An electric power engineering project management and control method and system based on an internet of things

Abstract

The application belongs to the technical field of power internet of things, and particularly relates to a power engineering project management and control method and system based on internet of things, wherein the method comprises the following steps: firstly, collecting and fusing internet of things monitoring data, building information model data and project management data on site to generate a unified fusion feature tensor; then, combining the tensor with a pre-constructed power engineering field knowledge graph, using a spatio-temporal graph neural network with embedded causal attention mechanism to identify project process abnormalities, and using uncertainty reasoning to trace the root cause to generate a root cause analysis report; finally, inputting the report into a reinforcement learning model, which combines transfer learning and a multi-objective optimization mechanism based on cooperative game to automatically generate and output an optimal process adjustment scheme to drive dynamic adjustment and closed-loop optimization of business processes. The application realizes intelligent perception, deep analysis and adaptive optimization of the whole project process, and significantly improves the predictability, accuracy and automation level of management and control.

Description

Technical Field

[0001] This invention relates to the field of power Internet of Things (IoT) technology, and in particular to a method and system for the management and control of power engineering projects based on IoT. Background Technology

[0002] The design, construction, and operation and maintenance management of power engineering projects involve multi-disciplinary collaboration, massive equipment monitoring, and strict safety regulations, placing extremely high demands on project schedule, quality, and cost control. Traditional power engineering project management systems are mostly based on Business Process Management (BPM) platforms, digitizing offline processes such as project initiation, design, review, and publication, and integrating office management functions (as described in the handover document). While these systems improve process efficiency, they still have significant shortcomings when facing the unique complexity of power engineering: 1) The systems are typically process-driven, lacking deep integration and intelligent analysis capabilities for the physical objects flowing through the process (such as specific design drawings, equipment parameters, and field sensor data), resulting in a disconnect between process management and the physical state; 2) The identification and response to project risks are delayed, often relying on post-event responses, lacking predictive early warning capabilities based on multi-source data; 3) Process adjustments and optimization heavily depend on human experience, making it difficult to achieve adaptive and optimized dynamic scheduling when facing complex and ever-changing field conditions.

[0003] Existing technologies have already attempted to apply the Internet of Things (IoT), big data, and artificial intelligence to process management. For example, Chinese patent application CN111385273B discloses an IoT business process identification method, which identifies business processes by parsing the payload field of CoAP data packets and using a machine learning model. The shortcomings of this technology are: 1) Its technical scenario focuses on general, standardized communication between IoT devices and platforms, failing to adapt to the diverse device protocols (such as IEC61850, Modbus, etc.) and highly heterogeneous data (including telemetry, remote signaling, BIM models, and documents) characteristics of power engineering scenarios; 2) Its identification purpose is to classify existing business processes, which is a post-event analysis and cannot predict or proactively intervene in risks such as equipment failures and schedule delays that may occur in power engineering projects; 3) The method is independent of specific business knowledge (such as power design specifications and equipment operation and maintenance procedures), and cannot make in-depth decisions that conform to industry standards.

[0004] Chinese patent application CN111860960A proposes a smart supervision platform integrating big data and blockchain, emphasizing multi-source data collection and event-triggered suggestion push. The shortcomings of this technology are: 1) Its system architecture is complex, with high coupling between modules, and the "perception-analysis-push" mechanism relies on a large number of predefined rules and event types. When facing numerous non-standard and sudden working conditions in power engineering projects (such as construction risks caused by extreme weather and quality hazards in concealed works), its generalization and real-time response capabilities are insufficient; 2) Its intelligent core relies on "deep self-learning" based on expert evaluation feedback, resulting in a long learning loop and difficulty in quickly responding to dynamically changing project sites; 3) It does not fully consider the security and reliable evidence storage requirements of power engineering data. Although blockchain is mentioned, it is not closely integrated with specific engineering data (such as test reports and acceptance certificates) and operational procedures.

[0005] Chinese patent application CN118394018B discloses an IoT process control method based on a directed acyclic graph (DAG), which can automatically orchestrate processes according to business needs and adjust them when nodes are abnormal. The shortcomings of this technology are: 1) Its process orchestration and anomaly handling logic are mainly based on equipment availability and general node relationships, without incorporating professional knowledge from the power engineering field (such as the process requirements and safety operation specifications in the "Power Construction Quality Acceptance Regulations"), which may lead to technically infeasible processes or safety risks; 2) The anomaly handling strategy is relatively mechanical (replacing or removing nodes), lacking multi-dimensional diagnostic capabilities for the root causes of anomalies (such as equipment performance degradation, environmental factors, and human error), and unable to provide radical optimization suggestions; 3) System initialization and optimization heavily rely on historical data, lacking effective startup and transfer learning mechanisms for new power engineering projects (such as offshore wind power and distributed photovoltaic clusters).

[0006] In summary, existing technologies are insufficient to meet the deep-seated needs of intelligent power engineering project management. The core issues are: "fragmented perception"—IoT monitoring data, BIM design data, and management process data are isolated from each other; "superficial decision-making"—the lack of in-depth data analysis integrating domain knowledge makes it impossible to achieve predictive early warning and root cause optimization; and "rigid system"—it is difficult to adapt to the complex and ever-changing environment of power engineering sites, and cannot achieve adaptive intelligent scheduling and closed-loop optimization. Summary of the Invention

[0007] This invention proposes a power engineering project management and control method and system based on the Internet of Things, aiming to solve the technical problems of the shortcomings of existing power engineering project management and control systems in terms of data fusion, intelligent analysis and adaptive optimization.

