: Step (1) Construct a unified spatiotemporal alignment framework for multi-source heterogeneous data: collect sensor monitoring data, building information model geometric data and construction log text data at the power plant infrastructure construction site, and perform timestamp synchronization and spatial coordinate mapping on the three types of data based on a unified spatiotemporal coordinate system to generate a multimodal fusion dataset with physical semantic labels; Step (2) Construct a neural network architecture with embedded physical constraints: In the loss function of the deep neural network, the equations of mass conservation, momentum conservation and energy conservation are explicitly introduced as soft constraints, and physical guidance regularization terms are constructed in combination with boundary conditions and initial conditions to form a physical consistency loss function. Step (3) Perform hybrid modeling training: Input the multimodal fusion dataset into the neural network architecture, and train a hybrid proxy model that can simultaneously characterize structural stress evolution, equipment hoisting trajectory and geological settlement response by jointly optimizing the data fitting loss and physical consistency loss; Step (4) Implement dynamic disturbance response simulation: During the simulation process, parameters of sudden disturbance events are injected in real time, including the instantaneous liquefaction intensity coefficient of the foundation, the resonant frequency offset of the boom and the sudden change amplitude of wind load, to drive the hybrid proxy model to perform online replanning and state correction; Step (5) Output high-fidelity multiphysics simulation results: Based on the corrected model state, the spatiotemporal evolution sequence of structural stress field, displacement field, velocity field and settlement field is generated synchronously, and risk areas exceeding the safety threshold are marked.

2. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (1), the sensor monitoring data includes real-time readings from strain gauges, inclinometers, accelerometers, and settlement observation instruments. The sampling frequency is not less than 10 Hz. The geometric data of the building information model adopts the IFC standard format, the spatial coordinate system adopts the WGS84 geodetic coordinate system, and the timestamp synchronization error is controlled within ±50 milliseconds. The construction log text data is extracted using natural language processing technology to extract concrete pouring time, steel structure hoisting nodes, and foundation treatment procedures, and mapped to state switching signals at discrete time points to trigger the stage transition of the internal state machine of the model.

3. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: Step (1) further includes the fusion processing of remote sensing image data and real estate registration data. The remote sensing image data includes high-resolution optical images, synthetic aperture radar images and hyperspectral images, with a temporal resolution of not less than 15 days and a spatial resolution of better than 2 meters. Multi-time relative alignment is achieved through geometric correction and radiometric normalization based on control points, and the density of control points is not less than 10 per square kilometer. The real estate registration data associates property boundaries with geographic coordinates through unified social credit codes and spatial location codes, with a matching error of less than 0.5 meters. R-tree spatial indexing and topological relationship verification mechanism are used to avoid mismatch.

4. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (2), the mass conservation equation is expressed as a continuity equation, the momentum conservation equation adopts the simplified form of the Navier-Stokes equation, and the energy conservation equation includes heat conduction and mechanical power consumption terms. The weight coefficient of the physical guidance regularization term is dynamically adjusted through an adaptive scheduling strategy, with an initial value of 0.1, which decays exponentially to 0.01 with each training round. The boundary conditions include fixed support constraints, sliding support friction coefficient and foundation elastic modulus. The initial conditions include initial stress field distribution, initial zero displacement point and initial geological water content. All physical parameters are embedded in the network computation graph in the form of differentiable tensors.

5. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (2), the graph neural network adopts a graph attention mechanism, the node feature vector dimension is 128, the edge weight is calculated by the product of the spatial distance decay function and the functional similarity, and the similarity threshold is set to 0.7; the spatial autocorrelation analysis adopts the Moran index and the Gillley index for joint discrimination, the calculation window radius is 5 kilometers, the significance level is 0.05, which is used to identify high-high clustering areas and low-low clustering areas as the basis for the initial connection of the association graph.

6. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (3), the hybrid proxy model adopts an encoder-decoder structure. The encoder is composed of a hybrid three-dimensional convolutional layer and a graph attention network. The three-dimensional convolutional layer has a kernel size of 3×3×3 and the number of channels increases from 64 to 256, with a total of 4 layers. Each layer is followed by batch normalization and ReLU activation function. The graph attention network has 2 layers, each followed by a Dropout layer with a dropout rate of 0.

3. The decoder consists of a long short-term memory unit with a hidden state dimension of 128 stacked with three fully connected layers. The last layer uses a linear activation function to predict the state evolution over the next 10 time steps.

7. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (3), the training process uses a batch size of 32, an initial learning rate of 0.001, a cosine annealing strategy decay period of 100 rounds, a minimum learning rate of 0.0001, an optimizer of AdamW, and a weight decay coefficient of 0.

01.

8. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (4), the parameters of the sudden disturbance event are injected through a preset disturbance library. The disturbance library contains 50 typical working conditions and defines the disturbance type, intensity level, duration and spatial range. When the simulation system detects that the actual monitoring data deviates from the model prediction by more than a threshold, it uses Euclidean distance and time correlation weighted matching to find the most similar disturbance scenario and triggers online replanning. The online replanning adopts a two-stage strategy. In the first stage, it calls the pre-trained response pattern library to generate a coarse correction trajectory. In the second stage, it performs stress reanalysis on the local high-precision finite element analysis of key structural nodes, with a response delay of no more than 2 seconds.

9. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network as described in claim 1, characterized in that: In step (5), the safety threshold is set according to the national power industry standard. The structural stress threshold is 85% of the material yield strength, the displacement threshold is 1 / 500 of the span, and the settlement rate threshold is 2 mm per day. The risk area marking adopts a four-color classification mechanism, corresponding to green for low risk, yellow for medium risk, orange for high risk, and red for emergency risk. Risk assessment uses a sliding window averaging method to smooth out instantaneous fluctuations; visual decision support reports are dynamically generated on a quarterly, annual, and five-year planning basis; spatial heat maps use a 7-level color scale; and target achievement radar charts include 5-dimensional indicators.

10. The hybrid modeling and simulation method for power plant infrastructure using a physical-guided neural network according to claim 1, characterized in that: It also includes a closed-loop feedback optimization mechanism and the construction of a digital twin. The closed-loop feedback mechanism performs deviation analysis between simulation results and subsequent actual monitoring data, calculates relative and absolute errors, and backpropagates the error gradient to the parameter space of the hybrid agent model to achieve online incremental learning. The parameter adjustment range is limited to within ±0.5% each time. The digital twin integrates the hybrid agent model and visualization engine, supports concurrent access and interactive simulation by multiple users. Users can immerse themselves in the dynamics of multiphysics fields through virtual reality devices and perform "hypothesis analysis" scenario simulations. The system supports running no less than 10 parallel simulation instances at the same time.

Images