Power plant operation management method based on three dimensions

Abstract

The invention discloses a power plant operation management method based on three dimensions, and belongs to the technical field of power plant operation management, and the management method comprises the following specific steps: Q1, building a power plant three-dimensional model and dividing functional regions according to the civil engineering structure and equipment layout of a power plant; q2, accessing video monitoring, personnel positioning, equipment data and operation information into the three-dimensional model, and performing information linkage; according to the method, the prediction accuracy is improved, the false report rate and the missing report rate are reduced, dynamic data can be analyzed in real time, a prediction result can be quickly generated, early warning can be given out, the deployment cost is reduced, the response efficiency of the system is improved, and the complexity of multi-model development and maintenance is reduced; resource waste or risk omission caused by local optimization can be effectively avoided, potential problems caused by improper resource scheduling are reduced, and the overall resource utilization rate and team cooperation efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power plant operation management, and in particular to a three-dimensional-based power plant operation management method. Background Art

[0002] As an important infrastructure for energy production, modern power plants have complex and high-risk operation and maintenance operations, involving multi-faceted collaboration of personnel, equipment, environment, etc. Traditional operation management relies on two-dimensional drawings, text information, and a single system, lacking real-time perception and intuitive display of the overall operation state, and it is difficult to comprehensively control the on-site operation state, optimize resource scheduling, and reduce operation risks. Especially in scenarios with frequent multi-disciplinary collaboration and cross-operation, it is difficult for managers to dynamically optimize resource allocation and scheduling, increasing operation risks and management costs. Therefore, constructing a three-dimensional-based power plant operation management method has become a solution, which can not only achieve visual display of operation states, but also provide a data basis for dynamic optimization and intelligent decision-making.

[0003] Existing three-dimensional-based power plant operation management methods have relatively high false alarm and missed alarm rates, and cannot analyze dynamic data in real time, reducing the response efficiency of the system and increasing the complexity of multi-model development and maintenance; in addition, existing three-dimensional-based power plant operation management methods are prone to resource waste or risk omission caused by local optimization, and the improper resource scheduling increases the probability of potential problems, reducing the overall resource utilization rate and team collaboration efficiency; therefore, we propose a three-dimensional-based power plant operation management method. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects existing in the prior art, and to propose a three-dimensional-based power plant operation management method.

[0005] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A three-dimensional-based power plant operation management method, and the specific steps of the management method are as follows:

[0007] Q1: Establish a three-dimensional model of the power plant and divide functional areas according to the civil structure and equipment layout of the power plant;

[0008] Q2: Connect video monitoring, personnel positioning, equipment data, and operation information to the three-dimensional model, and perform information linkage;

[0009] Q3: Simulate the equipment operation state, personnel operation behavior, and emergency situations, and analyze the risk points in different operation states through dynamic scene adjustment;

[0010] Q4: Predict equipment failures, personnel operation risks, and work order completion situations, issue early warnings in a timely manner, and optimize the operation paths of personnel and equipment;

[0011] Q5: Build a dynamic optimization model, analyze the benefits and risks of different strategies during the operation process, select the optimal strategy, and provide suggestions for cross-departmental collaboration.

[0012] Optionally, the specific steps for establishing the 3D model of the power plant and dividing the functional areas described in Q1 are as follows:

[0013] S101: Obtain the civil engineering drawings and equipment layout drawings of the power plant, use a laser scanner to scan the power plant site, capture the 3D spatial data of the buildings and equipment, generate point cloud data, and then align and fuse the drawing data and point cloud data;

[0014] S102: Remove the noise points through Gaussian filtering. After the noise points are removed, downsample the point cloud data using the meshing method. Through the Delaunay triangulation algorithm, convert the point cloud into a continuous 3D mesh model, and then smooth the mesh surface;

[0015] S103: According to the room information marked in the power plant building drawings, divide the 3D mesh model into different functional areas, attach attribute labels to each area according to the drawing data, compare the position and size of the 3D model with the building drawings, and verify the accuracy and integrity of the 3D model through the matching of on-site survey data and the model.

