A distributed photovoltaic cooperative regulation method based on multi-source data fusion

Abstract

The application relates to a distributed photovoltaic cooperative regulation method based on multi-source data fusion, and belongs to the technical field of new energy grid connection and intelligent regulation. In view of the problems of inconsistent multi-source data and poor regulation reliability, semantic features are extracted from multi-source heterogeneous data, and a dynamic semantic credibility graph is constructed. Through capability-oriented feature hierarchical distillation and collaborative semantic aggregation of a graph neural network, the power adjustable capacity of a photovoltaic cluster is dynamically quantified, and the inverter power distribution and voltage control strategy are adaptively optimized. The method realizes dynamic feature fusion and feedback correction driven by data, improves the accuracy and robustness of the evaluation result, and improves the rapid response and adaptive regulation capability of the distributed photovoltaic system to the operation change of the power grid.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic coordinated regulation and multi-source data fusion evaluation technology, and in particular to a distributed photovoltaic coordinated regulation method based on multi-source data fusion. Background Technology

[0002] The current assessment and control decisions for the adjustable power of distributed photovoltaic (PV) systems face the practical need for highly integrated multi-source heterogeneous data. Distribution network dispatching requires the collaborative processing of multi-dimensional information such as PV inverter power response, grid connection voltage, meteorological irradiance, and feeder load. Current mainstream technologies fall into two categories: one uses a cleaning-weighting-fusion process as its core, employing statistical noise reduction and linear weighting for feature trimming. While this alleviates outlier sensitivity, it struggles to discern semantic differences within the data stream, easily leading to redundant accumulation or conflict amplification. The other approach uses a deep learning end-to-end fitting route, directly inputting multiple measurement streams to train the evaluation model. While this can uncover complex spatiotemporal correlations, it treats each input as a homogeneous signal, failing to dynamically respond to fluctuations in data source confidence. Furthermore, the model's interpretability is weak under disturbances such as rapid cloud cover and sudden load changes, making it difficult to robustly guarantee control reliability.

[0003] The aforementioned solutions generally suffer from the following prominent problems: they treat multi-source inputs merely as numerical noise, failing to identify the semantic consistency and conflicting nature that affect power adjustability modeling; redundant signals are not effectively filtered, and conflicting signals are often indiscriminately suppressed or excluded, thus losing early warning information during operating condition switching. Furthermore, they lack an explicit dynamic modeling mechanism for semantic credibility, making it impossible to extrapolate the trust levels of various data sources describing the same physical capability boundary under different spatiotemporal and operating environments in real time. This results in significant model fluctuations during scenario switching or abnormal disturbances, limiting the security and economy of control commands. In actual engineering, inverter types and access points are widely distributed, data protocols vary across regions, sampling delays are uneven, and factors such as topological coupling between devices, timing response misalignments under sudden weather changes, and control closed-loop feedback delays further constrain the predictive boundaries of existing multi-source fusion models and affect their stability.

[0004] Existing technologies urgently need to overcome the following bottlenecks: They should be able to dynamically identify semantic redundancy and conflict in multi-source heterogeneous data streams, proactively mine valuable precursor features for operational condition switching and control strategies during periods of information inconsistency, and update the semantic credibility of each data source in real time based on physical topology, temporal correlation, and historical control feedback. This guides feature selection and adaptive adjustment of model structure, achieving robustness, interpretability, and closed-loop collaborative optimization of evaluation results. This patent addresses these issues by proposing a data conflict identification and capability-oriented feature distillation method based on a dynamic semantic credibility map. Under typical operating conditions, this method proactively identifies and utilizes the synergies and discrepancies between data from different sources, transforming information conflicts into interpretable identifiers for control strategy optimization. This provides a theoretical foundation and engineering implementation path for the collaborative control of photovoltaic clusters. Summary of the Invention

[0005] This application provides a distributed photovoltaic coordinated control method based on multi-source data fusion, which aims to solve the problems of lack of semantic credibility dynamic modeling in multi-source data fusion, suppression of conflict signals instead of utilization, and poor evaluation reliability during operating condition switching.

[0006] This application provides a distributed photovoltaic coordinated control method based on multi-source data fusion, specifically including: 1. A distributed photovoltaic coordinated control method based on multi-source data fusion, characterized in that it includes: Acquire multi-source heterogeneous raw data streams such as active and reactive power of photovoltaic inverters, grid connection point voltage, meteorological irradiance and distribution network feeder load, and extract implicit capability characterization semantic tags from the multi-source heterogeneous raw data streams to generate semantic anchoring feature sets; Based on the semantic anchoring feature set and historical data of equipment topology relationships, meteorological evolution paths and control command closed-loop feedback, a dynamic semantic credibility map is constructed. The capability-guided feature distillation operation is performed using the current credibility score of each node in the dynamic semantic credibility graph. High credibility sources are mapped to the capability boundary anchoring layer to extract the voltage reactive sensitivity slope, medium credibility sources are mapped to the trend guidance layer to extract the irradiation attenuation gradient, and low credibility but high conflict source is mapped to the operating condition switching detection layer to extract the cross-domain correlation pattern, thereby generating a hierarchical distillation feature subset. The graph neural network with a fixed input structure of the hierarchical distillation feature subset but with edge weights reconfigured in real time by the dynamic semantic credibility graph outputs the photovoltaic cluster power adjustability assessment result with semantic source tracing label; Based on the distribution network operation status, dispatch instructions and time-of-use electricity price signals, and combined with the photovoltaic cluster power adjustability assessment results, the active power allocation scheme and reactive voltage control strategy for each inverter are calculated and generated collaboratively to form a joint optimization control instruction set. The joint optimization control instruction set is decomposed and distributed to each photovoltaic inverter for execution. The actual operating trajectory of the inverter is collected and compared with the theoretical capability curve to generate the control closed-loop verification update quantity used to reverse the adjustment of the edge weights of the dynamic semantic credibility graph in the next round.

[0007] This application provides a distributed photovoltaic coordinated control method based on multi-source data fusion, which has the following beneficial effects: By introducing semantic anchoring units and a dynamic semantic credibility graph mechanism, this approach abandons traditional numerical normalization and static weight allocation strategies. Instead, it extracts the implicit capability representation semantic labels from each data stream and dynamically constructs a real-time evolving credibility graph based on three dimensions: topological coupling, temporal compatibility, and control closed-loop verification. This adaptively identifies the main control signal source and suppresses the misleading influence of conflict noise. Furthermore, a hierarchical feature distillation and graph neural network collaborative architecture is adopted, designing a three-stage distillation path: a capability boundary anchoring layer, a trend guidance layer, and a condition switching detection layer. These layers extract hard constraint features, trend features, and patterns, respectively. The switching signal uses high-confidence sources to lock the equipment's adjustment limits, medium-confidence sources to provide the direction of output evolution, and low-confidence but abrupt signal characteristics to be captured as the basis for triggering model structure switching. The graph neural network only learns the semantic cooperative relationship between features and includes a semantic source tracing path to ensure interpretability. The overall solution replaces the redundant calculations at the traditional numerical level with a structured organization at the semantic level. Without significantly increasing the algorithm's complexity, it can effectively cope with typical disturbance conditions such as rapid cloud cover and load jumps, providing reliable technical support for the accurate perception and autonomous collaborative control of distributed resources in new power systems. Attached Figure Description

[0008] Figure 1 This is the main flowchart of a distributed photovoltaic collaborative control method based on multi-source data fusion.

[0009] Figure 2 This is a sub-flowchart of a distributed photovoltaic coordinated control method based on multi-source data fusion.

[0010] Figure 3 This is another sub-flowchart of a distributed photovoltaic coordinated control method based on multi-source data fusion. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1As shown, this application provides a distributed photovoltaic coordinated control method based on multi-source data fusion, specifically including: S1: Acquire multi-source heterogeneous raw data streams such as active and reactive power of photovoltaic inverters, grid connection point voltage, meteorological irradiance and distribution network feeder load, and extract implicit capability characterization semantic tags from the multi-source heterogeneous raw data streams to generate a semantic anchoring feature set including AGC response ready state Boolean signal, irradiance attenuation persistence level and rigid load ratio.

[0014] S2: Based on the semantic anchoring feature set and historical data of equipment topology, meteorological evolution path and control command closed-loop feedback, a dynamic semantic credibility graph is constructed. The graph nodes are defined as data source types, and the edge weights are dynamically generated by the three-dimensional scores of topological coupling degree, temporal compatibility degree and control closed-loop verification degree calculated in real time. The output is a dynamic semantic credibility graph with dynamic edge weights.

[0015] S3: Utilize the current credibility score of each node in the dynamic semantic credibility graph to perform capability-oriented feature distillation operation. Map high credibility sources to the capability boundary anchoring layer to extract the voltage reactive power sensitivity slope, map medium credibility sources to the trend guidance layer to extract the irradiance attenuation gradient, and map low credibility but high conflict source to the operating condition switching detection layer to extract the cross-domain correlation pattern, generating a hierarchical distillation feature subset.

[0016] S4: A graph neural network with a fixed input structure of hierarchical distillation feature subsets but whose edge weights are reconfigured in real time by a dynamic semantic credibility graph is trained to learn the semantic cooperative relationship between features rather than the numerical fitting relationship, and outputs the photovoltaic cluster power adjustable capability assessment result with semantic source tracing label.

[0017] S5: Based on the distribution network operation status, dispatch instructions and time-of-use electricity price signals, combined with the photovoltaic cluster power adjustability assessment results with semantic traceability tags, collaboratively calculate and generate active power allocation schemes and reactive voltage control strategies for each inverter, forming a joint optimization control instruction set.

[0018] S6: Decompose the joint optimization control instruction set and send it to each photovoltaic inverter through the coordination control device to execute actions. At the same time, collect the actual action trajectory of the inverter and compare it with the theoretical capability curve to generate the control closed-loop verification update quantity used to reverse correct the edge weights of the dynamic semantic credibility graph in the next round.

[0019] Step S1: Acquire multi-source heterogeneous raw data streams, including active and reactive power of the photovoltaic inverter, grid connection point voltage, meteorological irradiance, and distribution network feeder load. Extract implicit capability characterization semantic tags from each of the multi-source heterogeneous raw data streams to generate a semantic anchoring feature set containing AGC response ready state Boolean signals, irradiance attenuation persistence levels, and the proportion of rigid loads. Specifically, this includes: S1.1: Acquire multi-source heterogeneous raw data streams such as active and reactive power of photovoltaic inverters, grid connection point voltage, meteorological irradiance, and distribution network feeder load, and perform protocol parsing and timing alignment processing on the multi-source heterogeneous raw data streams to generate multi-source heterogeneous synchronous data streams with a unified time base, ensuring the timing compatibility basis in the subsequent semantic extraction process.

[0020] It should be noted that the semantic anchoring feature set is a set of structured feature vectors formed by standardizing and encapsulating the discretized capability representation semantic tags extracted from multi-source heterogeneous raw data streams according to a unified data structure. It includes device-side capability anchoring components, environment-side trend guidance components, and load-side acceptance capability components, and is used as direct input for constructing a dynamic semantic credibility map.

[0021] The smart gateway collects raw data streams of photovoltaic inverter active / reactive power, grid connection point voltage, meteorological irradiance, and distribution network feeder load from heterogeneous communication protocols such as Modbus and IEC104. It performs protocol parsing on the raw data streams, extracts the original timestamp tags and physical measurement values ​​of each data source, and establishes an initial data tuple set containing source identifiers, timestamps, and measurement values.

[0022] To address the issue of inconsistent time bases in multi-source data, the high-precision time synchronization signal of the distribution network dispatching master station is selected as the global unified time base. The linear offset between the local clock of each data source and the global time base is calculated, and the clock drift curve is fitted using the least squares method to generate dynamic time compensation parameters for each data source.

[0023] Based on dynamic time compensation parameters, the original timestamps of each data stream are linearly corrected to eliminate systematic time deviations caused by communication delays or device clock drift. The corrected timestamps are then mapped to a unified discrete time grid with a grid resolution set to milliseconds to match the high-frequency sampling requirements of the inverter.

[0024] Furthermore, a confidence-based weighted interpolation algorithm (parameters: historical stability score, grid resolution) is used to achieve grid alignment of asynchronous sampled data. For cases where multiple sampled values ​​exist but are mapped to the same time grid point, weights are assigned according to the historical stability score of the data source, and the unique valid measurement value at that time grid point is calculated to fill the data gaps caused by packet loss or asynchronous acquisition.

