Distributed photovoltaic grid-connected point electric energy quality dynamic monitoring system

Abstract

A distributed photovoltaic grid-connected point electric energy quality dynamic monitoring system relates to the field of distributed photovoltaic grid-connected points, and comprises a knowledge graph construction module used for constructing a dynamic knowledge graph based on a dynamic database and updating the dynamic knowledge graph by using real-time data and integrated dynamic knowledge; the causal analysis module is used for establishing a causal relationship model of the power quality problem based on the dynamic knowledge graph; and the backtracking and positioning module is used for performing causal chain backtracking based on the causal relationship model, positioning a root cause influencing the power quality and generating an analysis report. According to the method, the causal chain backtracking is carried out along the weighted directed acyclic graph by taking the abnormal index as a starting point based on the causal relationship model, so that the root cause node, such as a specific equipment fault or health deterioration, which causes the power quality problem can be quickly and automatically positioned, and the automation degree and the positioning precision of fault diagnosis are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic grid connection points, specifically a dynamic power quality monitoring system for distributed photovoltaic grid connection points. Background Technology

[0002] With the ongoing global energy transition and the deepening implementation of the "dual carbon" target, distributed photovoltaic (PV) power generation has been widely adopted and connected to the grid due to its advantages such as cleanliness, flexibility, and local consumption. However, the intermittent and random output characteristics of a large number of PV power sources, as well as their grid connection through power electronic devices, have a significant impact on the power quality of the distribution network. This is mainly manifested in increasingly prominent problems such as voltage fluctuations and deviations, frequency offsets, harmonic pollution, and three-phase imbalance. These power quality degradations not only affect the normal operation of sensitive loads in the region but may also trigger chain reactions such as grid protection malfunctions and equipment overheating damage, thus restricting the grid connection and consumption of high-proportion distributed PV and the safe and stable operation of the grid. Therefore, real-time, accurate, and intelligent monitoring, analysis, and management of the power quality at distributed PV grid connection points have become a key technical requirement for ensuring reliable power supply and high-quality operation of the new power system.

[0003] Currently, power quality monitoring technology has evolved from traditional periodic inspections and handheld instrument measurements to routine monitoring based on online monitoring devices and data acquisition systems. Existing systems typically deploy intelligent terminals such as power quality analyzers at grid connection points to achieve real-time acquisition and over-limit alarms for key indicators such as voltage, current, harmonics, and flicker, and use data platforms for centralized monitoring and historical data analysis. Some advanced solutions have further introduced big data storage, cloud platform computing, and simple trend analysis functions, improving the scale and efficiency of data processing. However, traditional monitoring systems or existing solutions focus primarily on data acquisition, storage, and display. Their analytical capabilities often remain at the level of indicator statistics, threshold comparison, and post-event recapitulation, lacking deep integration and intelligent correlation of massive amounts of multi-source operational data (such as grid status, photovoltaic power generation parameters, and equipment health information). Furthermore, they struggle to locate the root causes of power quality problems in real-time and automatically from the complex and dynamic grid-photovoltaic interaction system.

[0004] The core problem facing current technological development lies in the widespread "data silos" phenomenon in existing monitoring systems, where the analysis process "knows what happens but not why." Specifically: First, the system struggles to effectively integrate and correlate multi-dimensional information such as grid-side events, real-time operating conditions of photovoltaic power generation units, and equipment health degradation trends, resulting in a singular and one-sided perspective in problem analysis. Second, it lacks a knowledge system capable of dynamically evolving and integrating physical topology and operational status, failing to reflect the system's internal relationships and state changes in real time, and exhibiting insufficient predictive ability for potential risks. Third, most systems lack the ability to automatically mine and quantify causal relationships based on data-driven approaches; when power quality events occur, maintenance personnel still rely on experience for tedious manual troubleshooting, which is inefficient and lacks accuracy. Finally, the closed-loop chain from diagnosis to control is broken, making it difficult to directly and reliably guide rapid regulation or optimization of operations and maintenance, and the entire process lacks tamper-proof and reliable evidence, hindering responsibility definition and experience accumulation.

[0005] Therefore, it is necessary to invent a dynamic power quality monitoring system for distributed photovoltaic grid connection points to solve the above problems. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a dynamic monitoring system for power quality at a distributed photovoltaic grid-connected point, comprising: The data acquisition module is used to collect operational data from the grid side and the photovoltaic side. The operational data includes power quality indicators, grid operation status, power generation system parameters, and equipment health status. The data integration module is used to integrate the collected operational data, extract power characteristics, and build a dynamic database; The knowledge graph construction module is used to construct a dynamic knowledge graph based on the dynamic database, and to update the dynamic knowledge graph using real-time data and integrated dynamic knowledge. The causal analysis module is used to establish a causal relationship model for power quality problems based on the dynamic knowledge graph. The backtracking and positioning module is used to trace the causal chain back based on the causal relationship model, locate the root cause affecting power quality, and generate an analysis report; The reporting and feedback module is used to execute power quality control or equipment operation and maintenance instructions based on the analysis report, and to store the analysis report and related operation logs on the blockchain.