[0016] Optionally, the specific steps for analyzing the risk points in different operation states through dynamic scene adjustment described in Q3 are as follows:

[0017] S201: Set the simulation objectives, scene parameters, and initial state. Based on the physical characteristics and working principles of the equipment, construct a dynamic operation model of the equipment, and through path planning and task allocation, simulate the operation behavior of personnel to establish a personnel behavior model. Finally, establish an environmental dynamic model based on each physical process;

[0018] S202: According to the dynamic operation model of the equipment, the personnel behavior model, and the environmental dynamic model, define emergency events and event trigger conditions, analyze the changes in equipment status and personnel behavior, and calculate the emergency response time to simulate the emergency response;

[0019] S203: Record the time series data of equipment status, personnel behavior, and environmental changes. According to the recorded data of each group, evaluate the risk indicators based on the probability of equipment overload and the proportion of dangerous areas in the personnel path, and mark the location and risk level of the risk points;

[0020] S204: According to the simulation results, optimize the equipment operation status and the personnel operation path. Through the 3D power plant model after information linkage, conduct multiple iterative adjustments until the change value of the risk function value of the scenario converges to the optimal solution within the preset range, and output the operation process with reduced risk and the equipment operation parameters.

[0021] Optionally, the specific steps for predicting equipment failures, personnel operation risks, and work order completion described in Q4 are as follows:

[0022] S301: Obtain comprehensive equipment operation data, personnel behavior data, and work order records from the power plant historical database, including time series features and static features. Remove the missing values and outliers in the historical data, encode the non-numerical data, normalize the processed data to the interval [0, 1], extract features from the time series data, screen the feature data according to the task objective, and then divide the data into a training set and a validation set;

[0023] S302: The central server selects a bidirectional GRU network as the teacher model and sets the network parameters, including the number of units in the hidden layer, the number of bidirectional layers, and the input data dimension. According to the power plant operation task category, set the fully connected layer and the Softmax activation function, input the training set into the teacher model for forward propagation, and generate the predicted probability distribution of each data through the fully connected layer of the teacher model;

[0024] S303: Use the cross-entropy loss function to evaluate the classification performance of the model, and use the gradient descent algorithm to minimize the loss function, iteratively update the model parameters. At the same time, evaluate the performance metrics of the teacher model on the validation set. If the model performance does not reach the preset metrics, adjust the hyperparameters or the model structure and retrain until the preset metrics are reached, and record the predicted probability distribution of the teacher model for the training set data and use it as the soft label;

[0025] S304: Based on the trained teacher model, obtain a lightweight bidirectional GRU model as the student model by reducing the number of hidden layer units and the network depth. Then use the same input data as the teacher model as the input of the student model and output the classification probability distribution. At the same time, the student model uses the true label of the training data, uses the cross-entropy loss function to measure the classification error as the hard label loss, and calculates the distillation loss between the predicted probability of the student model and the soft label through the KL divergence;

[0026] S305: Combine the hard label loss and the distillation loss with weights to obtain the total loss function. Randomly initialize the network weights and biases. The student model outputs a probability distribution through forward propagation. Then, use backpropagation and the gradient descent algorithm to minimize the total loss function, and dynamically adjust the learning rate according to the performance on the validation set. Repeatedly train and update the architectures and parameters of the student model and the teacher model until the preset requirements are met;

[0027] S306: The central server distributes the teacher model and the student model to the corresponding edge devices according to the computing resources of each edge device in the power plant. Then, input the real-time monitoring data into the teacher model and the student model to generate prediction results for the device status, operation risk, and work order completion probability. Set a threshold for the prediction results. If the predicted value exceeds the threshold, trigger an alarm, notify the management to take measures, and at the same time collect the accuracy feedback of the alarm and re-update the model based on the new training data.

