Thermal power plant operation digital twin

Abstract

The invention belongs to the field of digital twinning and energy industry intellectualization, relates to a thermal power plant three-dimensional operation state digital twinning system, and solves the technical problems of traditional operation and maintenance data islands, visualization deficiency and prediction lag. The system constructs a'multi-source data fusion-dynamic physical deduction-virtual-real interaction control 'system: a multi-source data fusion engine gets through a DCS / SIS barrier to realize 200,000 + data point / millisecond-level synchronization; a GPU (Graphics Processing Unit) drives a VFX Graph particle system and an LOD rendering technology to achieve 90fps (at) 4K flue gas flow simulation on RTX 4080; an AR instruction is directly connected with a PLC to construct a virtual-real bidirectional control chain, and a'monitoring-deduction-intervention 'closed loop is formed. Key indexes are as follows: fault early warning is greater than or equal to 72 h, fault positioning is 0.5 h / time, data delay is less than or equal to 10 ms, and model precision is + / -2 cm. And an edge-cloud collaborative architecture is adopted, and lightweight deployment of a Unity model is realized. According to actual measurement of a Jingfeng-Dalberg power plant, the emergency response is shortened from 26 minutes to 4.3 minutes, and a predictable intervention driven innovative paradigm foundation is laid for thermal power digitization.

Description

I. TECHNICAL FIELD

[0001] The present application relates to the field of digital twinning and intelligent technology for the energy industry, and in particular to a three-dimensional operation state digital twinning system for a thermal power plant, which realizes panoramic monitoring of the plant site, equipment state diagnosis and virtual inspection functions through multi-source data fusion and dynamic three-dimensional modeling. II. BACKGROUND

[0002] As a pillar of the national energy system, coal-fired power plants accounted for 35.7% of installed capacity and 54.8% of power generation in 2024, contributing more than half of the power generation. Its core process chain covers continuous production processes such as coal transportation, boiler combustion, steam power generation, and flue gas treatment, and a single 600MW unit needs to monitor more than 200,000 data points in real time. However, the traditional operation and maintenance mode relies on scattered SCADA systems and manual inspection, exposing serious technical defects:

[0003] 1. Data silos cause information fragmentation between subsystems, such as a boiler pipe explosion accident in a certain power plant, where the DCS system failed to associate the feedwater pump vibration data (stored in the independent TSI system), delaying the disposal for 6 hours and causing a direct loss of 2.8 million yuan;

[0004] 2. Insufficient visualization results in low fault location efficiency, and two-dimensional monitoring interfaces cannot express the spatial relationship of equipment, such as when the induced draft fan bearing is damaged, the average time for the operation and maintenance personnel to locate the fault point is 2.3 hours, and the actual fault location is only 35 meters from the operation station;

[0005] 3. Weak prediction ability, relying on fixed threshold alarms, cannot capture early features of gradual faults, and data from a certain group shows that the coal mill bearing wear has reached 30% when the threshold is triggered, missing the 72-hour warning window and causing maintenance costs to increase by 320%.

[0006] Existing digital twinning technologies also face serious challenges in the context of continuous production in thermal power plants:

[0007] Modeling staticity causes virtual deduction to be disconnected from physical entities, a certain domestic platform has a data update cycle of more than 1 hour, and the deduction power deviation under AGC frequency modulation conditions reaches 12.7%; low rendering efficiency restricts large-scale scene applications, the frame rate of a million-unit full-plant model using traditional CPU particle rendering is only 22fps (4K resolution), and special effects such as steam flow require pre-baking processing; lack of closed-loop control causes the system to remain in one-way monitoring, more than 90% of solutions cannot adjust equipment parameters in the reverse direction, and virtual debugging results need to be manually entered into the DCS system. These problems point to a core contradiction - the current technical system cannot support the real-time, spatial correlation and predictive intervention needs of thermal power plants.

