Photovoltaic optimizer control system based on edge computing

Abstract

The present application relates to the technical field of automatic control and photovoltaic power generation control, in particular to a photovoltaic optimizer control system based on edge computing. The system comprises an edge computing node and an edge coordination end. The edge computing node aligns the voltage, current, power and duty cycle sequence of each photovoltaic optimizer in the same photovoltaic string according to the control quantity change time, forming a control response dynamic fingerprint. The edge coordination end extracts the same direction response component across nodes, determines the local mismatch candidate node, distributes the distinguishable encoding perturbation sequence, and decodes the control coupling sensitivity between nodes according to the response summary, generates the collaborative target working point and staggered execution sequence within the string current constraint and the adjustable interval of each optimizer. The present application can distinguish the common irradiation change, local mismatch change and node control disturbance, reduce the control response misattribution, and inhibit the repeated reversing and working point oscillation in multi-node optimization.

Description

Technical Field

[0001] This invention relates to the fields of automatic control and photovoltaic power generation control technology, specifically a photovoltaic optimizer control system based on edge computing. Background Technology

[0002] Existing photovoltaic optimizer control systems typically collect input voltage, input current, output voltage, output current, and duty cycle data within the optimizer corresponding to each photovoltaic module. Based on power changes in adjacent sampling periods, they execute disturbance observation control or incremental conductance control to shift the operating point of a single module towards its local maximum power point. String-level systems generally upload the operating data of each optimizer to a centralized control terminal via a communication link, where the centralized control terminal performs operation monitoring, fault alarms, and target voltage distribution. Some solutions set up edge computing nodes on-site to filter, summarize, and perform short-cycle control on the sampled data to reduce remote communication latency. For situations with partial shading or inconsistent module performance, conventional solutions typically adjust the disturbance step size based on the power change amplitude of a single node, or sequentially change the control quantities of each optimizer according to fixed time slots, and then compare the string power before and after the adjustment. These solutions use single-node power response or sequential trial results as the optimization basis, failing to establish a distinguishable correlation between control commands, node local responses, string current changes, and other node responses. They also fail to utilize the common changing components of each node's response to identify irradiance fluctuations. Therefore, the control process of each optimizer still relies primarily on independent local judgments.

[0003] In a photovoltaic string, optimizers are electrically coupled through string current. Changes in the duty cycle or target operating point of any node can alter the output state of other nodes. Simultaneously, factors such as cloud movement can cause power changes in multiple nodes to occur in the same direction at similar times. Existing control methods struggle to distinguish between changes in common irradiance, local mismatches, and cascading changes caused by node control disturbances in the mixed power response. This leads to individual optimizers easily misinterpreting the control effects of other nodes or changes in the common environment as their own disturbances. When multiple nodes optimize simultaneously, incorrect attribution can cause conflicting adjustment directions in their calculations. Even if the centralized control unit collects all data, transmission and computation lags may result in control quantities generated based on already altered operating states. The main technical problem arising from this is that, under conditions of simultaneous string coupling and common power fluctuations, existing systems cannot obtain inter-node coupling relationships suitable for group control, making it difficult to establish a coordinated optimization direction consistent with the total string power change. This leads to repeated control reversals, continuous oscillations in the operating point, or stagnation at local power peaks. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic optimizer control system based on edge computing, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The edge computing-based photovoltaic optimizer control system includes edge computing nodes that establish control data channels with each photovoltaic optimizer in the same photovoltaic string, and an edge coordination terminal that collects state summaries of each edge computing node. The edge computing nodes align the voltage, current, power, and duty cycle sequences according to the timing of control quantity changes, forming a dynamic fingerprint characterizing the control response. The edge coordination terminal extracts cross-node in-direction response components from each dynamic fingerprint as common power changes, and determines local mismatch candidate nodes based on the remaining local difference components, and assigns distinguishable coded perturbation sequences to the local mismatch candidate nodes. The edge computing node executes the encoded perturbation sequence and returns a response summary containing the encoded position and power response. The edge coordinator decodes to obtain the inter-node control coupling sensitivity, constructs the group feasible upward direction of string power based on the inter-node control coupling sensitivity, and generates a cooperative target operating point and staggered execution order within the string current constraint and the adjustable range of each photovoltaic optimizer.