[0008] In a first aspect, the present invention provides a power engineering project management and control method based on the Internet of Things, comprising: Collect and integrate IoT monitoring data, building information modeling data, and project management data from power engineering sites to generate a unified fusion feature tensor; Based on the fusion feature tensor and the pre-constructed knowledge graph of the power engineering field, project process anomalies are identified and root cause reasoning is performed to generate a root cause analysis report. The root cause analysis report is input into the reinforcement learning model to generate and validate the process adjustment strategy, and output the optimal process adjustment scheme to drive dynamic process adjustment.

[0009] The technical effect of the IoT-based power engineering project management and control method disclosed in this invention is as follows: This method integrates multi-source heterogeneous data, combines domain knowledge graphs for intelligent anomaly detection and root cause analysis, and utilizes reinforcement learning to achieve adaptive dynamic optimization of processes, forming an end-to-end intelligent management and control closed loop from data perception to intelligent decision-making and execution, which significantly improves the predictability, accuracy and adaptability of power engineering project management.

[0010] Furthermore, the reinforcement learning model is constructed and trained in the following manner: Define a state space, which is formed by concatenating the directed acyclic graph representation of the current process node with the fused feature tensor; Define an action space, which includes atomic operations for adding, deleting, replacing, merging, and adjusting resource allocation of process nodes; Define a reward function, which is a weighted sum of the reduction in time consumption of the critical path of the process, the degree of resolution of resource conflicts, the amount of cost savings, and the knowledge graph compliance verification score, and impose negative reward penalties on actions that violate the mandatory security specifications in the knowledge graph; Using historical power engineering project data, a near-end policy optimization algorithm is used to train the reinforcement learning model offline, and the policy network parameters are continuously adjusted by collecting actual adjustment effect data during the online operation phase.

[0011] Furthermore, when the reinforcement learning model is applied to new power engineering scenarios or in the early stages of project initiation, transfer learning optimization is performed: The initial strategy is to retrieve the historical project process pattern with the highest similarity to the current scenario from the knowledge graph of the power engineering field. By comparing online interactive data with historical prior strategies, the model's exploration weights for new data are dynamically adjusted, accelerating the model's convergence speed and strategy optimization effect in the target scenario.

[0012] Furthermore, the processing of the fused feature tensor includes: Establish a mapping relationship between IoT monitoring data and components in the building information model, and associate monitoring points with 3D model entities; Unify the timestamps of all data sources and align data of different frequencies to the same time granularity using an interpolation algorithm; Event information is extracted from engineering documents using natural language processing techniques and aligned with spatiotemporal benchmarks.

[0013] Furthermore, the generation of a unified fusion feature tensor aligns data of different frequencies to the same temporal granularity using an interpolation algorithm. Specifically, a fusion alignment algorithm based on dynamic time warping and adaptive quality assessment is employed, including: For any two time-series data sequences with different sampling frequencies to be aligned and Where M and N are the sequence lengths; Calculate the dynamic time warping distance between sequences X and Y, and obtain the optimal warping path P. Where K is the path length. Indicates the first in X The point and the first in Y Align the points; The time axis is divided into K alignment intervals along the optimal regularization path P. Within each alignment interval, the local information entropy of the data segments of sequences X and Y is calculated. and And confidence scores based on neighboring interval consistency. and ; Based on the local information entropy and confidence score, calculate the fusion weights of sequences X and Y within each alignment interval. and ,in: ; ; Within each alignment interval, the data of sequence X and Y are weighted and averaged using the calculated fusion weights to generate fusion data points with a uniform time granularity within that interval. By traversing all K aligned intervals, a complete fused data sequence with uniform time granularity is obtained.

[0014] Furthermore, the identification of abnormal project processes and the performance of root cause reasoning specifically include: Anomaly identification is performed using a spatiotemporal graph neural network model based on a multi-head attention mechanism; The input to the spatiotemporal graph neural network model is a dynamic attribute graph composed of process nodes, equipment and environmental elements, and its attributes are provided by the fused feature tensor. The spatiotemporal graph neural network model outputs anomaly detection results by aggregating neighborhood information and capturing spatiotemporal dependencies.

[0015] Furthermore, the spatiotemporal graph neural network model based on the multi-head attention mechanism includes an adaptive spatiotemporal graph construction layer and a causal attention pooling layer; The adaptive spatiotemporal graph construction layer calculates and updates the weights of edges in the graph structure in real time based on the dynamic similarity of node features in the fused feature tensor. When aggregating neighborhood information, the causal attention pooling layer introduces a time lag operator to construct the following causal relationship constraint function: ; Where Q, K, and V are the corresponding matrices for query, key, and value, respectively. The bias matrix characterizing time lag, Let be the dimension of the key vector, and softmax(·) be the normalization exponential function.

[0016] Furthermore, the reinforcement learning model incorporates a Nash equilibrium solution mechanism based on cooperative game theory during the verification process: The process of generating and verifying the process adjustment strategy is constructed into a cooperative game model, where each agent represents an optimization objective or constraint. Define the reward function for each agent as the degree of improvement of the corresponding objective or constraint; By solving for the Nash equilibrium point of this cooperative game, we obtain the set of optimal strategies that simultaneously satisfy multiple objectives and constraints, and select a Pareto optimal solution from this set as the optimal process adjustment scheme.