[0007] Combining all the above technical solutions, the positive effects of this invention are as follows: 1. This invention integrates and collects multi-dimensional operational data from the power grid and photovoltaic sides, including power quality indicators, power grid operation status, power generation system parameters and equipment health status. Based on three types of characteristics—steady-state, transient, and dynamic fluctuation—it extracts power features and constructs a dynamic database, achieving deep fusion and standardized representation of multi-source heterogeneous data, effectively breaking the "data silo" problem in traditional monitoring systems. 2. This invention constructs a dynamic knowledge graph with a physical layer, a state layer, a rule layer, and a prediction layer, and integrates real-time data streams, causal rules, and digital twin simulations for continuous graph updates. This enables a unified and dynamic expression of the topology, real-time status, domain knowledge, and prediction information of photovoltaic grid-connected systems, significantly improving the cognitive depth of system status and the ability to predict potential risks. 3. This invention constructs a quantifiable weighted causal relationship model based on dynamic knowledge graphs, mutual information entropy, causal discovery algorithms, and structural causal models. It can automatically and accurately mine and quantify the causal dependencies between power quality indicators and equipment status and operating parameters, achieving a leap from "association analysis" to "causal inference". 4. This invention, based on the aforementioned causal relationship model, uses abnormal indicators as a starting point and traces back the causal chain along a weighted directed acyclic graph. This enables the rapid and automatic identification of the root cause nodes leading to power quality problems, such as specific equipment failures or health deterioration, significantly improving the automation level and location accuracy of fault diagnosis. 5. By combining causal analysis reports with closed-loop control execution, this invention can automatically trigger rapid reactive power compensation, adjust power plant control parameters, or generate long-term optimization schemes based on the analysis results, realizing an intelligent closed loop from "monitoring and diagnosis" to "control and optimization," thereby improving the timeliness and effectiveness of power quality management. Attached Figure Description

[0008] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0009] Figure 1 This is a system framework diagram of the present invention.

[0010] Figure 2 A flowchart illustrating the steps involved in establishing the causal relationship model for this invention.

[0011] Figure 3 This is a flowchart illustrating the causal chain backtracking process of the present invention. Detailed Implementation

[0012] 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.

[0013] See Figure 1 As shown, this invention proposes a dynamic power quality monitoring system for distributed photovoltaic grid-connected points, including a data acquisition module, a data integration module, a knowledge graph construction module, a causal analysis module, a backtracking and positioning module, and a reporting and feedback module.

[0014] The data acquisition module is used to collect operational data from the grid side and the photovoltaic side. The operational data includes power quality indicators, grid operation status, power generation system parameters, and equipment health status. Furthermore, in the above technical solution, the power quality indicators include voltage parameters, current parameters, frequency deviation, three-phase imbalance, and power factor. The power grid operating status includes fault recording data and abnormal event timestamps; The parameters of the power generation system include DC current and voltage, AC output power, conversion efficiency, temperature, and environmental parameters. The health status of the equipment includes the degree of component aging and its operating status.

[0015] It should be noted that the data acquisition module achieves comprehensive collection and preprocessing of operational data through the following specific implementation methods: Power quality index collection: Voltage and current parameters: Real-time synchronous acquisition is performed using high-precision voltage transformers (PTs) and current transformers (CTs) installed at the grid connection point and key branch circuits, with a sampling frequency of no less than 6.4 kHz to capture subharmonic and high-frequency harmonic components. The acquired raw waveform data is then converted from analog to digital (ADC) and calculated in real time by a local monitoring unit (such as a power quality analyzer) to obtain parameters such as the effective value of voltage, the effective value of current, the content of each harmonic, and the total harmonic distortion (THD).

[0016] Frequency deviation: The system frequency is calculated in real time from the fundamental voltage signal at the grid connection point using a phase-locked loop (PLL) circuit or a digital algorithm based on zero-crossing detection, and its deviation from the rated frequency (e.g., 50Hz) is calculated.

[0017] Three-phase unbalance: Based on the collected instantaneous values ​​of three-phase voltage and current, calculate the three-phase voltage unbalance (ratio of negative sequence voltage to positive sequence voltage) and the three-phase current unbalance.

[0018] Power factor: The instantaneous power factor is obtained by calculating the ratio of total active power to total apparent power, and the average power factor over the statistical period is calculated.

[0019] Data collection on power grid operating status: Fault waveform data: Configure a dedicated fault waveform recording device or use a protection and control device with waveform recording function. When an event such as voltage surge, drop, interruption or current over-limit is detected, it will automatically trigger the recording of full waveform data (including three-phase voltage and current) for several cycles before and after the event, and generate a waveform file containing information such as amplitude, phase and duration.

[0020] Abnormal event timestamps: All detected abnormal events (such as over-limit alarms and fault recording triggers) are stamped with high-precision millisecond-level timestamps provided by a unified time synchronization system (such as GPS / BeiDou) to ensure the time synchronization of multi-source data and the traceability of event sequences.

[0021] Data collection of power generation system parameters: DC side parameters: DC voltage sensors and DC current sensors are deployed at the output of the photovoltaic module combiner box and the DC input terminal of the inverter to collect the DC output voltage and current of the photovoltaic array for calculating DC power.

[0022] AC output parameters: Collect the voltage, current, and power (active, reactive, and apparent power) of the inverter's AC output side, and calculate the inverter's conversion efficiency (AC output power / DC input power).

[0023] Temperature parameters: The temperatures of key components such as the backsheet temperature of photovoltaic modules, the heat sink temperature of inverters, and the winding temperature of transformer substations are collected through infrared temperature sensors or patch temperature sensors.

[0024] Environmental parameters: Data such as light intensity, ambient temperature, humidity, wind speed, and wind direction at the installation point are collected through a weather station.

[0025] Equipment health status data collection and assessment: Module aging degree: By comparing the measured output power and nominal power of the same model of modules under the same environment, or by periodically analyzing the degradation of the module's IV characteristic curve, the power degradation rate is evaluated as an indicator of the aging degree. Alternatively, the module insulation impedance data recorded by the inverter can be collected to help determine the insulation aging status.

[0026] Operating status: Collect self-test status signals, alarm information, cumulative operating hours, and number of switching actions of major equipment such as inverters, transformers, and circuit breakers to form a snapshot of the equipment's operating status.

[0027] Data preprocessing and uploading: Each local data acquisition unit performs preliminary verification (such as range checks and abrupt change detection) and filtering on the raw data. Following a unified communication protocol (such as IEC 61850 or Modbus TCP) and a preset acquisition cycle (seconds / minutes for steady-state data, event-triggered for transient data), the data is packaged and uploaded to the station-level or cloud-based data integration module. Each uploaded data packet includes the device identifier, parameter type, value, quality code, and precise timestamp.