[0006] Preferably, the dynamic fingerprint establishes a response start point according to the command timestamp corresponding to each duty cycle change, and a long-term baseline window is captured before the response start point and a short-term response window is captured after the response start point; The edge computing node obtains the power evolution baseline when no new control quantity is applied from the long-term reference window, and obtains the voltage offset trajectory, current offset trajectory and power offset trajectory after the control quantity is changed from the short-term response window. The power evolution baseline and each offset trajectory are normalized to the same scale according to the control increment magnitude and the reference power. The dynamic fingerprint includes a control direction identifier, a response hysteresis segment, an initial response slope, a power increment symbol sequence, a recovery trajectory, and a corresponding operating point interval, and is stored separately according to the operating point interval and the control direction.

[0007] Preferably, the edge coordination end forms a group response matrix by combining the power increment symbol sequence of each dynamic fingerprint in the same time interval and the normalized offset trajectory, extracts a common change basis vector based on the consistent change direction of most nodes in the group response matrix, and projects each dynamic fingerprint to a common perturbation subspace composed of the common change basis vector. The edge coordination end forms a common power change sequence with projection components, forms a residual envelope according to the distribution of projection residuals within the long-term reference window, and forms a local difference sequence with projection residuals that exceed the residual envelope. The edge coordination terminal compares the start time of the local difference sequence with the response hysteresis segment corresponding to the control command of this node, and determines the node whose start time does not fall into the response hysteresis segment and still retains the local difference sequence after the short-term response window ends as a local mismatch candidate node.

[0008] Preferably, the edge coordination end selects any two sequences from the preset coding set whose cross-correlation peak is less than their respective autocorrelation main peak, and configures staggered start time slots for each local mismatch candidate node according to the order of the start times of the local difference sequences. Each coded sequence consists of pairs of control increments with the same amplitude, opposite direction, and arranged according to the coded symbols. The edge computing node executes the pairs of control increments during the detection period when the cumulative value of the control increment is 0. The edge coordination terminal performs correlation decoding on the string current response, the local power response of the candidate node, and the power response of the non-candidate node with each coded sequence to generate a directed control coupling sensitivity matrix indexed by the scrambling node and the scrambling node.

[0009] Preferably, the edge computing node maintains a baseline fingerprint version and a candidate fingerprint version for each working point interval, performs time scaling alignment and control amplitude conversion on the newly acquired short-term response window and the baseline fingerprint version in the same working point interval, and forms a version residual envelope based on the residual distribution of the recovery trajectory in the baseline fingerprint version relative to the power evolution baseline. When the aligned response hysteresis segments overlap, the power increment symbol sequences are consistent, and the recovery trajectory returns to the version residual envelope, the newly acquired response is written into the candidate fingerprint version. After the candidate fingerprint versions simultaneously contain opposite control directions and the endpoints of the recovery trajectories corresponding to each control direction fall into the residual envelope of the version, the candidate fingerprint versions are merged into the baseline fingerprint version. Response windows that do not meet the conditions are written to a separate attribution sample area, which is not involved in the data writing of the baseline fingerprint version.

[0010] Preferably, the edge coordination terminal constructs a directed adjacency graph according to the node sequence of the photovoltaic optimizer in the string and the response propagation order of the previous control cycle, and decomposes the group response matrix into a graph smooth common response component and a graph sparse differential response component, wherein the graph smooth common response component satisfies the minimum total response difference between adjacent nodes, and the graph sparse differential response component satisfies the minimum set of non-zero support nodes when reconstructing the group response matrix. The edge coordination terminal updates the common change basis vector based on the directional consistency of the graph smooth common response components within a continuous control cycle, and matches the propagation starting point of the graph sparse differential response components with the control command timestamp chain, and determines the non-zero support node whose propagation starting point does not fall into any control command corresponding to the response hysteresis segment as the local mismatch candidate node.