[0017] Furthermore, the knowledge graph in the field of power engineering adopts an uncertainty reasoning mechanism; Assign confidence weights to entity relationships in a knowledge graph; When performing root cause reasoning, a probabilistic graphical model is used to calculate the overall confidence level of each potential root cause path, and this is presented in the root cause analysis report.

[0018] Secondly, the present invention provides an Internet of Things-based power engineering project management and control system, the system being used to execute the method, the system comprising: The data acquisition and fusion module is configured to collect and fuse IoT monitoring data, building information model data and project management data from power engineering sites to generate a unified fusion feature tensor. The intelligent analysis and decision-making module is configured to identify project process anomalies and perform root cause reasoning based on the fused feature tensor and a pre-built knowledge graph in the field of power engineering, and generate a root cause analysis report. The execution and feedback control module is configured to input the root cause analysis report into the reinforcement learning model, generate and verify the process adjustment strategy, and output the optimal process adjustment scheme to drive dynamic process adjustment.

[0019] The technical advantages of the system disclosed in this invention are as follows: The system constructs a closed-loop control architecture with three cascaded layers of data fusion, intelligent analysis, and optimized execution. Through modular design and feedback loops, it organically integrates multi-source data perception, domain knowledge reasoning, and reinforcement learning decision-making, thereby realizing real-time, adaptive, and continuously optimized intelligent control of the entire process of power engineering projects. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a power engineering project management and control method based on the Internet of Things proposed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the core intelligent analysis and decision-making process provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of adaptive process optimization and multi-objective game decision-making provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Traditional power engineering project management systems, as mentioned in the background, are mostly based on Business Process Management (BPM) platforms, typically driven by processes. They lack deep integration and intelligent analysis capabilities for the entities flowing through the processes (such as specific design drawings, equipment parameters, and field sensor data), resulting in a disconnect between process management and the actual status of these entities. Therefore, this invention aims to overcome the shortcomings of existing power engineering project management systems in data fusion, intelligent analysis, and adaptive optimization, providing an Internet of Things (IoT)-based power engineering project management method. Figures 1 to 3 As shown, the specific steps include: S100, Multi-Source Data Fusion Step: Collect and fuse IoT monitoring data, Building Information Modeling (BIM) data, and project management data from the power engineering site to generate a unified fusion feature tensor. The goal of this step is to generate a high-quality, unified fusion feature tensor to provide reliable input for subsequent intelligent analysis. Specifically, it includes the following sub-steps: S110, Data Acquisition and Preliminary Processing: The system continuously captures equipment status monitoring data (such as transformer oil temperature and circuit breaker open / close status) and environmental sensor data (such as temperature and humidity) from the IoT gateway. Simultaneously, it parses the current version of the 3D model from the BIM server, extracting component attributes and relationships. It also retrieves process instance data and document metadata from the project management database according to time windows. For unstructured documents (such as design change notices and site inspection reports), OCR and NLP technologies are used for preprocessing.

[0023] S120, Spatiotemporal Alignment and Feature Extraction, includes spatial alignment, temporal alignment, and feature extraction.

[0024] Spatial alignment: For each IoT monitoring point (identified by device code or geographic coordinates), the corresponding component (such as "#1 main transformer" or "10kV incoming line cabinet") is located in the BIM model, and a "monitoring point - BIM component" mapping table is established. This gives the sensor data a clear engineering semantic.

[0025] Time alignment is fundamental: all data is timestamped based on a unified clock source. Because data sampling frequencies vary (sensor data may be at the second level, BIM updates at the day level, and process status changes at the minute level), time granularity alignment is necessary.

[0026] Feature extraction: Extract statistical features (mean, variance, slope) from numerical sensor data; convert component type and parameters into embedded vectors for BIM data; and convert node type, handler, and time consumption into features for process data.

[0027] S130, Fusion Alignment Based on Dynamic Time Warping and Adaptive Quality Assessment. This is crucial for aligning data of different frequencies to the same analytical granularity. Traditional linear interpolation leads to severe distortion in power engineering scenarios with sudden events. Sensor data from power engineering sites (such as temperature and vibration) may experience sudden and drastic fluctuations (such as equipment startup or instantaneous failure), while BIM data and management logs are discrete events or low-frequency updates. Traditional uniform interpolation forcibly generates data at non-event points, severely smoothing or distorting the true physical event characteristics, causing subsequent traditional AI models to fail to identify key transient anomalies. Therefore, this invention proposes a further solution: in the multi-source data fusion step, an interpolation algorithm aligns data of different frequencies to the same time granularity, specifically employing a fusion alignment algorithm based on dynamic time warping and adaptive quality assessment, including: (1) Dynamic Time Warping Path Finding: For two time series data sequences that need to be aligned, such as the high-frequency "cable joint temperature" sequence X and the low-frequency "regional load dispatching instruction" event sequence Y, calculate the DTW distance between them and find the optimal bending path P. This path allows a temperature peak in X to be flexibly matched with a dispatching instruction in Y in time, rather than being forced to be aligned point-to-point, effectively preserving the temporal form of physical events.