[0008] The system is born to solve the above structural defects. By building a multi-source data fusion engine to break through the barriers of DCS, SIS, and TSI systems, it realizes millisecond-level data synchronization and cross-parameter correlation analysis (such as real-time coupling of boiler temperature, coal sulfur content, and vibration signals). Using GPU-driven VFX Graph particle systems and hierarchical LOD rendering technology, it dynamically simulates smoke flow and ash accumulation in the ash hopper at 90fps frame rate on an RTX4080 graphics card. It establishes a virtual-real two-way control chain, supports AR instructions directly connected to PLC actuators through OPC UA protocol, and forms a "monitoring-inference-intervention" closed loop. These innovations enable the system to achieve a breakthrough in core indicators such as fault early warning lead time (72 hours), positioning efficiency (0.5 hours / instance), and operation and maintenance cost (29.6 yuan / MWh), providing a new paradigm for the digital transformation of the thermal power industry.

[0009] The system constructs a full-life-cycle three-dimensional running state digital twin of a thermal power plant, deeply integrates physical entities and virtual models, and realizes closed-loop management from data perception to decision execution. Based on the three-dimensional twin of the plant area, the system reconstructs high-fidelity geometric models of key equipment such as boilers and steam turbines through registration of unmanned aerial vehicle laser scanning and BIM models (accuracy ±2 cm), and dynamically maps equipment states using PBR physical materials, such as generating a heat map (low temperature blue to high temperature red gradient) in real time through custom Shader Graph nodes when the pipe wall temperature exceeds the threshold, making the running risk visualized. In the data-driven layer, the multi-source fusion engine integrates DCS control signals, SIS historical data, infrared temperature measurement, and vibration monitoring, etc. heterogeneous data, and realizes millisecond-level synchronization of 200,000+ data points / second through OPC UA protocol and UniRx asynchronous stream processing framework, completely breaking down the information silos between traditional subsystems.

[0010] Dynamic working condition inference is the core innovation of the system. To address the complexity of continuous production in thermal power, a load response model based on physical rules is developed: under high load conditions of the generator (coal supply ≥120 t / h), the GPU parallel computing capability of Unity VFX Graph is used to simulate the flow of flue gas and the effect of ash accumulation in the ash hopper in real time: flue gas particles collide and flow in the directed distance field (SDF), and the height of ash accumulation in the ash hopper is dynamically raised with the falling ash frequency (increasing 1 m model height per 10% capacity). The engine supports custom variable inference, such as automatically triggering a coking risk warning when the input coal sulfur content is >1.8%, and generating diffusion particles to simulate the development path of the accident chain.

[0011] At the application layer, the system releases value through a three-layer architecture: at the global monitoring level, the digital large screen integrates XCharts dynamic charts and three-dimensional plant models, and clicking on the boiler building can penetrate to view the internal superheater temperature curve (582°C / historical average 540°C); at the predictive maintenance level, the vibration trend model fused with LSTM algorithm can provide early warning of gradual failure 72 hours in advance, reducing maintenance costs by 320% compared to traditional threshold alarms; at the emergency response level, the accident chain deduction engine combines infrared thermal imaging positioning and SDF particle diffusion simulation to generate a disposal plan within 10 minutes and push it to the terminal, setting a technical benchmark for the evolution of thermal power digital twins from "visualization" to "intervention". III. SUMMARY

[0012] The present application provides a three-dimensional operation state digital twin system for thermal power plants, aiming to systematically solve the problems of data silos, visualization gaps, and prediction lags in traditional thermal power operation and maintenance, and break through the bottlenecks of existing digital twin technology in real-time performance, rendering efficiency, and closed-loop control. The system, through the triple technical innovation of multi-source data fusion engine, dynamic physical deduction model, and virtual-real interactive control chain, builds a full-process digital management system from equipment-level monitoring to plant-level decision-making. The core technical solution first establishes a millimeter-precision plant three-dimensional twin body: based on unmanned aerial vehicle laser scanning (precision ±2cm) and BIM model reverse reconstruction technology, geometric modeling is performed on key equipment such as boilers and steam turbines, and a hierarchical dynamic LOD management mechanism is adopted to achieve efficient rendering of million-level equipment scenes - within 1 meter, 500,000 high-detail surfaces (such as bolt textures) are presented, and at 50 meters, they are simplified to bounding box tracking, ensuring stable performance of 90fps@4K on RTX 4080 graphics cards. The physical properties of the equipment are deeply bound with real-time data, for example, temperature sensor data is mapped to material color change (blue below 300°C, red above 550°C) through Shader Graph nodes, and OPC UA signals from the DCS system (such as valve opening >80%) are analyzed to trigger pre-defined animations of the three-dimensional model, realizing synchronized visualization of the state.