[0025] Through the above protocol parsing, clock drift correction and grid interpolation processing, the multi-source heterogeneous raw data stream is transformed into a multi-source heterogeneous synchronous data stream with a unified time base and spatial alignment, ensuring the strict temporal correspondence and compatibility of each physical quantity in the subsequent semantic tag extraction process.

[0026] For example, the global time base is set to UTC time, and the grid resolution is 10ms. Photovoltaic inverter data is acquired via the Modbus TCP protocol, which has an average communication delay of 50ms; meteorological data is acquired via the IEC104 protocol, which is subject to random jitter. The smart gateway collects the inverter's data at its local time t. inv =1000m The active power at time t is 15.0kW. inv The power output is 15.4 kW at 1010 ms. The inverter clock drift slope k is obtained by fitting using the least squares method. inv =1.2×10 6 intercept b inv = 48ms. After linear compensation of the inverter timestamp, it is mapped to a 10ms grid. At UTC time T=1000ms, if the inverter reports two neighboring sampling points, weighted interpolation is performed based on their historical stability scores w1=0.9 and w2=0.7. Assuming a time decay coefficient σ=5ms, the unique active power value of 15.2kW at that moment is calculated. In the final generated synchronous data stream, the inverter power, grid connection point voltage, irradiance, and load current all have definite corresponding values ​​every 10ms, with timing deviation controlled within ±2ms, significantly improving the timing consistency of multi-source data fusion.

[0027] S1.2: Based on the active and reactive power and grid connection point voltage data of photovoltaic inverters in the multi-source heterogeneous synchronous data stream, the built-in inverter control logic state machine is used to perform response characteristic analysis and processing to extract the AGC response ready state Boolean signal that characterizes the inverter's real-time adjustment readiness state, and map continuous electrical measurement values ​​into discrete capability characterization semantic tags.

[0028] Read the active power sequence, reactive power sequence, and grid connection point voltage amplitude sequence of the photovoltaic inverter under a unified time reference, and construct a multi-dimensional electrical state vector containing the current time and historical sampling points within the forward sliding window, which serves as the basic input data for the reverse analysis of the inverter's internal control logic.

[0029] Based on the voltage-reactive power droop control curve and active power ramp-up rate limit parameters defined in the photovoltaic inverter grid connection standard, a theoretical response boundary model of the inverter is established, and the theoretical maximum regulation margin and minimum regulation dead zone range of the inverter reactive power output are calculated under the current grid connection point voltage conditions.

[0030] The actual rate of change of active power within a preset time window is calculated using a sliding window differential algorithm. This actual rate of change is then compared with the maximum active power ramp rate specified on the inverter's nameplate to generate an active power regulation capability confidence factor, which is used to quantify the inverter's dynamic response potential under the current thermal stability state.

[0031] The system detects whether the instantaneous voltage value at the grid connection point falls within the operating threshold range of the inverter's low voltage ride-through or overvoltage protection. If the voltage value is within the critical range of protection operation, the inverter is determined to be in transient protection priority mode, and the AGC response ready state is forcibly marked as unavailable to avoid conflicts between control commands and protection logic.

[0032] By combining reactive power regulation margin, active power ramping capability confidence factor, and voltage protection status, a multi-condition logic decision function is constructed. The AGC response ready state Boolean signal is output as true if and only if the reactive power margin is greater than the preset minimum control step size, the active power ramping capability confidence factor is higher than the threshold, and the voltage protection is not triggered; otherwise, the output is false.

[0033] Through the reverse analysis based on the control logic state machine and the multi-condition logic decision, the continuously fluctuating electrical measurement data is transformed into discrete binary semantic labels, which clearly identify whether the inverter has the physical and logical conditions to receive the upper-level coordinated control command at the current moment, thus realizing the mapping from numerical monitoring to capability semantics.

[0034] For example, the sliding window length is set to 50ms and the sampling frequency is 1kHz. The inverter's nominal maximum active power ramp rate is 10kW / s, and the minimum reactive power control step size is 0.5kVar. At a certain time t, the grid connection point voltage is 230V, which is within the normal operating range (220V-240V). The calculated maximum theoretical reactive power adjustment margin is 2.0kVar, which is greater than the 0.5kVar threshold. The active power change in the first 50ms is 0.4kW, and the calculated actual change rate is 8kW / s, which is less than the nominal value of 10kW / s. The active power ramp capability confidence factor is set to 1.0. The voltage does not trigger the protection logic. After logical judgment, all three conditions are met, and the AGC response ready state Boolean signal is True (1). If the voltage suddenly changes to 255V at this time, triggering the overvoltage protection critical judgment, then regardless of the power margin, False (0) is directly output, effectively preventing the risk of mis-regulation under abnormal grid conditions.

[0035] S1.3: Based on historical meteorological irradiance sequence data in multi-source heterogeneous synchronous data streams, the irradiance change trend is extrapolated using cloud motion vector inversion algorithm to generate irradiance attenuation persistence level, which characterizes the degree of light resource attenuation in the short term, and transforms meteorological observation values ​​into environmental constraint semantic labels with time dimension.

[0036] Read the time-aligned meteorological irradiance synchronous data stream, extract the global horizontal irradiance (GHI) sequence within the current time and the previous sliding window, and the cloud image matrix data captured by the sky imager to construct a meteorological state input tensor containing spatiotemporal dimensions.

[0037] The cloud map matrix is ​​preprocessed with grayscale conversion and Gaussian filtering for noise reduction. The pixel displacement field of adjacent frames of cloud maps is calculated using the optical flow method. The motion velocity components of the cloud clusters in the east-west and north-south directions are extracted, and the cloud cluster motion vectors are synthesized to quantify the physical trend of cloud layer movement.

[0038] By combining the geographical location of the photovoltaic array with the solar altitude angle, a cloud projection mapping model is established. The motion vector of the two-dimensional cloud map is projected onto the photovoltaic panel plane, the theoretical time delay of the cloud shading edge reaching the photovoltaic array is calculated, and the time window of the irradiance mutation is determined.

[0039] Within the theoretical time window, the GHI sequence is subjected to first-order difference operation to calculate the slope of the rate of change of irradiance, identify the start and end points of the irradiance decrease phase, and define the duration range of a single cloud cover event.

[0040] An exponential decay memory factor is introduced. Based on the product of the slope of the irradiance change rate and the duration of cloud cover, combined with a preset sensitivity adjustment coefficient and a benchmark threshold, an irradiance decay persistence level index is generated. The calculated index value is mapped to a preset discrete semantic interval and divided into three levels: "rapid recovery", "slow decay" and "continuous shading". Each level corresponds to a different confidence weight and control response strategy.

[0041] Through the above-mentioned cloud motion vector inversion and attenuation persistence quantification, continuous meteorological irradiance values ​​are transformed into time-dimensional environmental constraint semantic tags with clear physical meaning, realizing early perception and structured characterization of photovoltaic power output fluctuation trends.

[0042] For example, the sliding window length is set to 5 minutes and the sampling interval to 1 second. Cloud image data at a certain moment is acquired, and the cloud movement vector is calculated using the optical flow method. After being projected onto the photovoltaic panel plane, the theoretical time delay for the cloud shading edge to arrive is calculated to be 30 seconds. Within this time window, the GHI sequence shows that the irradiance continuously decreases from 800 W / m² to 400 W / m², lasting for approximately 120 seconds, with a relatively large negative average rate of change. After mapping using the exponential decay memory model, the semantic label for the irradiance decay persistence level is output as "continuous shading," triggering the trend guidance layer in the subsequent feature distillation module, thereby increasing the weight of meteorological data in power prediction.

[0043] S1.4: Based on the load current data of the distribution network feeder in the multi-source heterogeneous synchronous data stream, the wavelet packet decomposition algorithm is used to extract the frequency domain energy distribution features to calculate and output the proportion of rigid loads that characterize the rigidity and fluctuation characteristics of the load. The complex load waveform data is abstracted into load semantic labels that reflect the grid's acceptance capacity.

[0044] The synchronous data stream of distribution network feeder load current, after time-aligned processing, is received and used as the input for calculating the proportion of rigid loads.

[0045] Discrete wavelet packet decomposition is performed on the load current synchronization data stream. The Daubechies-4 wavelet basis function is selected, and the number of decomposition layers is set to 3. The original time domain signal is mapped to the multi-scale frequency domain space to generate a node coefficient matrix containing low-frequency approximate components and high-frequency detail components.

[0046] The energy distribution characteristics of each decomposition node are calculated based on the Shannon entropy principle. The normalized energy ratio of each node is obtained by calculating the proportion of the sum of squares of wavelet packet coefficients of each node to the sum of squares of total coefficients.

[0047] The set of low-frequency nodes is defined as the rigid load characteristic domain, and the set of high-frequency nodes is defined as the fluctuating load characteristic domain. The normalized energy proportions of all nodes in the low-frequency band are summed to obtain the rigid load energy concentration index.

[0048] By introducing a load fluctuation penalty factor, the severity of fluctuations is calculated based on the energy entropy value of high-frequency nodes. This nonlinear correction is applied to the rigid load energy concentration index to eliminate the influence of instantaneous harmonic interference on rigidity determination.

[0049] The modified rigid load energy concentration index is mapped to the [0,1] interval to generate a semantic label representing the rigid load ratio, which indicates that the load is less sensitive to voltage fluctuations and has a stronger capacity to accept them.

[0050] By extracting wavelet packet frequency domain energy distribution features and performing nonlinear correction processing, complex load waveform data is transformed into load semantic labels that reflect the grid's acceptance capacity, thereby enabling load characteristics to provide refined support for the power adjustable capability model.

[0051] For example, feeder load current data with a sampling frequency of 1kHz is subjected to 3-level Daubechies-4 wavelet packet decomposition, resulting in 8 frequency band nodes. The energy of each node is calculated; the total energy of low-frequency nodes (0-62.5Hz) accounts for 85% of the total energy, while the energy of high-frequency nodes accounts for 15%. A fluctuation penalty factor of 0.9 is set, resulting in a corrected rigid load proportion of 0.765. This indicator accurately reflects that the current load is mainly composed of rigid loads such as motors, possessing strong voltage support capabilities, and provides a quantitative basis for assigning high confidence values ​​to the load data source in the subsequent dynamic semantic credibility map.

[0052] S1.5: Based on the generated AGC response ready state Boolean signal, irradiance attenuation persistence level and rigid load ratio, a structured assembly process is performed using a feature vector encapsulation protocol to generate a semantic anchoring feature set containing multi-dimensional capability representation information, which serves as the direct input object for constructing a dynamic semantic credibility map.

[0053] The AGC response ready state Boolean signal, irradiance attenuation duration level, and discrete semantic tags of rigid load ratio output from steps S1.2 to S1.4 are received as the raw input data for feature encapsulation.

[0054] Define a standardized multidimensional semantic vector structure, which includes a device identifier field, a timestamp field, a semantic type field, and a confidence weight field, and establish a unified data schema specification for the feature set.

[0055] The AGC response ready state Boolean signal is mapped to a binary value of 0 or 1 and assigned a high-priority weight coefficient to characterize the inverter’s current hard regulation capability boundary, forming a device-side capability anchoring component.

[0056] The irradiance attenuation persistence level is quantified into a continuous value within the normalized interval [0,1]. Dynamic weights are assigned according to the level to characterize the environmental constraint strength on power fluctuations and form an environmental trend guidance component.

[0057] The calculated value of the rigid load ratio is directly embedded into the corresponding dimension of the vector, retaining its original physical dimensions or undergoing Z-score standardization, to characterize the grid side's acceptance elasticity to power backfeed, forming a load-side acceptance capacity component.

[0058] The component data of the above three dimensions are packaged and assembled using a serialization protocol, and a check code is added to ensure the integrity of the evidence chain, generating a compact semantic anchoring feature vector.

[0059] By using a feature vector encapsulation protocol for structured assembly, the multi-source heterogeneous discrete semantic tags extracted in the previous step are transformed into a semantic anchoring feature set with a unified format, clear physical meaning, and direct parsing by graph algorithms. This achieves a standardized leap from the numerical level to the semantic level for multi-source data, providing high-quality structured input objects for constructing dynamic semantic credibility graphs.

[0060] For example, the semantic vector dimension is set to 3. At a certain sampling time t, S1.2 outputs the AGC response ready state as True, mapped to the value 1, and the weight is set to 0.9; S1.3 outputs the irradiance attenuation persistence level as "moderate attenuation", quantized to 0.6, and the weight is set to 0.7; S1.4 outputs the rigid load proportion as 0.45, which is standardized to 0.5, and the weight is set to 0.8. The system assembles these three components into a vector V=[1,0.6,0.5], and attaches metadata {source_ids:[inv_01,met_02,load_03],timestamp:t,weights:[0.9,0.7,0.8]}. This feature set eliminates the dimensional differences and protocol barriers of the original data, enabling the subsequent graph construction module to directly calculate edge weights based on semantic consistency, significantly reducing the complexity of data preprocessing and improving the efficiency and interpretability of feature fusion.

