Photovoltaic power station early warning and platform construction method based on digital twinning

Abstract

The invention discloses a photovoltaic power station early warning and platform construction method based on digital twinning, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: building a digital twinning body through employing a three-dimensional modeling engine based on a design drawing, geographic information system data and equipment asset information of a power station; establishing a unified data middle station, accessing multi-source heterogeneous data, and binding real-time data with entities in the digital twin through a data mapping rule to realize virtual-real linkage; an AI analysis model is integrated, a multi-stage early warning rule engine is configured, and fault early warning and risk grading are realized based on model output and historical data; on the basis of the early warning information, an operation and maintenance work order is automatically generated in combination with an operation and maintenance knowledge base, processing suggestions are provided, the whole work order execution process is tracked, and closed-loop management from early warning to verification is achieved; functions are modularized by adopting a micro-service architecture and packaged into independent services, a visual interaction portal is developed, and an intelligent operation and maintenance platform is formed through deployment.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method for early warning and platform construction of photovoltaic power plants based on digital twins. Background Technology

[0002] With the continuous expansion of photovoltaic power plant scale, especially against the backdrop of rapid construction of large-scale bases in desert and Gobi areas, traditional operation and maintenance (O&M) models are facing increasingly severe challenges. Various system data are isolated; for example, drone inspection data, SCADA operation data, and environmental monitoring data are stored in different formats and are difficult to integrate, analyze, and utilize uniformly, forming information silos. O&M response is generally lagging, relying heavily on reactive repairs, resulting in untimely fault detection, which not only increases power generation losses but also drives up maintenance costs. The lack of trend prediction and early diagnostic capabilities for equipment status makes maintenance planning lack forward-looking basis and decision-making insufficiently scientific. Furthermore, O&M processes rely on manual experience, resulting in low efficiency and difficulty in closing loops from work order dispatch and on-site execution to result acceptance, thus restricting the level of refined management.

[0003] Chinese invention patent CN119784970B discloses a method, device and equipment for generating a digital twin simulation model of a photovoltaic power station. It uses UAV orthophotos and 3D modeling technology to mark equipment units, extract geographic information and create a hierarchical model to achieve high-precision digital reconstruction of the photovoltaic power station. However, its functional boundaries are basically limited to visualization and basic information query, and it lacks deep integration with real-time data streams, which makes it impossible for the virtual model to dynamically map the operating status of the physical power station.

[0004] Digital twin technology, by constructing virtual mappings of physical entities, provides a feasible path for the intelligent transformation of power plant operation and maintenance. However, most current digital twin applications remain at the level of 3D visualization or basic data monitoring, failing to deeply integrate mechanistic models, AI algorithms, real-time data, and operation and maintenance business processes, thus failing to form an integrated intelligent closed loop encompassing perception, diagnosis, early warning, decision-making, and execution. Therefore, there is an urgent need to construct a systematic methodology for building digital twin platforms, promoting their evolution from visual to a higher level of knowability, predictability, and decision-making, to support the intelligent operation and maintenance and refined management of photovoltaic power plants. Summary of the Invention

[0005] This application provides a method for early warning and platform construction of photovoltaic power plants based on digital twins, which solves the problems of information silos, passive response, lack of equipment status trend prediction and early diagnosis capabilities, and low efficiency of operation and maintenance processes faced by photovoltaic power plants in the prior art. It realizes the technical effects of full-element digital mapping through the construction of digital twins and real-time fusion of multi-source heterogeneous data, early fault warning and risk classification based on AI models and multi-level early warning rule engines, automatic generation of work orders driven by operation and maintenance knowledge base and closed-loop management of the whole process, and scalable platform deployment supported by microservice architecture, thereby improving the level of intelligent and refined operation and maintenance management.

[0006] This application provides a method for early warning and platform construction of photovoltaic power plants based on digital twins, including: S1: Based on the power plant's design drawings, geographic information system data, and equipment asset information, a digital twin is constructed using a 3D modeling engine; S2: Establish a unified data platform to access and integrate multi-source heterogeneous data from drone inspection systems, data acquisition and monitoring control systems, weather stations and video surveillance. Bind real-time data to entities in the digital twin through data mapping rules to achieve virtual-real linkage. S3: Integrates AI analysis models, including visual defect detection models and performance prediction models, and is equipped with a multi-level early warning rule engine to achieve early warning of faults and risk classification based on model output and historical data; S4: Based on early warning information and combined with the operation and maintenance knowledge base, it automatically generates operation and maintenance work orders, provides handling suggestions, and tracks the entire process of work order execution to achieve closed-loop management from early warning to verification; S5: Adopts a microservice architecture to modularize functions and encapsulate them as independent services. It develops a visual interactive portal, integrates data dashboards, 3D scenes, early warning centers and work order management functions, and deploys them to form an intelligent operation and maintenance platform.

[0007] Furthermore, the 3D modeling engine adopts the Unity3D engine and uses level of detail technology to perform hierarchical processing of the model; among them, for key equipment such as inverters and transformer substations, the number of faces in the high-precision model is no less than 500,000, the number of faces in the medium-precision model is [10,500,000, 50,000, and the number of faces in the low-precision model is less than 10,000, and the building information model and geographic information system data are integrated; at the same time, PBR material is used to simulate the optical properties of the equipment surface, and a dynamic lighting system is constructed to simulate the changes in the angle of sunlight and shadows at different times in real time.