[0013] At the data fusion and dynamic deduction layer, the system builds an asynchronous stream processing engine supporting throughput of 200,000+ data points / second, completely breaking through the barriers of DCS, SIS, and TSI systems. Through the sliding time window algorithm, the timestamps of heterogeneous data are aligned (error ≤10ms), and two-dimensional data is mapped to the corresponding positions of the three-dimensional model through a spatial coordinate conversion matrix, realizing global data space-time alignment. The equipment health assessment uses a quantitative degradation index:

[0014] (α, β, γ are coal quality calibration coefficients), when DI > 0.85, the pre-warning is automatically triggered. The working condition deduction engine builds a load response model based on physical rules: in the high load working condition (coal supply > 120 t / h), the GPU parallel computing capability of Unity VFX Graph is used to simulate the flow motion of flue gas in the pre-baking oriented distance field (SDF), and the flue gas concentration is dynamically related to the sulfur content of coal (the calculation formula is 0.8 x sulfur content + 0.2 x air volume). The fault chain prediction function can input infrared temperature measurement > 500℃ and vibration > 7mm / s signal, deduce the coking area within 24 hours through SDF particle diffusion model, and the positioning accuracy is ± 5cm.

[0015] The virtual inspection and fault pre-control module of the system deeply integrates spatial perception and physical simulation technology, realizes holographic monitoring and risk active intervention of the state of thermal power plant equipment. The module first builds an adaptive inspection path planning engine: based on the equipment health index (DI), the priority route is dynamically generated, when DI > 0.7, the equipment is automatically added to the key inspection queue, and the shortest inspection path in three-dimensional space is optimized through A* algorithm (such as the route from the boiler room console to the high-risk superheater is shortened by 42%). The operator can roam the plant area through the first person perspective, combine VR handle displacement and head tracking to realize six degrees of freedom space navigation, and reproduce real walking, looking up the pipe gallery, climbing the escalator and other operations in the virtual environment, and immerse in checking the equipment state.

[0016] The AI defect recognition function breaks through the barrier between physical monitoring and digital model: the real-time temperature field data captured by the infrared thermal imager is accurately superimposed on the surface of the three-dimensional equipment model through the spatial coordinate conversion matrix. When a local overheating area (such as the wall of the economizer > 550℃) is detected, the system automatically triggers a three-way alarm - the area flashes in high light red in the three-dimensional model, generates a temperature gradient heat map in the digital large screen, and marks the positioning coordinates on the screen (error ≤ 3cm). At the same time, the vibration spectrum analysis (FFT transform) and visual anomaly detection (YOLOv5 model) are fused to identify mechanical defects such as bearing cracks and loose bolts, which improves the efficiency by 400% compared with traditional manual inspection, and the defect detection rate reaches 98.7%.

[0017] The fault deduction engine predicts the development path of the accident chain based on a multi-physical field coupling model. Taking pipe rupture as an example: input the initial rupture point coordinates and steam pressure parameters (such as 8.5 MPa), the system builds a fluid dynamics model through pre-baked directed distance field (SDF), simulates the diffusion process of high-temperature steam in the pipeline system - uses GPU-driven VFX Graph particle system to calculate the steam injection angle, diffusion radius and temperature decay curve (12°C per meter) in real time, and dynamically renders the expansion range of the dangerous area (accuracy ± 0.5 meters) with red particles. The deduction result synchronously outputs the accident influence quantitative report: such as “diffuse to No. 3 unit control room within 30 seconds, suggest emergency isolation of A3 valve”, and automatically generates a disposal animation guide superimposed on the large screen. In the actual measurement of Jingneng Chifeng Power Plant, this module shortens the emergency response time from an average of 26 minutes to 4.3 minutes, and avoids secondary accidents by 93%.