[0061] like Figure 2 As shown, step S2: Based on the semantic anchoring feature set and historical data on equipment topology relationships, meteorological evolution paths, and closed-loop feedback of control commands, a dynamic semantic credibility graph is constructed. Graph nodes are defined as data source types, and edge weights are dynamically generated from the real-time calculated three-dimensional scores of topological coupling, temporal compatibility, and control closed-loop verification. The output is a dynamic semantic credibility graph with dynamic edge weights. Specifically, this includes: S2.1: Based on the semantic anchoring feature set and the physical connection relationship of the equipment, the electrical distance impedance matrix and power flow path topology between the photovoltaic inverter grid connection point, the meteorological monitoring station and the feeder terminal are analyzed to obtain the initial value of the topological coupling degree that characterizes the spatial correlation strength, and the initial value of the topological coupling degree is mapped to the static basic edge weight between nodes in the dynamic semantic credibility graph.

[0062] It should be noted that the dynamic semantic credibility graph is a dynamic evolution graph structure with multi-source heterogeneous data source types as nodes and real-time quantitative scores of semantic consistency strength between data sources as edge weights. It is used to characterize the relative credibility of each data source under the current operating conditions and the semantic correlation strength between them in each regulation cycle.

[0063] Read the node coordinate data and line parameter table from the distribution network GIS geographic information system database, extract the physical location index of the photovoltaic inverter grid connection point, meteorological monitoring station and feeder terminal in the distribution network topology map, and construct a basic topology node set containing node ID, latitude and longitude coordinates and the corresponding feeder identifier.

[0064] Based on the distribution network line impedance parameter table, a breadth-first search algorithm is called to traverse the minimum electrical path from the node where the meteorological monitoring station is located to the grid connection point of each photovoltaic inverter, calculate the cumulative resistance and reactance of all series branches on the path, and generate a path impedance vector that represents the tightness of electrical connection.

[0065] The Euclidean distance formula is used to calculate the straight-line geographical distance between the meteorological monitoring station and the grid connection point of each photovoltaic inverter. Combined with the path impedance vector, a spatial correlation strength index is constructed using a weighted fusion mechanism. The geographical distance reflects the spatial correlation of irradiance distribution, and the path impedance reflects the electrical conduction characteristics of voltage fluctuation.

[0066] The initial value of the topological coupling degree between the i-th photovoltaic inverter node and the meteorological monitoring station node is calculated using the following formula. : in, For the first The straight-line geographical distance between the inverter and the weather station The minimum electrical path impedance magnitude between the two. and These are the geographical distance weighting factor and the electrical impedance weighting factor, respectively. To prevent tiny constants with a denominator of zero.

[0067] The initial values ​​of the calculated topological coupling between each node pair are normalized and mapped to the interval between 0 and 1. These values ​​serve as the static basic edge weights connecting the "meteorological data source node" and the "inverter data source node" in the dynamic semantic credibility graph, thus establishing the initial trust benchmark for multi-source data in the spatial dimension.

[0068] Through the above-mentioned topological analysis and impedance weighted calculation processing method, the physical device location information on which the semantic anchoring feature set generated in the previous step depends is transformed into a quantified spatial correlation strength index, realizing the accurate initialization of the static basic edge weights in the dynamic semantic credibility graph, and providing stable spatial structure support for the dynamic superposition of subsequent temporal compatibility and closed-loop verification.

[0069] For example, suppose a 10kV distribution network feeder includes three photovoltaic inverters (Inv1, Inv2, and Inv3) and one meteorological monitoring station (Met1). Inv1 is 500 meters away from Met1 with a path impedance of 0.5 ohms; Inv2 is 1200 meters away from Met1 with a path impedance of 1.2 ohms; and Inv3 is 2000 meters away from Met1 with a path impedance of 2.5 ohms. Set the geographical distance weighting factor α1 to 0.001, the electrical impedance weighting factor α2 to 0.1, and the constant ε to 0.0001. Calculate the topological coupling degree C of Inv1. topo_1The formula is 1 / (0.001500+0.10.5+0.0001)=1 / 0.5501≈1.817, which is normalized to 0.95; C in Inv2 topo_2 The formula is 1 / (0.0011200+0.11.2+0.0001)=1 / 1.3201≈0.757, which is normalized to 0.65; Inv3's C topo_3 The value is 1 / (0.0012000+0.12.5+0.0001)=1 / 2.2501≈0.444, which is normalized to 0.35. This result accurately reflects the high spatial correlation between Inv1 and meteorological data, and the low correlation between Inv3, significantly improving the spatial rationality of the map construction.

[0070] S2.2: Using the time series data of the irradiance attenuation persistence level and AGC response ready state Boolean signal contained in the semantic anchor feature set, the sliding correlation calculation is performed on the time offset window between the start time of meteorological irradiance mutation and the start time of inverter reactive power response to generate a time series compatibility score characterizing the time synchronization accuracy, and the time series compatibility score is superimposed on the static basic edge weight to form the preliminary dynamic edge weight.

[0071] Extract the irradiance attenuation persistence level sequence and AGC response ready state Boolean signal sequence from the semantic anchor feature set, and construct a sliding time window of length L, where L is set as the number of sampling points covering the duration of a typical cloud cover process, ensuring that the window contains the complete meteorological abrupt event and the equipment response process it triggers.

[0072] Within the sliding window, identify the starting time t when the irradiance attenuation persistence level undergoes a step change. irr And the starting time t when the AGC response ready state Boolean signal flips from 0 to 1. inv Calculate the time offset Δt between the two, which reflects the sum of the physical delay and communication delay of the weather disturbance being transmitted to the inverter control system.

[0073] The cross-correlation function algorithm is used to calculate the correlation coefficient between the irradiance variation sequence and the inverter response sequence under lag τ, by traversing the preset lag interval [-T]. max ,T max [Find the optimal lag time τ that causes the correlation coefficient to reach its peak] opt This optimal lag time is the theoretical benchmark for the inherent response delay of the system.

[0074] Calculate the actual time offset Δt and the optimal lag time τ opt The absolute value of the deviation between |Δt-τ optThe deviation value characterizes the degree of inconsistency in the timing of the current data source. The larger the deviation, the worse the timing alignment of the two data sources when describing the same physical event.

[0075] Introducing a Gaussian kernel function maps timing bias to timing compatibility score S. temp The specific calculation formula is as follows: in, The time tolerance coefficient is set based on the inverter control dead zone range and is used to adjust the penalty for timing deviations on the compatibility score.

[0076] The calculated timing compatibility score S temp The static basic edge weights W generated in step S2.1 static By performing linear superposition, the preliminary dynamic edge weights W are obtained. dynamic_pre The overlay rule is W dynamic_pre =W static +λ·S temp , where λ is the temporal weight adjustment factor, used to balance the contribution ratio between topological stability and temporal dynamic consistency.

[0077] Through the above sliding window cross-correlation analysis and Gaussian kernel mapping, the static topological weights of the previous step are transformed into preliminary dynamic edge weights that fuse time-series dynamic features. This enables a quantitative assessment of the semantic consistency of multi-source data in the time dimension and significantly improves the map's adaptability to rapidly changing conditions.

[0078] For example, the sliding window length L is set to 50 sampling points (corresponding to 500ms), and the time tolerance coefficient σ is 10ms. In a certain cloud cover event, the irradiance attenuation start time t is detected. irr The inverter AGC response start time is t = 10:00:00.100. inv If the time interval is 10:00:00.135, then the time offset Δt is 35ms. The optimal lag time τ is obtained through cross-correlation calculation. opt The time series compatibility score is calculated by substituting the values ​​into the formula, assuming λ is 0.2 and the static basic edge weights W are... static If the initial dynamic edge weight W is 0.8, then... dynamic_pre The expression is 0.8 + 0.2 × exp(-(35-30)). 2 / (2×10 2 ))≈0.8+0.2×0.969=0.9938. This result shows that despite the additional 5ms timing skew, the two data sources remain highly compatible in timing, thus preserving strong connection weights in the graph and effectively avoiding misjudgments caused by minute timing jitter.

[0079] S2.3: Based on the residual sequence of the deviation between the actual operation trajectory of the inverter and the theoretical capability curve recorded in the closed-loop feedback data of the historical control command, the accuracy of the contribution of each data source to the power adjustable capability assessment in the previous control cycle is verified by reverse tracing. In order to calculate the control closed-loop verification degree correction coefficient that characterizes the reliability of historical decisions, the control closed-loop verification degree correction coefficient is applied to the preliminary dynamic edge weight to complete the dynamic iterative update of the weight.

[0080] This sub-step is the third stage in the dynamic semantic credibility graph construction process. It aims to use the actual physical response feedback after the regulation is executed to quantitatively evaluate the real contribution value of each data source in the previous cycle decision, thereby achieving closed-loop correction of the graph edge weights.

[0081] Read the historical control command closed-loop feedback database, extract the active / reactive setpoint sequence issued in the previous control cycle T-1, the actual output measurement value of the inverter at the corresponding time, and the theoretical capability boundary curve based on the multi-source data at that time, and construct a four-dimensional comparison dataset containing timestamps, command values, measured values ​​and predicted values.

[0082] A dynamic time warping algorithm is used to align the measured response waveform with the theoretical expected waveform in time, eliminating phase deviation caused by communication delay or equipment inertia. The absolute error value at each sampling point is calculated, and a motion trajectory deviation residual sequence characterizing the equipment response accuracy within a single cycle is generated.

[0083] Based on the deviation residual sequence, the mean and standard deviation of the residuals are calculated using the sliding window statistical method. A Gaussian distribution probability density function is constructed to quantify the degree of dispersion of the measured values ​​from the theoretical prediction values, which serves as a basic indicator for evaluating the prediction accuracy of the data source.

[0084] The Shapley value decomposition algorithm is introduced to inversely distribute the total prediction error to each data source node that participated in the previous round of power adjustability assessment. By arranging and combining different subsets of data sources to change the marginal contribution of the model output, the responsibility share of each data source for the final prediction error is calculated.

[0085] The correction coefficient for the closed-loop control is defined as the normalized value of the data source responsibility share after Sigmoid mapping. The larger the error responsibility, the smaller the correction coefficient.

[0086] The calculated control closed-loop verification correction coefficient is multiplied element-wise with the preliminary dynamic edge weights generated in step S2.2 to achieve dynamic iterative updates of the weights, reduce the connection strength of high-error responsibility data sources in the graph, and increase the confidence weight of low-error responsibility data sources.

[0087] By using the aforementioned reverse tracing verification and weight iterative update processing methods, historical control execution deviations are transformed into quantitative correction indicators of data source credibility, achieving adaptive optimization of the edge weights of the dynamic semantic credibility graph, and significantly improving the prediction stability and decision reliability of the model under complex working conditions.

[0088] For example, with a control cycle of 5 minutes, the maximum deviation between the measured and theoretically predicted active power of the inverter in the previous cycle was 2.5 kW. Through Shapley value decomposition, the error responsibility value for the meteorological data source was 0.6, for the inverter's local data source it was 0.1, and for the load data source it was 0.3. The calculated correction coefficients for the meteorological data source were approximately 0.35, for the inverter data source it was approximately 0.47, and for the load data source it was approximately 0.42. Multiplying these coefficients by the preliminary dynamic edge weights, the weight of the edge related to the meteorological data was significantly reduced, while the weight of the edge related to the inverter's local data remained relatively high. This effectively suppressed the interference of meteorological forecast fluctuations on the current control decision, reducing the root mean square error of the next round of power adjustability assessment results by approximately 15%.

[0089] S2.4: Based on the updated dynamic edge weights and predefined three-dimensional score fusion rules, normalized weighted fusion processing is performed on topological coupling degree, temporal compatibility degree and control closed-loop verification degree to generate a comprehensive edge weight matrix that finally represents the semantic consistency strength of multi-source data under the current working condition. The comprehensive edge weight matrix is ​​then assigned to the connection edges of the dynamic semantic credibility graph to complete the graph construction.