[0028] Optionally, the specific steps for optimizing the operation paths of personnel and equipment described in Q4 are as follows:

[0029] S401: According to the personnel locations, equipment locations, and operation areas in the 3D model of the power plant, model the operation scenario as a weighted graph G=(V, E), where V represents the set of nodes, indicating equipment or key locations, and E represents the set of edges, indicating the passage paths between locations. Assign initial weights to each edge according to information such as path length, operation priority, and cross-operation risk, and dynamically introduce constraint conditions in path planning. At the same time, dynamically update the graph structure and edge weights according to real-time data;

[0030] S402: Set the initial number of ant colonies m according to the number of paths to be planned. Assign initial pheromones to each edge E, and calculate the heuristic value of each edge according to the edge weight, where η represents the heuristic information of the edge between node i and node j, and the larger the value, the better the path. d ij represents the weight of the edge between node i and node j. Then, each ant starts from a random starting position, calculates the selection probability according to the pheromone concentration and heuristic information of each edge, and selects the next node based on the calculated probability; ij

[0031] S403: Repeat the calculation of the selection probability and the selection of nodes until each ant constructs a complete path planning scheme. Then, update the pheromone concentration of the corresponding path according to the total path length found by each ant. After the update is completed, re-perform path selection and pheromone concentration update;

[0032] ​S404: Repeatedly perform path planning and pheromone concentration update until the improvement value of the path quality converges within a preset range, and output the path planning scheme with the highest pheromone concentration as the optimal path. At the same time, display the path through the 3D model of the power plant for operators' reference.

[0033] Optionally, the specific steps for analyzing the benefits and risks of different strategies during the operation process described in Q5 are as follows:

[0034] S501: Model the resource status and operation status in the operation management system as the resource status space R and the operation status space M respectively. The resource status space R represents different statuses of various resources such as equipment and personnel, and the operation status space M represents the execution progress of the current operation. Then, based on the two sets of status spaces, establish a set of operation management system status spaces S = {s t |s t =(r t ,m t )}, r t ∈R, m t ∈M, where s t represents the status of the operation management system at time t, r t represents the resource status at time t, and m t represents the operation status at time t;

[0035] S502: Collect all executable operations and decisions, and construct the corresponding action space A. Each action a t ∈A will cause the system to transfer from state s t to another state s t+1 . Calculate the transition probability P(s t |s t , a t+1 ) of transferring to another state s t+1 after taking action a t in state s t . Then, set the corresponding reward function based on various factors such as operation progress, resource utilization efficiency, and risk control. At the same time, calculate the reward value of any action a t in state s t through this reward function;

[0036] S503: Based on the current system state s t , the available action a t , the reward function, and the transition probability, calculate the value function V(s t ) of each state s t using the value iteration algorithm to obtain the expected long-term benefit of executing the optimal strategy starting from state s t , and recursively update the value function using the Bellman equation. By calculating each state s tThe maximum expected reward V(s t ) is determined, and the optimal action a t to be taken in each state s t is determined;

[0037] S503: Continuously update the policy using the Q-learning algorithm until it converges to the optimal policy, which is the optimal action a t selected in each state s t , maximizing the job efficiency and reducing risks, and optimizing resource scheduling and risk control according to the generated optimal policy. At the same time, it monitors the changes in job progress, equipment status, and personnel location information in actual operations in real time, and dynamically adjusts the job policy according to each real-time feedback.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This three-dimensional power plant operation management method transfers knowledge from a more powerful teacher model to a lightweight student model, enabling the student model to obtain high-performance prediction capabilities close to those of the teacher model in an environment with limited resources, and condensing the knowledge of complex models with high computational costs into the teacher model and the student model, improving prediction accuracy, reducing false alarm and missed alarm rates, and being able to analyze dynamic data in real time, quickly generate prediction results and issue early warnings, reducing deployment costs, improving the response efficiency of the system, and reducing the complexity of multi-model development and maintenance.

[0040] 2. The present invention can quantify and analyze the uncertainty of the operation state, and adjust the current policy according to the state-action pairs dynamically updated according to the real-time operation state, quickly adapting to changes in the operation plan. By setting reward mechanisms and penalty mechanisms with different risk levels, it dynamically adjusts the priority of resource allocation, can effectively avoid resource waste or risk omission caused by local optimization, reduce potential problems caused by improper resource scheduling, and improve the overall resource utilization rate and team collaboration efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

[0042] Figure 1 is a flowchart of a three-dimensional power plant operation management method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] Embodiment 1

[0044] Referring to Figure 1 , a three-dimensional power plant operation management method, the specific steps of the management method are as follows:

[0045] According to the civil engineering structure and equipment layout of the power plant, establish a 3D model of the power plant and divide it into functional areas.