[0028] (2) Local quality assessment: Calculate the local information entropy of the two data segments along each alignment interval K of path P. and (Measures the information content / uncertainty of the segment) and confidence score based on consistency between preceding and following intervals. and (This measures the reliability of the data segment). For example, a segment of sensor data with a stable signal and a consistent trend has a high confidence level (C).

[0029] (3) Adaptive weighted fusion: The fusion weight is dynamically calculated based on the quality assessment results. and ,in: ; ; Specifically, when sensor data has high reliability (low entropy, high confidence), the weights are tilted towards the sensor; when the sensor signal is disturbed (high entropy, low confidence), the weights automatically shift to more stable BIM or management data. This achieves noise-resistant, context-aware intelligent fusion.

[0030] (4) Within each alignment interval, the calculated fusion weights are used to perform a weighted average of the data in sequences X and Y, ultimately resulting in a sequence with unified timestamps, where each data point incorporates multi-source information and contains quality weights. This process is repeated for all data pairs requiring alignment, and all features are integrated to form the final fusion feature tensor. This tensor is the foundation for all subsequent intelligent analyses.

[0031] This solution addresses the "information distortion" problem inherent in traditional interpolation methods (such as linear and spline interpolation). Dynamic Time Warping (DTW) does not enforce uniform point-to-point alignment but instead seeks the "optimal nonlinear bending path" between two sequences on the time axis. This allows it to reasonably correlate a transient peak (which may be very short-lived) from an IoT sensor with a longer "equipment operating status" event in the BIM model, perfectly preserving the physical event form of the original data.

[0032] Weight and The weighting changes dynamically with the time interval K. In intervals where sensor signals are stable, there may be greater reliance on IoT data; in intervals with network interruptions or high data noise, the weighting of more stable data sources such as BIM or management data is automatically increased. This achieves adaptive fusion of "context-aware" data.

[0033] S200, Anomaly Detection and Root Cause Analysis Step: This step proactively identifies deviations in the project process using the fused, high-quality data. It specifically includes the following sub-steps: S210, Construct a dynamic attribute graph: Nodes in the graph are project process nodes (e.g., "preliminary design," "structural calculation," "drawing verification"), physical equipment (e.g., "GIS room circuit breaker," "main control panel"), and environmental elements (e.g., "work area"). Node attributes are derived from the corresponding feature vectors in the fused feature tensor. Edges between nodes represent data flows, physical connections, logical dependencies, or process relationships, forming a dynamic attribute graph.

[0034] S220, anomaly recognition based on a spatiotemporal graph neural network model, using the aforementioned dynamic attribute graph as input, includes: Adaptive Spatiotemporal Graph Construction Layer: This layer does not rely on a fixed adjacency matrix, but instead calculates the feature similarity between nodes in real time based on the fused feature tensor, dynamically updating the edge weights. For example, when the features of the "transformer temperature" node and the "cooling system status" node show a strong correlation, the connection weight between them will increase.

[0035] Causal attention pooling layer: This is the core of the model. A time lag operator is introduced when aggregating neighbor node information. This refers to the bias matrix representing the time lag. Its attention mechanism is as follows: ; Where Q, K, and V are the corresponding matrices for query, key, and value, respectively. The bias matrix characterizing time lag, Let be the dimension of the key vector, and softmax(·) be the normalized exponential function. The bias matrix forces the model to consider the influence of earlier nodes (causes) on subsequent nodes (effects) more when calculating attention scores. This allows the model to learn to identify temporal causal relationships such as "yesterday's cable laying delay (cause)" leading to "today's electrical installation being unable to begin (effect)," rather than just spatial correlations, thus more accurately locating the source of anomalies.

[0036] Abnormal output: The model outputs an abnormal score for each node or the entire graph, identifying process nodes, devices, or links that deviate from the normal pattern (learned from historical normal data).

[0037] S230, Uncertainty Root Cause Reasoning Based on Knowledge Graphs, specifically includes the following four parts: Knowledge Graph Construction: The graph includes "entities" (e.g., 110kV transformer, IEC61557 standard, Zhang Gong, lightning overvoltage), "relationships" (e.g., belong to, conform to, responsible for, may cause), and "attributes". Relationships such as "may cause", "affected by", etc. are assigned confidence weights (e.g., 0.8).

[0038] Reasoning Process: When a "drawing verification process timeout" anomaly is detected, the system uses this anomaly node as a starting point to perform multi-hop queries and reasoning in the knowledge graph. For example, the path might be: drawing verification timeout <- related to high drawing complexity <- conforms to - special terrain tower design specifications -> may lead to (0.7) design rework. Meanwhile, another path might be related to the current excessive workload of the responsible engineer, Mr. Li.

[0039] Uncertainty synthesis: Using a probabilistic graphical model, the confidence weights and evidence support of all potential reasoning paths are combined to calculate the overall confidence of each root cause hypothesis (such as "complex design specifications" or "excessive engineer workload").

[0040] Report Generation: The root cause analysis report lists the ranked root cause hypotheses and their confidence levels, and may include links to relevant case studies from the knowledge graph. Its technical advantage lies in providing managers with in-depth, interpretable decision support with quantifiable uncertainty, rather than just alerts.