[0028] The data integration module is used to integrate the collected operational data, extract power characteristics, and build a dynamic database; Furthermore, in the above technical solution, the extraction of electrical features includes the following steps: Extracting steady-state characteristics: Based on the operating data, voltage deviation, frequency deviation, three-phase unbalance, and power factor are calculated as steady-state indicators; Extracting transient features: Based on the amplitude, duration, and occurrence time of voltage swell, voltage drop, or short-term interruption, instantaneous disturbance features are determined as transient indicators; Extracting dynamic fluctuation characteristics: Based on voltage fluctuations and flicker values, the periodic variation characteristics of voltage amplitude are determined as dynamic fluctuation indicators.

[0029] It should be noted that the specific implementation method of "extracting power characteristics" in the data integration module is as follows: The specific steps for extracting steady-state features are as follows: obtaining the grid connection point voltage measurement value (U) from the operating data. meas ) and current measurement value (I meas Voltage deviation (ΔU) is expressed by the formula ΔU=(U meas -U N ) / U N Calculate by multiplying by 100%, where U N This is the system's rated voltage. The frequency deviation (Δf) is calculated using the formula Δf = f meas -f N Calculate, where f meas f is the measured frequency. N The rated frequency is 50Hz. The three-phase unbalance is calculated using the symmetrical component method, and the voltage unbalance (ε) is... U ε is the percentage of the ratio of the negative-sequence voltage component (U2) to the positive-sequence voltage component (U1). U =(U2 / U1)×100%; Current imbalance (ε) I The calculation is similar. The power factor (λ) is calculated using the formula λ=P / S, where P is the total active power and S is the total apparent power. The calculated voltage deviation, frequency deviation, three-phase unbalance, and power factor are used as steady-state indicators and stored in a database in time series.

[0030] The specific steps for extracting transient features are as follows: real-time monitoring of the effective voltage value (U) rms ). When U rms Transient event identification is initiated when a sudden change occurs within a timeframe of 0.5 cycles to 1 minute. A voltage sag is defined as U. rms A voltage dip to below 90% but above 10% of the rated value for a duration between 0.5 cycles and 1 minute is defined as a voltage spurt. rms Rise to 110% or more of the rated value, for the same duration as above; define a short interruption as U. rms The voltage drops to below 10% of the rated value, and the duration does not exceed 1 minute. For the identified event, its event type is recorded, and three key features are extracted: first, the disturbance amplitude, i.e., the maximum percentage deviation of the effective voltage value from the rated value; second, the duration, i.e., the total time from the start of exceeding the limit to recovery to the threshold range; and third, the precise timestamp of the occurrence. These features together constitute a transient index describing the instantaneous disturbance.

[0031] The specific steps for extracting dynamic fluctuation characteristics are as follows: Statistical analysis of the voltage time series is performed to calculate the difference (ΔU) between the maximum and minimum values ​​of the effective voltage change within a short time period (e.g., 10 minutes). short The percentage of this difference relative to the rated voltage is the voltage fluctuation value. Flicker value (P) st The measurement of voltage fluctuations is based on the international standard IEC 61000-4-15. It is obtained by weighting and statistically analyzing the voltage fluctuation signal using a model simulating the human eye's sensitivity to light flicker. Based on the calculated voltage fluctuation and flicker value sequence, further spectral analysis or wavelet analysis is used to identify whether there is a dominant periodic variation component, and the frequency and amplitude characteristics of this periodic variation are extracted as a dynamic fluctuation index describing the periodic or quasi-periodic changes in voltage amplitude.

[0032] Furthermore, in the above technical solution, the construction of the dynamic database includes: Use a time-series database to store second-level runtime data; Use graph databases to store device relationships and historical failure cases; A digital twin database is used to store real-time mirror files of the digital twin, which include a 3D model of the photovoltaic power station and threshold ranges for equipment parameters.

[0033] It should be noted that the specific implementation method of "building a dynamic database" is as follows: The dynamic database is composed of three types of specialized databases: time-series database, graph database, and twin database. Data exchange and synchronization are achieved through a unified data bus and API interface.

[0034] Time-series database construction and data storage: InfluxDB or TimescaleDB is selected as the core of the time-series database. A storage structure is designed for the second / millisecond-level runtime data from the data acquisition module. Taking "power quality" as an example, a measurement named `power_quality` is created, whose tags include `station_id` (power station number), `device_id` (equipment identifier, such as "INV_01"), and `point_of_common_coupling` (grid connection point identifier). Its fields store specific indicator values, such as voltage, THD (harmonic distortion rate), and frequency. Each data entry is automatically appended with a high-precision timestamp. The database is partitioned by time range (e.g., by month), and indexes are created for frequently queried fields (such as `device_id` and `point_of_common_coupling`) to achieve efficient writing and fast retrieval of historical and real-time data streams.

[0035] Graph database construction and relational storage: Neo4j or JanusGraph are chosen as the core of the graph database. In the graph database, the entities of the photovoltaic power station are abstracted as nodes, and the relationships between entities are abstracted as edges.

[0036] Node Definition: Two main types of nodes are created. First, device nodes, whose properties include device_type (e.g., photovoltaic module, inverter, transformer), model, location, manufacturer, etc. Second, fault case nodes, whose properties include fault_id, fault_time, fault_phenomenon (e.g., "voltage sag"), root_cause (e.g., "module string failure"), solution, etc.

[0037] Edge definition: Between device nodes, a PHYSICALLY_CONNECTED_TO (physical connection) relationship is established based on the electrical wiring diagram, forming a topology network. Between device nodes and fault case nodes, a HAS_HISTORICAL_CASE (with historical cases) relationship is established; the edge attributes record the frequency of that device's occurrence in this type of fault. Furthermore, based on data analysis, a SIMILAR_TO (similar to) relationship can be established between different fault case nodes for case reasoning.

[0038] The construction and mirror storage of the twin database: A database management system based on object storage (such as Amazon S3 or MinIO) is used to store the full mirror files of the digital twin.