[0011] Preferably, the edge coordination end establishes a detection conflict map based on the directed control coupling sensitivity matrix formed in the previous detection cycle, connects local mismatch candidate nodes that have bidirectional non-zero coupling or have non-zero matrix elements pointing to the same disturbed node as conflict edges, and divides the coded perturbation batches according to the independent node set of the detection conflict map. The coded sequences within the same batch cover the cyclic shift position of the string and satisfy the cross-correlation peak constraint. Isolation time slots covering the current recovery section of the string are set between different batches. The edge computing node executes the encoding sequence of the current batch only when the working point interval corresponding to the dynamic fingerprint in this section has not changed. When the working point interval changes, the remaining encoding symbols are terminated and the executed encoding position is submitted to the edge coordination terminal.

[0012] Preferably, the edge coordination end arranges the coding sequences within the same coding perturbation batch according to the coding position to form a coding matrix, arranges the corresponding string current response and node power response to form a response matrix, and synchronously deletes the corresponding rows in the coding matrix and the response matrix according to the executed coding position, and uses the column space decomposition of the retained coding matrix to obtain the direct response coefficients corresponding to each perturbation node. The edge coordination terminal eliminates the common power change sequence from the power response of non-candidate nodes, performs time-series layering according to the response start time of each non-candidate node relative to the harassing node, and calculates the indirect response coefficient after deducting the determined direct response component layer by layer. The direct response coefficient and the indirect response coefficient are written into the directed control coupling sensitivity matrix. Matrix elements corresponding to coding columns that cannot maintain linear independence from other reserved coding columns are marked as pending confirmation.

[0013] Preferably, the edge coordination end maps the matrix elements in the directed control coupling sensitivity matrix that are in the unconfirmed state to candidate coupling edges, sets the matrix elements corresponding to the candidate coupling edges to zero one by one, and re-solves the cooperative target working point. The difference vector between the re-solved result and the original cooperative target working point represents the sensitive direction of the candidate coupling edge. The edge coordination end includes the local mismatch candidate nodes that are linearly correlated by the difference vector into the same supplementary detection set, reselects the encoding sequence that keeps the encoding column of the local mismatch candidate node linearly independent, and restricts the supplementary detection to the working point interval and control direction corresponding to the dynamic fingerprint before the supplementary detection; After the supplementary response is decoded, the edge coordination end replaces the pending state with the newly added direct response coefficient, and restores the matrix elements that cannot be reconstructed from the same set of matrix elements to the pending state before redistributing the encoded perturbation batches.