[0041] S300, Adaptive Process Optimization and Execution Steps: This step automatically generates and executes optimization solutions based on root cause analysis, specifically including the following sub-steps: S310, reinforcement learning model decision-making, including state-action and reward function design, as detailed below: State and Action: State s t It is the encoding of the current process DAG graph and the fused feature tensor. Action a t These include atomic operations such as splitting the structural calculation task into two parallel sub-tasks, adding a senior engineer to the drawing verification process, and bringing forward the equipment procurement node to run in parallel with the preliminary design.

[0042] Reward function design: Reward It is a weighted sum of multiple items: ; in, Reduce the amount of time spent on the critical path. For resource conflict resolution, To estimate cost savings, The score is used to validate the knowledge graph. To penalize violations of mandatory safety regulations, a reward function is designed to guide the model to find a solution that is both efficient and safe and compliant.

[0043] Nash Equilibrium Solving Based on Cooperative Game Theory: Multi-objective optimization is modeled as a cooperative game. Each objective (e.g., "shortest construction period," "lowest cost," "most balanced resources") or constraint (e.g., "safety procedure A must be satisfied") represents an agent. By solving for the Nash equilibrium, a set of strategies is obtained, in which no agent can benefit by unilaterally changing its strategy. This ensures that the final solution is not an extreme solution for a single objective, but rather a Pareto optimal solution that achieves a good balance among multiple objectives, thus better meeting practical management needs.

[0044] S320, employing transfer learning and cold start optimization, addresses the limited historical data available for new offshore wind power projects. The system retrieves successful process patterns from similar historical projects such as "onshore wind power" or "large substations" in the knowledge graph, using these patterns as the initial strategy for the reinforcement learning model. During online learning, the model compares current interaction data with historical priors, dynamically adjusting the exploration rate and quickly converging to an effective strategy adapted to the new scenario, thus solving the "cold start" problem in the early stages of project launches or new business areas.

[0045] S330, Solution Execution and Feedback: The optimal process adjustment solution is converted into instructions recognizable by the BPM engine (such as modifying process definitions, reassigning tasks, and inserting approval nodes), driving dynamic process adjustments. Simultaneously, the system promptly notifies relevant personnel (such as designers, reviewers, and project managers) of process changes and urges pending tasks via SMS notifications and mobile app pushes. All decision-making basis, execution actions, and results are recorded and fed back to the data acquisition layer for online fine-tuning of the reinforcement learning model, forming a closed loop of continuous improvement. This fully achieves the goals outlined in the disclosure document: "improving business process efficiency, enhancing the company's business management level, and increasing work efficiency."

[0046] Based on the same inventive concept, this invention also provides an Internet of Things-based power engineering project management and control system, which can be generally divided into three levels, forming a complete closed-loop management and control circuit: Data Acquisition and Fusion Module: This layer is the system's "sensory nerves," responsible for acquiring raw data from the physical and digital worlds. It includes: Internet of Things (IoT) sensor networks: Deployed at design sites, laboratories, and processing workshops, these networks collect physical world state data through various sensors (such as temperature, humidity, vibration, and current sensors) and smart devices. The data may be transmitted via multiple protocols such as CoAP, MQTT, and Modbus.

[0047] Building Information Modeling Server: Stores and manages the 3D BIM design model of a project, including structured and geometric information such as equipment parameters, spatial relationships, and pipeline routes.

[0048] Project Management Database: Stores process data from traditional BPM platforms, such as project initiation applications, planning tasks, design drawing review processes, finished product publication status, performance records, contract information, etc. It also includes automated office management process data such as work contacts, attendance, and material requisition.

[0049] Fusion processing module: This is the core processing unit. It receives the above-mentioned multi-source heterogeneous data, performs cleaning, alignment, feature extraction and fusion, and finally outputs the fused feature tensor.

[0050] Intelligent Analysis and Decision-Making Module: This layer is the "brain" of the system, responsible for cognition and decision-making. It includes: Knowledge Graph Management Unit: Stores and maintains knowledge graphs in the field of power engineering; Anomaly detection unit: Built-in spatiotemporal graph neural network model, responsible for identifying anomalies; Root Cause Reasoning Unit: Utilizes knowledge graphs and probability models for root cause analysis; This layer takes a fused feature tensor as input and outputs a root cause analysis report.

[0051] Adaptive Optimization and Execution Module: This layer is the "limb" of the system, responsible for decision execution and learning evolution. It includes: Reinforcement learning decision engine: Receives root cause analysis reports and generates process adjustment strategies; Process Engine: Drives the execution of business processes and can be integrated with or modified from existing BPM process engines; Feedback closed-loop unit: collects execution results for model updates.

[0052] This layer outputs the optimal process adjustment plan and executes the adjustments by driving the business system through the process engine and sending instructions or notifications (such as SMS or mobile app push notifications). Mobile terminals (such as phones and tablets) can install a dedicated app for real-time monitoring. Figure 1 The system allows users to monitor the status of all processes, receive SMS messages or in-app notifications to expedite tasks, and view the 3D model and data fusion dashboard, thus realizing the "real-time monitoring of mobile terminal information and SMS notification function" emphasized in the handover document.

[0053] The entire system is built on a microservice architecture, with modules communicating with each other through an API gateway and message queue, ensuring high cohesion, low coupling, and scalability.

