A method and system for intelligent optimization of three-dimensional scene arrangement of a thermal power plant

Abstract

This invention discloses an intelligent optimization method and system for the layout of a three-dimensional scene in a thermal power plant, belonging to the field of intelligent optimization technology for industrial facilities. It includes: integrating thermodynamic simulation and image recognition technology to capture images of equipment and site conditions in the target thermal power plant; automatically labeling equipment types and dimensions through image classification; and constructing a three-dimensional scene data model based on the input requirements of a deep Q-network. This invention, by integrating thermodynamic simulation and a deep Q-network, constructs a predictive model for equipment thermal deformation, which can accurately simulate the dynamic displacement of equipment under operating conditions. Based on this, reinforcement learning is used to iteratively adjust the initial static spacing of equipment, controlling the net distance between equipment within 1.05–1.15 times the design specification requirement. This effectively eliminates the waste of land resources caused by excessive safety margins in traditional static layouts, achieving refined compression of land area and improving land resource utilization efficiency while ensuring dynamic operational safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology for industrial facilities, specifically to an intelligent optimization method and system for the layout of a three-dimensional scene in a thermal power plant. Background Technology

[0002] As a key component of the power system, the design quality of thermal power plants directly affects power generation efficiency, construction costs, and the safety of operation and maintenance. Traditional two-dimensional design has significant limitations in expressing complex spatial relationships, achieving multi-disciplinary collaboration, and identifying design conflicts in advance, making it difficult to meet the high-efficiency and refined design requirements of modern power plants. With the development of digital technology, three-dimensional scene layout provides power plant design with intuitive and comprehensive technical means, which can effectively improve design quality and efficiency.

[0003] In the 3D scene layout of thermal power plants, traditional static optimization methods have inherent defects. Because they do not consider the dynamic deformation of equipment caused by thermal effects during operation, the traditional static layout model will have redundant equipment spacing, resulting in a waste of land resources. Secondly, while saving land, the length and path complexity of process pipeline connections between equipment will increase significantly, leading to a rebound in construction costs. After pipeline path optimization, the excessive density of equipment in some areas will result in obstructed operation and maintenance channels, increased maintenance risks, and potential long-term safety and efficiency hazards. Therefore, an intelligent optimization method and system for the 3D scene layout of thermal power plants is proposed to solve the above-mentioned problems. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for intelligent optimization of the layout of a three-dimensional scene in a thermal power plant, comprising the following steps: S1. Integrating thermodynamic simulation and image recognition technologies, it captures images of equipment and sites in the target thermal power plant, automatically labels equipment type and size through image classification, and constructs a three-dimensional scene data model based on the input requirements of a deep Q-network. S2. Construct a device thermal deformation prediction model based on a deep Q-network to simulate the dynamic displacement of the device under operating conditions. Iteratively adjust the initial static spacing through reinforcement learning to generate a minimum redundancy spacing scheme that meets safety specifications. S3. Image classification and recognition of equipment connection relationships are adopted, and the decision-making ability of deep Q network is combined to optimize pipeline route. The connection length and complexity are evaluated through value function to balance land saving and construction cost rebound risk. S4. Analyze densely populated equipment areas, adjust equipment layout and plan maintenance channels through the sequential decision-making capabilities of deep Q networks, ensure the width of maintenance channels and accessibility of fire lanes, and reduce long-term operational risks; S5. Construct a multi-objective deep Q-network optimization framework that includes land area, construction cost and operation and maintenance convenience, and realize the automatic selection and visualization output of the Pareto optimal solution set through the priority weight dynamic allocation mechanism. S6. Monitor the actual layout effect in real time, analyze the existing deviations, and input the deviation data into the deep Q network for strategy update and correction, thereby outputting the optimal layout scheme of the thermal power plant in a three-dimensional scene that takes into account dynamic safety, land use efficiency and operation and maintenance convenience.

[0005] Preferably, in step S1, the process of constructing the three-dimensional scene data model is as follows: By integrating thermodynamic simulation parameters with on-site image data of equipment and site of the target thermal power plant, multi-source data input is formed to achieve a unified digital representation of physical properties and spatial morphology. By using a pre-trained image classification convolutional neural network, devices in multi-source data input can be automatically identified and classified, and their geometric dimensions and device type attributes can be labeled, thereby improving device identification accuracy and modeling efficiency and reducing human intervention errors. By integrating the geometric dimensions and equipment type attributes after classification and labeling, as well as thermodynamic simulation parameters, a structured 3D scene data model conforming to the input format of a deep Q-network is constructed, providing a complete and computable dynamic scene foundation for intelligent optimization.

[0006] Preferably, in step S2, the process of generating the minimum redundancy spacing scheme that meets safety specifications is as follows: Based on the three-dimensional scene data model, using the thermodynamic simulation parameters of the equipment operating conditions as input, a prediction model for equipment thermal deformation is constructed to achieve high-precision prediction of equipment thermal deformation with an error controlled within ±5mm. By simulating the dynamic displacement of equipment under different operating conditions using deep Q-networks, the state space of which includes the initial position of the equipment and the temperature field distribution, the dynamic displacement of the equipment is simulated, thereby improving the adaptability of the layout scheme to the operating conditions. By using reinforcement learning to iteratively adjust the initial static spacing of the equipment, and using dynamic displacement as a constraint, the system outputs a set of minimum redundant spacings that meet safety specifications, thereby generating minimum redundant spacings that meet safety specifications, saving land use and ensuring dynamic operational safety.

[0007] Preferably, in step S3, the process of optimizing the pipeline route and evaluating the connection length and complexity is as follows: Image classification technology is used to identify the process connection relationships and interface spatial coordinates between equipment, automatically establish a precise equipment connection topology, reduce manual annotation errors, and construct a value function with pipeline connection length, number of bends and path intersection complexity as evaluation indicators to quantify pipeline economy and safety and provide clear guidance for optimization. By leveraging the decision-making capabilities of deep Q-networks, the three-dimensional routing of pipelines is optimized, the optimal path is intelligently searched, and the complexity of pipeline layout is reduced. Under the guidance of the value function, the goals of shortening land use and controlling pipeline costs are balanced, achieving synergistic optimization of land saving and cost control.

[0008] Preferably, the value function is expressed as: ; In the formula: Value function; This represents the total length of the pipe connection. The number of elbows required by the system; The complexity is the path intersection complexity; This is the weighting coefficient for the total connection length of the pipeline; The weighting coefficient for the number of elbows required by the system; The weighting coefficients for path intersection complexity; weighting coefficients Based on the engineering economic analysis, and ; The design of the weight configuration for the value function includes three configuration schemes: The first configuration option, , , It focuses on direct economic benefits and is suitable for projects that are extremely sensitive to control of initial investment; The second configuration option, , , It focuses on operation and maintenance safety and reliability, and is suitable for renovation and expansion projects with compact factory space, dense equipment layout or extremely high requirements for system reliability; The third configuration option , , It focuses on comprehensive balance and is suitable for exploring and comparing schemes when there is no clear single tendency in the early stage of a project. It provides decision-makers with a benchmark layout scheme that performs well in multiple indicators.