[0008] Furthermore, the data in the data platform is built on a high-throughput data bus based on Apache Kafka, which accesses and integrates heterogeneous data from multiple sources; the incoming data is cleaned and standardized in format, including filtering outliers, unifying the timestamp to ISO8601 format, and completing the device's unique identifier; the cleaned data is stored in a time-series database and a relational database according to its characteristics, and the real-time data is bound to the corresponding entity in the digital twin through the device's globally unique identifier (UUID); the WebSocket protocol is used to actively push data, and the data refresh frequency is no more than 5 seconds to ensure the real-time dynamic mapping between the virtual model and the physical world.

[0009] Furthermore, the data mapping rules include: establishing a precise matching table between the device's unique identifier (UUID) and the model instance in the digital twin; specifying that data points in the physical world are mapped to the coordinate positions and attribute fields of the corresponding model instance in the virtual world through the UUID; when real-time data flows in, mapping the electrical parameters of the data acquisition and monitoring control system to the numerical panel displayed next to the model according to the matching table, mapping the defect types identified by the UAV to the model's color status, and mapping meteorological data to scene lighting parameters; and ensuring that the spatial position matching accuracy error between the data and the model does not exceed 0.5 meters through a virtual-real mapping consistency verification mechanism.

[0010] Furthermore, the visual defect detection model is a vision-based computationally intensive model that processes high-definition images captured by the UAV inspection system. The input is image data, and the output is a structured recognition result containing defect type, location, and confidence level. It adopts an asynchronous calling mode and achieves non-blocking processing through task ID and callback mechanism. The performance prediction model is a data-driven time series analysis model. It processes multi-dimensional numerical sequences provided by data acquisition and monitoring control systems and meteorological stations. The input is historical and real-time data, and the output is the performance degradation trend and loss prediction value of photovoltaic modules. It can provide early warning of typical faults ≥72 hours in advance and adopts a synchronous calling mode to meet the low-latency response requirements.

[0011] Furthermore, the multi-level early warning rule engine includes: establishing a three-level early warning system of prompt level, warning level, and severe level, with each level corresponding to different response strategies and notification methods; supporting compound logical judgments based on threshold conditions, trend conditions, and correlation conditions; the rule engine, as a background resident service, listens to data bus messages and uses the Rete algorithm to match incoming AI analysis results with real-time data; introducing a debouncing mechanism, requiring abnormal states of volatile data to persist for a set time before triggering an early warning, avoiding false alarms due to instantaneous noise; after an early warning is triggered, a structured early warning record is generated and pushed to the visual portal early warning center in real time, and relevant personnel are notified via SMS, email, and mobile terminals according to the level.

[0012] Furthermore, the early fault warning and risk classification includes: based on the structured recognition results output by the visual defect detection model and the performance degradation trend prediction value generated by the performance prediction model, combined with historical operating data, a multi-level warning rule engine is used to achieve early fault detection; the warning rules set differentiated thresholds according to the equipment type, setting a prompt-level warning when the visible defect area of ​​the photovoltaic module surface is ≥5%, a warning-level warning when the string-level performance loss exceeds 5% for three consecutive prediction values, and a severe-level warning when the inverter's critical equipment experiences abnormal temperature and a sudden power drop of ≥20%; the risk classification is based on a comprehensive assessment of the warning level, the criticality of the equipment, and the scope of impact, with the prompt level corresponding to locally observable anomalies, the warning level corresponding to potential faults that may affect power generation efficiency, and the severe level corresponding to emergency situations that lead to equipment shutdown and safety accidents.

[0013] Furthermore, the operation and maintenance knowledge base integrates static asset data, dynamic operation data, early warning work order records, and external knowledge documents to build a structured knowledge base. It adopts a hybrid database architecture to store device attributes, performance history, fault paths, and unstructured data respectively. It supports intelligent retrieval and weighted optimization recommendation based on device type and early warning keywords, and implements version control for knowledge updates.

[0014] Furthermore, the maintenance work order includes: when an early warning instance is created and its level is warning, the system automatically starts the work order generation program. The new work order inherits the associated early warning ID, target device information, and problem description from the early warning instance to form an initial data skeleton; the system calls the diagnostic interface of the maintenance knowledge base using the early warning information as the query condition to obtain a structured processing solution containing processing suggestions, security specifications, risk levels, required tools, and spare parts; the work order system automatically fills in the work order title, detailed work instructions, safety precautions, and resource list based on the solution returned by the knowledge base; after the work order content is completed, it enters the dispatch stage and pushes a notification to the designated maintenance personnel.

[0015] Furthermore, the visual interactive portal includes: using the Vue.js framework as the foundation for front-end development, combined with the Three.js library to achieve real-time rendering and interaction of the 3D scene on the Web; integrating the ECharts chart library to build a 2D data dashboard, dynamically displaying the power plant's total real-time power, daily cumulative power generation, defect statistics, early warning information, and key performance indicators; the portal interface integrates four major functional modules: a global situation overview dashboard, a 3D digital twin main scene, an early warning center, and a work order management center; supporting multi-terminal adaptive access, and communicating with the backend microservices in real time via the WebSocket protocol to achieve data-driven visual updates.