[0014] Preferably, the edge coordination terminal establishes a string power increment prediction relationship based on the directed control coupling sensitivity matrix after completing the supplementary detection, and defines the cooperative target operating point by the adjustable range of each photovoltaic optimizer, the current change boundary of each string, and the cumulative change boundary of the control quantity. Each edge computing node applies corresponding control increments according to the staggered execution order, and sends back the actual string power increment and the dynamic fingerprint of this section at each execution stage; The edge coordination terminal forms a stage tolerance range based on the historical prediction residuals of the dynamic fingerprint within the same working point interval. It temporarily places the corresponding matrix elements of the execution stage that are inconsistent with the predicted trend or exceed the stage tolerance range into a state awaiting recovery. For the direction of the recovery instruction of the previous target working point, it freezes the corresponding matrix elements and suspends the writing of the response of the corresponding execution stage into the dynamic fingerprint. The recovered working point is used as the starting state of the next round of the encoded perturbation sequence.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Within the edge computing nodes, the voltage, current, power, and duty cycle sequences are aligned according to the timing of control quantity changes. A dynamic fingerprint is formed using a long-term reference window and a short-term response window. The edge coordination end then extracts common change components from the dynamic fingerprints of multiple nodes and identifies mismatch candidate nodes by removing local differences after common changes. This process separates common irradiance changes from the control response, reducing interference from environmental synchronization fluctuations on node optimization judgment. For candidate nodes, coded perturbation sequences with limited cross-correlation and staggered start time slots are assigned. Correlation decoding is performed on the string current response, candidate node power response, and non-candidate node power response to obtain control coupling sensitivity with perturbation and perturbation directions. Based on this, the edge coordination end constructs a group feasible upward direction of string power and generates target operating points and staggered execution sequences within the string current boundary, node adjustable range, and cumulative control quantity change boundary, ensuring that the control increment of each node is constrained by the same string power increment relationship. This reduces mutual misjudgment between independent disturbances, keeps the control direction consistent with the change in total string power, suppresses repeated commutation and continuous oscillation, and reduces local peak dwell caused by local judgment.

[0016] 2. Dynamic fingerprints are versioned according to operating point intervals and control directions. Newly acquired responses are only written to the baseline version when the hysteresis segment, power increment sign, and recovery trajectory meet the consistency conditions. Other responses are entered into the attribution sample area, which avoids abnormal responses directly changing the subsequent control basis. The edge coordination end uses the node sequence to construct a directed adjacency graph, decomposing the group response into smooth common response components and sparse differential response components. It also uses the command timestamp chain to determine the starting point of differential propagation, keeping the local mismatch candidate range consistent with the string response propagation relationship. A detection conflict graph is established based on the existing coupling sensitivity, which can allocate candidate nodes with mutual influence to different perturbation batches, reducing the overlap of detection responses in the same batch. The encoding matrix and response matrix are synchronously deleted at the encoding position of execution interruption, and supplementary detection is performed on the coupling edge to be confirmed, which can maintain the identifiable conditions required for coupling solution. During the execution of cooperative control, the actual string power increment and the predicted increment are matched stage by stage. For inconsistent stages, the previous target operating point is restored, the corresponding coupling elements are frozen, and the writing of dynamic fingerprints is stopped, which can limit the erroneous coupling estimation to continue to enter the subsequent control loop. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of collaborative control of a photovoltaic optimizer based on edge computing. Figure 2 For the normalized power offset baseline plot; Figure 3 A decomposition diagram of common changes and local differences; Figure 4 This is a smooth sparse decomposition graph. Detailed Implementation

[0018] In one embodiment, a photovoltaic optimizer control system based on edge computing is provided, including edge computing nodes that establish control data channels with each photovoltaic optimizer in the same photovoltaic string, and an edge coordination terminal that collects state summaries of each edge computing node. The edge computing nodes align the voltage, current, power, and duty cycle sequences according to the timing of control quantity changes to form a dynamic fingerprint characterizing the control response. The edge coordination terminal extracts the cross-node in-direction response components from each dynamic fingerprint as the common power change, and determines local mismatch candidate nodes based on the remaining local difference components, and assigns distinguishable coded perturbation sequences to the local mismatch candidate nodes. The edge computing nodes execute the coded perturbation sequences and return a response summary containing the coded position and power response. The edge coordination terminal decodes the control coupling sensitivity between nodes, constructs the group feasible upward direction of string power based on the control coupling sensitivity between nodes, and generates a cooperative target operating point and a staggered execution sequence within the string current constraint and the adjustable range of each photovoltaic optimizer.