[0009] Preferably, in step S4, the process of adjusting the equipment layout and planning the operation and maintenance channels is as follows: By analyzing the equipment layout after spacing and pipeline optimization, local areas where the equipment density exceeds the preset density threshold are identified, and high-density areas are accurately located to provide target areas for subsequent channel optimization. Integrating channel accessibility metrics into the reward mechanism of deep Q networks, whose state space includes device coordinates and channel topology relationships, guides the network to proactively ensure channel accessibility during layout optimization; By leveraging the network's sequential decision-making capabilities, the positions of devices within densely populated areas can be fine-tuned, and basic maintenance channel space can be planned, achieving efficient collaboration between compact device layout and unobstructed channels.

[0010] Preferably, in step S4, planning the basic operation and maintenance channel space further includes: Clearly define the minimum width requirements for maintenance access routes and the network accessibility standards for fire access routes, and quantify them as hard constraints to prevent violations of access width regulations from the decision-making stage and ensure basic safety. When selecting actions, deep Q-networks prioritize satisfying channel constraints, and their reward function imposes high penalties on behaviors that violate channel criteria, thus strengthening the channel protection strategy. The system outputs a 3D equipment layout scheme that ensures unobstructed access for maintenance and fire escape routes even in densely populated areas, guaranteeing unobstructed life-saving passages within a compact layout and eliminating long-term maintenance risks.

[0011] Preferably, in step S5, the process of automatically filtering and visualizing the Pareto optimal solution set is as follows: Construct a multi-objective deep Q-network optimization framework that integrates land area, construction cost, and operation and maintenance convenience to achieve simultaneous optimization and automatic trade-off of the three objectives; A dynamic weight allocation mechanism is introduced into this multi-objective deep Q-network optimization framework to adaptively adjust the priority of objectives and avoid local optimization stagnation. The priority weights of each optimization objective are adjusted in real time according to decision preferences. The initial priority weights are set to 0.4 for land area, 0.4 for construction cost, and 0.2 for operation and maintenance convenience. During training, the improvement of each objective is automatically evaluated after every 100 iterations, and the weight allocation is dynamically adjusted according to decision preferences: when the improvement rate of an objective is less than 5% for 20 consecutive generations, its weight is automatically increased by 0.05; conversely, when the improvement rate exceeds 15%, the weight is decreased by 0.03 to ensure balanced optimization of each objective. By exploring and utilizing the network, the system automatically generates and filters out Pareto optimal layout solutions that achieve Pareto optimality across multiple objectives. It intuitively displays the layout and performance indicators of each solution and provides 3D visualization output, supporting decision-makers to compare and select the optimal layout from multiple perspectives.

[0012] Preferably, in step S6, the process of outputting the optimal layout scheme of the three-dimensional scene of the thermal power plant is as follows: After the layout schemes included in the Pareto optimal solution set are implemented, the actual location and operating status of the equipment are monitored in real time based on the deployed sensors and image systems, layout deviation data are collected, and real-time monitoring ensures that the layout scheme is consistent with the actual site conditions. The collected bias data is used as new training samples and input into a deep Q-network for online reinforcement learning and policy updates. Online learning enables the model to continuously optimize and adapt to changes in the environment. The layout scheme is dynamically modified based on the updated strategy to form a closed-loop optimization system that continuously adapts to actual operating conditions. This results in the output of the optimal layout scheme for the three-dimensional scenario of the thermal power plant. The closed-loop optimization ensures that the scheme remains optimal throughout its entire life cycle.

[0013] Secondly, a three-dimensional scene layout intelligent optimization system for thermal power plants is provided to implement the aforementioned intelligent optimization method for three-dimensional scene layout of thermal power plants, comprising the following modules: The multi-source sensing dynamic modeling module integrates thermodynamic simulation and high-precision visual data to construct a three-dimensional scene data model that combines the geometric properties and thermodynamic behavior of equipment, enabling fully automatic identification and digitization of equipment shape, interface location, and thermal deformation parameters. The deformation constraint spacing optimization module predicts the thermal displacement of the equipment under operating conditions based on a deep Q-network, and iteratively adjusts the equipment spacing with dynamic deformation as a constraint to generate a minimum redundancy layout, reducing the equipment spacing from 1.3–1.5 times that of the traditional static specification to 1.05–1.15 times. The pipe gallery and passageway collaborative optimization module is used to simultaneously optimize the layout of pipeline routes and maintenance passageways, taking into account multiple optimization objectives such as pipe length, bends, intersection complexity and passageway width. While controlling pipeline costs, it ensures that fire lanes are ≥4.0 meters and maintenance lanes are ≥2.5 meters, and eliminates maintenance conflicts in densely populated equipment areas. The closed-loop evolution and real-time control module is used to monitor deployment deviations in real time through sensor networks, drive deep Q-networks to learn online and update strategies, and achieve self-optimization throughout the entire lifecycle.

[0014] This invention provides a method and system for intelligent optimization of the layout of a three-dimensional scene in a thermal power plant. It has the following beneficial effects: (I) The intelligent optimization method and system for three-dimensional scene layout of thermal power plants integrates thermodynamic simulation and deep Q network to construct equipment thermal deformation prediction model, which can accurately simulate the dynamic displacement of equipment under operating conditions. On this basis, reinforcement learning is used to iteratively adjust the initial static spacing of equipment, and control the net distance between equipment within 1.05-1.15 times the design specification requirement. This effectively eliminates the waste of land resources caused by excessive safety margin in traditional static layout, and achieves refined compression of land area under the premise of ensuring dynamic operation safety, thereby improving the efficiency of land resource utilization.