[0018] Its technical value lies in the deep integration of industrial Internet of Things, physical simulation and immersive interaction, promoting the leap of thermal power digital twin from “visual monitoring” to “predictable intervention”: through multi-source data spatio-temporal alignment and health index calculation, early warning of faults such as coking and wear (DI>0.85, 48 hours in advance) is realized; based on the simplified flue gas fluid model of Navier-Stokes equation, the prediction error of ash hopper ash accumulation height is ≤3%; the closed loop of “AR instruction → digital deduction → real machine regulation” constructed can pre-visualize the influence of valve opening adjustment on boiler efficiency in virtual environment (such as opening degree increase of 10% → efficiency +1.2%), and after verification, it is issued to the field equipment in seconds, providing a highly reliable digital foundation for the construction of a new power system. IV. BRIEF DESCRIPTION OF DRAWINGS

[0019] The present application relates to a thermal power plant operation digital twin system, and the technical implementation involves the cooperative work of multiple modules, including three-dimensional modeling, data acquisition, digital twin engine, simulation and visualization, intelligent analysis and user interaction, etc. To more clearly show the structure and function of the present application, the drawings are as follows:

[0020] Figure 1 : System overall architecture diagram

[0021] This diagram shows the overall architecture of the thermal power plant operation digital twin system, including the application interaction layer, the cloud intelligent layer, the edge computing layer, and the physical perception layer. Through this diagram, the logical relationship and cooperative working mode between the levels of the system can be intuitively understood, providing technical reference for subsequent implementation.

[0022] Figure 2 : Multi-source data fusion flowchart

[0023] The figure shows the implementation process of data acquisition and fusion calculation to transmission in the digital twin system of thermal power plant operation, including data acquisition, data transmission and preprocessing, feature fusion and decision output. The figure shows the complete process of data acquisition and fusion calculation to transmission, providing data support for system operation. V. DETAILED DESCRIPTION

[0024] The three-dimensional operation state digital twin system of the thermal power plant of the present application has the following full-process technical solutions in its specific embodiments:

[0025] In the system design stage, the entire physical environment of the thermal power plant needs to be deeply mapped, and the spatial topological relationship of the equipment of the boiler island, the steam turbine island and the ash conveying system is collected. Based on the unmanned aerial vehicle laser scanning (precision ±2 cm) and BIM reverse modeling technology, a millimeter-level precision three-dimensional model is constructed using the Unity engine, with special attention paid to the geometric restoration of the internal grid structure of the ash hopper, the piston motion trajectory of the warehouse pump and the stress concentration area of the pipe elbow. The model uses PBR physical material shader to accurately simulate the optical properties of the device surface, such as the oxidation and rust texture of high-temperature pipes at 550°C and the mirror reflection effect of valve metal parts. All models support FBX / OBJ / glTF universal format import, and through the layered dynamic LOD mechanism, the rendering performance is optimized - within 1 meter from the camera, 500,000 face bolt-level high models are loaded, and 50 meters away, simplified to bounding box colliders, ensuring that a million face scene maintains a 90fps 2K frame rate on an RTX 4080 graphics card.

[0026] In the data acquisition stage, a multi-source sensor network is deployed, temperature sensors (range 0-800°C) are embedded in the ash hopper wall, pressure transmitters (accuracy ±0.1% FS) are installed at the inlet and outlet of the warehouse pump, and high-frequency vibration probes (sampling rate 10 kHz) are installed at key nodes of the pipeline. Through the OPC UA gateway, real-time aggregation of DCS control signals, SIS performance data and TSI vibration spectrum is realized, and the UniRx responsive framework is used to build an asynchronous stream processing channel, processing 200,000+ data points per second. To cope with high concurrency transmission, a dynamic Retrofit configuration layer is designed: Delta-ZigZag encoding is used to compress the data stream (bandwidth reduction of 70%), multi-threaded data cleaning is realized through LINQ extension operators, and based on the sliding time window algorithm, the timestamps of heterogeneous data are aligned (deviation ≤10 ms). Spatial positioning data is mapped to the three-dimensional model coordinates through a conversion matrix, enabling sub-millimeter-level matching between vibration monitoring points and steam turbine bearing models.