[0090] The system receives the preliminary dynamic edge weight matrix updated in step S2.3. This matrix contains the original score data for three dimensions: topological coupling, temporal compatibility, and control loop verification. Min-Max normalization is performed on the scores of each dimension to eliminate the influence of dimensional differences on the fusion result, mapping the values ​​of each dimension to the [0,1] interval to generate a standardized score vector. Based on the current operating condition of the distribution network, the corresponding three-dimensional score fusion rules are called from the preset strategy library to determine the dynamic weight coefficients of topological coupling, temporal compatibility, and control loop verification. Under steady-state operating conditions, a higher weight is assigned to topological coupling to emphasize physical connection constraints; under transient disturbance operating conditions, the weights of temporal compatibility and control loop verification are increased to strengthen real-time response consistency. The weighted geometric average algorithm is used to calculate the comprehensive edge weights, and the comprehensive edge weight value representing the semantic consistency strength of multi-source data is generated using the following formula: in, To incorporate edge weights, , , These represent the normalized topological coupling degree, temporal compatibility, and control loop validation degree, respectively. , , These are the dynamic weight coefficients for the corresponding dimensions. The calculated comprehensive edge weight values ​​are assigned to the connecting edges between corresponding nodes in the dynamic semantic credibility graph, completing the final construction of the graph weights. Through a weighted geometric average fusion processing method, the multi-dimensional dynamic scores from the previous step are transformed into a comprehensive edge weight matrix representing the semantic consistency strength of multi-source data under the current working condition. This achieves quantitative suppression of data source conflicts and redundancy, improving the stability and interpretability of the graph structure.

[0091] For example, assuming the current operating condition is a scenario of abrupt changes in irradiance caused by rapid cloud movement, the dynamic weighting coefficients are configured as α=0.3, β=0.4, and γ=0.3. The normalized score of a certain inverter node and the weather station node is: W top =0.85, W tim =0.60, W fb =0.75. Substituting into the formula: Wcom=(0.85) 0.3 0.60 0.4 0.75 0.3 ) 1 ≈0.732. This comprehensive weight is significantly lower than the static weight that relies solely on topological relationships, accurately reflecting the decrease in confidence caused by temporal asynchrony, and effectively avoiding interference from erroneous data sources on the power adjustability model.

[0092] S2.5: Perform connectivity verification and abnormal edge pruning on the constructed dynamic semantic credibility graph with dynamic edge weights to eliminate false high-confidence connections caused by transient communication interference. Finally, output a dynamic semantic credibility graph with a stable topology as the direct input object for capability-oriented feature distillation operation.

[0093] Receive the dynamic semantic credibility graph with dynamic edge weights generated by S2.4, extract the connection edge weight values ​​of all node pairs in the graph and the corresponding topological path information, and construct the set of edges to be verified.

[0094] For each connecting edge in the edge set, calculate the deviation of its weight from the average weight within adjacent historical time windows. Use the sliding window statistical method to obtain the local mean and standard deviation of the weight as the basic statistical measure for judging instantaneous anomalies.

[0095] An anomaly detection algorithm based on Mahalanobis distance is adopted. The edge weight vector at the current moment is mapped to a multidimensional statistical space, and the Mahalanobis distance between it and the center of the normal distribution is calculated. Outlier edges that exceed the preset confidence interval are identified. The calculated Mahalanobis distance is thresholded. If it is greater than the preset anomaly threshold θ, the connection edge is marked as a suspected false high confidence connection and included in the list to be pruned.

[0096] For the connecting edges in the pruning list, further connectivity checks are performed to verify whether the graph still maintains strong connectivity or satisfies the minimum spanning tree constraint after removing the edge, thus preventing graph structure breakage due to erroneous pruning.

[0097] If removing a suspected abnormal edge causes the graph to split into disconnected subgraphs, then the edge is retained but its weight is reduced to a safe lower bound; otherwise, a pruning operation is performed to permanently remove the edge from the graph.

[0098] The pruned graph is normalized, and the relative weight ratio of the remaining edges is recalculated to ensure that the overall semantic consistency score of the graph is distributed within a reasonable range, and a dynamic semantic credibility graph with a stable topological structure is output.

[0099] Through the above connectivity verification and abnormal edge pruning, the noisy dynamic graph generated in the previous step is transformed into a structurally robust and pure graph without false connections, thus achieving the technical effect of eliminating the impact of transient communication interference on feature distillation decision-making.

[0100] For example, the sliding window size is set to N=10 sampling periods, and the anomaly threshold θ=3.5. At a certain time t, the edge weight between the weather station and the inverter suddenly increases to 0.95, while the historical mean μ=0.6 and standard deviation σ=0.1. The Mahalanobis distance D is calculated. t =4.2, exceeding the preset anomaly threshold, and marked as an anomaly. Connectivity check revealed that the graph remained connected after removing the edge, so pruning was performed. After pruning, the contribution of this data source to feature distillation was suppressed, avoiding model misjudgments caused by instantaneous data jumps, and significantly improving the stability of the power adjustability assessment.

[0101] Step S3: Utilize the current credibility score of each node in the dynamic semantic credibility graph to perform capability-guided feature distillation. Map high-credibility sources to the capability boundary anchoring layer to extract the voltage reactive power sensitivity slope; map medium-credibility sources to the trend guidance layer to extract the irradiance attenuation gradient; and map low-credibility but high-conflict sources to the operating condition switching detection layer to extract cross-domain correlation patterns, generating a hierarchical distillation feature subset. Specifically, this includes: S3.1: Based on the node credibility score output by the dynamic semantic credibility graph, a threshold screening process is performed on the high credibility data source to generate a high credibility source set. The high credibility source set is then mapped to the capability boundary anchoring layer. The voltage and reactive power sensitivity slope calculation algorithm is used to perform differential transformation processing on the grid connection point voltage and reactive power data to obtain the voltage and reactive power sensitivity slope characteristic parameter that characterizes the hard constraint characteristics of the inverter.

[0102] It should be noted that capability-oriented feature distillation refers to a hierarchical feature extraction process that uses the current credibility score of each data source node in the dynamic semantic credibility graph to divide multi-source heterogeneous data into layers according to their credibility levels. This allows data sources at different credibility levels to enter their respective feature extraction layers, thereby extracting physical hard constraint features from high credibility data, trend evolution features from medium credibility data, and working condition switching detection features from low credibility and high conflict data.

[0103] Receive node credibility scores and high-credibility source raw data streams output by the dynamic semantic credibility map, including grid connection point voltage time series and reactive power response series.

[0104] A threshold filtering process is performed on the set of high-confidence sources. The confidence threshold is set to 0.85. Node data with scores below the threshold are removed, and local measurement data of inverters with high confidence are retained as the input basis for the capability boundary anchoring layer.

[0105] The selected high-confidence source data is mapped to the capacity boundary anchoring layer, and the instantaneous voltage value V(t) and reactive power output value Q(t) at the grid connection point are extracted to construct a voltage-reactive power coupled dataset.

[0106] The central difference method is used to perform differential transformation on the voltage-reactive power coupling data, and the ratio of voltage change rate to reactive power change rate is calculated to quantify the voltage support sensitivity of the inverter at a specific operating point.

[0107] The voltage reactive sensitivity slope K is calculated using the following formula. VQ : in The slope of voltage reactive power sensitivity. and They are respectively Before and after time Reactive power value within the time window and This represents the grid connection point voltage value at the corresponding time. The value is set to 10ms to match the inverter control cycle.

[0108] For the calculated K VQ The sequence is processed by moving average filtering with a window length of 5 control cycles to suppress the interference of high-frequency measurement noise on sensitivity evaluation and extract steady-state sensitivity characteristic parameters that characterize the hard constraint characteristics of the inverter.

[0109] The filtered voltage reactive power sensitivity slope characteristic parameters are encapsulated into a structured vector with a high-confidence source identifier, which serves as the final output of the capability boundary anchoring layer.

[0110] By using differential transformation and filtering, the high-reliability raw measurement data from the previous step is transformed into voltage and reactive power sensitivity slope characteristic parameters that characterize the physical regulation limit of the inverter. This achieves accurate anchoring of the hard boundary of active / reactive power regulation of the photovoltaic cluster and significantly improves the prediction stability of the model in voltage over-limit risk scenarios.

[0111] For example, an inverter node with a confidence score of 0.92 in a distributed photovoltaic cluster is selected, and its grid-connected voltage and reactive power data are collected. A time window Δt is set to 10ms. The voltage is read as 220.5V 10ms before time t and 220.8V 10ms after time t; the corresponding reactive power is -5kVar and -4.2kVar, respectively. Substituting these values ​​into the formula, K is calculated. VQ The value is 26.67 kVar / kV. A 5-point moving average was applied to the calculation results over 100 consecutive cycles, yielding a smoothed sensitivity slope of 26.5 kVar / kV. This parameter accurately reflects the inverter's voltage support capability under current operating conditions. As a hard constraint input to the subsequent optimization model, it effectively avoids the problem of exceeding regulation limits due to data conflicts.

[0112] S3.2: Based on the node credibility score output by the dynamic semantic credibility graph, interval matching processing is performed on the medium credibility data source to generate a medium credibility source set, and the medium credibility source set is mapped to the trend guidance layer. The time series gradient analysis algorithm is used to perform sliding window difference processing on the meteorological irradiance data to obtain the irradiance attenuation gradient characteristic parameters that characterize the direction of power evolution.

[0113] The node credibility scores and medium-credibility meteorological data source set output by the dynamic semantic credibility graph are received as input objects for feature extraction of the trend guidance layer.

[0114] For the meteorological irradiance time series data in the medium confidence source set, a sliding window truncation process is performed, with the window length set to the short-term prediction time domain and the step size set to the data acquisition period, to construct a local time series analysis segment.

[0115] The first derivative of the irradiance data within the window is estimated using the central difference algorithm. The rate of change of irradiance at adjacent time points is calculated to eliminate the interference of measurement noise on the gradient direction and obtain the characteristic parameters of the irradiance attenuation gradient.

[0116] The calculated instantaneous gradient value is compared with the historical baseline gradient for the same period to identify the acceleration or deceleration trend of irradiance change and generate a state label characterizing the direction of power evolution.

[0117] The above gradient calculation is performed on all meteorological nodes within the medium confidence source set to form a multidimensional irradiance attenuation gradient feature vector, which reflects the spatiotemporal evolution of regional light resources.

[0118] By using time-series gradient analysis, the medium-confidence meteorological data from the previous step is transformed into irradiance attenuation gradient characteristic parameters that characterize the direction of power evolution. This enables precise guidance of photovoltaic power output trends and improves the predictive foresight of the model under conditions of rapid cloud movement.

[0119] For example, the sliding window length is set to 5 minutes, and the step size is 15 seconds. When a weather station node is detected to have a confidence score of 0.75, which falls within the medium confidence range, the irradiance sequence of that node over the past 5 minutes is extracted. Assuming the irradiance at time t is 800 W / m², the previous time is 820 W / m², and the next time is 770 W / m², with a sampling interval of 15 seconds, the irradiance is calculated to show a rapid decreasing trend after central difference analysis, with a change rate of approximately -1.67 W / (m²·s). This negative value indicates that the irradiance is rapidly decaying, and the system marks it as a "power decreasing trend" and assigns it a corresponding trend-guiding weight. Compared to the traditional mean filtering method, this gradient feature can capture the cloud cover effect 30 seconds earlier, significantly improving the dynamic response speed of power adjustability assessment.

[0120] S3.3: Based on the node credibility score and conflict identifier output by the dynamic semantic credibility graph, perform abnormal pattern recognition processing on low credibility and high conflict data sources to generate a set of operating condition switching detection sources. The set of operating condition switching detection sources is then mapped to the operating condition switching detection layer. The cross-domain correlation mining algorithm is used to calculate the mutual information entropy of the load current harmonic components and the inverter IGBT junction temperature data to obtain cross-domain correlation mode feature parameters that characterize the sudden change in system state.

[0121] It should be noted that the operating condition switching detection layer is a feature extraction layer specifically designed for handling low-confidence and high-conflict data sources in the capability-oriented feature distillation operation. Its function is to extract mutual information entropy features from load current harmonic components and inverter IGBT junction temperature data through cross-domain correlation mining algorithms to identify whether the system has experienced an operating condition switching event, and to transform data conflicts from noise into semantic signals of operating condition evolution.

[0122] Receive node credibility scores and conflict identifiers from the dynamic semantic credibility graph output, filter out low credibility and high conflict data sources with credibility below a preset threshold and significant semantic conflicts with other data sources, and generate a set of working condition switching detection sources.