[0039] 3D Model Storage: The 3D model of the photovoltaic power station is saved using a lightweight glTF format file, which contains the location, geometry, and hierarchical structure information of all equipment in the power station.

[0040] Parameter threshold ranges and real-time mapping: Each digital twin device object is associated with a parameter configuration file (e.g., JSON format), which defines the normal threshold ranges, alarm threshold ranges, and danger threshold ranges for various operating parameters (e.g., voltage, temperature, efficiency). The twin database obtains or subscribes to the latest operating data of key equipment from a time-series database in real time through a predefined interface and updates it to the corresponding device twin in the 3D model, thus achieving "real-time mirroring." When a virtual fault data injection instruction is received from the knowledge graph construction module, the twin database creates and saves a simulation mirror copy under a specific fault scenario.

[0041] In one specific embodiment: All power characteristic data are stored in a time-series database (such as InfluxDB). Corresponding data models (measurements) are designed for different types of characteristics. Steady-state metrics are stored in a Measurement named steady_state_metrics. Its tags include station_id, device_id, and metric_type (e.g., voltage_deviation). Fields store the specific calculated values ​​(e.g., value=2.1).

[0042] Transient event characteristics are stored in a Measurement named transient_events. Each transient event (such as a voltage sag) is recorded as an independent record, with a label identifying the event type (event_type) and associated device, and fields recording the quantitative characteristics of the event, including the disturbance magnitude (depth), duration, and the precise timestamp of the occurrence.

[0043] Dynamic fluctuation characteristics are stored in a Measurement named dynamic_fluctuation, recorded periodically (e.g., every second) or by event triggering. Tags identify the feature type (e.g., flicker, voltage_fluctuation), and fields record information such as short-term flicker value (Pst), voltage fluctuation amplitude (ΔU), and the dominant periodic frequency (dominant_freq) extracted through spectral analysis.

[0044] All records include a high-precision timestamp, data quality code, and device identifier assigned by the data source, ensuring the timeliness, traceability, and query capabilities of the data.

[0045] Specific feature sequences or events stored in the time-series database can be associated with corresponding device nodes and historical fault case nodes in the graph database through shared device identifiers (device_id) and timestamps. For example, a record of a severe voltage sag event can establish an INSTANCE_OF (belongs to) relationship with the case rule node in the graph database that describes "a voltage sag caused by a photovoltaic string fault," thereby enriching the instance data in the case library.

[0046] During simulation, digital twins can read historical or real-time feature data from a time-series database as input boundary conditions. Simultaneously, the virtual feature data output from the simulation, such as output fluctuations under extreme lighting conditions, can be captured by the twin database to generate corresponding virtual feature records. The storage structure is consistent with the real data, but they are distinguished by special tags (such as source=simulation) to supplement the knowledge graph.

[0047] The data integration module provides a unified query interface. When the system needs to analyze power quality anomalies during a certain period, it can simultaneously retrieve relevant steady-state, transient, and dynamic fluctuation characteristic sequences from the time series database, retrieve associated equipment topologies and historical rules from the graph database, and obtain equipment thresholds and simulation images from the twin database, thereby realizing the fusion analysis and visualization of multi-dimensional data.

[0048] Inter-database collaboration: The data integration module acts as the central hub, coordinating the three databases. Raw operational data is first continuously written to the time-series database. Based on equipment ledgers and network topology, the nodes and physical connections in the graph database are initialized. The 3D model and initial thresholds of the digital twin database are loaded. When a fault event occurs, the final analysis report generated by the reporting and feedback module is extracted as a new "fault case node," along with its related "equipment nodes" and "causal relationship chains," and persistently stored in the graph database, thus achieving self-enrichment of the knowledge base. The initial data and boundary conditions required for digital twin simulation are obtained from the time-series database and the graph database.

[0049] The knowledge graph construction module is used to construct a dynamic knowledge graph based on the dynamic database, and to update the dynamic knowledge graph using real-time data and integrated dynamic knowledge. Furthermore, in the above technical solution, the construction of the dynamic knowledge graph includes: Physical layer construction: Based on device connection relationships, a topology network including photovoltaic modules, inverters, transformers and grid nodes is constructed; Constructing a state layer: dynamically associating and mapping real-time parameters to form real-time power quality nodes, equipment operating status nodes, and equipment health nodes; Building a rules layer: Integrating rule nodes based on national standards and historical cases; Construct a prediction layer: Based on graph neural networks, learn the embedded representations of nodes in the state layer and physical layer to predict the range of impact of equipment failure on the power quality of the grid connection point.

[0050] It should be noted that the specific implementation method for "constructing a dynamic knowledge graph" is as follows: The knowledge graph construction module is based on a graph database (such as Neo4j) and builds and integrates the physical layer, state layer, rule layer and prediction layer by defining a unified graph schema.

[0051] The physical layer is constructed based on the device nodes and PHYSICALLY_CONNECTED_TO relationship edges stored in the graph database, creating a topology network representing electrical connections. In this layer, each device node (such as a photovoltaic module, inverter, or transformer) has fixed attributes such as device ID, device type, and rated parameters. Relationship edges have attributes such as connection type (e.g., "DC series," "AC parallel") and impedance parameters. This layer forms the static skeleton of the knowledge graph, clearly defining the physical paths of power flow and device influence.

[0052] The state layer is constructed and dynamically associated: it overlays dynamic running states onto the physical layer skeleton. It creates three types of dynamic nodes: Real-time power quality node: Associated with the "grid connection point" device node in the physical layer. The node's attribute values ​​(such as RMS voltage and harmonic distortion rate) are directly mapped and updated from the latest data obtained from the time-series database through a real-time API interface.

[0053] Device operation status node: Associated with each critical device node in the physical layer (such as inverters and transformers). Its attributes (such as output power, internal temperature, and alarm codes) are also mapped in real time from the time-series database.

[0054] Equipment Health Node: Associated with the equipment operating status node. Its health score attribute is calculated by a predefined evaluation model. The input data for this model comes from trend analysis of the equipment's historical performance data in the time series database (such as efficiency decay curves) and real-time status data.