[0027] The digital twin engine uses the Unity HDRP pipeline to achieve physical-level rendering. The lighting system simulates the factory environment light shading effect, and the global illumination is calculated in real time through ray tracing to determine the shadow projection relationship of the device. The device state driving engine associates real-time data with model behavior: when the ash bucket level is greater than 80%, the vertex shader displacement (height is lifted by 8 meters) is triggered, and when the valve opening is greater than 75%, the hydraulic rod extension animation is played. The dynamic particle system is developed based on VFX Graph, which constrains the smoke particle trajectory through pre-baked signed distance field (SDF), and realizes the physical simulation of fly ash concentration in the ash conveying pipeline - for every 0.1% increase in coal sulfur content, the smoke particle density increases by 1200 units per cubic meter.

[0028] The simulation and visualization module integrates a multi-physics coupling engine. The ash bucket ash falling frequency model (Q_coal is the coal supply) drives the particle generation system, which dynamically renders 2000+ dust particles per second; the compressed air ash blowing process uses the Navier-Stokes equation to solve, and under GPU parallel computing, it simulates the airflow vortex form in real time (velocity field accuracy ± 0.3 m / s). The visualization interface supports three-dimensional penetrating navigation: click on the dust collector model, the camera automatically flies to the target position, and the right panel dynamically displays the import and export pressure difference curve (such as the current value 2.4 kPa / warning threshold 3.0 kPa) through the XCharts plug-in. The overheated equipment automatically triggers a triple alarm - the three-dimensional model surface is covered with a fiery red heat map, the AR glasses project the positioning coordinates, and the digital large screen generates a maintenance work order.

[0029] The intelligent decision-making module deploys an LSTM time series prediction model. Input the vibration spectrum (feature vector after FFT transformation) and temperature gradient data, output the device degradation index DI (calculation formula), and push the bearing replacement suggestion 72 hours in advance when DI > 0.85. The fault deduction engine builds an accident chain model based on the SDF physical field: input the steam pipe rupture point coordinates (X = 28.4 m, Y = 7.1 m) and pressure parameters 8.5 MPa, real-time solve the high-temperature steam diffusion vector (update particle motion trajectory every second), predict the radius of the dangerous area within 30 seconds (accuracy ± 0.5 meters), and generate an AR guide animation to assist emergency disposal.

[0030] The user interaction layer supports multi-modal control. The operation and maintenance personnel control the first-person perspective inspection through the VR handle, and the head tracking technology realizes the overhead inspection of the coal conveying pipe gallery; the voice command "increase the fan speed to 1500 rpm" is analyzed by NLP to generate OPCUA control signals, which are directly connected to the PLC actuator (response delay ≤200 ms). The digital large screen uses modular UI design: the column chart dynamically compares the before and after coal consumption (such as 142 g / kWh → 135 g / kWh), the scatter plot shows the device health state distribution, and the three-dimensional model and data charts are real-time linked through the event bus.

[0031] The system deployment phase adopts an edge-cloud collaborative architecture. The local server in the plant runs the real-time control engine (processing delay ≤ 50 ms), and the cloud trains the LSTM prediction model for daily incremental updates. The security system implements triple protection: OPC UA communication uses AES-256 encryption, and historical operation logs are stored on the chain to ensure traceability. In the maintenance phase, sensors are calibrated every month, and the device digital gene library is updated every quarter (integrating the latest maintenance records and failure modes), and model optimization packages are pushed through OTA, such as improving the YOLOv5 bolt loosening recognition accuracy to 99.2%.

[0032] The extended upgrade phase supports process iteration. When a desulfurization tower is added, 3D modeling is completed within 72 hours through point cloud scanning-BIM fusion technology. After connecting to online carbon emission monitoring data, the system dynamically optimizes the ash conveying frequency and air compressor start-stop strategy, reducing carbon emissions by 1.8 kg per megawatt-hour. After verification by Jingneng Chifeng Power Plant, the system reduces fault location time to 0.5 hours (an 83.9% improvement) and reduces annual operation and maintenance costs by 12 million yuan, providing a full-life-cycle management paradigm for thermal power digital twins.