[0054] In a specific implementation scenario, this case study will use a 220kV smart substation expansion project undertaken by a power design company as an example to illustrate in detail how the method of this invention operates in a real project, solving the entire process from problem identification to closed-loop optimization. The company needs to expand a 220kV smart substation in an industrial park, adding a 180MVA main transformer and supporting equipment. The project adopts a BPM-based management system, and the process includes: feasibility study -> project initiation -> preliminary design -> construction drawing design -> material procurement -> civil construction -> electrical installation -> commissioning. Simultaneously, IoT sensors are deployed on-site, and the design utilizes BIM 3D collaborative design.

[0055] The first stage involves multi-source data fusion and anomaly detection, including data acquisition, fusion alignment, and feature generation.

[0056] (1) Data acquisition specifically includes: settlement observation point (sensor S1) installed on the main transformer foundation, ambient temperature and humidity sensor (S2), and associated cooling system status monitoring data, which are uploaded in CoAP / Modbus protocol.

[0057] BIM data includes a 3D model of the main transformer, foundation dimensions, installation space, and clearance requirements with adjacent equipment (such as GIS and surge arresters).

[0058] Process data: The currently active process is "Electrical Installation - Main Transformer Placement" (process node P1). Its predecessor node "Civil Engineering - Main Transformer Foundation Maintenance" (node ​​P0) is shown as completed, and the subsequent node "Main Transformer Accessory Installation" (node ​​P2) is pending.

[0059] Document data: The construction log shows "The foundation curing period has ended and the strength test report is qualified"; the supervision liaison form mentions "the scheduling of large cranes on site is tight".

[0060] Resource data: The project resource pool shows that "Engineer Li," a certified special hoisting operator, is currently being used by another adjacent project.

[0061] (2) Fusion alignment and feature generation.

[0062] Spatiotemporal alignment: The system binds the settlement sensor S1 to the unique ID of the "main transformer foundation" component in the BIM model. All data timestamps are unified.

[0063] Dynamic fusion: Scenario: Settlement data (high frequency, every 10 minutes) shows recent minor but continuous settlement changes. The supervisor's contact form (low frequency event) records "crane resources are scarce." The construction log (text) uses NLP to extract the "curing completed" event.

[0064] Processing: The fusion processing module invokes an algorithm based on DTW and adaptive quality assessment. It discovers a non-linear temporal correlation (DTW path) between the small change sequence (X) of settlement data and the event state (Y) of "crane resource shortage". The algorithm calculates that the local information entropy of the settlement data fragments is low (stable change), but the confidence level is slightly lower due to slight differences from historical settlement patterns; while the "resource shortage" event comes from the official supervision documents and has extremely high confidence.

[0065] Result: A fused feature tensor was generated. One key feature vector represents that "within time window T, there is a continuous small settlement in the main transformer foundation area, which is highly spatiotemporally correlated with the event of limited hoisting resources, with a comprehensive confidence level of 0.75".

[0066] The second phase involves anomaly detection and root cause analysis, including the construction and detection of anomalies and knowledge-driven root cause reasoning.

[0067] (1) Dynamic graph construction and anomaly detection.

[0068] The system constructs an attribute graph containing nodes {P0-basic maintenance, P1-main transformer in place, P2-accessory installation, equipment-main transformer, sensor-S1, resources-large crane, personnel-Engineer Li}, with node attributes derived from the fused feature tensor.

[0069] The causal attention pooling layer of the Spatiotemporal Graph Neural Network (STGNN) begins to work. It calculates the attention scores for the P1-main variable in-situ node and the sensor-S1 (settlement) and resource-large crane nodes.

[0070] Because of the introduction of a time lag bias τ, the model pays more attention to the impact of the "cause" (P0 - settlement data after basic maintenance is completed, and historical resource - large crane scheduling records) on the "effect" (P1 - main transformer in place and currently stagnant).

[0071] Model output: The anomaly score of the P1-main transformer in-place node is significantly higher than the threshold and is marked as a "high-risk delay" anomaly. At the same time, the sensor-S1 and resource-large crane nodes are also marked as associated anomaly sources.

[0072] (2) Knowledge-driven root cause reasoning.

[0073] Knowledge graph query: The system starts with "main transformer installation delay" and traverses the knowledge graph in the field of power engineering.

[0074] Path A: Installation delay <- possibly caused by... - unstable foundation <- manifested as - continuous settlement. Path confidence: 0.70 (from the specification: the foundation must be stable before installation).

[0075] Path B: Installation delay <- Possibly caused by... - Missing critical resources <- Example: - Large crane <- Requires - Specialized personnel. Path confidence: 0.85 (from project management knowledge).

[0076] Path C: Unstable foundation <- Need to be checked - Curing plan <- Basis - Concrete curing specifications. Path confidence level: 0.90.

[0077] Uncertainty Synthesis and Report Generation: A probabilistic graphical model integrates multiple paths and calculates the following: Root cause 1 (confidence level 0.81): Conflict in the coordination of hoisting resources (special cranes and personnel) is the most significant bottleneck at present.

[0078] Root cause 2 (confidence level 0.65): There is continuous minor settlement in the foundation, and it is necessary to verify whether the installation conditions are met.

[0079] Additional information: The graph is linked to a historical case, "A 110kV project suffered foundation rework after installation due to neglect of early settlement," and relevant regulatory clauses are pushed out.