[0123] High-frequency harmonic components are extracted from the load current data of the distribution network feeder in the set of operating condition switching detection sources. The time-domain current signal is converted into frequency-domain spectrum using fast Fourier transform. The total distortion rate of the 3rd to 19th odd harmonics is calculated as an electrical characteristic index reflecting the sudden change in local impedance of the power grid and the change in load characteristics.

[0124] Simultaneously acquire real-time monitoring data of IGBT junction temperature inside the photovoltaic inverter corresponding to low-reliability and high-conflict data sources, perform noise reduction and smoothing processing on the junction temperature time series to eliminate measurement noise interference and retain the temperature change trend reflecting the thermal accumulation effect of power devices.

[0125] The cross-domain correlation mining algorithm is used to calculate the mutual information entropy between the total harmonic distortion rate sequence of the load current and the junction temperature change sequence of the IGBT, quantify the degree of information sharing between the two heterogeneous physical quantities under nonlinear coupling, and identify potential state change modes of the system.

[0126] When the calculated mutual information entropy value exceeds the dynamically set mutation detection threshold, it is determined that there is a system operating condition switching event caused by severe load fluctuation or power grid failure at the current moment, and the mutual information entropy value and its corresponding time position are marked as cross-domain correlation mode feature parameters.

[0127] By using cross-domain association mining, the low-confidence conflict data from the previous step is transformed into cross-domain association pattern feature parameters that characterize the sudden changes in system state, thereby achieving the expected technical effect of transforming data conflicts from noise into semantic information about the evolution of operating conditions.

[0128] For example, a confidence threshold of 0.4 is set to filter out a data source from a feeder terminal. The load current collected from this source is subjected to an FFT transformation with a window length of 20ms, resulting in a total distortion rate of 5.2% for the 3rd to 19th harmonics. Inverter IGBT junction temperature data is simultaneously read and filtered using a moving average to obtain a temperature change sequence. A histogram is constructed to statistically analyze the joint probability distribution with 10 bins, yielding a mutual information entropy value of 0.85 bits. A sudden change detection threshold of 0.6 bits is set; since 0.85 is greater than 0.6, a change in operating condition is determined. The mutual information entropy value of 0.85 and the timestamp are output as cross-domain correlation mode feature parameters, triggering the model to switch from steady-state assessment to transient response assessment mode, significantly improving the response speed and accuracy of power adjustability assessment under sudden load disturbance scenarios.

[0129] S3.4: Based on the voltage reactive power sensitivity slope feature parameter, irradiation attenuation gradient feature parameter and cross-domain correlation mode feature parameter, perform multi-dimensional feature vector concatenation and label attachment processing to generate a hierarchical distillation feature subset, where each feature vector is accompanied by a semantic source tag indicating its source level and credibility path, which serves as the standardized input object for subsequent graph neural networks.

[0130] The voltage reactive sensitivity slope characteristic parameters, irradiance attenuation gradient characteristic parameters, and cross-domain correlation mode characteristic parameters output from steps S3.1 to S3.3 are received as the original input data for concatenating multi-dimensional feature vectors.

[0131] Dimensional standardization is performed on the three types of heterogeneous feature parameters, and the Z-Score normalization algorithm is used to eliminate the difference in dimensions, mapping feature values ​​with different physical meanings to a unified numerical range to ensure the weight balance in the subsequent splicing process.

[0132] A hierarchical semantic source tracing label structure is constructed, and a unique hierarchical identifier is assigned to each feature parameter, which is marked as the capability boundary anchoring layer ID, trend guidance layer ID, and working condition switching detection layer ID, respectively, to establish the source attribution of the feature.

[0133] Extract the real-time credibility score of the corresponding data source node in the dynamic semantic credibility graph, encode it into a confidence weight vector, and bind it to the corresponding feature parameters at the element level to form a weighted feature tuple.

[0134] Using a serialization splicing protocol, high-confidence hard constraint features, medium-confidence trend features, and low-confidence operating condition switching features are arranged in sequence according to a preset feature priority order to generate a comprehensive feature vector of a unified dimension.

[0135] A global timestamp and topological location index are appended to the head of the comprehensive feature vector, and a semantic source tracing metadata block consisting of a hierarchical identifier and a confidence weight vector is embedded at the tail to complete the label attachment process.

[0136] By using the above-mentioned multi-dimensional feature vector concatenation and label attachment processing method, the heterogeneous features extracted in the previous step are transformed into a hierarchical distillation feature subset with a unified structure and interpretability, so as to realize the efficient learning and accurate source tracing of the semantic collaborative relationship of features by the subsequent graph neural network.

[0137] For example, data from inverter PV-01 in a distributed photovoltaic cluster, along with its associated meteorological station and feeder terminal, are selected. The voltage reactive power sensitivity slope is 0.05 pu / Mvar, which is 0.8 after normalization; the irradiance attenuation gradient is -200 W / m² / min, which is -0.6 after normalization; and the cross-domain correlation mutual information entropy is 0.9 bits. The corresponding confidence scores are 0.95, 0.75, and 0.4, respectively. A feature vector of [0.8, -0.6, 0.9] is constructed, with the source tag {Layer: [Anchor, Trend, Switch], Confidence: [0.95, 0.75, 0.4]} attached. This structured subset is directly input into the graph neural network, significantly improving the model's response speed to sudden cloud cover conditions, reducing the delay in issuing control commands to the millisecond level, and effectively avoiding erroneous adjustments caused by data conflicts.

[0138] like Figure 3As shown, step S4: Input a graph neural network with a fixed structure but whose edge weights are reconfigured in real time by a dynamic semantic credibility graph using a hierarchical distillation feature subset as input. Train the network to learn semantic cooperative relationships between features rather than numerical fitting relationships, and output the photovoltaic cluster power adjustability assessment result with semantic source tracing labels. Specifically, this includes: S4.1: Construct a static graph neural network topology based on the hierarchical distillation feature subset, map the features of the capability boundary anchoring layer, the trend guidance layer, and the working condition switching detection layer to the node attribute vectors of the graph neural network, initialize the adjacency matrix with a fixed structure to define the potential connection relationship between nodes, and generate a static graph neural network model with initial node attribute vectors.

[0139] It should be noted that the semantic source tag is a type of structured metadata attached to the power adjustability assessment results. It includes the identifier of the dominant data source combination, the confidence weight of each data source, and the complete backtracking path information from the original data source to the output results. It is used to indicate the data sources on which the assessment results depend and the credibility support links, thereby achieving the interpretability of the assessment results.

[0140] The system receives a subset of features from a hierarchical distillation process, including feature vectors and semantic source labels from the capacity boundary anchoring layer, trend guidance layer, and operating condition switching detection layer. A static graph neural network topology is constructed, defining node sets corresponding to the three feature subsets and initializing node attribute vectors with the feature parameters extracted from each layer. The adjacency matrix dimension is set to equal the total number of nodes, with an initial value of an identity matrix to represent potential fully connected relationships between nodes. The voltage reactive power sensitivity slope of the capacity boundary anchoring layer is mapped to a high-confidence node attribute and assigned a high initial state weight. The irradiance attenuation gradient of the trend guidance layer is mapped to a medium-confidence node attribute and assigned a medium initial state weight. The cross-domain correlation pattern of the operating condition switching detection layer is mapped to a low-confidence but high-sensitivity node attribute and assigned a dynamic gating initial weight. Tensor concatenation operations are used to integrate the three feature vectors into a unified node feature matrix, ensuring feature dimension alignment. A fixed-structure adjacency matrix is ​​initialized using a sparse matrix storage format to reduce memory usage and improve subsequent computational efficiency. A static graph neural network model with initial node attribute vectors and fixed topological connections is generated as the baseline framework for dynamic weight reconfiguration. The above processing method transforms the hierarchical distillation features from the previous step into structured graph data input, realizing the mapping from discrete features to topological association models, and providing a stable computational skeleton for subsequent message passing based on dynamic semantic credibility.

[0141] For example, for a photovoltaic cluster in a certain region, the capacity boundary anchoring layer extracts the voltage and reactive power sensitivity slopes of 10 inverters, forming a 10-dimensional vector; the trend guidance layer extracts the irradiance attenuation gradients of 5 weather stations, forming a 5-dimensional vector; and the operating condition switching detection layer extracts the cross-domain association patterns of 3 feeder terminals, forming a 3-dimensional vector. The total number of nodes is 18. An 18x18 unit adjacency matrix is ​​constructed. The 10-dimensional sensitivity vector is filled into the attributes of the first 10 nodes, the 5-dimensional gradient vector is filled into the middle 5 nodes, and the 3-dimensional association pattern is filled into the last 3 nodes. The node attributes are Z-score standardized, with a mean of 0 and a variance of 1. The static graph model initialization is completed, awaiting the injection of dynamic edge weights in step S4.2.

[0142] S4.2: The adjacency matrix of the static graph neural network model is reconfigured in real time using the dynamic edge weights in the dynamic semantic credibility graph. The three-dimensional scores of topological coupling degree, temporal compatibility degree and control closed-loop verification degree are converted into message passing weight coefficients of the graph neural network, generating a reconfigurable graph neural network instance with dynamic message passing weight coefficients.

[0143] It receives the initial adjacency matrix of the static graph neural network model and the three-dimensional score data output by the dynamic semantic credibility graph, including topological coupling degree, temporal compatibility degree and control loop verification degree.

[0144] Normalization mapping is performed on the three-dimensional score data to linearly transform the score values ​​of each dimension to the [0,1] interval, eliminating the interference of dimensional differences on weight calculation and generating standardized semantic consistency coefficients.

[0145] A message passing weight calculation model is constructed based on standardized coefficients. A weighted product fusion strategy is used to calculate the dynamic weight coefficients of edges between nodes. The calculated dynamic message passing weight coefficients are filled into the corresponding positions in the static adjacency matrix, replacing the original fixed connection weights, thereby realizing the dynamic reconfiguration of the adjacency matrix.

[0146] The reconfigured adjacency matrix is ​​sparsified by removing weak connections with weight coefficients below a preset threshold, thereby reducing the computational complexity of the graph neural network and suppressing noise propagation.

[0147] Generate a reconfigurable graph neural network instance with dynamic message passing weight coefficients. This instance retains a fixed node attribute structure but has the ability to adjust the information aggregation path in real time as the operating conditions change.

[0148] By using dynamic weight mapping and matrix reconfiguration, the static topology structure from the previous step is transformed into a reconfigurable network instance that adapts to real-time semantic consistency, thereby achieving the expected technical effect of dynamic perception and adaptive aggregation of multi-source data conflicts and complementary relationships in graph neural networks.

[0149] For example, the adjustment factors are set to α=0.4, β=0.3, and γ=0.3. At a certain moment, the topological coupling degree between the inverter node and the meteorological node is 0.85, the temporal compatibility degree is 0.72, and the control closed-loop validation degree is 0.90. After weighted product fusion calculation, the dynamic weight coefficient is approximately 0.824. If the preset sparsity threshold is 0.1, the connection is retained and participates in subsequent message passing as a strong semantic association. Compared with the fixed weight scheme, this dynamic reconfiguration mechanism significantly reduces the standard deviation of the prediction error of the model under sudden cloud cover conditions, improving the robustness of the power adjustability assessment.

[0150] S4.3: Perform multi-layer semantic collaborative aggregation operation based on reconfigurable graph neural network instances. The hard constraint features of high confidence sources and the trend guidance features of medium confidence sources are fused through a weighted message passing mechanism. At the same time, the cross-domain association pattern of low confidence sources is used to activate nonlinear gating units to suppress noise propagation, thereby generating a hidden layer state representation containing multi-dimensional feature semantic collaborative relationships.

[0151] Receive a reconfigurable graph neural network instance with dynamic message passing weight coefficients and a hierarchical distillation feature subset, map the feature vectors of the capability boundary anchoring layer, trend guidance layer and operating condition switching detection layer to graph node attributes respectively, and initialize the hidden layer state matrix.

[0152] The first-level semantic collaborative aggregation operation is performed. Based on the topological coupling degree component in the dynamic edge weight, the voltage reactive power sensitivity slope characteristics of the high-confidence source are weighted and summed to generate a basic state representation that reflects the physical hard constraints of the equipment.

[0153] By introducing the irradiance attenuation gradient characteristics of medium-confidence sources, adjusting the message transmission intensity through temporal compatibility weights, and using a linear transformation matrix to fuse meteorological trend information into the basic state representation, the prediction bias of power evolution direction is corrected.

[0154] For cross-domain association patterns with low confidence but high conflict, the mutual information entropy value between the pattern and the current hidden layer state is calculated and used as the activation signal of the nonlinear gating unit. The gating coefficient is dynamically adjusted to suppress noise propagation and retain the working condition switching indication information.