[0015] (II) The intelligent optimization method and system for the layout of a three-dimensional scene of a thermal power plant establishes a multi-objective deep Q-network optimization framework that integrates land area, construction cost and operation and maintenance convenience. It introduces a dynamic weight allocation mechanism to realize the real-time adaptive adjustment of the weights of each objective. It can automatically adjust the priority weights according to the improvement of each objective during the training process. It also combines external archives and congestion calculation to maintain the diversity of Pareto solution sets. It can automatically generate and select the layout scheme that achieves Pareto optimality on multiple objectives, providing designers with multiple trade-off schemes and supporting scientific decision-making. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the workflow of an intelligent optimization method and system for three-dimensional scene layout of a thermal power plant according to the present invention. Figure 2 This is a schematic diagram of the process flow of an intelligent optimization method for the layout of a three-dimensional scene in a thermal power plant according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: an intelligent optimization method for the layout of a three-dimensional scene in a thermal power plant, comprising the following steps: S1. Integrating thermodynamic simulation and image recognition technologies, it captures images of equipment and sites of the target thermal power plant, automatically labels equipment type and size through image classification, constructs a three-dimensional scene data model in combination with the input requirements of the deep Q network, integrates thermodynamic simulation parameters and on-site collected image data of equipment and sites of the target thermal power plant, forms multi-source data input, and realizes a unified digital representation of physical characteristics and spatial morphology. By using a pre-trained image classification convolutional neural network, devices in multi-source data input can be automatically identified and classified, and their geometric dimensions and device type attributes can be labeled, thereby improving device identification accuracy and modeling efficiency and reducing human intervention errors. By integrating the geometric dimensions and equipment type attributes after classification and labeling, as well as thermodynamic simulation parameters, a structured 3D scene data model conforming to the input format of a deep Q-network is constructed, providing a complete and computable dynamic scene foundation for intelligent optimization; The specific work involves: in practical operation, integrating multi-source data required for the optimization of the target thermal power plant layout, including exporting key parameters of equipment under typical operating conditions from the thermodynamic simulation system, covering steady-state temperature field distribution data of the main equipment of the thermal power plant under rated load (temperature range 200°C–600°C), and collecting color images and depth point cloud data of the plant area and key equipment areas through high-resolution industrial cameras deployed on-site (resolution not less than 1920×1080 pixels). The image data must meet the requirements of the "Technical Specification for Design of Thermal Power Plants" (DL). Based on the accuracy requirements for engineering documents in GB 5000-2016, and utilizing a pre-trained convolutional neural network model (ResNet-50) on ImageNet and a self-built power plant equipment image dataset, pixel-level semantic segmentation and instance recognition are performed on the input image data. The outlines of main equipment, including boilers, turbines, generators, and transformers, are automatically annotated. Based on the image scale and point cloud depth information, the dimensions of their outer envelope rectangles (length, width, and height, with accuracy controlled within ±0.1 meters) are calculated, and equipment type codes are assigned. After completing equipment recognition and basic geometric annotation, the geometric attributes of the equipment are fused with their thermodynamic behavior characteristics. Specifically, the three-dimensional dimensions of the equipment output from image recognition are associated and bound with thermodynamic simulation parameters, according to the "Fire Protection Design Standard for Thermal Power Plants and Substations" (GB 5000-2016). Based on the requirements for equipment safety distances in 50229-2019 and the consideration of thermal stress release space in engineering mechanics, a comprehensive data object is constructed for each equipment instance. This object includes static geometric attributes (position coordinates X, Y, Z, attitude angle, dimensions), dynamic thermal deformation parameters (thermal expansion coefficient, reference temperature, predicted displacement vector), and equipment function type. All data objects are organized according to the plant area coordinate system to construct a hierarchical, uniformly formatted 3D scene data model, stored in JSON format. The generated structured 3D scene data model is then encapsulated to adapt to the training and inference process of deep Q-networks. Each device object in the 3D scene data model is represented as part of a state vector. This state vector contains the device's normalized 3D coordinates, size, type unique thermal encoding, and associated thermodynamic state variables. The state of the entire plant scene is composed of the set of all device state vectors. In this process, it is ensured that all physical parameters are normalized according to the actual scale of the plant area, so that they fall into the range of [-1, 1] or [0, 1] to meet the numerical stability requirements of neural network training. Finally, a complete, machine-readable digital power plant layout example is output, which is a 3D scene data model that integrates visual geometric information and physical simulation data. S2. A thermal deformation prediction model for equipment is constructed based on a deep Q-network to simulate the dynamic displacement of equipment under operating conditions. The initial static spacing is adjusted iteratively through reinforcement learning to generate a minimum redundancy spacing scheme that meets safety specifications. Based on a three-dimensional scene data model, the thermal deformation prediction model for equipment is constructed using the thermodynamic simulation parameters of the equipment operating conditions as input, achieving high-precision prediction of equipment thermal deformation with an error controlled within ±5mm. The dynamic displacement of the equipment under different operating conditions is simulated through a deep Q-network. Its state space includes the initial position of the equipment and the temperature field distribution, dynamically simulating equipment displacement and improving the adaptability of the layout scheme to operating conditions. The initial static spacing of the equipment is adjusted iteratively using reinforcement learning, and the dynamic displacement is used as a constraint to output a set of minimum redundancy spacings that meet safety specifications, generating a minimum redundancy spacing that meets safety specifications, saving land use and ensuring dynamic operation safety. The specific work involves integrating thermodynamic simulation parameters under typical operating conditions of thermal power plants, including but not limited to steady-state temperature field distribution data of main equipment under rated load. The temperature range is controlled between 200°C and 600°C, meeting the requirements for thermal system simulation accuracy in the "Technical Specification for Design of Thermal Power Plants" (DL 5000-2016). The model input dimensions cover the thermal expansion coefficient of the equipment materials (1.2 × 10⁻⁶ for carbon steel equipment). -5 / °C, alloy equipment takes 1.5×10 -5 The temperature field is mapped to an equivalent thermal stress field using the finite element method, and the theoretical deformation is calculated using Hooke's law. A temperature-displacement transfer function is established to construct a thermal deformation prediction model for the equipment. The dynamic displacement vector of the equipment under operating conditions is output, and the prediction accuracy meets the requirement of engineering error control within ±5mm. The expression for the dynamic displacement vector of the equipment during operation is as follows: ; In the formula: It is a dynamic displacement vector, and the engineering error is controlled within ±5mm (±0.005m); The coefficient of thermal expansion of the equipment material; This represents the three-dimensional temperature field distribution vector of the equipment under operating conditions. The reference temperature (initial temperature when the equipment is installed or under no stress) is taken as 20°C. The feature dimension vectors of the device (original dimensions in the X, Y, Z directions) are provided by the 3D scene data model; The state space design of the deep Q-network includes the initial coordinates (X, Y, Z) of the equipment and temperature field distribution parameters. The coordinate data adopts the plant coordinate system, and the temperature parameters are normalized to the [0, 1] interval according to the actual measured values. The action space is defined as the fine-tuning displacement of the equipment in three dimensions, with the single-step displacement limited to ±0.5 meters, which complies with the requirements for equipment adjustment accuracy in the "Fire Protection Design Standard for Thermal Power Plants and Substations" (GB 50229-2019). The network structure adopts a combination of fully connected layers and long short-term memory networks, with the hidden layer dimension set to 256. The initial learning rate is 0.001 and an exponential decay strategy is adopted. The reward function design comprehensively considers the compliance of spacing (meeting the minimum safe distance requirement) and land use efficiency (the proportion of land area reduction). When the equipment spacing is lower than the standard requirement, a penalty value of -10 is applied, and a reward of +2 is given for every 1% reduction in land area. An experience replay mechanism is adopted during training, and the buffer capacity is set to 10,000 state-action pairs. During the reinforcement learning iteration, the equipment layout state is initialized, and an ε-greedy strategy (initial ε=0.9, decay) is adopted. The adjustment action is selected with a reduction coefficient of 0.995. Each training round contains 200 steps. After each step, the Q network parameters are updated. The target network synchronizes the main network weights every 100 steps. The dynamic displacement is used as a hard constraint and is handled by the constraint domination method: when the predicted displacement causes the equipment spacing to be lower than the standard requirement, the current action is immediately terminated and a new action is selected. The optimization process continues until the convergence condition is met, that is, the reward value fluctuation is less than 5% for 20 consecutive rounds. Finally, the minimum redundant spacing set is output to ensure that the net distance between equipment is always maintained within 1.05-1.15 times the design specification value under various operating conditions, eliminating excessive redundancy in static layout and ensuring safety margin in dynamic operation. The expression for the