[0019] In this embodiment, please refer to Figure 1 The control data channel transmits voltage, current, power, duty cycle, target operating point, command generation time, and command execution time on both the input and output sides. Edge computing nodes rearrange the sampling sequence based on the command execution time alignment. For missing positions, piecewise linear interpolation of adjacent valid sample values ​​is used. Out-of-order positions are reset according to timestamps. Samples spanning the next control command are truncated. The dynamic fingerprint can be represented as: ; Where i is the node index and r is the control event index. In response to delay duration, , and Let be the initial response slope of voltage, current, and power relative to the control increment. It is a sequence of power increment symbols. To recover the trajectory; for example, when the power changes continuously in the positive direction after a brief stable region and returns to the original fluctuation band after the command, the hysteresis duration is taken as the interval from the command time to leaving the fluctuation band, the initial power response slope is taken as a positive value, the symbol sequence transitions from positive symbols to zero symbols, and the recovered trajectory records the distance of each sampled value to the original fluctuation band. Based on this, the edge coordination end separates the control response from the power changes that existed before the command. After the edge coordination end completes the elimination of common changes and encoding and decoding, it combines the local power slope of each node with the coupling sensitivity into the string power increment relationship, calculates the control increment of each node in a unified constraint space, and arranges the execution order according to the direction of the transient change of the string current by the control increment.

[0020] In this embodiment, each edge computing node establishes an event record for each control event, including a node identifier, event index, working point interval identifier, control direction identifier, command version number, sampling start and end positions, and data integrity identifier. The edge coordination end only performs group comparison on event records with consecutive command version numbers and overlapping time coverage intervals. When the state summary of a node is later than the current coordination period, the edge coordination end maintains the previous valid working point of the node and expands its state uncertainty interval, and does not directly use the expired summary for common change extraction or coupled decoding. When the state summary becomes continuous again, the dynamic fingerprint is re-established with the first complete response window after recovery. The response summary uses the encoding position as the primary key, and binds and stores the control increment, local power change, string current change, and response start and end time corresponding to the same encoding position. This allows the probe data to be reconstructed according to the actual execution position even when the encoding sequence is interrupted, delayed, or partially executed, avoiding the inference of control responses that did not actually occur based on the issued sequence.

[0021] The feasible upward direction of the group is as follows: , , ; Where u is the current control vector. Let g be the control increment to be executed, g be the slope vector of the local power response of each node, and H be the power coupling sensitivity matrix. To control the influence coefficient vector of the increment on the string current, b is the boundary of the string current change in this round. To limit the non-negative coefficients of continuous control increment mutations; for example, when the local slopes of the three candidate nodes are all positive, and node 1 forms a negative coupling with the other nodes and has a large current influence coefficient, the solution results in node 1 obtaining a small positive increment and the other nodes obtaining a large positive increment. The edge coordination end then arranges the execution in the order of increasing overlap of the current response hysteresis segment corresponding to each increment.

[0022] Before solving for the feasible upward direction of the population at the edge coordination end, the local power response slope is calculated based on the most recent positive and negative control events of each node. If the slopes in the two directions have opposite signs, the slope consistent with the current control direction is selected. If both directions lack a complete recovery trajectory, the control increment of the corresponding node is fixed to 0 and its coupling column to other nodes is retained. The string current influence coefficient is obtained from the decoding result of the current response to the control increment during the coded perturbation. The adjustable interval of the node is determined by the intersection of the interval where the current operating point is located and the boundary of the adjacent interval. After the solution is completed, the control increments with the same direction and acting on the same node in two consecutive stages are merged, and the control increments with opposite directions are retained as independent stages. The staggered order prioritizes the node whose predicted current response recovers first. If a data version change occurs in any stage, the subsequent stages that have not yet been executed are canceled.

[0023] In a preferred embodiment, the dynamic fingerprint establishes a response start point according to the command timestamp corresponding to each duty cycle change. A long-term reference window is truncated before the response start point, and a short-term response window is truncated after the response start point. The edge computing node obtains the power evolution baseline when no new control quantity is applied from the long-term reference window, and obtains the voltage offset trajectory, current offset trajectory, and power offset trajectory after the control quantity change from the short-term response window. The power evolution baseline and each offset trajectory are normalized to the same scale according to the control increment amplitude and the reference power. The dynamic fingerprint includes a control direction identifier, a response hysteresis segment, an initial response slope, a power increment symbol sequence, a recovery trajectory, and a corresponding operating point interval, and is stored separately according to the operating point interval and the control direction.