[0016] The aforementioned photovoltaic power plant early warning and platform based on digital twins includes: Digital Twin Module: Stores and manages high-fidelity 3D virtual models and equipment asset data of photovoltaic power plants; Data Platform and Fusion Module: Responsible for the access, cleaning, fusion, and real-time push of multi-source data; Intelligent Early Warning Engine Module: Built-in AI analysis models and early warning rules, responsible for fault diagnosis and risk warning; Operation and Maintenance Decision and Work Order Management Module: Generates operation and maintenance strategies based on early warning results and manages the entire lifecycle of work orders; Visual Interactive Portal: Provides users with a unified graphical operation interface, displaying the overall status of the power plant and operation and maintenance information.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: By employing 3D modeling and multi-source data fusion technologies, the system achieves digital mapping and virtual-real linkage of all elements of the photovoltaic power station, enabling intuitive understanding of the power station's operating status through 3D scenes. By integrating AI analysis models and a multi-level early warning rule engine, it achieves early detection and accurate early warning of faults. By building an operation and maintenance knowledge base and an automatic work order generation system, it achieves fully automated closed-loop management from early warning to handling, significantly improving operation and maintenance efficiency. Through the development of a microservice architecture and a visual interactive portal, it achieves high scalability and user-friendliness of the platform. Attached Figure Description

[0018] Figure 1 This is a flowchart of a photovoltaic power plant early warning and platform construction method based on digital twins in an embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Example 1: As Figure 1 As shown, a method for early warning and platform construction of photovoltaic power plants based on digital twins is presented.

[0022] S1: Based on the power plant's design drawings, geographic information system data, and equipment asset information, a digital twin is constructed using a 3D modeling engine; Specifically, the process involves acquiring multimodal data, including collecting original CAD drawings, equipment layout diagrams, electrical wiring diagrams, and BIM models of the power plant; obtaining high-precision GIS maps and on-site survey data to determine the absolute coordinates and topography of the power plant and equipment; compiling an equipment asset list, including the unique code, model, specifications, supplier, installation date, and technical parameters of each piece of equipment; and for existing power plants, using UAV oblique photogrammetry technology for aerial photography, generating realistic 3D point cloud models and orthophotos through photogrammetry algorithms for verifying design drawings and supplementing details. The collected multimodal data undergoes standardization processing, including coordinate system unification, data format conversion, and data association, forming a structured master data package that can be directly used by 3D modeling software.

[0023] The 3D modeling engine uses the Unity3D engine and employs level of detail technology to process the model in a hierarchical manner. For key equipment such as inverters and transformer substations, the high-precision model has no less than 500,000 faces, the medium-precision model has [10,000,000,000] faces, and the low-precision model has less than 10,000 faces. It also integrates building information model and geographic information system data. At the same time, PBR material is used to simulate the optical properties of the equipment surface, and a dynamic lighting system is constructed to simulate the changes in sunlight angle and shadow at different times in real time.

[0024] Specifically, the Unity3D engine is used as the 3D modeling and rendering platform to support smooth interaction and high-quality visualization of large-scale scenes. Based on the GIS and site plan, the site topography and infrastructure such as roads and walls are established. Modeling is refined level by level according to a hierarchical structure: overall site, power generation unit area, photovoltaic array, string, and photovoltaic module / inverter / substation. Each level of model is given a clear parent-child hierarchy, and each device model exists as an independent instance, bound to its globally unique identifier. To address the performance bottleneck of large-scale scene rendering, multiple levels of detail models are created for the same object, using LOD (Level of Detail) technology. For example, for an inverter, three levels of models are created: high, medium, and low. When the camera zooms in, the high-poly model is displayed, with no less than 500,000 faces, accurately reproducing all details such as the device's appearance, buttons, and labels; the medium-poly model is displayed at mid-range distances, with 10,000 to 50,000 faces, retaining the main structure; and the low-poly model is displayed in the distance, with fewer than 10,000 faces, representing only simple geometry. Using PBR (Physically Based Rendering) materials, the system accurately simulates the gloss, roughness, and reflectivity of different surfaces such as metal, glass, and plastic. Simultaneously, a dynamic lighting system is constructed to simulate changes in sunlight angle and shadows at different times of day.

[0025] The pre-processed equipment asset information is associated with the corresponding entities in the 3D model. In the model database, a data table is created for each model instance, with fields including equipment ID, name, model, rated power, and commissioning date. When developing the visualization interaction module, logic code is written to implement a click-to-query function. When a user clicks on a equipment model in the 3D scene, the system can query the database based on the bound unique identifier and display all its static attributes and dynamic operating data in real time. This generates a structured virtual power plant model containing geometric information, topological relationships, and semantic attributes.

[0026] S2: Establish a unified data platform to access and integrate multi-source heterogeneous data from drone inspection systems, data acquisition and monitoring control systems, weather stations and video surveillance. Bind real-time data to entities in the digital twin through data mapping rules to achieve virtual-real linkage. The data platform is built on a high-throughput data bus using Apache Kafka, which integrates and merges heterogeneous data from multiple sources. The incoming data is cleaned and standardized, including filtering outliers, unifying timestamps to ISO8601 format, and completing the device's unique identifier. The cleaned data is stored in a time-series database and a relational database according to its characteristics, and real-time data is bound to the corresponding entity in the digital twin through the device's globally unique identifier (UUID). The WebSocket protocol is used for proactive data push, with a data refresh frequency of no more than 5 seconds, to ensure real-time dynamic mapping between the virtual model and the physical world.