minimum redundancy spacing set is as follows: ; ; In the formula: The final optimized output is the set of minimum redundant spacings, containing the optimal spacing between all device pairs; For equipment With equipment The minimum redundant spacing between them is the set One of the elements is to ensure that the net distance between equipment is always maintained within 1.05-1.15 times the design specification requirement under various operating conditions; The distance between devices under static layout obtained through reinforcement learning optimization; The minimum safe static distance required by design specifications; For equipment and The dynamic displacement vector is output by the thermal deformation prediction model, with a prediction accuracy of ±5mm; The resultant displacement of the displacement vectors of the two devices is given, taking into account the worst case, i.e., the two devices move towards each other and their displacement effects are superimposed. The cosine of the angle between the displacement vectors of two devices is used to calculate the effective component of the resultant displacement along the line connecting the centers of the devices. This is useful when the devices move towards each other. ; To maximize the function, the final spacing must simultaneously satisfy the constraints of both the static layout optimization result and the dynamic safety envelope; S3. Image classification is used to identify the connection relationships between equipment. Combined with the decision-making ability of deep Q-networks, the pipeline route is optimized. The connection length and complexity are evaluated through a value function to balance land saving and the risk of construction cost rebound. Image classification technology is used to identify the process connection relationships and interface spatial coordinates between equipment, automatically establish a precise equipment connection topology, reduce manual annotation errors, and construct a value function with pipeline connection length, number of bends, and path intersection complexity as evaluation indicators to quantify pipeline economy and safety and provide clear guidance for optimization. The decision-making ability of deep Q-networks is used to optimize the three-dimensional route of the pipeline, intelligently search for the optimal path, reduce the complexity of pipeline layout, and balance the goals of shortening land use and controlling pipeline costs under the guidance of the value function, so as to achieve synergistic optimization of land saving and cost control. The specific work involves: based on high-resolution color images and depth point cloud data of the plant area's panoramic view and key equipment areas collected on-site, using a convolutional neural network model (ResNet-50) pre-trained on ImageNet and a self-built power plant equipment image dataset, performing pixel-level semantic segmentation and instance recognition on the images, automatically identifying the main equipment, and accurately labeling the process interface locations on its body, such as the boiler's steam outlet, the turbine's steam inlet, and the condenser's circulating water inlet, etc. By combining the three-dimensional spatial information provided by the depth point cloud data, each identified process interface is assigned a coordinate system within the plant area. Spatial coordinates (X, Y, Z), with an accuracy controlled within ±0.1 meters, are used to construct a complete topology map of the process connections between equipment. Based on this, a value function is constructed to evaluate the piping layout scheme. This value function consists of three core indicators linearly weighted: total pipe connection length L (unit: meters), the number of elbows required by the system N (unit: number), and path intersection complexity C (quantified by statistically analyzing the number of intersections with other facilities whose minimum distance is less than 0.5 meters). After completing the digitization of process connections and the definition of the value function, a deep Q-network is used to automatically optimize the three-dimensional piping path, covering the entire plant. The 3D space of the area is discretized into a 3D grid with a resolution of 0.5 meters. The pipeline path is planned on this grid. The state space of the deep Q-network is defined as the node sequence of the currently planned pipeline path, the coordinates of the target interface, and the spatial distribution of obstacles (other equipment, buildings, and planned pipelines) within the plant area. The action space is defined as the next extension direction of the pipeline in the grid, including movement along the six positive and negative directions of the X, Y, and Z axes. The network structure adopts a combination of fully connected layers and long short-term memory networks, with the hidden layer dimension set to 256 to effectively handle the sequential decision-making characteristics of path planning. The initial learning rate is 0.0. 01. An exponential decay strategy is adopted. During training, an ε-greedy strategy (initial ε=0.9, decay coefficient 0.995) is used to explore the action space. For each successful planning of a complete path from the starting device interface to the ending device interface, the reward value R is calculated according to the value function, where R=-V, to transform the path optimization problem into a reward maximization problem. The target network synchronizes the main network weights every 100 steps to ensure training stability. During path optimization, relevant safety specifications in the "Technical Code for Design of Thermal Power Plants" (DL5000-2016) and the "Fire Protection Standard for Design of Thermal Power Plants and Substations" (GB 50229-2019) are integrated into the decision-making process as hard constraints. These constraints include, but are not limited to: the minimum parallel clearance between pipes and cable trays is not less than 0.3 meters, the perpendicular clearance when pipes cross is not less than 0.15 meters, and a minimum distance of 1 meter must be maintained between pipes and fire lanes.With a maintenance space of 0 meters, the Deep Q-Network, after each decision, predicts whether the new path point violates the above constraints or conflicts with obstacles. Once a conflict is detected, the current path exploration is immediately terminated, a high penalty reward (-50) is applied, and the network is guided to reselect an action. Through a combination of constraint domination and value function guidance, the Deep Q-Network gradually learns a 3D pipeline layout scheme that can minimize the total pipeline length, reduce the number of bends, avoid complex intersections, and meet all safety regulations and spatial constraints during iterative training. The final output is a pipeline direction coordinate sequence that achieves the optimal balance between land saving and pipeline construction costs (including materials, installation, and subsequent operating costs caused by flow resistance). This sequence can be directly input into 3D design software to generate construction drawings. Furthermore, the value function is expressed as: ; In the formula: Value function; This represents the total length of the pipe connection. The number of elbows required by the system; The complexity is the path intersection complexity; This is the weighting coefficient for the total connection length of the pipeline; The weighting coefficient for the number of elbows required by the system; The weighting coefficients for path intersection complexity; weighting coefficients Based on the engineering economic analysis, and ; The design of the weight configuration for the value function includes three configuration schemes: The first configuration option, , , This approach prioritizes direct economic efficiency, allocating core weights to the total pipeline connection length and the number of elbows required by the system. It prioritizes controlling direct construction costs, as pipeline materials and installation costs account for a significant portion of power plant construction. Shortening the total length can directly reduce material procurement costs. At the same time, elbows are critical and weak points in the pipeline system's resistance. Reducing their number not only reduces procurement and welding costs but also helps reduce system pressure loss and improve long-term operational economics. Setting the weight of path intersection complexity to a relatively low 0.1 means that, while ensuring compliance with the "Technical Code for Design of Thermal Power Plants" regarding the mandatory safety standard of a minimum cross-crossing clearance of 0.15 meters, a certain degree of path intersection is allowed in exchange for greater optimization of the total length and number of elbows. This approach is suitable for projects that are extremely sensitive to initial investment control. The second configuration option, , , Prioritizing operational safety and reliability, this configuration increases the weight of path intersection complexity to generate a clearer layout and more convenient maintenance pipeline solution. In densely populated areas, complex pipeline intersections can severely compress maintenance space, increasing the difficulty and risk of daily inspections and emergency repairs. It will guide the planning of paths with fewer intersections and clearer spatial hierarchy, but at the cost of slightly increasing the total length of pipelines or individual bends. It follows the spirit of the "Fire Protection Design Standard for Thermal Power Plants and Substations" regarding the unobstructed access of operation and maintenance channels, and is suitable for renovation and expansion projects with compact plant space, dense equipment layout, or extremely high requirements for system reliability. The third configuration option , , This configuration, which emphasizes comprehensive balance, is a compromise solution that gives relatively balanced consideration to the three objectives. While pursuing the reduction of direct costs (L and N), it also pays considerable attention to the simplicity and security (C) of the layout. The weight setting guides the algorithm to find a globally optimal solution. Its output solution achieves a good balance between construction cost, operational energy efficiency, and long-term operation and maintenance convenience, avoiding overly extreme layout results. This configuration is highly versatile and suitable for exploring and comparing solutions in the early stages of a project when there is no clear single tendency. It provides decision-makers with a benchmark layout solution that performs well in multiple indicators. S4. Analyze densely populated equipment areas, adjust equipment layout and plan maintenance channels through the sequential decision-making capabilities of deep Q networks, ensure the width of maintenance channels and accessibility of fire lanes, and reduce long-term operational risks; S5. Construct a multi-objective deep Q-network optimization framework that includes land area, construction cost and operation and maintenance convenience, and realize the automatic selection and visualization output of the Pareto optimal solution set through the priority weight dynamic allocation mechanism. S6. Monitor the actual layout effect in real time, analyze the existing deviations, and input the deviation data into the deep Q network for strategy update and correction, thereby outputting the optimal layout scheme of the thermal power plant in a three-dimensional scene that takes into account dynamic safety, land use efficiency and operation and maintenance convenience.