[0055] The state layer links the aforementioned dynamic nodes with the static device nodes of the physical layer through relation edges such as HAS_REAL_TIME_STATUS (which has real-time status) and HAS_HEALTH_INDEX (which has health indicators), thereby achieving the binding of state and entity.

[0056] Construction of the rule layer: This layer solidifies domain knowledge and experience into rule nodes and their relationships.

[0057] National Standard Rule Node: Create a node representing the national power quality standard (e.g., "GB / T 12325-Voltage Deviation"), whose attributes include the standard number, limit parameters (e.g., "220V single-phase power supply, +7%, -10%)", and measurement method. Link such nodes to the "Real-time Power Quality Node" in the state layer via the APPLIES_TO (Applicable) relation edge.

[0058] Case rule nodes: Historical fault case nodes stored in the graph database are introduced into the knowledge graph as case rule nodes. Through relational edges such as CAUSED_BY (caused by...) and MANIFESTED_AS (manifested as), causal or association rule chains are constructed from "device nodes", "device health nodes" to "power quality anomaly nodes".

[0059] Construction of the prediction layer and application of graph neural networks (GNN): The prediction layer is a GNN-based inference model. It does not store static nodes, but uses the graph composed of the first three layers as input for computation.

[0060] Node embedding representation learning: The attributes of physical layer nodes (such as device type), attributes of state layer nodes (such as real-time values ​​and health status), and their connections in the multi-layer graph are input into a GNN model (such as GraphSAGE or GAT). This model learns a low-dimensional, dense vector representation (i.e., embedding representation) for each node in the graph through message passing and aggregation. This vector captures the structural information and contextual state of the node.

[0061] Impact Scope Prediction: When a device node (e.g., "Inverter_01") displays a fault in its state layer attribute (e.g., "Fault Code: Over-temperature"), the prediction layer model is activated. The algorithm takes the embedded representation of the faulty device and the current embedding state of the entire knowledge graph as input, and outputs a list of affected nodes and their probabilities through a trained feedforward neural network layer. For example, it predicts that the "Voltage Fluctuation" node at the grid connection point has a 95% probability of being affected, while another unrelated inverter node at a distance has only a 5% probability of being affected. The prediction results are dynamically added to the graph in the form of "PREDICTED_IMPACT (Predicted Impact)" relationship edges, where the edge weights are the predicted probabilities.

[0062] At this point, a multi-layered dynamic knowledge graph integrating static topology, dynamic states, expert rules, and intelligent prediction has been constructed, providing a data and relational foundation for subsequent causal analysis.

[0063] Furthermore, in the above technical solution, the updating of the dynamic knowledge graph includes: Process real-time data streams and update the association mapping relationship between real-time power quality nodes, equipment operation status nodes, and equipment health nodes in the state layer; Based on the causal relationship model, corresponding association rule nodes are generated or updated in the rule layer; Virtual fault data generated by simulating extreme working conditions using digital twins is used to supplement the missing links and nodes in the knowledge graph.

[0064] It should be noted that the specific implementation of the "dynamic knowledge graph update" is as follows: The knowledge graph is updated as a continuous and automated process, coordinated by an update engine, and encompasses three modes: data-driven, knowledge-driven, and simulation-driven.

[0065] State layer updates based on real-time data streams: The update engine continuously monitors data streams from time-series databases (such as Kafka message queues). When it receives a running data packet with a new timestamp, it triggers the state layer update process. Based on the device ID and parameter type in the data packet, it locates the corresponding state layer node in the knowledge graph (such as the "Device Running Status Node" for "Inverter_01"). Then, it updates the corresponding attribute values ​​(such as output power and internal temperature) of that node with the latest data. At the same time, the update engine has pre-defined "state-relationship" mapping rules. For example, when the "internal temperature" attribute of an inverter continuously exceeds the threshold, the system not only updates the value but also automatically creates or strengthens an INDICATES_HEALTH_DEGRADATION (indicating a decrease in health) relationship edge between the inverter's "Device Running Status Node" and its "Device Health Node" in the knowledge graph, and assigns a weight value positively correlated with the degree and duration of exceeding the limit.

[0066] Rule layer iteration based on causal relationship model: When the causal analysis module outputs a new, verified causal relationship, the update engine instantiates it as a new "association rule node" in the knowledge graph rule layer. The node attributes include cause description, result description, strength value, confidence level, discovery time, etc. Subsequently, a CAUSES relationship path is created in the knowledge graph from the device node corresponding to the physical layer, through the "association rule node", pointing to the corresponding result node in the state layer. If the newly discovered causal relationship points to the same causal pair as an existing rule in the rule layer, the update engine compares the old and new data, updates the strength value and confidence attribute of the rule node according to a predefined strategy (such as taking a weighted average, prioritizing the one with higher confidence), and records the version iteration history.

[0067] Knowledge graph completion based on digital twin simulation: When a sparse data region is identified in the knowledge graph (such as missing fault cases under certain extreme conditions), the update engine sends an instruction to the twin database to drive the digital twin to load the corresponding extreme condition parameters and run the simulation model. After the virtual fault data generated during the simulation is recorded by the twin database, the update engine formats this data and adds it to the rule layer of the knowledge graph as a special "simulation case node". At the same time, corresponding virtual device nodes and state nodes are created or associated in the physical layer and state layer, and labeled with SIMULATED_UNDER_CONDITION (simulated under ... conditions) relationship edges to distinguish them from nodes and relationships formed by real data, thereby supplementing fault modes and impact paths that have not yet been observed in actual operation.

[0068] Update Coordination Mechanism: The three update methods mentioned above are managed uniformly by the scheduler in the update engine. State layer updates have the highest priority and are performed almost in real time; rule layer updates are performed in batches after causal analysis is completed; graph completion simulations are triggered when the system load is low or as planned. All update operations are logged to ensure the traceability of the knowledge graph's evolution process.