[0155] The cross-domain association pattern features filtered by the gating mechanism are superimposed onto the hidden layer state, and a multilayer perceptron is used for nonlinear mapping to generate the final hidden layer state representation containing multidimensional feature semantic collaborative relationships.

[0156] By using weighted message passing and nonlinear gating collaborative processing, the reconfigurable graph network input from the previous step is transformed into a high-dimensional hidden layer state representation containing semantic collaborative relationships. This achieves effective suppression of conflicts in multi-source heterogeneous data and deep fusion of complementary information, significantly improving the stability and robustness of the power adjustable capability assessment model under complex operating conditions.

[0157] For example, the graph neural network is set to contain two hidden layers, with 64 nodes in each layer. For a certain photovoltaic cluster node, the slope of the voltage reactive power sensitivity of the high-confidence source is 0.05 pu / Mvar, the irradiance attenuation gradient of the medium-confidence source is -200W / m² / h, and the cross-domain mutual information entropy of the low-confidence source is 0.8 bits. The topological coupling degree in the dynamic edge weights is 0.9, and the temporal compatibility degree is 0.7. After the first layer aggregation, the basic state vector element values ​​are updated to 0.045. After introducing meteorological trend correction, the state vector is adjusted to 0.042. The mutual information entropy of 0.8 triggers the gating coefficient of 0.9, effectively suppressing load harmonic noise interference. The final output hidden layer state representation accurately reflects the actual adjustment margin of the inverter under the current cloud cover conditions. Compared with the traditional weighted average method, the prediction error standard deviation is reduced by about 40%, significantly improving the reliability of control decisions.

[0158] S4.4: Perform semantic sourcing reverse tracing processing on the hidden layer state representation, identify the dominant data source combination and its confidence evolution path that each output dimension depends on based on the weight contribution on the message passing path, and generate semantic sourcing metadata tags containing the dominant data source combination identifier and confidence evolution path information.

[0159] The system receives the hidden layer state representation vector and the real-time edge weight matrix of the dynamic semantic credibility graph generated by S4.3 as the initial input data for semantic source tracing and reverse tracking.

[0160] For each output dimension node in the hidden layer state representation, based on the message passing mechanism of the graph neural network, the set of its predecessor neighbor nodes that it depends on in the aggregation operation is traversed in reverse.

[0161] The dynamic edge weight coefficients on each message transmission path are extracted. These coefficients are formed by fusing the three-dimensional scores of topological coupling degree, temporal compatibility degree, and control closed-loop verification degree, and represent the confidence strength of information flow.

[0162] The gradient contribution decomposition algorithm is used to calculate the absolute value of the partial derivative of each preceding neighbor node feature with respect to the current output dimension state, thereby quantifying the actual contribution ratio of each data source feature in the semantic collaboration process.

[0163] The dynamic edge weight coefficients on the path are multiplied element-wise with the gradient contribution to construct a weighted confidence propagation chain that traces back from the output layer to the original input layer.

[0164] The calculated cumulative confidence contribution value is normalized to eliminate numerical bias caused by different dimensional features, and a standardized semantic source tracing weight distribution vector is generated.

[0165] A confidence threshold filtering mechanism is set up to identify and mark the top three dominant data source combinations that contribute the most to the current output dimension, forming a list of dominant data source identifiers.

[0166] Record the complete backtracking path from the dominant data source to the output dimension, including the intermediate level nodes along the path and their corresponding dynamic edge weight evolution sequences, which constitute the confidence evolution path information.

[0167] The list of dominant data source identifiers and confidence evolution path information are structured and encapsulated to generate semantic source tracing metadata tags with timestamps and scene labels.

[0168] Through the aforementioned reverse tracing and weighted calculation processing method, the black-box hidden layer state of the previous step is transformed into interpretable metadata indicators with a clear physical meaning, thereby realizing the expected technical effect of making the power adjustable capability assessment results credible and transparent and visualizing the decision-making basis.

[0169] For example, for a 10kV distribution network area, when the output dimension is "active power upscaling margin of inverter A", the system reverse tracing reveals that it mainly relies on two sources: "local voltage of inverter A" and "meteorological station irradiance". The calculated gradient contribution of the inverter local voltage is 0.65, the gradient contribution of meteorological irradiance is 0.25, and the contribution of the feeder load is 0.10. Combining the dynamic edge weights (inverter-meteorological edge weight 0.9, inverter-load edge weight 0.4), the calculated cumulative confidence contribution of the inverter local voltage is 0.585, and the meteorological irradiance is 0.225. The system determines the dominant data source combination as {inverter local voltage, meteorological irradiance} and records the path weight evolution sequence [0.9, 0.85]. The generated semantic source tags clearly indicate that the margin assessment is mainly based on high-confidence local electrical measurements and medium-confidence meteorological trends, eliminating the interference of low-confidence load data and significantly improving the pertinence and reliability of control instructions.

[0170] S4.5: Combine the hidden layer state representation with semantic source traceability metadata tags to perform regression mapping decoding operation, project the high-dimensional semantic collaborative feature space onto the physical power adjustment domain, and output the photovoltaic cluster power adjustability assessment result with semantic source traceability tags, thus completing the final quantitative expression from semantic collaborative relationship to specific adjustability value.

[0171] It receives the hidden layer state representation vector generated by S4.3 and the semantic source metadata tag output by S4.4 as the input basis for the regression mapping decoding operation.

[0172] A multilayer perceptron decoder network is constructed, and the hidden layer state representation vector is input into the first fully connected layer. The high-dimensional semantic collaborative features are mapped to the intermediate hidden space through linear transformation. The ReLU activation function is selected to introduce non-linear expressive power.

[0173] Dropout regularization is applied to the intermediate latent space features, randomly discarding some neuron connections to suppress overfitting and enhance the model's generalization ability under varying conditions.

[0174] The regularized features are input into the second fully connected layer. By combining the confidence evolution path information in the semantic source traceability metadata tags, the weight bias of each output node is dynamically adjusted to achieve adaptive decoding based on the confidence of the data source.

[0175] The mean square error loss function is used to constrain the deviation between the decoded output and the true value of the physical power adjustment domain. The decoder network parameters are updated through the backpropagation algorithm to optimize the mapping accuracy from semantic space to physical space.

[0176] Physical constraint projection processing is performed on the decoder output. Based on the inverter's rated capacity, voltage safety range, and equipment response rate limit, the predicted adjustable capability value is truncated and smoothed to ensure that the output results conform to the physical laws of power grid operation.

[0177] Generate photovoltaic cluster power adjustability assessment results with semantic traceability labels, including active power up / down margin, reactive power compensation range and corresponding confidence level labels.

[0178] By using regression mapping decoding and physical constraint projection processing, the high-dimensional semantic collaborative features of the previous step are transformed into specific numerical indicators of the adjustable power capability of photovoltaic clusters. This achieves the final quantitative expression from abstract semantic relationships to specific physical control quantities, providing accurate and interpretable boundary conditions for subsequent collaborative optimization.

[0179] For example, the hidden layer state is represented as a 128-dimensional vector, and the semantic source label indicates a high-confidence source weight of 0.85. The first fully connected layer of the decoder maps the 128-dimensional input to a 64-dimensional latent space, which is then activated by ReLU and regularized with a dropout rate of 0.2. The second fully connected layer adjusts the bias based on the confidence weight, and outputs a 3-dimensional vector corresponding to the active power up-adjustment margin, active power down-adjustment margin, and reactive power adjustment range, respectively. After physical constraint projection, if the predicted active power up-adjustment margin exceeds the inverter's remaining capacity, it is truncated to the maximum allowable value. The final output active power up-adjustment margin is 5.2kW, with a confidence level of A, significantly improving the physical feasibility of the control command.

[0180] Step S5: Based on the distribution network operating status, dispatch instructions, and time-of-use pricing signals, and combined with the photovoltaic cluster power adjustability assessment results with semantic traceability tags, collaboratively calculate and generate active power allocation schemes and reactive power voltage control strategies for each inverter, forming a joint optimization control instruction set. Specifically, this includes: S5.1: Obtain the photovoltaic cluster power adjustability assessment results, distribution network operation status data, dispatch instruction information and time-of-use electricity price signals with semantic traceability labels, perform spatiotemporal alignment and standardization processing on multi-source input information, and generate a collaborative optimization basic dataset containing the voltage amplitude of each node, line load rate, adjustment margin boundary and electricity price time period identifier.

[0181] The system receives photovoltaic cluster power adjustability assessment results with semantic traceability tags and parses the active / reactive power adjustment margin boundaries and confidence path information contained therein. It simultaneously acquires real-time operating status data uploaded from the distribution network SCADA system, including voltage amplitude at each node, branch current, and transformer load rate. It reads day-ahead / intra-day dispatch instructions issued by the dispatch master station and extracts the total output limit threshold and voltage control target range. It accesses time-of-use pricing signal data and identifies the current time period and corresponding economic incentive coefficient.

[0182] A unified spatiotemporal reference alignment process is performed on multi-source input information. A linear interpolation algorithm is used to resample meteorological data and electrical measurement data with different sampling frequencies to a unified timestamp, eliminating feature misalignment caused by time-series asynchrony. Based on the distribution network topology connection matrix, a power flow tracing algorithm is used to map load data collected by feeder terminals to corresponding photovoltaic grid-connected points, establishing a spatial correlation mapping relationship to ensure the consistency of physical location.

[0183] Data standardization and dimensionless processing are performed. For continuous variables such as voltage amplitude and power, Z-Score standardization is used to eliminate dimensional differences. For discrete Boolean signals such as AGC response ready state, they are directly mapped to binary values ​​of 0 or 1. For electricity price time period identifiers, one-hot encoding is used to convert them into multi-dimensional vectors, preserving their category characteristics without introducing magnitude order bias.

[0184] A foundational dataset for collaborative optimization is constructed. Standardized node voltage amplitudes, line load rates, adjustment margin boundaries verified by semantic source tracing labels, and electricity price time period identifier vectors are structurally concatenated. Based on the confidence weights in the semantic source tracing labels, anomaly markers are added to low-confidence data items for robustness handling in subsequent optimization models. A structured data table containing multidimensional state variables and constraints is generated as the direct input to the collaborative optimization model.

[0185] Through the above spatiotemporal alignment, standardized mapping, and structured assembly processing methods, heterogeneous multi-source input information is transformed into a collaborative optimization basic dataset with unified dimensions and unified benchmarks, achieving semantic consistency and numerical standardization at the data level, and providing high-quality data support for subsequent multi-objective collaborative optimization.

[0186] For example, a data acquisition cycle of 5 seconds is set to process 20 photovoltaic inverter nodes on a 10kV feeder. The active power regulation margin reported by the inverters is obtained as ±5kW, with a semantic confidence level of 0.95; the SCADA system node voltage is 10.3kV, and the load factor is 65%. Using Z-Score normalization, assuming a voltage mean μ = 10.0kV and a standard deviation σ = 0.2kV, the normalized voltage value is (10.3kV). (10.0) / 0.2 = 1.5. Electricity price periods are identified as "peak" and encoded as [0,1,0]. All features are concatenated into a vector of length N, with low-confidence data marked as 1. The final generated collaborative optimization dataset contains 20 sets of standardized state vectors, each with a semantic source index, ensuring the reliability of data sources in subsequent optimization processes and significantly improving the accuracy and robustness of control command generation.

[0187] S5.2: Based on the collaborative optimization dataset, a collaborative optimization model is constructed with multiple objective functions, namely minimizing network loss, maximizing photovoltaic absorption, and smoothing voltage fluctuations. The multi-objective function is transformed into a single-objective fitness function using the weighted summation method, generating a global optimization objective function expression to guide the search for inverter action strategies.

[0188] Receive the basic dataset for collaborative optimization and extract node voltage amplitude, line load rate, and regulation margin boundary as input constraints.

[0189] A target function for minimizing network losses is constructed based on the sum of the products of the squares of the currents and the resistances of each branch; a target function for maximizing photovoltaic absorption is constructed with the negative value of the difference between the actual output power of the inverter and the theoretical maximum available power as the optimization direction; and a target function for smoothing voltage fluctuations is constructed based on the sum of the squares of the voltage deviations of each node from the rated value.

[0190] The weighted summation method is introduced to transform the multi-objective function into a single-objective fitness function, and the weight coefficients α, β and γ are set to correspond to the loss, absorption and voltage indicators, respectively.

[0191] Generate a global optimization objective function expression in the form F=αf1+β(-f2)+γf3, ensuring dimensional consistency and performing normalization.