[0019] Example 2, as Figure 1 , Figure 2As shown, based on Embodiment 1, the present invention provides a technical solution: In S4, the process of adjusting the equipment layout and planning the operation and maintenance channel is as follows: Analyze the equipment layout after spacing and pipeline optimization, identify local areas where the equipment density exceeds the preset density threshold, accurately locate high-density areas, provide target areas for subsequent channel optimization, integrate the channel unobstructedness index into the reward mechanism of the deep Q network, whose state space includes the equipment coordinates and channel topology relationship, guide the network to actively ensure channel unobstructedness in layout optimization, fine-tune the positions of equipment in dense areas through the network's sequential decision-making capability, and plan the basic operation and maintenance channel space to achieve efficient coordination between compact equipment layout and unobstructed channel; The specific work content is as follows: After completing the initial optimization of equipment spacing and piping system, a spatial density analysis of the entire plant's equipment layout is conducted to identify potential operation and maintenance conflict areas. Specifically, a three-dimensional meshing method is used to divide the plant area into 1.0m × 1.0m × 1.0m cubic units. The density of the area is quantified by calculating the proportion of the projected area of ​​equipment within each unit. Based on the provisions regarding operation and maintenance passages in the "Technical Code for Design of Thermal Power Plants" (DL 5000-2016), a 40% equipment density threshold is set. When the proportion of the projected area of ​​equipment in a certain area exceeds this density threshold, it is automatically marked as a high-density area. Simultaneously, this is combined with the "Code for Fire Protection Design of Buildings" (GB According to the standard (GB 50016-2014) requiring a minimum net width of 4.0m for fire lanes and 2.5m for maintenance lanes, accessibility verification is performed on marked areas. This analysis is based on the coordinates of the equipment's outer envelope rectangle (accuracy ±0.1m) and thermal deformation displacement vectors (prediction accuracy ±5mm) to ensure the identification results match the actual operating status of the equipment. The accessibility requirements are deeply integrated into the decision-making mechanism of a deep Q-network, constructing a multi-dimensional state space containing equipment spatial coordinates and channel topology. The state vector specifically includes: normalized equipment coordinates (interval [-1, 1]), equipment outer envelope dimensions, channel network node coordinates, and channel width status indicators. The reward function adopts a hierarchical design: the basic reward is based on equipment spacing compliance (compliant with GB 50016-2014). The system adheres to the safety distance requirements (50229-2019) and optimizes the floor space. Accessibility rewards are dynamically adjusted based on access width compliance: +5 for main fire access routes reaching 4.0m and +3 for maintenance routes reaching 2.5m. A -10 penalty is applied to access obstruction. The network structure combines a fully connected layer (256-dimensional) with LSTM, with an initial learning rate of 0.001 decaying at 0.95 per thousand steps to ensure accessibility is considered simultaneously when optimizing equipment layout. Based on reinforcement learning's sequential decision-making capabilities, fine-tuning of equipment positions is implemented in identified high-density areas. The action space is defined as the three-dimensional displacement adjustment of the equipment, with single-step adjustment limited to ±0.5m, complying with the "Fire Protection Design Standard for Thermal Power Plants and Substations" (GB). According to the specifications of 50229-2019 regarding the accuracy of equipment adjustment, in each decision step, it is verified whether the hot-state spacing of the adjusted equipment meets the safety factor of 1.05-1.15 times the specification requirement. At the same time, it is checked whether the passage width meets the design standard. During the training process, the constraint dominance method is used. When the equipment adjustment causes the passage width to be lower than the minimum requirement of 4.0m for fire lane or 2.5m for maintenance lane, the action is immediately terminated and a new selection is made. After iterative optimization, the final output is an equipment layout scheme that ensures reasonable equipment spacing while ensuring that all maintenance lanes meet the specification requirements. The width of fire lane is controlled within the optimized range of 4.0-4.5m, and the width of maintenance lane is controlled within the optimized range of 2.5-3.0m. Furthermore, in S4, the basic maintenance access space planning includes: clearly defining the minimum width requirements for maintenance access and the network accessibility standards for fire access, and quantifying them as hard constraints to prevent access width violations from the decision-making source, ensuring basic safety. When selecting actions, the deep Q network prioritizes meeting access constraints, and its reward function imposes high penalties on behaviors that violate access standards, strengthening access protection strategies. It outputs a three-dimensional equipment layout scheme that can still ensure unobstructed maintenance and fire access in densely populated equipment areas, ensuring unobstructed life-saving access under a compact layout and eliminating long-term maintenance risks. The specific work content includes: Based on the "Code for Fire Protection Design of Buildings" (GB 50016-2014) and the "Technical Specification for Design of Thermal Power Plants" (DL... The mandatory provisions of GB / T 5000-2016 explicitly set the minimum clear width of maintenance passages at 2.5 meters and the minimum clear width of fire passages at 4.0 meters, integrating these as inviolable hard constraints into the optimization system. Based on the outer envelope rectangle coordinates of the equipment (positioning accuracy controlled within ±0.1 meters) and the thermal deformation displacement vector (prediction accuracy ±5 mm), dynamic envelope calculation of the passage space is performed in a 3D scene. When adjusting the layout, the minimum Euclidean distance between the equipment and the virtual passage boundary is detected in real time. Any equipment displacement that causes the passage clear width to fall below the above standard values ​​is directly marked as an illegal action by the system's underlying logic and excluded, thus ensuring the basic security of all generated schemes from the source. The decision machine of the deep Q-network is designed to prioritize satisfying the hard constraints of the passage. Its state space includes the normalized coordinates of the equipment (mapped to the [-1, 1] interval), the outer envelope size of the equipment, and the coordinates of the key nodes of the passage network and the real-time width identifier. The excitation function adopts a hierarchical structure: the basic layer rewards compliance of equipment spacing and optimization of floor space; the accessibility layer uses dynamic rewards and penalties, awarding +5 when the width of the main fire lane reaches the standard of 4.0 meters and +3 when the width of the maintenance lane reaches 2.5 meters. Simultaneously, any violation of the accessibility standard incurs a high penalty of -10. The network structure combines a fully connected layer (256 dimensions) with a long short-term memory network to handle the sequential decision-making characteristics of layout optimization. The initial learning rate is set to 0.001 and an exponential decay strategy is adopted (decay coefficient 0.95 / thousand steps) to ensure that the policy gradient strongly favors maintaining accessibility during space exploration. Based on the sequential decision-making capability of reinforcement learning, fine-tuning of the positions of identified high-density equipment areas is implemented. The action space is defined as the displacement of the equipment on the three-dimensional coordinate axis, with the single-step adjustment limited to ±0.5 meters, complying with the "Fire Protection Design Standard for Thermal Power Plants and Substations" (GB). According to the specifications of 50229-2019 regarding the accuracy of equipment adjustment, two key indicators are verified simultaneously in each decision step: first, whether the hot-state spacing of the adjusted equipment meets the safety factor of 1.05 to 1.15 times the specified value; and second, whether the width of the passageway continuously meets the minimum requirements of 4.0 meters for fire protection and 2.5 meters for maintenance. During the training process, the constraint domination method is used. If any action causes any constraint to be unmet, the action is immediately terminated and the network is guided to reselect. After iterative optimization, the final three-dimensional equipment layout scheme is output. In areas with dense equipment, the width of the fire passageway is stably controlled within the optimized range of 4.0 to 4.5 meters, and the width of the maintenance passageway is controlled within the optimized range of 2.5 to 3.0 meters, ensuring land use efficiency while ensuring the absolute unobstructed access for maintenance and fire protection. In S5, the process of automatically selecting