[0027] Specifically, a unified data platform is established to access and integrate multi-source heterogeneous data from UAV inspection systems, SCADA systems (data acquisition and monitoring control systems), weather stations, and video surveillance. Through data interfaces and mapping rules, real-time data is bound to corresponding entities in the digital twin, enabling real-time dynamic mapping of the virtual model to the physical world.

[0028] After each inspection mission, the drone inspection system automatically receives a defect report packaged in JSON format. The report includes a URL link to the defect image, precise GPS coordinates, defect type, and confidence level. The SCADA system interfaces via an industry-standard protocol, acquiring electrical parameters such as current, voltage, power, and frequency from inverters, combiner boxes, and other equipment at a high frequency of 1-5 seconds. The weather station accesses environmental data such as irradiance, ambient temperature, wind speed, wind direction, and humidity from meteorological sensors within the site via 4G / 5G networks or a local area network, with an update frequency typically 1-5 minutes. The video surveillance system accesses real-time video streams from key equipment areas via the GB / T28181 national standard protocol and supports PTZ control. Apache Kafka is used as a high-throughput, scalable data bus; all incoming data is first published to the corresponding topics on the bus.

[0029] Following the data bus, a stream processing module is deployed to perform data cleaning and format standardization. This filters out outliers that clearly exceed physical limits, such as negative power values, and handles null values ​​caused by data packet loss. Timestamps from different data sources are standardized to the ISO 8601 standard format, all numerical units are unified, and unique device identifiers are added.

[0030] Establish a data association model and assign a globally unique identifier (UUID) to each device. Using the UUID, data from different systems describing the same device can be linked together. For example, real-time power data of string A01 in SCADA, defect records on that string discovered by a drone, and meteorological data for the area can be correlated at the data level to form a complete data profile of the device.

[0031] Data is stored in different databases based on its characteristics to optimize query efficiency: time-series databases are used to store time-stamped and frequently written sequence data such as SCADA data and meteorological data, supporting efficient time range queries and aggregation analysis; relational databases are used to store structured data such as equipment static attributes, inspection report metadata, early warning rules, and work order information; and object storage is used to store unstructured big data such as high-definition images and video recordings taken by drones.

[0032] The data mapping rules include: Specifically, a data mapping rule table is set up to define the correspondence between data points in the physical world and model instances in the virtual world through precise matching of device UUIDs. The low-latency WebSocket communication protocol is used, and when the data governance layer generates new valid data, it is proactively and quickly pushed to the front-end browser via the protocol. Upon receiving the data, the front-end visualization engine finds the corresponding model instance in the scene according to the preset mapping rules and dynamically updates its status, including updating the real-time power and temperature display panel next to the model; mapping the defect types identified by the drone to model color status (green for normal, yellow for warning, and red for severe alarm); mapping meteorological data to scene lighting parameters; and ensuring that the spatial location matching accuracy between data and the model does not exceed 0.5 meters through a virtual-real mapping consistency verification mechanism.

[0033] The system uses Vue.js as the front-end framework, Three.js as the core engine for 3D rendering on the web, and the ECharts charting library for building 2D data dashboards. The Socket.io client library is used to establish a WebSocket connection with the backend data service, enabling real-time data push. The front-end interface mainly includes a global situation overview dashboard, a 3D digital twin main scene, and an early warning and work order management center.

[0034] The overall situation overview dashboard serves as the homepage, providing managers with an overview of the power plant's operating status through two-dimensional charts and key performance indicators. It displays core indicators such as the power plant's total real-time power, daily cumulative power generation, equivalent utilization hours, and system efficiency; and presents statistics on the number of defects, real-time early warning information, and environmental parameters in chart form.

[0035] The 3D digital twin main scene serves as the visual and interactive center of the portal, occupying the main screen area. Its functions include providing smooth scene loading progress indicators and offering multiple browsing modes such as first-person roaming, God's-eye view, and path navigation; enabling mouse clicks to select models, with clicking any device immediately displaying an information card that dynamically shows the device's static attributes (such as model and capacity) and real-time operating data (such as current, voltage, and temperature) from the data platform; mapping data to visual representations, for example, using a color gradient from green (high efficiency) to red (low efficiency) to render the string model based on the string's real-time power or efficiency value, making abnormal areas immediately apparent; and providing a search box or device tree, where the camera automatically pans and zooms to the device and highlights it after the user enters its ID.

[0036] The alert and work order management center integrates the alert center and work order management list in the form of a sidebar or pop-up window. Clicking on an alert in the alert list will automatically locate and highlight the corresponding faulty device in the 3D scene. Conversely, clicking on a device in an alarm state in the scene will also select the corresponding alert or work order details in the list.

[0037] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application achieves high-precision digital mapping of all elements of a photovoltaic power station by systematically integrating multimodal data acquisition, high-fidelity 3D modeling, and real-time driving of multi-source heterogeneous data. This enables operation and maintenance personnel to intuitively see the overall situation and details of the power station through an immersive 3D scene, changing the limitations of information silos and static monitoring in traditional operation and maintenance. Through multi-source data fusion based on the Apache Kafka data bus and WebSocket protocol, and virtual-real linkage with a refresh rate of ≤5 seconds, it ensures that the virtual model dynamically and synchronously reflects the state of the physical world, providing a real-time data foundation for early fault detection. Relying on the visual interactive portal built with Vue.js and Three.js, data-driven visualization and multi-terminal linkage interaction lower the threshold for operation and maintenance decision-making and improve efficiency.