[0080] Output: Generate a root cause analysis report, push it to the project manager and supervising engineer via mobile app, highlight the two root causes and their confidence levels, and suggest "prioritize resource coordination and synchronously verify basic data".

[0081] The third stage: adaptive process optimization and closed-loop execution, including intelligent optimization decision-making, transfer learning assistance, solution execution and closed-loop feedback.

[0082] (1) Intelligent optimization decision-making.

[0083] Reinforcement learning model initiation: The P1-major variable in-situ anomalous state, the fused feature tensor, and the root cause report are used as input state s. t .

[0084] Strategy Generation and Game Optimization: The model generates multiple candidate actions (strategies): A1: Forcefully recall "Engineer Li" and urgently allocate a large crane from other projects of the company (high cost, which may affect other projects).

[0085] A2: Disassemble and position the P1-main transformer. First, use existing small equipment to complete part of the positioning work. Then, complete the remaining work when the crane is available 3 days later (extending the construction period).

[0086] A3: Pause P1, insert a parallel subprocess "P1a - Foundation Settlement Verification and Reinforcement", and coordinate resources. Continue after verification is completed and resources are available.

[0087] Cooperative game theory solution: The system models "shortest construction period," "lowest cost," "minimum resource impact," and "compliance with safety standards (foundation must be stable before installation)" as multiple agents engaging in a game. The solution is obtained through Nash equilibrium. A1 violates the objectives of "minimizing resource impact" and "lowest cost".

[0088] A2 violated the principles of "shortest construction period" and "safety regulations" (continuing installation before the foundation was clearly stable).

[0089] A3 achieves a balance across multiple metrics: although it slightly increases the construction period, it completely eliminates fundamental risks and buys time for resource coordination, making it the Pareto optimal solution.

[0090] (2) Transfer learning assistance. The system found that the current "resource conflict" scenario had an 85% similarity to a historical "GIS installation resource conflict" case in the knowledge graph. The model used the successful resolution pattern of this historical case ("inserting technical review process and coordinating resources in parallel") as the initial strategy preference, which accelerated the evaluation and selection process of the A3 strategy.

[0091] (3) Implementation of the plan and closed-loop feedback.

[0092] (a) Dynamic process adjustment: The system automatically sends instructions to the BPM engine: 1) Insert a new node P1a - foundation settlement special review before P1 - main transformer is in place; 2) Set P1a and the task of "coordinating cranes and personnel" to run in parallel; 3) Assign tasks to structural engineers for P1a and push relevant specifications and historical cases.

[0093] (b) Real-time notification and monitoring: SMS / APP push notification: The structural engineer received the task: "Please conduct a special review of the main transformer foundation settlement data. The data has been synchronized to the BIM model." The equipment coordinator received the task: "Please coordinate a 180T crane and Engineer Li. The latest required time has been updated to 3 days later." Mobile terminal monitoring: The project manager can see on the mobile APP that the flowchart has been automatically updated, the P1a node is highlighted, and can click to view the BIM 3D view that integrates real-time settlement data.

[0094] (c) Closed-loop learning: Three days later, the review conclusion, "settlement is within the allowable range, installation is possible," was entered into the system, and the crane coordination was completed.

[0095] Node P1 executed successfully. The entire event, from anomaly detection to recovery, including all data (status, actions, and impact on final project schedule), was used as a new experience sample. This data was fed back to the reinforcement learning model for online fine-tuning and then structured and incorporated into a knowledge graph for future decision-making in similar scenarios.

[0096] Through the above complete scenario demonstration, the method of this invention demonstrates its significant advantages over traditional BPM systems: From "passive response" to "proactive early warning": The system, through multi-source fusion (IoT + BIM + process) and STGNN model, provides early warning of the complex risk of "installation delay" before the crane is completely absent and before settlement causes substantial damage, rather than recording the problem afterward.

[0097] From "surface treatment" to "root cause management": Knowledge graphs and uncertainty reasoning guide analysis to directly address the dual root causes of "resource conflict" and "foundation settlement concerns," avoiding one-sided decisions that simply urge cranes to operate while ignoring potential quality risks.

[0098] From "manual scheduling" to "intelligent optimization": Solution A3, automatically generated by reinforcement learning and cooperative game theory models, achieves a multi-objective balance of safety, schedule, cost, and resources, and provides the optimal solution of "interpolation review and parallel coordination" that human managers may overlook.

[0099] End-to-end closed-loop and continuous evolution: The entire process flows, executes, records, and learns automatically within the system, forming a positive cycle of management capabilities. Deep integration with mobile terminals enables seamless and efficient collaboration between on-site operations and management.

[0100] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A power engineering project management and control method based on the Internet of Things, characterized in that, include: Collect and integrate IoT monitoring data, building information modeling data, and project management data from power engineering sites to generate a unified fusion feature tensor; Based on the fusion feature tensor and the pre-constructed knowledge graph of the power engineering field, project process anomalies are identified and root cause reasoning is performed to generate a root cause analysis report. The root cause analysis report is input into the reinforcement learning model to generate and validate the process adjustment strategy, and output the optimal process adjustment scheme to drive dynamic process adjustment.