[0046] Specifically, obtain the civil engineering drawings and equipment layout drawings of the power plant, use a laser scanner to scan the power plant site, capture the 3D spatial data of the buildings and equipment, generate point cloud data, then align and fuse the drawing data and point cloud data, remove noise points through Gaussian filtering. After the noise points are removed, use the meshing method to downsample the point cloud data. Through the Delaunay triangulation algorithm, convert the point cloud into a continuous 3D mesh model, then smooth the mesh surface. According to the room information marked in the power plant building drawings, divide the 3D mesh model into different functional areas, and attach attribute labels to each area according to the drawing data. Compare the position and size of the 3D model with the building drawings, and verify the accuracy and integrity of the 3D model through the matching of on-site survey data and the model.

[0047] Connect video monitoring, personnel positioning, equipment data, and operation information to the 3D model and perform information linkage.

[0048] Simulate the equipment operation status, personnel operation behavior, and emergency situations, and analyze the risk points under different operation states through dynamic scene adjustment.

[0049] Specifically, set the simulation objectives, scene parameters, and initial states. Based on the physical characteristics and working principles of the equipment, construct a dynamic operation model of the equipment, and through path planning and task allocation, simulate the operation behavior of personnel to establish a personnel behavior model. Finally, based on each physical process, establish an environmental dynamic model. According to the equipment dynamic operation model, personnel behavior model, and environmental dynamic model, define emergency events and event trigger conditions, analyze the changes in equipment status and personnel behavior, and calculate the emergency response time to simulate emergency response. Record the time series data of equipment status, personnel behavior, and environmental changes. According to the recorded data of each group, evaluate the risk indicators based on the probability of equipment overload and the proportion of dangerous areas in the personnel path, and mark the location and risk level of the risk points. According to the simulation results, optimize the equipment operation status and personnel operation path, and through the information-linked 3D model of the power plant, perform multiple iterative adjustments until the change value of the risk function value of the scene converges to the optimal solution within the preset range, and output the operation process and equipment operation parameters with reduced risks.

[0050] Embodiment 2

[0051] Refer to Figure 1 , a 3D-based power plant operation management method, and the specific steps of this management method are as follows:

[0052] Predict equipment failures, personnel operation risks, and work order completion status, issue early warnings in a timely manner, and optimize the operation paths of personnel and equipment

[0053] Specifically, obtain comprehensive equipment operation data, personnel behavior data, and work order records from the power plant historical database, including time series features and static features. Remove missing values and outliers from the historical data, encode non-numerical data, normalize the processed data to the interval [0, 1], extract features from the time series data, and screen the feature data according to the task objectives. Then divide the data into a training set and a validation set. The central server selects a bidirectional GRU network as the teacher model and sets the network parameters, including the number of units in the hidden layer, the number of bidirectional layers, and the input data dimension. According to the power plant operation task category, set the fully connected layer and the Softmax activation function. Input the training set into the teacher model for forward propagation, and generate the predicted probability distribution of each data through the fully connected layer of the teacher model. Use the cross-entropy loss function to evaluate the classification performance of the model, and use the gradient descent algorithm to minimize the loss function and iteratively update the model parameters. At the same time, evaluate the performance indicators of the teacher model on the validation set. If the model performance does not reach the preset indicators, adjust the hyperparameters or the model structure and retrain until the preset indicators are reached. Record the predicted probability distribution of the teacher model for the training set data and use it as the soft label. Based on the trained teacher model, obtain a lightweight bidirectional GRU model as the student model by reducing the number of hidden layer units and the network depth. Then use the same input data as the teacher model as the input of the student model and output the classification probability distribution. At the same time, the student model uses the true labels of the training data, and uses the cross-entropy loss function to measure the classification error as the hard label loss. Calculate the distillation loss between the predicted probability of the student model and the soft label through the KL divergence. Combine the hard label loss and the distillation loss with weights to obtain the total loss function. Randomly initialize the network weights and biases. The student model outputs the probability distribution through forward propagation, and then uses backpropagation and the gradient descent algorithm to minimize the total loss function and dynamically adjust the learning rate according to the performance on the validation set. Repeatedly train and update the architectures and parameters of the student model and the teacher model until the preset requirements are met. The central server distributes the teacher model and the student model to the corresponding edge devices according to the computing resources of each edge device in the power plant. Then input the real-time monitoring data into the teacher model and the student model to generate the prediction results of the equipment status, operation risk, and work order completion probability. Set a threshold for the prediction results. If the predicted value exceeds the threshold, trigger an early warning and notify the management to take measures. At the same time, collect the accuracy feedback of the early warning and re-update the model based on the new training data.