[0038] Example 2: Example 1 achieved panoramic monitoring and data fusion of photovoltaic power plants, but it still lacked intelligent early warning and diagnostic capabilities. This example further supplements and explains the content of Example 1.

[0039] S3: Integrates AI analysis models, including visual defect detection models and performance prediction models, and is equipped with a multi-level early warning rule engine to achieve early warning of faults and risk classification based on model output and historical data; Specifically, the platform integrates two core AI analysis models: a visual defect detection model and a performance prediction model. The visual defect detection model is a computationally intensive model that processes high-resolution images captured by drones. Its input is image data, and its output is structured defect identification results. It adopts an asynchronous call model; after submitting the image URL, the platform receives the task ID first, and the results are retrieved via callback or query after the model completes processing. The performance prediction model is a data-driven model that processes time-series data. Its input is a multi-dimensional numerical sequence, and its output is a predicted future performance degradation trend or loss value. It can provide early warnings of typical faults (such as hot spots and string faults) at least 72 hours in advance, providing ample time for proactive maintenance. It adopts a synchronous call model, requiring low response latency for rapid prediction. A unified API interface specification is designed for both models to simplify the calling process.

[0040] To ensure that other parts of the platform can easily and reliably call these AI microservices, an API gateway is deployed in front of the AI ​​service cluster. All external requests first reach the gateway, which is responsible for request routing, load balancing, authentication, authorization, rate limiting, and circuit breaking to prevent malicious requests or traffic surges from overwhelming the AI ​​services. A dedicated model scheduler microservice is created on the platform backend. Its functions include: subscribing to the data platform's message queue; retrieving the specific data required for the current analysis task from the data platform; calling the corresponding AI service interface through the API gateway according to the task type; formatting the returned results from the AI ​​service and publishing them as a new message to the data bus for consumption by the early warning rule engine.

[0041] Integrate a monitoring agent into each AI microservice to collect metrics such as service response latency and error rate. Simultaneously, record detailed inference logs to facilitate troubleshooting and model performance analysis. Include version numbers in the API interface design to provide a smooth transition for subsequent model iterations and upgrades.

[0042] The multi-level early warning rule engine includes: establishing a three-level early warning system of prompt level, warning level, and severe level, with each level corresponding to different response strategies and notification methods; Specifically, the platform management backend provides a graphical rule configuration interface, supporting the creation, modification, and enabling / disabling of alert rules. A three-tiered alert system is established, with each tier corresponding to different response strategies and notification methods: The alert level (blue) is for minor anomalies or trend reminders, such as a slight decrease in component efficiency, requiring only recording and monitoring without immediate intervention. The warning level (yellow) indicates a clear failure risk, such as performance loss predictions exceeding 5% multiple times consecutively accompanied by visible defects, requiring an inspection within 24 hours. The critical level (red) indicates a major failure that has occurred or is about to occur, such as SCADA data interruption and AI identifying hotspots, requiring immediate response. Multiple condition types are supported for combined judgments: threshold conditions (directly comparing data values ​​with preset thresholds); trend conditions (analyzing data changes over time); and correlation conditions (cross-data source correlation judgments).

[0043] The rules engine, running as a persistent background service, listens for messages on the data bus. When new AI analysis results or real-time SCADA data flow in, the engine automatically treats these data as "events" to trigger the evaluation of relevant rules. The engine loads all enabled rules and matches the incoming event data with the rule conditions. High-efficiency algorithms such as Rete are used to optimize matching speed, ensuring that complex rule judgments are completed in milliseconds. Once the conditions are met, the engine immediately instantiates an alert record, including the alert ID, trigger time, device information, and rule details. A debouncing mechanism is introduced; for example, for fluctuating data, an alert is only triggered after the abnormal state persists for a certain period (e.g., 5 minutes), avoiding false alarms caused by instantaneous noise.

[0044] When the rules engine triggers an alert, the system automatically generates a structured alert record containing the following key fields: Basic Information: Alert ID (unique identifier), trigger time, alert level (hint / warning / critical), alert type. Spatial and Device Information: Associated device UUID, device name, location / array / string, and spatial coordinates in the 3D scene. Diagnostic Information: Specific rule description, associated AI model output, and relevant real-time data snapshot. Processing Status: Initially a new alert, updated as the processing flow progresses.

[0045] Based on the alert level and preset strategies, real-time push notifications are sent through multiple channels, refreshed in real-time as a list in the alert center of the visualization portal, and alerted with unread count badges and sound prompts. For warnings and critical alerts, core information is pushed to the mobile terminals of relevant operations and maintenance personnel through integration with SMS gateways, email, etc. Push frequency and escalation rules are configurable.

[0046] In the 3D digital twin master scene, visual icons are dynamically overlaid on the device models that issue warnings, based on the warning level, and periodically highlighted and flashed to achieve rapid location. For regional issues, a heatmap layer of warning distribution can be provided to quickly identify high-incidence areas of warnings.