Claims

1. A three-dimensional operating state digital twin system for a thermal power plant, characterized in that A "multi-source data fusion-dynamic physical deduction-virtual-real interactive control" full-process digital management system is constructed, through edge computing and cloud collaborative architecture, to realize holographic monitoring and predictive intervention of power plant equipment status, including the following levels: (1) Three-dimensional modeling layer: based on unmanned aerial vehicle laser scanning and BIM model reverse reconstruction technology, a millimeter-level precision three-dimensional twin of the plant area is constructed, using hierarchical dynamic LOD management mechanism and PBR physical material, to realize visual mapping of equipment status; (2) Data fusion layer: integrate asynchronous stream processing engine of multi-source heterogeneous data, through OPC UA protocol and sliding time window algorithm, realize 200,000+ data points / second millisecond-level synchronization and space-time alignment; (3) Dynamic deduction layer: physical rule model based on GPU parallel computing, real-time simulation of smoke flow, ash accumulation in ash bucket and other working conditions, support accident chain diffusion deduction triggered by custom variables; (4) Virtual-real control layer: establish a two-way control chain from AR instructions to PLC execution mechanism, form a "monitoring-deduction-intervention" closed loop, support virtual debugging and real-time regulation of equipment parameters.

2. The digital twin system of claim (1), wherein, The three-dimensional modeling layer includes: (1) Unmanned aerial vehicle laser scanning module, scanning accuracy ±2cm, realizing high-fidelity geometric modeling of boiler, steam turbine and other equipment; (2) Hierarchical dynamic LOD rendering module, presenting 500,000 high-detail surfaces within 1m, simplified to bounding box tracking at 50m, ensuring 90fps@4K frame rate under RTX 4080 graphics card; (3) Material dynamic mapping module, mapping temperature, vibration and other data into material color change and heat map visualization through Shader Graph nodes.

3. The digital twin system of claim (1), wherein, The data fusion layer includes: (1) Multi-source data access module, fusion of DCS control signal, SIS historical data, infrared temperature measurement and vibration monitoring and other heterogeneous data; (2) Asynchronous stream processing module, realize millisecond-level synchronization of data by using UniRx framework, reduce 70% bandwidth occupancy by Delta-ZigZag encoding; (3) Space-time alignment module, map two-dimensional data to corresponding positions on three-dimensional model through conversion matrix, timestamp error ≤10ms.

4. The digital twin system of claim (1), wherein, The dynamic deduction layer includes: (1) Load response model, based on Unity VFX Graph to simulate the flow motion of smoke in the directed distance field (SDF), the smoke concentration calculation formula is 0.8×sulfur + 0.2×air volume; (2) Fault chain prediction module, input infrared temperature >500℃ and vibration >7mm / s signal, deduce the coking area within 24 hours through SDF particle diffusion model, positioning accuracy ±5cm; (3) Multi-physical field coupling engine, simulate the temperature decay curve (12℃ per meter) and fluid dynamics characteristics in the process of steam diffusion.

5. The digital twin system of claim (1), wherein, The virtual-real control layer includes: (1) AR instruction interaction module, support VR handle and voice control, instruction response delay ≤200ms; (2) Closed-loop control module, directly connect PLC execution mechanism through OPC UA protocol, realize "virtual debugging-real machine regulation" second-level delivery; (3) Emergency disposal module, generate accident disposal scheme within 10 minutes, deduce result error ≤0.5 meters.

6. The digital twin system of claim (1), wherein, Also includes intelligent analysis layer, deployment LSTM time series prediction model and YOLOv5 visual detection model: (1) Equipment degradation index DI calculation model, DI = α × vibration amplitude + β × temperature gradient + γ × load fluctuation rate, α, β, γ are coal quality calibration coefficients, DI > 0.85, 72 hours in advance warning; (2) AI defect identification module, fusion infrared thermal image and vibration spectrum analysis, defect detection rate of 98.7%, efficiency is improved by 400% compared with artificial inspection.

7. The digital twin system of claim (1), wherein, System core performance indicators include: (1) Fault location time ≤ 0.5 hours, operation and maintenance cost ≤ 29.6 yuan / MWh; (2) Ash hopper ash accumulation height prediction error ≤ 3%, smoke flow simulation frame rate 90 fps; (3) Emergency response time is shortened from 26 minutes to 4.3 minutes, and the rate of avoiding secondary accidents is 93%.

Images