and visualizing the Pareto optimal solution set is as follows: A multi-objective deep Q-network optimization framework integrating land area, construction cost, and operational convenience is constructed to achieve simultaneous optimization and automatic trade-offs among the three objectives. A dynamic weight allocation mechanism is introduced into this multi-objective deep Q-network optimization framework to adaptively adjust objective priorities, avoiding local optimization stagnation. The priority weights of each optimization objective are adjusted in real time according to decision preferences. The initial priority weights for land area, construction cost, and operational convenience are set to 0.4, 0.4, and 0.2, respectively. During training, the improvement of each objective is automatically evaluated after every 100 iterations, and the weight allocation is dynamically adjusted according to decision preferences: when the improvement rate of an objective is less than 5% for 20 consecutive generations, its weight is automatically increased by 0.05; conversely, when the improvement rate exceeds 15%, the weight is reduced by 0.03 accordingly, ensuring balanced optimization of each objective. Through network exploration and utilization, the Pareto optimal solution set is automatically generated and screened, intuitively displaying the layout and performance indicators of each scheme, and providing three-dimensional visualization output to support decision-makers in comparing the optimal layout from multiple perspectives. The specific work involves constructing a multi-objective deep Q-network optimization framework that integrates land area, construction cost, and operational convenience. The framework's state space design includes normalized equipment coordinates (mapped to the [-1, 1] interval), equipment envelope dimensions, equipment type codes, and thermodynamic state parameters. Equipment coordinate accuracy is controlled within ±0.1 meters, and thermodynamic parameters are normalized to the [0, 1] interval based on actual measurements. The action space is defined as the fine-tuning displacement of the equipment in three dimensions, with single-step displacement limited to ±0.5 meters, complying with the requirements for equipment adjustment accuracy in the "Fire Protection Design Standard for Thermal Power Plants and Substations." The network structure combines fully connected layers with long short-term memory networks, with the hidden layer dimension set to 256 to effectively handle the sequential decision-making characteristics of layout optimization. Each optimization objective establishes an independent evaluation function: land area is calculated through the union of equipment projection areas; construction cost is based on a model of equipment procurement, infrastructure construction, and pipeline connection costs; operational convenience is quantified through equipment accessibility and maintenance space sufficiency indicators, forming a complete mathematical description of the multi-objective optimization problem. Dynamics are introduced into the multi-objective deep Q-network optimization framework. The weight allocation mechanism enables real-time adjustment of the priorities of each optimization objective. Configurable priority weights [wa, wb, wc] correspond to the three objectives of land area, construction cost, and ease of operation and maintenance, respectively. Manual intervention in weight settings is also supported, allowing designers to temporarily adjust priorities based on specific project needs. Through network exploration and utilization, the Pareto optimal layout solution set is automatically generated and filtered. During training, an improved ε-greedy strategy is used for multi-objective optimization, with an initial exploration rate of 0.9, decaying by a coefficient of 0.995 every thousand steps. The system maintains an external archive with a capacity of 500 to store all non-dominated solutions and maintains solution set diversity through congestion calculation. Optimization is considered converged when the Pareto front improvement rate is below 2% for 50 consecutive generations. The final output Pareto optimal solution set is presented through a 3D visualization engine, allowing designers to interactively browse optimization schemes under different weight configurations. The visualization system can simultaneously display the 3D equipment layout, pipeline routing, and channel network of each scheme, and display key indicators such as land area, construction cost, and ease of operation and maintenance scores in real time, providing intuitive basis for scheme decision-making. In S6, the process of outputting the optimal layout scheme for the 3D scene of the thermal power plant is as follows: After the layout schemes contained in the Pareto optimal solution set are implemented, the actual position and operating status of the equipment are monitored in real time based on the deployed sensors and image systems, layout deviation data is collected, and real-time monitoring ensures that the layout scheme is consistent with the actual site. The collected deviation data is used as new training samples and input into the deep Q network for online reinforcement learning and policy updates. Online learning enables the model to continuously optimize and adapt to changes in the site. Based on the updated policy, the layout scheme is dynamically corrected to form a closed-loop optimization system that continuously adapts to the actual operating conditions, thereby outputting the optimal layout scheme for the 3D scene of the thermal power plant. The closed-loop optimization ensures that the scheme remains optimal throughout its entire life cycle. The specific work involves: after the implementation of the 3D scene layout scheme for the thermal power plant, continuously monitoring the actual status of the equipment through a pre-deployed high-precision sensor network and industrial imaging system. The sensor network includes a total station positioning module (accuracy ±2mm) and a thermal displacement monitoring sensor (accuracy ±0.5mm) to collect real-time data on equipment foundation settlement and thermal operating position. Simultaneously, industrial cameras (2592×2048 pixels resolution) deployed at key nodes in the plant area collect panoramic images of the equipment area every 30 seconds. A visual recognition algorithm based on the ResNet-101 architecture is used to perform sub-pixel-level positioning of the equipment's actual location, and the collected equipment coordinates, foundation settlement, and thermal expansion displacement data are then processed. By comparing the actual equipment location with the design coordinates, when the deviation exceeds the allowable tolerance (±10mm for static equipment, ±15mm for dynamic equipment) specified in the "Fire Protection Design Standard for Thermal Power Plants and Substations" (GB50229-2019), a deviation report is automatically generated and the system's early warning mechanism is triggered to identify equipment position shifts caused by foundation settlement and thermal stress deformation. The collected deviation data, after preprocessing, is used as new training samples in a multi-objective deep Q-network for online reinforcement learning. The system maintains an experience replay buffer with a capacity of 10,000 state-action pairs, employing a priority experience replay mechanism. Sampling priority is allocated based on the magnitude of the temporal difference error, and network parameters are updated accordingly. Stochastic gradient descent with momentum was used, with a learning rate of 0.0001 and a momentum parameter of 0.9. The policy network weights were updated every 1000 new samples. During training, multiple constraints, including equipment safety distances, pipeline connection status, and unobstructed access to maintenance channels, were considered. Actual monitored equipment thermal displacement data were integrated into the reward function as hard constraints. The target network synchronized with the main network every 200 steps to ensure the stability of the training process. Through a continuous online learning mechanism, the deep Q-network gradually adapted to the actual operating environment of the power plant, continuously optimizing its decision-making strategy to make the layout scheme more consistent with the behavior characteristics of the equipment under real operating conditions. Based on the updated policy network, a three-dimensional field was generated for the current operating conditions. The layout correction scheme adopts a gradual adjustment strategy, with the adjustment amount of each equipment position controlled within ±50mm to ensure operational safety during equipment adjustment. After each correction, key indicators such as equipment spacing, pipeline stress distribution, and accessibility are re-evaluated to ensure that all parameters meet the requirements of the "Technical Specification for Design of Thermal Power Plants" (DL5000-2016). At the same time, a database containing historical correction records and effect evaluations is established. By analyzing the long-term effects of different correction strategies, the adjustment algorithm is continuously optimized so that the three-dimensional layout of the power plant can continuously adapt to the impact of long-term factors such as equipment aging and changes in operating conditions. The final output is a three-dimensional scene layout scheme that maintains optimal performance throughout the entire life cycle of the thermal power plant.