[0047] The alert center list offers filtering and sorting capabilities by time, level, type, device, and status. Clicking on any alert in the list expands the details panel to view all structured information. A location button is provided; clicking it automatically pans and zooms the 3D scene camera to the device and highlights it. Clicking an alert icon in the 3D scene automatically scrolls the alert center list and selects the corresponding alert record.

[0048] The alert status has a clearly defined status transition path, including new alert, confirmed, processing, resolved, false alarm, and closed. Status changes require access control, and the operator and time of the operation must be recorded. Upon receiving an alert, operations personnel must first confirm receipt, adding a text explanation during confirmation. If the alert is determined to be a false alarm, the false alarm status can be selected, and the reason for the false alarm must be filled in, serving as an important basis for optimizing alert rules. For genuine alerts, after confirmation, they can be transferred to the work order system and enter the processing flow.

[0049] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application integrates AI models into microservices to build a configurable, multi-level early warning rule engine, and achieves visualization and closed-loop management of the early warning lifecycle, thereby improving the intelligence level of photovoltaic power plant operation and maintenance. The microservice architecture ensures the high availability and elastic scalability of AI analysis capabilities, supporting real-time response to defect detection and performance prediction. The rule engine achieves accurate fault diagnosis and early warning through composite logic judgment and risk classification. At the same time, the visual interface and state machine design enable early warning information to be intuitively perceived, quickly located, and traced throughout the entire process. Combined with automatic effect verification and data analysis, it not only reduces the false alarm rate and optimizes the efficiency of operation and maintenance decision-making, but also enhances the scalability and stability of the platform, providing a comprehensive and reliable solution for the transformation of photovoltaic power plants from passive monitoring to proactive early warning.

[0050] Example 3: Example 2 achieved intelligent early warning and diagnosis, but it still has shortcomings such as the failure to automatically translate early warning information into operational actions and the lack of an execution closed loop. This example further supplements and explains the content of Example 2.

[0051] The operation and maintenance knowledge base is a structured knowledge base built by integrating static asset data, dynamic operation data, early warning work order records and external knowledge documents. It adopts a hybrid database architecture to store device attributes, performance history, fault paths and unstructured data respectively. It supports intelligent retrieval and weight optimization recommendation based on device type and early warning keywords, and implements version control for knowledge updates.

[0052] Specifically, a knowledge base is constructed by acquiring multimodal data from the digital twin platform, including: static asset data, dynamic operational data, early warning and work order records, and external knowledge input. Static asset data consists of structured information extracted from the equipment asset list, such as unique equipment codes, models, and technical parameters. Dynamic operational data comprises real-time data streams from SCADA systems and weather stations, forming a historical sequence of equipment performance. Early warning and work order records store early warning events, processed work orders, and verification results, including fields such as trigger rules, handlers, and timestamps. External knowledge input includes industry standards, equipment manufacturer maintenance manuals, and expert experience documents, with key entities extracted using natural language processing technology. All data undergoes unified cleaning and formatting, including unifying timestamps to ISO8601 format to ensure temporal consistency; and mapping equipment identifiers to digital twin UUIDs to avoid information silos.

[0053] A hybrid storage strategy is adopted to optimize query efficiency based on the characteristics of knowledge. A relational database stores highly structured data such as device static attributes and work order process status, supporting complex SQL queries. A time-series database retains historical device performance data, facilitating trend analysis and pattern mining. A graph database stores network knowledge such as fault diagnosis paths and device relationships, enabling rapid graph traversal. Object storage archives unstructured data, such as drone inspection images and maintenance videos, and links them to corresponding devices through metadata indexes. B+ tree indexes are built for frequently queried fields (such as device ID and fault type); a full-text search engine is introduced, supporting keyword fuzzy matching and semantic search; for example, maintenance personnel can quickly locate relevant cases by inputting "string power anomaly".

[0054] When the early warning engine generates an alert (such as overvoltage on the AC side of inverter A-01), the system automatically calls the knowledge base API. Query criteria include: device type (inverter) and keywords (overvoltage) in the alert message. The knowledge base returns all possible causes, handling strategies, safety specifications, and spare parts lists matching the scenario. The knowledge base can record the processing effect of each piece of knowledge after it is adopted. Strategies that successfully resolve the fault will be given higher weight and prioritized for future recommendations.

[0055] For example, the early warning platform reports that the performance loss rate of string B-03 exceeds a threshold and associates it with a hot spot defect identified by a drone. The knowledge base receives a query request; the entity is a photovoltaic module, and the fault phenomena are hot spots and performance degradation. The knowledge graph quickly traverses to find nodes connected to the hot spot, potentially returning multiple paths: Path 1: Hot spot -> Cause: Bird droppings / dust obstruction -> Remedial strategy: Clean the surface. Path 2: Hot spot -> Cause: Internal cell damage -> Remedial strategy: Replace the module. The knowledge base combines the confidence levels of the two paths to generate a merged suggestion: Confirm on-site whether it is surface obstruction. If so, clean it; if not, check the module for physical damage and prepare to replace the module. Required tools: Cleaning equipment, IV tester. Safety regulations: Safety belts must be worn for working at heights. This structured suggestion is automatically populated into the newly generated work order to guide on-site operations.