[0054] Specifically, according to the personnel positions, equipment positions, and operation areas in the 3D power plant model, the operation scenario is modeled as a weighted graph G=(V, E), where V represents the set of nodes, indicating equipment or key positions, and E represents the set of edges, indicating the access paths between positions. According to information such as path length, operation priority, and cross-operation risk, an initial weight is assigned to each edge, and constraint conditions are dynamically introduced in path planning. At the same time, the graph structure and edge weights are dynamically updated according to real-time data. The initial number of ant colonies m is set according to the number of paths to be planned. Initial pheromones are distributed on each edge E, and according to the weight of the edge, the heuristic value of each edge is calculated through where η ij represents the heuristic information of the edge between node i and node j. The larger the value, the better the path. d ij represents the weight of the edge between node i and node j. Then each ant starts from a random starting position, calculates the selection probability according to the pheromone concentration and heuristic information of each edge, and selects the next node based on the calculated probability. The selection probability calculation and node selection are repeated until each ant constructs a complete path planning scheme. After that, the pheromone concentration of the corresponding path is updated according to the total path length found by each ant. After the update is completed, the path selection and pheromone concentration update are performed again. The path planning and pheromone concentration update are repeatedly performed until the path quality improvement value converges to the preset range, and the path planning scheme with the highest pheromone concentration is output as the optimal path. At the same time, the path is displayed through the 3D power plant model for operators to refer to.

[0055] Construct a dynamic optimization model, analyze the benefits and risks of different strategies during the operation process, select the optimal strategy, and provide cross-departmental collaboration suggestions.

[0056] Specifically, the resource status and operation status in the operation management system are respectively modeled as a resource status space R and an operation status space M. The resource status space R represents the different statuses of equipment and personnel resources, and the operation status space M represents the execution progress of the current operation. Then, based on the two sets of status spaces, a set of operation management system status spaces S={s t |s t =(r t ,m t )}, r t ∈R, m t ∈M, where s t represents the status of the operation management system at time t, r t represents the resource status at time t, and m t represents the operation status at time t. All executable operations and decisions are collected, and the corresponding action space A is constructed. Each action a t ∈A will cause the system to transfer from state s t to another state st+1 , calculate the transition probability P(s t '|s t , a t+1 ) of transferring to another state s t+1 ' after taking action a t in state s t . Then, set the corresponding reward function based on various factors such as job progress, resource utilization efficiency, and risk control. At the same time, calculate the reward value of any action a t in state s t through this reward function. Based on the current system state s t , the available action a t , the reward function, and the transition probability, calculate the value function V(s t ) of each state s t using the value iteration algorithm to obtain the expected long-term reward of executing the optimal policy starting from state s t . And use the Bellman equation to recursively update the value function. By calculating the maximum expected reward V(s t ) for each state s t , determine the optimal action a t to be taken in each state s t . Use the Q-learning algorithm to continuously update the policy until it converges to the optimal policy, which is the optimal action a t selected in each state s t , maximizing job efficiency and reducing risks. And optimize resource scheduling and risk control according to the generated optimal policy. At the same time, monitor the changes in information such as job progress, equipment status, and personnel location in the actual job in real time, and dynamically adjust the job policy according to each real-time feedback.

Claims

1. A three-dimensional power plant operation management method, characterized in that: The specific steps of this management method are as follows: Q1: According to the civil structure and equipment layout of the power plant, establish a 3D model of the power plant and divide the functional areas; Q2: Connect video surveillance, personnel positioning, equipment data, and operation information to the 3D model and perform information linkage; Q3: Simulate the equipment operation status, personnel operation behavior and emergency situations, and analyze the risk points under different operation conditions through dynamic scene adjustment; Q4: Predict equipment failures, personnel operation risks and work order completion status, issue early warnings in a timely manner, and optimize personnel and equipment operation paths; Q5: Build a dynamic optimization model to analyze the benefits and risks of different strategies during the operation, select the optimal strategy, and provide cross-departmental collaboration recommendations.