[0192] By using a weighted fusion processing method, multidimensional physical constraints are transformed into scalar fitness values, thereby achieving a global optimal strategy search orientation under multi-objective conflict.

[0193] S5.3: Using the global optimization objective function expression and the adjustment margin boundary as a hard constraint, the improved particle swarm optimization algorithm is used to iteratively search for optimization in the solution space, and calculate the initial active power setpoint sequence and initial reactive power compensation sequence of each inverter that meet the grid safety operation constraints and have the best economic benefits.

[0194] Receive the global optimization objective function expression generated by S5.2 and the cooperative optimization basic dataset output by S5.1, and extract the adjustment margin boundary contained therein as hard constraints for the particle swarm search space.

[0195] Initialize the population parameters of the improved particle swarm optimization algorithm, set the number of particles to N, and each particle represents a multidimensional solution vector containing all inverter active power setpoints and reactive power compensations.

[0196] The upper and lower limits of the search in each dimension are determined based on the adjustment margin boundary. Initial particle positions and velocities are randomly generated within the feasible region to ensure that the initial solution satisfies the physical limit constraints of the device.

[0197] An fitness evaluation mechanism is constructed, which substitutes the current particle position into the global optimization objective function and calculates the weighted sum of the corresponding network loss, photovoltaic absorption capacity and voltage fluctuation index.

[0198] A dynamic inertia weight adjustment strategy is introduced, which linearly decreases the inertia weight coefficient according to the number of iterations to balance the search capability of the algorithm in the global exploration and local development stages.

[0199] An adaptive learning factor update rule is adopted to dynamically adjust the cognitive and social learning coefficients based on the Euclidean distance between the individual particle's historical best position and the group's global best position.

[0200] Perform velocity update calculations, combining inertia, individual cognition, and group social terms to calculate the flight velocity vector of the next generation of particles, preventing velocity divergence from causing search failure.

[0201] The updated particle positions are subjected to boundary absorption processing. If a particle exceeds the adjustment margin boundary, it is forcibly mapped to the nearest valid boundary point and the velocity of that dimension is reset to zero.

[0202] Compare the current particle fitness value with its individual historical best value. If the current value is better, update the individual extreme value. At the same time, compare all individual extreme values ​​to update the global optimal solution for the population.

[0203] Determine whether the maximum number of iterations or the fitness convergence threshold has been reached. If not, return to the velocity update step to continue iterating; otherwise, terminate the optimization process.

[0204] By improving the particle swarm optimization algorithm to iteratively search for solutions in the solution space, the multi-objective collaborative optimization problem is transformed into a sequence of initial active power setpoints and initial reactive power compensation values ​​for each inverter that satisfies grid security constraints and has the best economic benefits, thereby achieving a balance between the economy and security of control commands.

[0205] For example, the particle swarm size is set to 50, the maximum number of iterations is 100, and the inertia weight decreases linearly from 0.9 to 0.4. For a photovoltaic cluster with 20 inverters, each inverter's solution vector has a dimension of 2 (active and reactive power), for a total dimension of 40. Initially, 50 particles are randomly generated and substituted into the objective function to calculate the fitness. In the 30th iteration, the learning factors c1=1.5 and c2=2.5 are dynamically adjusted. After 100 iterations, the network loss corresponding to the global optimum is reduced to 12.5kW, the voltage deviation standard deviation is reduced to 0.02pu, and a determined active and reactive power command sequence is output.

[0206] S5.4: Based on the initial active power setpoint sequence and the initial reactive power compensation sequence, and combined with the AGC response ready state Boolean signal of each inverter, perform feasibility verification and dynamic correction, eliminate invalid command components that exceed the physical limits of the equipment or have mismatched response delays, and generate a corrected active power allocation scheme and a corrected reactive voltage control strategy that have been verified for feasibility.

[0207] Obtain the initial active power setpoint sequence and initial reactive power compensation sequence of each inverter output from step S5.3 as the candidate set of instructions to be verified. Read the AGC response ready state Boolean signal of each inverter generated in step S1.2 to construct the real-time availability state vector of the equipment. Compare the initial active power setpoint with the inverter's rated active power upper limit and the current maximum power point tracking value to identify over-limit instruction components that exceed the physical regulation boundary. Use the ramp rate limiting function to smooth the active power change to ensure that the power regulation rate does not exceed the maximum ramp rate allowed by the inverter hardware, preventing DC side voltage oscillation caused by sudden instruction changes. Based on the AGC response ready state Boolean signal, perform an instruction masking operation on inverters in the non-ready state, forcibly resetting their active and reactive power setpoints to zero or maintaining the steady-state value of the previous cycle, avoiding issuing invalid control instructions to offline or faulty equipment. The power factor angle corresponding to the reactive power compensation is calculated, and it is verified whether it falls within the PQ feasible region of the inverter's four-quadrant operating capability. Invalid reactive power commands that cause apparent power to exceed the rated capacity are eliminated. For command components that are corrected due to feasibility verification, the correction magnitude and correction reason label are recorded to form a command correction log. Through the above multi-dimensional physical constraint verification and state matching correction, the theoretical optimization solution is transformed into executable commands that conform to the actual operating capability of the equipment. A corrected active power allocation scheme and a corrected reactive voltage control strategy with feasibility verification are generated, significantly improving the success rate of control command issuance and the safety of equipment response.

[0208] S5.5: The modified active power allocation scheme and the modified reactive voltage control strategy are encapsulated and format converted according to the protocol, and timestamps, execution priorities and semantic traceability index information are added to generate a joint optimization control instruction set that can be directly issued to the coordination control device.

[0209] Step S6: Decompose the joint optimization control instruction set and distribute it to each photovoltaic inverter through the coordination control device for execution. Simultaneously, collect the actual operation trajectory of the inverters and compare it with the theoretical capability curve to generate the control closed-loop verification update quantity used for reverse correction of the control closed-loop verification value edge weights in the next round of dynamic semantic credibility graph. Specifically, this includes: S6.1: Based on the joint optimization control instruction set, the instruction data packet containing timestamp, execution priority and semantic source index information is parsed and processed by the protocol parsing and instruction decomposition algorithm to generate a distributed inverter control command sequence adapted to different communication protocol interfaces, which serves as the direct input object to drive the action of the underlying device.

[0210] It receives a joint optimization control instruction set data packet containing timestamps, execution priorities, and semantic source index information, which serves as the initial input for protocol parsing and instruction decomposition.

[0211] The header of the instruction data packet is validated and fields are extracted to parse out the target inverter ID, active power setpoint, reactive power compensation amount and expected execution time, and to build a list of metadata for instructions to be processed.

[0212] Query the device registry based on the target inverter ID to obtain the communication protocol type identifier of each inverter, including ModbusTCP, IEC104 or private JSON interface, and establish a protocol mapping matrix.

[0213] For inverters using the ModbusTCP protocol, the active and reactive power values ​​are converted into 16-bit integer format according to a predefined register address mapping table, and a CRC checksum is calculated and encapsulated into a standard Modbus write-hold register message.

[0214] For inverters using the IEC104 protocol, information object addresses and information element values ​​are constructed, the type identifier is set to single-point remote control or normalized measurement value, and an APDU application service data unit conforming to the IEC60870-5-104 standard is generated.

[0215] For inverters with private JSON interface types, control parameters are serialized into key-value pairs, timestamps and signature tokens are embedded, and a UTF-8 encoded JSON string payload is generated.

[0216] A unified scheduling layer frame header is added to each protocol message, including a global synchronization clock flag and an instruction priority field, forming a distributed inverter control command sequence adapted to the underlying communication interface.

[0217] By using protocol adaptation and format conversion, the abstract control commands generated by upper-layer collaborative optimization are transformed into execution messages that can be recognized by lower-layer devices, thereby achieving unified and precise driving of heterogeneous inverter clusters.

[0218] For example, for the inverter with ID "INV_001", its protocol type is ModbusTCP. The parsed active power setting is 50.5kW, and the reactive power setting is -10.2kVar. Looking up the table, the active power register address is 40001, and the reactive power register address is 40003. Multiplying 50.5kW by 10 converts it to the integer 505, which is hexadecimal 0x01F9; multiplying -10.2kVar by 10 converts it to -102, which is 0xFF9A in two's complement. A Modbus message is constructed: 01109C4100020401F9FF9A[CRC]. For the Sungrow inverter with ID INV_002, the protocol is IEC104. An Application Service Data Unit (ASDU) is constructed, with a type identifier of 45 (single command), an information object address of 0x0001, a switch status of closed, and a timestamp of the current milliseconds. A binary stream is generated and a start character 68H and a length field are added. The final output control command sequence contains messages in both of these different formats, ensuring that each inverter receives and executes actions under a unified time base, eliminating control delays caused by protocol differences.

[0219] S6.2: Based on the distributed inverter control command sequence, the real-time scheduling engine of the coordination control device performs time synchronization and concurrent issuance of the command sequence to trigger each photovoltaic inverter to perform active power regulation and reactive voltage compensation actions, and simultaneously start the high-precision data acquisition unit to record the actual action trajectory data stream of the inverter.

[0220] Receive the distributed inverter control command sequence after protocol parsing and instruction decomposition. This sequence includes the target active power setpoint, reactive power compensation amount, execution timestamp, and priority identifier for each inverter.

[0221] By utilizing the real-time scheduling engine within the coordination and control device, the current system global clock reference is read, and the execution timestamps in the control command sequence are aligned and verified at the microsecond level to eliminate timing jitter caused by communication delays and generate time-synchronized control command packets.

[0222] Based on the distribution network topology diagram, the electrical connection location and communication link level of each photovoltaic inverter on the feeder are identified, a concurrent task queue is constructed, and time-synchronized control command packets are mapped to the corresponding communication channel interface.

[0223] A multi-threaded concurrent transmission mechanism is initiated, and control command packets are simultaneously pushed to each photovoltaic inverter gateway through the IEC104 or ModbusTCP protocol stack to ensure that the time deviation of the command arriving at each node is controlled within the millisecond threshold range.

[0224] At the instant the command is issued, the synchronous sampling interrupt signal of the high-precision data acquisition unit is triggered, locking the instantaneous values ​​of the grid connection point voltage, current and the inverter's internal status register, and starting to record the actual action trajectory data stream.

[0225] Through the aforementioned timing synchronization and concurrent delivery processing, discrete control commands are transformed into coordinated action responses from physical devices, and high-time-resolution actual operating data is obtained, thereby achieving accuracy and traceability in the execution of control commands.

[0226] For example, the coordination and control unit receives a data packet containing instructions from 50 inverters, with a global clock reference of T0. The scheduling engine aligns the instruction timestamps, allowing a maximum jitter of 2ms. A concurrent queue is constructed, mapping the instructions to 5 communication subnets. Multi-threaded transmission is initiated, with all instructions sent within T0+5ms. The data acquisition unit samples at a frequency of 10kHz, recording voltage and current waveforms from T0 to T0+1s. If the communication delay of an inverter exceeds 10ms, it is marked as an abnormal node and logged. Finally, the actual motion trajectory data of 50 synchronized inverters is obtained for subsequent deviation analysis.

[0227] S6.3: Based on the actual action trajectory data stream of the inverter and the pre-stored theoretical capability curve dataset, the dynamic time warping algorithm is used to perform spatiotemporal alignment and deviation residual calculation on the actual response waveform and the theoretical expected waveform to generate an action trajectory deviation residual sequence that characterizes the device response accuracy within a single cycle.

[0228] It receives the actual operating trajectory data stream of the inverter and the pre-stored theoretical capability curve dataset as input objects for spatiotemporal alignment and deviation calculation.

[0229] The actual response waveform and the theoretical expected waveform are resampled on the time axis. A linear interpolation algorithm is used to map the non-uniformly sampled actual action data to a unified time reference grid, eliminating the sampling point misalignment caused by communication delay.

[0230] A dynamic time warping cost matrix is ​​constructed, and the Euclidean distance between the actual waveform sequence and the theoretical waveform sequence at each time step is calculated to form an initial local distance matrix, which quantifies the amplitude difference at a single point in time.

[0231] The dynamic programming algorithm is applied to search for the optimal matching path with the minimum cumulative distance, and the search window range is constrained to prevent pathological distortion, thus determining the best time alignment mapping relationship between the actual response and the theoretical curve.

[0232] Based on the optimal matching path, the theoretical capability curve is transformed by time warp to generate an aligned theoretical curve that is strictly synchronized with the actual action trajectory in the time dimension, thus eliminating the evaluation error caused by phase lag.