2. The method according to claim 1, characterized in that, The reinforcement learning model is constructed and trained in the following manner: Define a state space, which is formed by concatenating the directed acyclic graph representation of the current process node with the fused feature tensor; Define an action space, which includes atomic operations for adding, deleting, replacing, merging, and adjusting resource allocation of process nodes; Define a reward function, which is a weighted sum of the reduction in time consumption of the critical path of the process, the degree of resolution of resource conflicts, the amount of cost savings, and the knowledge graph compliance verification score, and impose negative reward penalties on actions that violate the mandatory security specifications in the knowledge graph; Using historical power engineering project data, a near-end policy optimization algorithm is used to train the reinforcement learning model offline, and the policy network parameters are continuously adjusted by collecting actual adjustment effect data during the online operation phase.

3. The method according to claim 2, characterized in that, When the reinforcement learning model is applied to a new power engineering scenario or at the initial stage of project initiation, transfer learning optimization is performed: The initial strategy is to retrieve the historical project process pattern with the highest similarity to the current scenario from the knowledge graph of the power engineering field. By comparing online interactive data with historical prior strategies, the model's exploration weights for new data are dynamically adjusted, accelerating the model's convergence speed and strategy optimization effect in the target scenario.

4. The method according to claim 1, characterized in that, The processing of the fused feature tensor includes: Establish a mapping relationship between IoT monitoring data and components in the building information model, and associate monitoring points with 3D model entities; Unify the timestamps of all data sources and align data of different frequencies to the same time granularity using an interpolation algorithm; Event information is extracted from engineering documents using natural language processing techniques and aligned with spatiotemporal benchmarks.

5. The method according to claim 4, characterized in that, The generation of a unified fusion feature tensor aligns data of different frequencies to the same temporal granularity using an interpolation algorithm. Specifically, a fusion alignment algorithm based on dynamic time warping and adaptive quality assessment is employed, including: For any two time-series data sequences with different sampling frequencies to be aligned and Where M and N are the sequence lengths; Calculate the dynamic time warping distance between sequences X and Y, and obtain the optimal warping path P. Where K is the path length. Indicates the first in X The point and the first in Y Align the points; The time axis is divided into K alignment intervals along the optimal regularization path P. Within each alignment interval, the local information entropy of the data segments of sequences X and Y is calculated. and And confidence scores based on neighboring interval consistency. and ; Based on the local information entropy and confidence score, calculate the fusion weights of sequences X and Y within each alignment interval. and ,in: ； ； Within each alignment interval, the data of sequence X and Y are weighted and averaged using the calculated fusion weights to generate fusion data points with a uniform time granularity within that interval. By traversing all K aligned intervals, a complete fused data sequence with uniform time granularity is obtained.

6. The method according to claim 1, characterized in that, The identification of abnormal project processes and the root cause reasoning specifically include: Anomaly identification is performed using a spatiotemporal graph neural network model based on a multi-head attention mechanism; The input to the spatiotemporal graph neural network model is a dynamic attribute graph composed of process nodes, equipment and environmental elements, and its attributes are provided by the fused feature tensor. The spatiotemporal graph neural network model outputs anomaly detection results by aggregating neighborhood information and capturing spatiotemporal dependencies.

7. The method according to claim 6, characterized in that, The spatiotemporal graph neural network model based on multi-head attention mechanism includes an adaptive spatiotemporal graph construction layer and a causal attention pooling layer; The adaptive spatiotemporal graph construction layer calculates and updates the weights of edges in the graph structure in real time based on the dynamic similarity of node features in the fused feature tensor. When aggregating neighborhood information, the causal attention pooling layer introduces a time lag operator to construct the following causal relationship constraint function: ； Where Q, K, and V are the corresponding matrices for query, key, and value, respectively. The bias matrix characterizing time lag, Let be the dimension of the key vector, and softmax(·) be the normalization exponential function.

8. The method according to claim 1, characterized in that, The reinforcement learning model incorporates a Nash equilibrium solution mechanism based on cooperative game theory during the verification process: The process of generating and verifying the process adjustment strategy is constructed into a cooperative game model, where each agent represents an optimization objective or constraint. Define the reward function for each agent as the degree of improvement of the corresponding objective or constraint; By solving for the Nash equilibrium point of this cooperative game, we obtain the set of optimal strategies that simultaneously satisfy multiple objectives and constraints, and select a Pareto optimal solution from this set as the optimal process adjustment scheme.

9. The method according to claim 1, characterized in that, The knowledge graph in the field of power engineering adopts an uncertainty reasoning mechanism; Assign confidence weights to entity relationships in a knowledge graph; When performing root cause reasoning, a probabilistic graphical model is used to calculate the overall confidence level of each potential root cause path, and this is presented in the root cause analysis report.

10. A power engineering project management and control system based on the Internet of Things, characterized in that, The system is used to perform the method according to any one of claims 1-9, the system comprising: The data acquisition and fusion module is configured to collect and fuse IoT monitoring data, building information model data and project management data from power engineering sites to generate a unified fusion feature tensor. The intelligent analysis and decision-making module is configured to identify project process anomalies and perform root cause reasoning based on the fused feature tensor and a pre-built knowledge graph in the field of power engineering, and generate a root cause analysis report. The execution and feedback control module is configured to input the root cause analysis report into the reinforcement learning model, generate and verify the process adjustment strategy, and output the optimal process adjustment scheme to drive dynamic process adjustment.

Images