2. A three-dimensional power plant operation management method according to claim 1, characterized in that: The specific steps for establishing a 3D model of a power plant and dividing functional areas as described in Q1 are as follows: S101: Obtain the civil engineering drawings and equipment layout drawings of the power plant, and use a laser scanner to scan the power plant site to capture the three-dimensional spatial data of the buildings and equipment, generate point cloud data, and then align and fuse the drawing data and point cloud data; S102: removing noise points by Gaussian filtering. After the noise points are removed, downsampling the point cloud data is performed using a gridding method. The point cloud is converted into a continuous three-dimensional grid model by a Delaunay triangulation algorithm, and then the grid surface is smoothed. S103: Divide the 3D grid model into different functional areas according to the room information marked in the power plant architectural drawings, and add attribute labels to each area according to the drawing data. Compare the position and size of the 3D model with the architectural drawings, and verify the accuracy and completeness of the 3D model by matching the model with the on-site survey data.

3. A three-dimensional power plant operation management method according to claim 2, characterized in that: The specific steps for analyzing risk points under different operating conditions through dynamic scenario adjustment as described in Q3 are as follows: S201: Set simulation targets, scenario parameters and initial states, build a dynamic operation model of the equipment based on the physical characteristics and working principles of the equipment, and build a personnel behavior model by simulating personnel's work behaviors through path planning and task allocation. Finally, build an environmental dynamic model based on various physical processes. S202: defining emergency events and event triggering conditions according to the equipment dynamic operation model, the personnel behavior model and the environment dynamic model, analyzing the changes in equipment status and personnel behavior, and calculating the emergency response time to simulate the emergency response; S203: Record the time series data of equipment status, personnel behavior and environmental changes, and evaluate the risk index based on the equipment overload probability and the proportion of dangerous areas in the personnel path according to the recorded data of each group, and mark the location of the risk point and the risk level; S204: According to the simulation results, the equipment operating status and personnel operation path are optimized, and multiple iterative adjustments are performed through the three-dimensional model of the power plant after information linkage until the change value of the risk function value of the scenario converges to the optimal solution within the preset range, and the operation process and equipment operation parameters after risk reduction are output.

4. A three-dimensional power plant operation management method according to claim 1, characterized in that: The specific steps for predicting equipment failure, personnel operation risks and work order completion described in Q4 are as follows: S301: Obtain comprehensive equipment operation data, personnel behavior data and work order records from the power plant historical database, including time series features and static features, remove missing values ​​and outliers in historical data, encode non-numerical data, normalize the processed data to the [0, 1] interval, extract features from time series data, filter feature data according to task objectives, and then divide the data into training set and validation set; S302: The central server selects a bidirectional GRU network as the teacher model and sets network parameters, including the number of units in the hidden layer, the number of bidirectional layers, and the input data dimension. According to the task category of the power plant operation, the fully connected layer and the Softmax activation function are set. The training set is input into the teacher model for forward propagation, and the predicted probability distribution of each data is generated through the fully connected layer of the teacher model. S303: Use the cross entropy loss function to evaluate the classification performance of the model, and use the gradient descent algorithm to minimize the loss function, iteratively update the model parameters, and evaluate the performance indicators of the teacher model on the validation set. If the model performance does not reach the preset indicators, adjust the hyperparameters or model structure, retrain until the preset indicators are reached, and record the predicted probability distribution of the teacher model for the training set data, and use it as a soft label; S304: Based on the trained teacher model, a lightweight bidirectional GRU model is obtained as a student model by reducing the number of hidden layer units and the network depth. Then, the same input data as the teacher model is used as the input of the student model, and the classification probability distribution is output. At the same time, the student model uses the real label of the training data and adopts the cross entropy loss function to measure the classification error as the hard label loss. The predicted probability of the student model and the distillation loss of the soft label are calculated by KL divergence. S305: The hard label loss and the distillation loss are weightedly combined to obtain the total loss function. The network weights and biases are randomly initialized. The student model outputs the probability distribution through forward propagation. Then, the backpropagation and gradient descent algorithms are used to minimize the total loss function. The learning rate is dynamically adjusted according to the performance on the validation set. The student model and the teacher model architecture and parameters are repeatedly trained and updated until the preset requirements are met. S306: The central server sends the teacher model and the student model to the corresponding edge devices based on the computing resources of each edge device in the power plant, and then inputs the real-time monitoring data into the teacher model and the student model to generate prediction results of equipment status, operation risk and work order completion probability. A threshold is set for the prediction result. If the prediction value exceeds the threshold, an early warning is triggered to notify the management personnel to take measures. At the same time, feedback on the accuracy of the early warning is collected, and the model is updated again based on the new training data.