[0056] A dedicated knowledge management specialist will regularly organize experts to review and update the knowledge base content. When a work order processing result differs from the knowledge base's recommended solution, or when a new failure mode is discovered, the system will mark the knowledge entry as requiring manual review. After confirmation, the new knowledge will be added to the knowledge base. Version control will be implemented for modifications to the knowledge base to ensure knowledge traceability and allow for rollback to previous versions in case of problems.

[0057] S4: Based on early warning information and combined with the operation and maintenance knowledge base, it automatically generates operation and maintenance work orders, provides handling suggestions, and tracks the entire process of work order execution to achieve closed-loop management from early warning to verification; Specifically, when an alert instance is created and its level is warning or severe, the system automatically initiates the work order generation process. The new work order automatically inherits all key information from the alert instance, forming an initial data skeleton, including: associated alert ID, target device information, and problem description. The system uses the alert information as a query condition to call the diagnostic interface of the knowledge base, which returns a structured processing solution. The work order system fills in the work order title, detailed work instructions, safety specifications and risk warnings, and spare parts and tool list based on the returned processing solution. After the work order content is completed, it enters the dispatch stage, updating the work order status to dispatched and sending a push notification to the target maintenance personnel's mobile APP or WeChat mini-program; the work order information is updated in the platform's work order management center, and relevant shift leaders or managers may be notified via email or SMS. Work orders have a complete lifecycle status, such as created, dispatched, received, processing, completed, under acceptance, and closed. Status changes are triggered automatically by the system or manually by maintenance personnel. A work order management center dashboard is set up in the platform's visual interactive portal. The status, progress, and responsible person information of all work orders can be displayed in real time in the form of lists, charts, or Gantt charts, which makes it convenient for managers to have a global grasp of the operation and maintenance dynamics and intervene in abnormal situations in a timely manner.

[0058] Once a work order associated with an alert is marked as completed, the system automatically initiates an effectiveness verification cycle (e.g., 24 hours). During this period, key performance indicators of the equipment are continuously monitored and compared with pre-maintenance data. If the indicators return to normal, the alert is automatically closed; if not, a prompt to re-investigate or escalate the alert may be displayed. The platform provides alert statistical analysis functions, such as generating alert quantity trend charts, alert percentages at each level, average response time, and false alarm rates. Closed alert cases are structured and stored in a case library. Regular reviews can be conducted to optimize alert rule thresholds, refine AI models, or enrich the operations and maintenance knowledge base.

[0059] S5: Adopts a microservice architecture to modularize functions and encapsulate them as independent services. It develops a visual interactive portal, integrates data dashboards, 3D scenes, early warning centers and work order management functions, and deploys them to form an intelligent operation and maintenance platform.

[0060] Specifically, the Spring Boot framework is used to develop various business microservices (user service, device service, early warning service, ticket service, etc.), and Spring Cloud is used for service governance. All services are containerized using Docker and orchestrated and deployed using Kubernetes to ensure high availability and elastic scaling. The Vue.js framework, combined with the Three.js library, is used for rendering 3D scenes on the web. The platform is deployed on public cloud, private cloud, or local servers. Nginx is used for load balancing and reverse proxying to ensure secure and stable access to the web interface. Simultaneously, a Role-Based Access Control (RBAC) model is adopted to achieve fine-grained management of multi-tenancy, regional, and role-based access, ensuring system data security and operational compliance.

[0061] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application successfully achieved full automation, standardization, and closed-loop management of the photovoltaic power plant operation and maintenance process by constructing a structured operation and maintenance knowledge base, realizing the automatic generation and dispatch of intelligent work orders, and promoting mobile on-site operations and execution feedback. By using knowledge graph technology to transform scattered operation and maintenance experience into machine-readable structured knowledge, early warning information can be automatically matched with the optimal handling solution, improving the scientific nature and accuracy of decision-making. Relying on a closed-loop verification mechanism, the system can automatically monitor the maintenance effect and feed back to optimize the knowledge base, forming a continuously self-learning autonomous operation and maintenance system, ultimately achieving a comprehensive improvement in operation and maintenance efficiency, management precision, and system self-adaptability.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a photovoltaic power plant early warning and platform based on digital twins, characterized in that, include: S1: Based on the power plant's design drawings, geographic information system data, and equipment asset information, a digital twin is constructed using a 3D modeling engine; S2: Establish a unified data platform to access and integrate multi-source heterogeneous data from drone inspection systems, data acquisition and monitoring control systems, weather stations and video surveillance. Bind real-time data to entities in the digital twin through data mapping rules to achieve virtual-real linkage. S3: Integrates AI analysis models, including visual defect detection models and performance prediction models, and is equipped with a multi-level early warning rule engine to achieve early warning of faults and risk classification based on model output and historical data; S4: Based on early warning information and combined with the operation and maintenance knowledge base, it automatically generates operation and maintenance work orders, provides handling suggestions, and tracks the entire process of work order execution to achieve closed-loop management from early warning to verification; S5: Adopts a microservice architecture to modularize functions and encapsulate them as independent services. It develops a visual interactive portal, integrates data dashboards, 3D scenes, early warning centers and work order management functions, and deploys them to form an intelligent operation and maintenance platform.

2. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The 3D modeling engine uses the Unity3D engine and employs level of detail technology to process the model in a hierarchical manner. For key equipment such as inverters and transformer substations, the high-precision model has no less than 500,000 faces, the medium-precision model has [10,000,000,000] faces, and the low-precision model has less than 10,000 faces. It also integrates building information model and geographic information system data. At the same time, PBR material is used to simulate the optical properties of the equipment surface, and a dynamic lighting system is constructed to simulate the changes in sunlight angle and shadow at different times in real time.

3. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The data platform is built on a high-throughput data bus using Apache Kafka, which integrates and merges heterogeneous data from multiple sources. The incoming data is cleaned and standardized, including filtering outliers, unifying timestamps to ISO8601 format, and completing the device's unique identifier. The cleaned data is stored in a time-series database and a relational database according to its characteristics, and real-time data is bound to the corresponding entity in the digital twin through the device's globally unique identifier (UUID). The WebSocket protocol is used for proactive data push, with a data refresh frequency of no more than 5 seconds, to ensure real-time dynamic mapping between the virtual model and the physical world.

4. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The data mapping rules include: establishing a precise matching table between the device's unique identifier (UUID) and the model instance in the digital twin; specifying that data points in the physical world are mapped to the coordinate positions and attribute fields of the corresponding model instance in the virtual world through the UUID; when real-time data flows in, mapping the electrical parameters of the data acquisition and monitoring control system to the numerical panel displayed next to the model according to the matching table, mapping the defect types identified by the UAV to the model's color status, and mapping meteorological data to scene lighting parameters; and ensuring that the spatial position matching accuracy error between the data and the model does not exceed 0.5 meters through a virtual-real mapping consistency verification mechanism.

5. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The visual defect detection model is a vision-based computationally intensive model that processes high-definition images captured by a drone inspection system. The input is image data, and the output is a structured recognition result containing defect type, location, and confidence level. It adopts an asynchronous call mode and achieves non-blocking processing through task ID and callback mechanism. The performance prediction model is a data-driven time series analysis model. It processes multi-dimensional numerical sequences provided by data acquisition and monitoring control systems and meteorological stations. The input is historical and real-time data, and the output is the performance degradation trend and loss prediction value of photovoltaic modules. It can provide early warning of typical faults ≥72 hours in advance and adopts a synchronous calling mode to meet the low-latency response requirements.

6. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The multi-level early warning rule engine includes: establishing a three-level early warning system of prompt, warning, and severity levels, with each level corresponding to different response strategies and notification methods; supporting composite logical judgments based on threshold conditions, trend conditions, and correlation conditions; the rule engine, as a background resident service, listens to data bus messages and uses the Rete algorithm to match incoming AI analysis results with real-time data; a de-jitter mechanism is introduced, requiring abnormal states of volatile data to persist for a set time before triggering an early warning, avoiding false alarms due to instantaneous noise; after an early warning is triggered, a structured early warning record is generated and pushed to the visual portal early warning center in real time, and relevant personnel are notified via SMS, email, and mobile terminals according to the level.

7. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The early warning and risk classification of faults includes: early detection of faults through a multi-level warning rule engine, based on the structured recognition results output by the visual defect detection model and the performance degradation trend prediction value generated by the performance prediction model, combined with historical operating data; the warning rules set differentiated thresholds according to equipment type, triggering a prompt-level warning when the visible defect area of ​​photovoltaic module surface defects is ≥5%, triggering a warning-level warning when string-level performance loss exceeds 5% for three consecutive prediction values, and triggering a severe-level warning when critical inverter equipment experiences abnormal temperature and a sudden power drop of ≥20%; the risk classification is based on a comprehensive assessment of the warning level, the criticality of the equipment, and the scope of impact, with the prompt level corresponding to locally observable anomalies, the warning level corresponding to potential faults that may affect power generation efficiency, and the severe level corresponding to emergency situations that lead to equipment shutdown and safety accidents.

8. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The operation and maintenance knowledge base is constructed by integrating static asset data, dynamic operation data, early warning work order records and external knowledge documents to build a structured knowledge base. It adopts a hybrid database architecture to store device attributes, performance history, fault paths and unstructured data respectively. It supports intelligent retrieval and weight optimization recommendation based on device type and early warning keywords, and implements version control for knowledge updates.

9. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The maintenance work order includes: when an early warning instance is created and its level is warning, the system automatically starts the work order generation process. The new work order inherits the associated early warning ID, target device information, and problem description from the early warning instance to form an initial data skeleton; using the early warning information as the query condition, the system calls the diagnostic interface of the maintenance knowledge base to obtain a structured processing solution containing processing suggestions, security specifications, risk levels, required tools and spare parts; the work order system automatically fills in the work order title, detailed work instructions, safety precautions, and resource list based on the solution returned by the knowledge base; after the work order content is completed, it enters the dispatch stage and pushes a notification to the designated maintenance personnel.

10. The method for early warning and platform construction of photovoltaic power plants based on digital twins as described in claim 1, characterized in that, The visualization and interactive portal includes: using the Vue.js framework as the foundation for front-end development, combined with the Three.js library to achieve real-time rendering and interaction of the 3D scene on the web; integrating the ECharts chart library to build a 2D data dashboard, dynamically displaying the power plant's total real-time power, daily cumulative power generation, defect statistics, early warning information, and key performance indicators; the portal interface integrates four major functional modules: a global situation overview dashboard, a 3D digital twin main scene, an early warning center, and a work order management center; supporting multi-terminal adaptive access, and communicating with the backend microservices in real time via the WebSocket protocol to achieve data-driven visualization updates.

Images