[0233] The point-by-point difference between the actual motion trajectory after alignment and the alignment theoretical curve is calculated to obtain the original residual sequence containing active power deviation and reactive voltage deviation.

[0234] The original residual sequence is processed by moving average filtering to remove high-frequency measurement noise interference and extract the smooth motion trajectory deviation residual sequence that characterizes the equipment response accuracy.

[0235] By using dynamic time warping and alignment and residual extraction processing, the asynchronous waveform data from the previous step is transformed into a motion trajectory deviation residual sequence that characterizes the device response accuracy within a single cycle, thus achieving a high-precision quantitative basis for the control closed-loop verification degree.

[0236] For example, the sampling frequency is set to 50Hz and the time window length is 2s. The actual active power sequence P act Contains 100 data points, theoretical sequence P theo It contains 100 data points. Construct a 100x100 cost matrix C, where each element C(i,j) is |P act (i)-P theo (j)|。 Calculate the cumulative distance matrix D using dynamic programming, D(i,j)=C(i,j)+min(D(i-1,j),D(i,j-1),D(i-1,j-1)). Backtrack to obtain the optimal path, and P theo Remapping to P aligned Calculate the residual Res=P act -P aligned After 5-point moving average filtering, the final bias residual sequence is output, which significantly improves the accuracy of subsequent confidence inversion.

[0237] S6.4: Based on the residual sequence of motion trajectory deviation, the statistical confidence inversion algorithm is used to reverse trace and quantify the accuracy of the contribution of each data source in the previous round of power adjustability assessment, so as to generate the original correction coefficient of the control closed loop verification degree that characterizes the reliability of historical decisions.

[0238] The residual sequence of the received action trajectory deviation and the edge weight matrix of the dynamic semantic credibility graph in the previous control cycle are used as the input basis for confidence inversion.

[0239] The deviation residual sequence is segmented by a sliding window, and the root mean square error and maximum absolute deviation within each time window are extracted to construct a local error vector characterizing the accuracy of the single-point response.

[0240] Based on the local error vector, the posterior probability distribution of each data source under specific working conditions is calculated using the Bayesian inference principle, quantifying its contribution to the uncertainty of the prediction results.

[0241] The Sigmoid mapping function is used to convert the normalized local error value into the original correction coefficient for regulating the closed-loop validation degree. The larger the error, the lower the correction coefficient.

[0242] The calculated correction coefficients from each data source are mapped to the corresponding nodes in the dynamic semantic credibility graph, forming a node-level credibility update vector.

[0243] The update vector is subjected to spatiotemporal smoothing filtering to eliminate coefficient abrupt changes caused by instantaneous communication jitter and ensure the stability of the correction process.

[0244] By using a statistical confidence inversion algorithm, the deviation of the action trajectory in the previous step is transformed into the original correction coefficient of the control closed-loop verification degree, which characterizes the reliability of historical decisions, thereby achieving the expected technical effect of model self-calibration.

[0245] For example, the sliding window length is set to 5 seconds, the sensitivity adjustment factor k is 2.0, and the historical average error threshold μ is 0.05. When the root mean square error of the actual active power response deviation of a certain inverter is 0.08, the correction coefficient obtained after Sigmoid mapping is approximately 0.35. This coefficient is significantly lower than the baseline value of 0.9 under high confidence conditions, indicating that the data source has a large prediction bias under the current operating conditions. The system applies this low correction coefficient to the edge weight update of the corresponding inverter node in the graph, reducing its influence in the next round of feature fusion. Meanwhile, if the prediction error of the meteorological data source is only 0.02, its correction coefficient remains above 0.85, maintaining a high weight. After smoothing filtering, the fluctuation range of the correction coefficient is controlled within ±0.05, effectively avoiding model oscillations caused by single measurement noise, and significantly improving the robustness and prediction stability of multi-source data fusion under complex operating conditions.

[0246] S6.5: Based on the original correction coefficient of the closed-loop verification degree, the edge weights in the current dynamic semantic credibility graph are dynamically corrected and normalized using the weight iteration update rule to generate the closed-loop verification degree update amount for the next round of feature fusion process, thus completing the closed-loop iteration from execution feedback to model optimization.

[0247] Receive the original correction coefficient of the control closed-loop verification degree generated in the previous sub-step. This coefficient represents the accuracy deviation of the contribution of each data source in the previous round of power adjustability assessment.

[0248] We construct a weight iteration update rule based on exponential decay memory, and set a forgetting factor to balance the influence weight of historical confidence and current measured feedback, so as to ensure that the dynamic evolution of graph edge weights has temporal continuity.

[0249] The weighted geometric average method is used to fuse topological coupling degree, temporal compatibility, and control loop validation degree. The dynamic edge weight between node i and node j at time t is calculated using the following formula: in, For the updated edge weights, The edge weights are from the previous time step. The time series compatibility score is given. To adjust the original correction coefficient for closed-loop validation, , , These are the corresponding balance coefficients and satisfy... .

[0250] The calculated dynamic edge weights are normalized using the L2 norm to eliminate numerical oscillations caused by differences in dimensions, ensuring that the weight values ​​of all connected edges in the graph converge to the [0,1] interval, thus maintaining the numerical stability of message passing in the graph neural network.

[0251] The normalized edge weight matrix is ​​written into the global state storage area of ​​the dynamic semantic credibility graph, replacing the old weight parameters, completing the closed-loop iteration from execution feedback to model optimization, and generating the control closed-loop verification update amount for the next round of feature fusion process.

[0252] Through the above weight iteration and normalization processing, the deviation residuals of the previous step are transformed into dynamic adjustment instructions for the spectral structure, realizing adaptive calibration of the semantic consistency strength of multi-source data, and significantly improving the predictive ability of the model under complex working conditions.

[0253] For example, the forgetting factor α is set to 0.6, the temporal compatibility weight β to 0.2, and the closed-loop validation weight γ to 0.2. If the edge weight between the inverter and the weather station was 0.85 at the previous time step, the current temporal compatibility score is 0.9, and the control closed-loop validation correction coefficient is 0.75, then the calculated new edge weight is 0.834. After L2 normalization, this weight is updated into the graph. This process causes the connection weights of high-biased data sources to decay rapidly, while the connection weights of low-biased data sources are maintained, effectively suppressing the interference of conflicting data on the power adjustability model and significantly improving the execution accuracy of subsequent collaborative optimization instructions.

[0254] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0255] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0256] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A distributed photovoltaic coordinated control method based on multi-source data fusion, characterized in that, include: Acquire multi-source heterogeneous raw data streams of active and reactive power of photovoltaic inverters, grid connection point voltage, meteorological irradiance and distribution network feeder load, and extract implicit capability characterization semantic tags from the multi-source heterogeneous raw data streams to generate semantic anchoring feature sets. Based on the semantic anchoring feature set and historical data of equipment topology relationships, meteorological evolution paths and control command closed-loop feedback, a dynamic semantic credibility map is constructed. The capability-guided feature distillation operation is performed using the current credibility score of each node in the dynamic semantic credibility graph. High credibility sources are mapped to the capability boundary anchoring layer to extract the voltage reactive sensitivity slope, medium credibility sources are mapped to the trend guidance layer to extract the irradiation attenuation gradient, and low credibility but high conflict source is mapped to the operating condition switching detection layer to extract the cross-domain correlation pattern, thereby generating a hierarchical distillation feature subset. The graph neural network with a fixed input structure of the hierarchical distillation feature subset but with edge weights reconfigured in real time by the dynamic semantic credibility graph outputs the photovoltaic cluster power adjustability assessment result with semantic source tracing label; Based on the distribution network operation status, dispatch instructions and time-of-use electricity price signals, and combined with the photovoltaic cluster power adjustability assessment results, the active power allocation scheme and reactive voltage control strategy for each inverter are calculated and generated collaboratively to form a joint optimization control instruction set. The joint optimization control instruction set is decomposed and distributed to each photovoltaic inverter for execution.

2. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, After the joint optimization control instruction set is decomposed and distributed to each photovoltaic inverter for execution, the following steps are also included: The actual operating trajectory of the inverter is collected and compared with the theoretical capability curve to generate the control closed-loop verification update quantity used to reverse the adjustment of the edge weights of the dynamic semantic credibility graph in the next round.

3. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, The construction of a dynamic semantic credibility map based on the semantic anchoring feature set, equipment topology relationships, meteorological evolution paths, and historical data of closed-loop feedback of control commands includes: Based on the semantic anchoring feature set and the physical connection relationship of the equipment, the electrical distance impedance matrix and power flow path topology between the photovoltaic inverter grid connection point, the meteorological monitoring station and the feeder terminal are analyzed to obtain the initial value of the topological coupling degree and map it as the static basic edge weight between the graph nodes. Using the time series data of the irradiance attenuation persistence level and the AGC response ready state Boolean signal contained in the semantic anchoring feature set, a sliding correlation calculation is performed to generate a time series compatibility score, which is then superimposed on the static basic edge weights to form preliminary dynamic edge weights. Based on the deviation residual sequence in the historical control command closed-loop feedback data, reverse tracing verification is performed, the control closed-loop verification degree correction coefficient is calculated and applied to the preliminary dynamic edge weight to complete dynamic iterative update. Based on the updated dynamic edge weights and predefined 3D score fusion rules, the topological coupling degree, temporal compatibility degree and control closed-loop verification degree are normalized and weighted to generate a comprehensive edge weight matrix and assign it to the graph connection edge to complete the graph construction. The constructed dynamic semantic credibility graph is subjected to connectivity verification and abnormal edge pruning to output the dynamic semantic credibility graph with a stable topology.

4. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, A graph neural network with a fixed input structure for the hierarchical distillation feature subset but whose edge weights are reconfigured in real time by the dynamic semantic credibility graph outputs a photovoltaic cluster power adjustability assessment result with semantic source tracing labels, including: Based on the hierarchical distillation feature subset, a static graph neural network topology is constructed, and the features of each layer are mapped to node attribute vectors and a fixed structure adjacency matrix is ​​initialized to generate a static graph neural network model. The adjacency matrix of the static graph neural network model is reconfigured in real time using the dynamic edge weights in the dynamic semantic credibility graph to generate a reconfigurable graph neural network instance with dynamic message passing weight coefficients. Based on the reconfigurable graph neural network instance, perform multi-layer semantic collaborative aggregation operation, fuse high-confidence source hard constraint features and medium-confidence source trend guidance features, and use low-confidence source cross-domain association patterns to suppress noise propagation to generate hidden layer state representation; Semantic tracing is performed on the hidden layer state representation to identify the dominant data source combination and its confidence evolution path based on the message passing path weight contribution, and semantic tracing metadata tags are generated. By combining the hidden layer state representation with the semantic traceability metadata tag to perform regression mapping decoding, the high-dimensional semantic collaborative feature space is projected onto the physical power adjustment domain, and the evaluation result of the photovoltaic cluster power adjustability with the attached semantic traceability tag is output.

5. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, The semantic anchoring feature set includes discrete semantic labels generated by the multi-source heterogeneous original data streams after being discriminated by the inverter control logic state machine, cloud motion vector inversion, and wavelet packet decomposition.

6. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, The cross-domain correlation pattern is obtained by the operating condition switching detection layer using a cross-domain correlation mining algorithm to calculate the mutual information entropy of the load current harmonic components and the inverter IGBT junction temperature data. The operating condition switching detection layer is a feature extraction layer that extracts cross-domain correlation patterns from low-confidence, high-conflict data sources.

7. A distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 6, characterized in that, In the mutual information entropy calculation, the moment when the mutual information entropy value exceeds the preset fluctuation threshold within a continuous time window is determined as the working condition switching moment.

8. The distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, Before generating the joint optimization control instruction set, the feasibility of the initially generated active power allocation scheme and reactive voltage control strategy is verified based on the AGC response ready state Boolean signal of each inverter, and invalid instruction components with mismatched response delays are eliminated.

9. A distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 8, characterized in that, In the feasibility verification, instruction components that correspond to the AGC response ready state Boolean signal as logical false are eliminated, and the remaining instruction components after elimination are limited and corrected according to the physical limits of each inverter.

10. A distributed photovoltaic coordinated control method based on multi-source data fusion according to claim 1, characterized in that, The voltage reactive power sensitivity slope is used to characterize the marginal impact of active power regulation on reactive power voltage support capability near the current operating point. The irradiance attenuation gradient is used to characterize the power decrease rate within a future preset time window. The cross-domain correlation mode is used to trigger the activation signal of the gating mechanism during the inference process of the generative graph neural network.

Images