5. A three-dimensional power plant operation management method according to claim 2, characterized in that: The specific steps for optimizing the operation paths of personnel and equipment described in Q4 are as follows: S401: Based on the personnel positions, equipment positions and operation areas in the three-dimensional model of the power plant, the operation scene is modeled as a set of weighted graphs G = (V, E), where V represents a node set, representing equipment or key positions, and E represents an edge set, representing the passage paths between positions. An initial weight is assigned to each edge based on the path length, operation priority and cross-operation risk information, and constraints are dynamically introduced in the path planning. At the same time, the graph structure and edge weights are dynamically updated according to real-time data; S402: Set the initial ant colony number m according to the number of paths to be planned, distribute initial pheromones on each edge E, and calculate the initial pheromone according to the weight of the edge. Calculate the heuristic value of each edge, where η ij Represents the heuristic information of the edge between node i and node j. The larger the value, the better the path. ij represents the weight of the edge between node i and node j. After that, each ant starts from a random starting position, calculates the selection probability according to the pheromone concentration of each edge and the heuristic information, and selects the next node based on the calculated probability; S403: Repeat the selection probability calculation and node selection until each ant constructs a complete path planning scheme, and then update the pheromone concentration of the corresponding path according to the total path length found by each ant. After the update is completed, re-select the path and update the pheromone concentration; S404: Repeat the path planning and pheromone concentration update until the path quality improvement value converges to a preset range, and output the path planning scheme with the highest pheromone concentration as the optimal path. At the same time, the path is displayed through the three-dimensional model of the power plant for reference by operators.

6. A three-dimensional power plant operation management method according to claim 1, characterized in that: The specific steps for analyzing the benefits and risks of different strategies in the operation process described in Q5 are as follows: S501: Model the resource status and job status in the job management system as resource status space R and job status space M respectively, where the resource status space R represents the different status of each resource of equipment and personnel, and the job status space M represents the execution progress of the current job. Then, based on the two sets of status spaces, a set of job management system state space S = {s t |s t =(r t ,m t )},r t ∈R,m t ∈M, where s t represents the state of the job management system at time t, r t represents the resource status at time t, m t Represents the job status at time t; S502: Collect all executable operations and decisions, and construct the corresponding action space A, where each action a t ∈A will cause the system to go from state s t Transfer to another state s t+1 , calculated in state s t Take action a t Then, transfer to another state s t+1 The transition probability P(s t+1 |s t ,a t ), then set the corresponding reward function based on the operation progress, resource utilization efficiency and risk control factors, and calculate any action a through the reward function t In status t The reward value of S503: Based on the current system status t , possible actions a t , reward function and transition probability, and calculate each state s through the value iteration algorithm t The value function V(s t ) to obtain the state s t Start executing the expected long-term return of the optimal strategy and recursively update the value function using the Bellman equation by calculating each state s t The maximum expected reward V(s) under t ), determine in each state s t The optimal action to take is a t ; S503: Use the Q-learning algorithm to continuously update the strategy until it converges to the optimal strategy, that is, in each state s t The best action a selected under t , maximize operational efficiency and reduce risks, and optimize resource scheduling and risk control based on the generated optimal strategy. At the same time, it monitors the changes in the actual operation progress, equipment status, and personnel location information in real time, and dynamically adjusts the operation strategy based on real-time feedback.

Images