[0069] The causal analysis module is used to establish a causal relationship model for power quality problems based on the dynamic knowledge graph. Furthermore, in the above technical solution, refer to Figure 2 The establishment of the causal relationship model includes: Extract node association data from the state layer and rule layer of the dynamic knowledge graph; Based on the aforementioned associated data, mutual information entropy is used to quantify the dependencies between parameters, and a high-dimensional parameter association matrix is ​​constructed. Set a threshold to filter strongly correlated parameter pairs; Determine the causal direction between variables using causal discovery algorithms; The impact of the intervention was simulated using a structural causal model, and the strength of the causal effect was calculated. Based on at least one of the dimensions of time lag, effect significance, and physical interpretability, each causal relationship is evaluated and assigned a corresponding weight to construct a causal relationship model that includes the causal relationships and their weights.

[0070] It should be noted that the specific implementation method of the "establishment of the causal relationship model" is as follows: The steps for establishing the causal relationship model are as follows: Step 1: Node Association Data Extraction: The causal analysis module first extracts structured association data from the state and rule layers of the dynamic knowledge graph. Specifically, it extracts the time-series attribute values ​​(such as voltage, power, temperature, and health scores) of all "real-time power quality nodes," "equipment operation status nodes," and "equipment health nodes" within a historical time window from the state layer, forming a multivariate time series dataset indexed by time. Simultaneously, it extracts the attributes (such as cause description, result description, and intensity value) of all "association rule nodes" from the rule layer and transforms them into a set of prior knowledge pairs of "cause variable - result variable." These two parts of data together constitute the input for subsequent causal analysis.

[0071] Step 2: Mutual Information Entropy Calculation and Correlation Matrix Construction: For each pair of parameters in the above multivariate time series (e.g., variable X: "inverter_01 temperature" and variable Y: "grid connection point harmonic distortion rate"), calculate its mutual information entropy I(X;Y). Mutual information entropy is calculated based on the Kullback-Leibler divergence between the joint probability distribution and the respective marginal probability distributions of the two variables. A larger value indicates a stronger statistical dependency between the two variables, potentially implying a causal relationship. After traversing all parameter pairs, construct an N×N high-dimensional parameter correlation matrix M, where the matrix elements M... ij That is, the mutual information entropy value I(i;j) between parameters i and j.

[0072] Step 3: Screening of Strongly Correlated Parameter Pairs: Set a mutual information entropy threshold θ (e.g., the value corresponding to p < 0.01 determined by a statistical significance test, or a fixed value set based on domain experience). Set all elements in the association matrix M that are less than the threshold θ to zero, thereby screening out all strongly correlated parameter pairs that satisfy I(i;j) ≥ θ. These parameter pairs constitute a preliminary undirected association network, serving as a candidate edge set for causal discovery.

[0073] Step 4: Causal Direction Discovery: Using constraint-based or score-based causal discovery algorithms, the causal direction between variables is determined on the network of strongly correlated parameter pairs. For example, using the PC algorithm or FCI algorithm, and utilizing conditional independence tests (such as partial correlation tests based on Gaussian distributions or conditional independence tests based on kernel methods), the direction of each edge is gradually determined while considering the conditional dependencies of other variables, ultimately outputting one or more possible causal directed acyclic graphs. If the data is a time series, the time-lag Granger causality test or variational autoencoder causal discovery method can be combined to use the temporal order to assist in determining the causal direction.

[0074] Step 5: Calculation of Causal Effect Strength: For each candidate causal relationship (e.g., X→Y) output from the causal discovery step, a structural causal model is constructed. In this model, an intervention is simulated using "do-calculus" or "counterfactual reasoning." Specifically, the expected change in outcome variable Y when a unit intervention (do(X=x+1)) is applied to causal variable X is calculated; that is, the average causal effect (ACE) or a similar indicator (e.g., ATE) is calculated. This effect value is calculated using observation-based estimation methods (e.g., inverse probability weighting, dual machine learning), and its magnitude quantifies the strength of the causal relationship, where x is a specific observation or value of causal variable X at a particular moment or under a certain baseline condition. For example, it could be the actual temperature measurement of the inverter at time t (e.g., "50℃"), or the historical average temperature value of the inverter under normal operating conditions.

[0075] Step 6, Causal Relationship Weighting and Model Building: For each causal relationship verified in the above steps, a comprehensive evaluation is performed from three dimensions and a comprehensive weight W is assigned: Time lag: Analyze the time lag relationship between causal variable X and outcome variable Y, and calculate the time delay τ that maximizes their correlation. A higher score is assigned to lag relationships that conform to the laws of physical propagation.

[0076] Effect significance: Based on the causal effect strength and its statistical confidence interval (e.g., 95% CI) calculated in step 5, the stronger the effect and the narrower the confidence interval, the higher the score.

[0077] Physical interpretability: Based on knowledge of the power system domain (such as circuit principles and equipment characteristics), determine whether the causal relationship has a reasonable physical explanation. This can be assessed by querying prior rules in the knowledge graph rule layer or consulting a database of domain experts.

[0078] Ultimately, the weight W of each causal relationship is synthesized from the scores of the three dimensions mentioned above through a weighted average or a rule-based mapping function. All causal relationships with direction, effect strength, and overall weight together constitute a complete and quantifiable causal relationship model, which is essentially a weighted, directed causal network.

[0079] The backtracking and positioning module is used to trace the causal chain back based on the causal relationship model, locate the root cause affecting power quality, and generate an analysis report; Furthermore, in the above technical solution, refer to Figure 3 The causal chain backtracking includes: when the backtracking and positioning module detects an abnormal power quality indicator, it automatically aggregates related data; and using the abnormal indicator as the starting node, it recursively traverses the upstream causal nodes in descending order of the weight of the causal edges in the directed acyclic graph constructed based on the causal relationship model, until the root cause node is located.