[0020] Example 3, as Figure 1 As shown, based on Embodiments 1-2, the present invention also provides an intelligent optimization system for the layout of a three-dimensional scene in a thermal power plant, used to implement the above-mentioned intelligent optimization method for the layout of a three-dimensional scene in a thermal power plant, comprising the following modules: The multi-source sensing dynamic modeling module integrates thermodynamic simulation and high-precision visual data to construct a three-dimensional scene data model that combines the geometric properties and thermodynamic behavior of equipment. This enables fully automatic identification and digitization of equipment shape, interface location, and thermal deformation parameters, overcoming the information loss problem of traditional two-dimensional design and static models. The deformation constraint spacing optimization module predicts the thermal displacement of equipment under operating conditions based on a deep Q-network, and iteratively adjusts the equipment spacing with dynamic deformation as a constraint to generate a minimum redundancy layout. This reduces the equipment spacing from 1.3–1.5 times that of the traditional static specification to 1.05–1.15 times, significantly saving land while ensuring safe operation in hot conditions. The pipe gallery and passageway collaborative optimization module is used to simultaneously optimize the layout of pipeline routes and maintenance passageways. It coordinates multiple optimization objectives such as pipe length, bends, intersection complexity and passageway width. While controlling pipeline costs, it ensures that fire lanes are ≥4.0 meters and maintenance lanes are ≥2.5 meters, eliminates maintenance conflicts in densely populated equipment areas, and introduces a hierarchical reward mechanism that combines value functions and hard constraints on passageways to achieve collaborative optimization of pipeline layout and passageway planning under the reinforcement learning framework. The closed-loop evolution and real-time control module is used to monitor deployment deviations in real time through sensor networks, drive deep Q-networks to learn and update strategies online, achieve self-optimization throughout the entire life cycle, form a monitoring-learning-correction closed loop, and enable the deployment system to continuously adapt to long-term changes such as foundation settlement and equipment aging, and always maintain the optimal state.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent optimization of three-dimensional scene layout in thermal power plants, characterized in that, Includes the following steps: S1. Integrating thermodynamic simulation and image recognition technologies, it captures images of equipment and sites in the target thermal power plant, automatically labels equipment type and size through image classification, and constructs a three-dimensional scene data model based on the input requirements of a deep Q-network. S2. Construct a thermal deformation prediction model for the equipment, simulate the dynamic displacement of the equipment under operating conditions, and iteratively adjust the initial static spacing through reinforcement learning to generate a minimum redundant spacing scheme that meets safety specifications. Specifically, this includes: based on the three-dimensional scene data model, using the thermodynamic simulation parameters of the equipment operating conditions as input, constructing a thermal deformation prediction model for the equipment. The dynamic displacement of the device under different operating conditions is simulated by a deep Q-network, and its state space includes the initial position of the device and the temperature field distribution. The initial static spacing of the equipment is adjusted iteratively using reinforcement learning, with dynamic displacement as a constraint, and the minimum redundant spacing set that meets the safety specifications is output. The expression for the minimum redundancy spacing set is as follows: ； ； In the formula: The set of minimum redundant spacing; For equipment With equipment Minimum redundancy spacing between them; The distance between devices under static layout obtained through reinforcement learning optimization; The minimum safe static distance required by design specifications; For equipment and The dynamic displacement vector; The magnitude of the resultant displacement of the two device displacement vectors; The cosine of the angle between the displacement vectors of two devices is used to calculate the effective component of the resultant displacement along the line connecting the centers of the devices. This is useful when the devices move towards each other. ; This is a function to find the maximum value. S3. Use image classification to identify the connection relationship of the equipment, optimize the pipeline route and evaluate the connection length and complexity; S4. Analyze areas with dense equipment, adjust equipment layout and plan maintenance access routes to ensure the width of maintenance access routes and accessibility of fire lanes; S5. Construct a multi-objective deep Q-network optimization framework that includes land area, construction cost and operation and maintenance convenience, and realize the automatic selection and visualization output of the Pareto optimal solution set through the priority weight dynamic allocation mechanism. S6. Monitor the actual layout effect in real time, analyze the existing deviations, and input the deviation data into the deep Q network for strategy update and correction, thereby outputting the optimal layout scheme for the three-dimensional scene of the thermal power plant.

2. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 1, characterized in that: In S1, the process of constructing the three-dimensional scene data model is as follows: Integrate thermodynamic simulation parameters with on-site image data of the target thermal power plant to form multi-source data input; Using a pre-trained image classification convolutional neural network, devices in multi-source data input are automatically identified and classified, and their geometric dimensions and device type attributes are labeled. By integrating the geometric dimensions and equipment type attributes after classification and annotation, as well as thermodynamic simulation parameters, a structured 3D scene data model conforming to the input format of a deep Q-network is constructed.

3. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 1, characterized in that: In S3, the process of optimizing the pipeline route and evaluating the connection length and complexity is as follows: Image classification technology is used to identify the process connection relationships and interface spatial coordinates between equipment, and a value function is constructed with pipeline connection length, number of bends and path intersection complexity as evaluation indicators. Optimize the three-dimensional routing of pipelines by leveraging the decision-making capabilities of deep Q-networks, and balance the goals of reducing land use and controlling pipeline costs under the guidance of the value function.

4. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 3, characterized in that: The value function is expressed as: ； In the formula: Value function; This represents the total length of the pipe connection. The number of elbows required by the system; The complexity is the path intersection complexity; This is the weighting coefficient for the total connection length of the pipeline; The weighting coefficient for the number of elbows required by the system; This represents the weighting coefficient for path intersection complexity; Based on the engineering economic analysis, and ; The design of the weight configuration for the value function includes three configuration schemes: The first configuration option, , , It focuses on direct economic benefits and is suitable for projects that are extremely sensitive to control of initial investment; The second configuration option, , , It focuses on operation and maintenance safety and reliability, and is suitable for renovation and expansion projects with compact factory space, dense equipment layout or extremely high requirements for system reliability; The third configuration option , , It focuses on comprehensive balance and is suitable for exploring and comparing solutions when there is no clear single tendency in the early stage of a project.

5. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 1, characterized in that: In step S4, the process of adjusting the equipment layout and planning the operation and maintenance channels is as follows: Analyze the equipment layout after spacing and pipe optimization to identify local areas where the equipment density exceeds a preset density threshold; The channel accessibility metric is integrated into the reward mechanism of the deep Q network, whose state space includes the relationship between device coordinates and channel topology. By leveraging the network's sequential decision-making capabilities, the locations of devices within densely populated areas can be fine-tuned, and basic maintenance channel space can be planned.

6. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 5, characterized in that: In S4, the basic operation and maintenance channel space is further planned, including: Clearly define the minimum width requirements for maintenance access routes and the network accessibility standards for fire lanes, and quantify them as hard constraints. Deep Q-networks prioritize channel constraints when selecting actions, and their reward function imposes high penalties on behaviors that violate channel criteria. Output a 3D equipment layout plan that ensures unobstructed access for maintenance and fire fighting in densely populated areas.

7. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 1, characterized in that: In S5, the process of automatically filtering and visualizing the Pareto optimal solution set is as follows: Construct a multi-objective deep Q-network optimization framework that integrates footprint, construction cost, and ease of operation and maintenance; A dynamic weight allocation mechanism is introduced into this multi-objective deep Q-network optimization framework. The priority weights of each optimization objective are adjusted in real time according to decision preferences. The initial priority weights are set to 0.4 for land area, 0.4 for construction cost, and 0.2 for operation and maintenance convenience. During training, the improvement of each objective is automatically evaluated after every 100 iterations, and the weight allocation is dynamically adjusted according to decision preferences: when the improvement rate of an objective is less than 5% for 20 consecutive generations, its weight is automatically increased by 0.05; conversely, when the improvement rate exceeds 15%, the weight is decreased by 0.03 accordingly. By exploring and utilizing the network, the system automatically generates and filters out Pareto optimal layout solutions that achieve Pareto optimality across multiple objectives, and outputs the results in 3D visualization.

8. The intelligent optimization method for three-dimensional scene layout of a thermal power plant according to claim 1, characterized in that: In step S6, the process of outputting the optimal layout scheme of the three-dimensional scene of the thermal power plant is as follows: After the layout schemes included in the Pareto optimal solution set are implemented, the actual position and operating status of the equipment are monitored in real time based on the deployed sensors and image systems, and layout deviation data are collected. The collected bias data is used as new training samples and input into the deep Q network for online reinforcement learning and policy updates; The layout scheme is dynamically modified based on the updated strategy to form a closed-loop optimization system that continuously adapts to actual operating conditions, thereby outputting the optimal layout scheme for the three-dimensional scenario of the thermal power plant.

9. A three-dimensional scene layout intelligent optimization system for thermal power plants, used to implement the three-dimensional scene layout intelligent optimization method for thermal power plants as described in any one of claims 1-8, characterized in that, Includes the following modules: The multi-source sensing dynamic modeling module integrates thermodynamic simulation and high-precision visual data to construct a three-dimensional scene data model that combines the geometric properties and thermodynamic behavior of the equipment. The deformation constraint spacing optimization module predicts the thermal displacement of the equipment under operating conditions based on a deep Q-network, and iteratively adjusts the equipment spacing with dynamic deformation as a constraint to generate a minimum redundancy layout. The pipeline corridor and passageway collaborative optimization module is used to simultaneously optimize the layout of pipeline routes and maintenance passageways, taking into account multiple optimization objectives such as pipe length, bends, intersection complexity, and passageway width. The closed-loop evolution and real-time control module is used to monitor deployment deviations in real time through sensor networks, drive deep Q-networks to learn online and update strategies, and achieve self-optimization throughout the entire lifecycle.

Images