[0080] It is important to know that the specific implementation method of the "causal chain backtracking" is as follows: When the backtracking and location module detects that any power quality indicator (such as voltage sag depth) exceeds a preset threshold through the real-time monitoring interface, it automatically triggers the backtracking process. First, data aggregation: The module immediately aggregates various types of data that are temporally related and topologically adjacent to the abnormal indicator from the dynamic database and knowledge graph. This includes the operating status of all relevant equipment within a time window before and after the abnormal event, historical similar cases, current environmental parameters, and all nodes and edges related to this indicator in the causal relationship model. Next, a query subgraph is constructed: Using the "abnormal indicator" as the starting node, in the globally weighted directed acyclic graph of the causal relationship model, all incoming edges ending at this node (i.e., the "result") and their source nodes (i.e., the "cause") are extracted to form a query subgraph for backtracking. Then, recursive backtracking traversal: Starting from the starting node, the algorithm selects the node with the highest weight causal edge among all upstream causal nodes pointing to it as the primary tracing object and backtracks along this edge to that upstream node. Subsequently, this upstream node is treated as the new "result" node, and the above process is repeated—that is, among all its upstream nodes, the edge with the highest weight is selected for tracing. This process proceeds recursively until one of the following termination conditions is met: a) Tracing back to the root node that no longer has upstream causal nodes (i.e., no other in-model variables are its cause). b) Trace back to a specific cause node. If the node type is a specific and actionable equipment failure state (such as "Inverter_01_IGBT overheating") or a severely deteriorated equipment health (such as "PV string_A_aging degree>30%)", this node is identified as the root cause node of the current abnormal event.

[0081] The entire tracing path was recorded as a "causal chain" from the root cause to the initial anomaly.

[0082] The backtracking and location module automatically generates a structured analysis report based on the located root cause node, the complete causal chain, and aggregated related data. The report includes: anomaly description, root cause diagnosis, causal path diagram, impact scope assessment, and handling recommendations. Upon receiving this report, the reporting and feedback module performs corresponding operations based on its conclusions: Power quality control—if the report recommends emergency control (e.g., "voltage drop, requiring rapid reactive power support"), the module issues a preset reactive power compensation command to the designated inverter or SVG device via standard protocols (e.g., IEC 61850 GOOSE or Modbus TCP); Equipment operation and maintenance instructions—if the report indicates long-term parameter deviations or equipment health issues, such as "control parameter drift of a certain inverter," the module can automatically, or after confirmation by operation and maintenance personnel, issue parameter adjustment commands through the power plant monitoring system, or generate a work order and push it to the operation and maintenance management system. Long-term optimization recommendations (e.g., energy storage configuration schemes) are output as a recommendation report for decision-making reference.

[0083] The reporting and feedback module is used to execute power quality control or equipment operation and maintenance instructions based on the analysis report, and to store the analysis report and related operation logs on the blockchain.

[0084] It's important to understand that the reporting and feedback module initiates a notarization process simultaneously with generating analysis reports and executing any regulatory instructions. Using a permissioned blockchain platform (such as Hyperledger Fabric), the analysis report's hash value, key evidence data (such as the original data fragments that triggered the anomaly, causal chain summaries), executed operation logs (including instruction content, time, and execution results), and operator digital signatures (if manual confirmation is required) are packaged into a single data block. This data block, after verification by consensus nodes in the blockchain network, is added to an immutable distributed ledger, generating a unique transaction receipt (Transaction ID). This receipt is archived along with the source data, completing trusted notarization of the entire process from "data collection - analysis and diagnosis - decision execution," for subsequent auditing, liability determination, or data sharing verification.

[0085] Furthermore, in the above technical solution, the execution of power quality regulation or equipment operation and maintenance instructions includes: quickly triggering the reactive power compensation function of the inverter when the voltage drops sharply; automatically adjusting the control parameters of the photovoltaic power station when the analysis report shows that the power quality indicators have long-term deviations; and generating a configuration optimization scheme for the distributed energy storage system based on long-term trend analysis.

[0086] It should be noted that the specific implementation method for "executing power quality control or equipment operation and maintenance instructions" is as follows: The reporting and feedback module performs three different levels of closed-loop operations based on the type and urgency of the analysis report: For rapid reactive power compensation in response to voltage dips: When the analysis report diagnoses a voltage dip at the grid connection point (e.g., the report indicates "effective voltage value below 0.9 pu, duration > 100ms") and the root cause is related to insufficient reactive power support in the system, the module immediately initiates a rapid control process. First, the module calculates the required dynamic reactive power compensation (Q_ref) based on real-time grid data (e.g., voltage dip depth, phase transition) and inverter operating status. Then, through a low-latency communication link (e.g., based on IEC 61850-8-1 GOOSE messages or high-speed substation internal bus), a reactive power compensation command containing the target reactive power value, response time requirement (e.g., ≤20ms), and control mode (e.g., constant voltage mode) is sent to a pre-set, fast-response grid-connected inverter group or dedicated static var generator (SVG). Upon receiving the command, the inverter or SVG control system immediately adjusts the trigger pulses of its power devices, outputting capacitive or inductive reactive current to support rapid grid voltage recovery. The status and effects of the entire control process are monitored and fed back in real time, forming a closed-loop control.

[0087] For self-adjustment of control parameters to address long-term deviations in power quality indicators: When an analysis report indicates that a certain power quality indicator (such as voltage deviation or harmonic content) exhibits a "long-term deviation" (defined as a pass rate below a set threshold, such as 95%, within a statistical period, such as one week), and the root cause points to improper power plant control strategies or parameter settings, the module initiates a parameter optimization process. The module calls upon its built-in optimization algorithm library (such as model predictive control or heuristic optimization algorithms) to iteratively calculate the control parameters of relevant equipment, aiming to improve long-term power quality indicators and constraining the safe operating range of the equipment. Parameters to be optimized may include: inverter power factor setting curves, low-voltage ride-through curve parameters, active power change rate limits, and PI parameters of the PLL control loop. After calculating the optimized parameter set, the module securely sends the new parameters to the target equipment via standard remote control or parameter download commands from the Supervisory Control and Data Acquisition (SCADA) system. After adjustment, the system will continuously track changes in relevant indicators to verify the optimization effect.

[0088] Energy storage configuration optimization scheme generation based on long-term trend analysis: The module periodically (e.g., quarterly or annually) performs in-depth trend analysis on long-term operational data stored in the knowledge graph to identify the regularity and vulnerability of power quality problems, as well as the interaction needs between the power plant and the grid. Based on this, energy storage configuration optimization analysis is initiated. This analysis establishes a multi-objective optimization model that includes investment costs, operational benefits, power quality improvement benefits, and grid constraints (such as power fluctuation limits). Model inputs include historical and predicted load curves, photovoltaic power generation curves, electricity price signals, and power quality event statistics. By solving this model (using linear programming, mixed integer programming, or genetic algorithms), the module automatically generates a detailed distributed energy storage system configuration optimization scheme report. The report includes: recommended rated power and capacity of the energy storage system, optimal installation location (e.g., grid connection point, key feeders), operational strategies (e.g., priority and timing of peak shaving, valley filling, fluctuation mitigation, and reactive power support), and expected investment payback period and power quality improvement indicators (e.g., percentage improvement in voltage compliance rate). This scheme report is used for power station planning and decision-making.

[0089] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic power quality monitoring system for distributed photovoltaic grid-connected points, characterized in that, include: The data acquisition module is used to collect operational data from the grid side and the photovoltaic side. The operational data includes power quality indicators, grid operation status, power generation system parameters, and equipment health status. The data integration module is used to integrate the collected operational data, extract power characteristics, and build a dynamic database; The knowledge graph construction module is used to construct a dynamic knowledge graph based on the dynamic database, and to update the dynamic knowledge graph using real-time data and integrated dynamic knowledge. The causal analysis module is used to establish a causal relationship model for power quality problems based on the dynamic knowledge graph. The backtracking and positioning module is used to trace the causal chain back based on the causal relationship model, locate the root cause affecting power quality, and generate an analysis report; The reporting and feedback module is used to execute power quality control or equipment operation and maintenance instructions based on the analysis report, and to store the analysis report and related operation logs on the blockchain.

2. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 1, characterized in that: The power quality indicators include voltage parameters, current parameters, frequency deviation, three-phase imbalance, and power factor. The power grid operating status includes fault recording data and abnormal event timestamps; The parameters of the power generation system include DC current and voltage, AC output power, conversion efficiency, temperature, and environmental parameters. The health status of the equipment includes the degree of component aging and its operating status.

3. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 1, characterized in that: The extraction of electrical characteristics includes the following steps: Extracting steady-state characteristics: Based on the operating data, voltage deviation, frequency deviation, three-phase unbalance, and power factor are calculated as steady-state indicators; Extracting transient features: Based on the amplitude, duration, and occurrence time of voltage swell, voltage drop, or short-term interruption, instantaneous disturbance features are determined as transient indicators; Extracting dynamic fluctuation characteristics: Based on voltage fluctuations and flicker values, the periodic variation characteristics of voltage amplitude are determined as dynamic fluctuation indicators.

4. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 1, characterized in that: The construction of the dynamic database includes: Use a time-series database to store second-level runtime data; Use graph databases to store device relationships and historical failure cases; A digital twin database is used to store real-time mirror files of the digital twin, which include a 3D model of the photovoltaic power station and threshold ranges for equipment parameters.

5. A dynamic power quality monitoring system for distributed photovoltaic grid-connected points as described in claim 1, characterized in that: The construction of the dynamic knowledge graph includes: Physical layer construction: Based on device connection relationships, a topology network including photovoltaic modules, inverters, transformers and grid nodes is constructed; Constructing a state layer: dynamically associating and mapping real-time parameters to form real-time power quality nodes, equipment operating status nodes, and equipment health nodes; Building a rules layer: Integrating rule nodes based on national standards and historical cases; Construct a prediction layer: Based on graph neural networks, learn the embedded representations of nodes in the state layer and physical layer to predict the range of impact of equipment failure on the power quality of the grid connection point.

6. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 5, characterized in that: The updating of the dynamic knowledge graph includes: Process real-time data streams and update the association mapping relationship between real-time power quality nodes, equipment operation status nodes, and equipment health nodes in the state layer; Based on the causal relationship model, corresponding association rule nodes are generated or updated in the rule layer; Virtual fault data generated by simulating extreme working conditions using digital twins is used to supplement the missing links and nodes in the knowledge graph.

7. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 1, characterized in that: The establishment of the causal relationship model includes: Extract node association data from the state layer and rule layer of the dynamic knowledge graph; Based on the aforementioned associated data, mutual information entropy is used to quantify the dependencies between parameters, and a high-dimensional parameter association matrix is ​​constructed. Set a threshold to filter strongly correlated parameter pairs; Determine the causal direction between variables using causal discovery algorithms; The impact of the intervention was simulated using a structural causal model, and the strength of the causal effect was calculated. Based on at least one of the dimensions of time lag, effect significance, and physical interpretability, each causal relationship is evaluated and assigned a corresponding weight to construct a causal relationship model that includes the causal relationships and their weights.

8. The distributed photovoltaic grid-connected point power quality dynamic monitoring system as described in claim 1, characterized in that: The causal chain backtracking includes: when the backtracking and positioning module detects an abnormal power quality indicator, it automatically aggregates related data; and using the abnormal indicator as the starting node, it recursively traverses the upstream causal nodes in descending order of the weights of the causal edges in the directed acyclic graph constructed based on the causal relationship model, until the root cause node is located.

9. A dynamic power quality monitoring system for distributed photovoltaic grid-connected points as described in claim 1, characterized in that: The execution of power quality control or equipment operation and maintenance instructions includes: quickly triggering the reactive power compensation function of the inverter when the voltage drops sharply; automatically adjusting the control parameters of the photovoltaic power station when the analysis report shows that the power quality indicators have long-term deviations; and generating a configuration optimization scheme for the distributed energy storage system based on long-term trend analysis.

Images