[0024] Specifically, the long-term reference window extends from the response start point to the stable region after the most recent control change. If continuous unidirectional drift exists within the window, a robust straight line is used to fit the power evolution baseline; otherwise, the median value within the window forms the horizontal baseline. The short-term response window extends from the response start point to when the power enters the stable region or the next control command arrives. The voltage offset trajectory, current offset trajectory, and power offset trajectory are obtained by subtracting the reference sequence from the measured sequence, respectively. The normalized power offset can be written as: ; Where k is the sampling position within the window, For actual measured power, As a baseline for power evolution, To control the increment, The reference power is used for the long-term reference window. When the control increment direction is opposite and the operating point interval is the same, the normalization results can be compared on the same scale. Forward control events and reverse control events are written to the corresponding storage partitions respectively. When the operating point crosses the interval boundary, a new fingerprint record is created without merging with the original interval record, thereby maintaining the correspondence between the hysteresis segment, response slope and recovery trajectory and the specific operating state.

[0025] Please refer to Figure 2 , Figure 2 The horizontal axis represents the sampling number within the short-term response window, ranging from 0 to 42, while the vertical axis represents the power reading in watts. The solid line represents the response power calculated based on the normalization result, and the dashed line represents the reference power determined by the long-term reference window. The legend corresponds to the two curves: the response power curve and the reference power curve.

[0026] from Figure 2 As can be seen, the reference power curve remains relatively stable around 148 watts, indicating that the power evolution within the long-term reference window is relatively smooth and can serve as a reference background for the current control event. The response power curve initially drops to approximately 137 watts between sampling numbers 0 and 16, then gradually recovers, exceeding the reference power after sampling number 32, reaching approximately 167 watts at the end. This change reflects that after the control increment is applied, the node power first undergoes a brief deviation and recovery process, subsequently forming an effective positive response.

[0027] In this embodiment, the end point of the long-term reference window is adjacent to the response start point. Its start point searches forward to the end point of the recovery trajectory of the most recent control command. If there are data gaps or multiple power direction flips within the search interval, it continues to expand forward until a continuous stable segment is obtained. The end point of the short-term response window is the position where the power is restored to the reference residual envelope, the next control command arrives, or the operating point interval changes. The response hysteresis segment is defined by the first position where the continuous sampled value leaves the reference residual envelope and the last position formed by the stabilization of the initial response slope. The recovery trajectory is stored in descending order of absolute distance from the power evolution baseline. If the recovery process leaves the envelope again, the second deviation position is retained. The operating point interval is divided according to the continuously adjustable range of the control quantity. Adjacent intervals are only allowed to be merged when their reference power evolution slope and response hysteresis distribution remain continuous, thereby avoiding compressing the control responses under different operating states into the same fingerprint.

[0028] When performing same-scale normalization on the long-term reference window and the short-term response window, edge computing nodes first remove invalid events with a power evolution baseline of 0 or a control increment of 0. Then, the product of the absolute value of the reference power and the absolute value of the control increment is used as the scale factor, and the voltage offset and current offset are supplemented with their respective reference absolute values ​​for normalization. When the power sequence within the short-term response window does not leave the reference residual envelope, the response hysteresis segment is recorded as an empty segment, the initial response slope is recorded as 0, and the entire symbol sequence is recorded as zero symbols. This record is only used to indicate that the current control increment has not formed an identifiable response and does not participate in the numerical update of the coupling sensitivity. When the recovery trajectory crosses the next command, the recovery trajectory is truncated at the next command time, and an incomplete flag is written in the event record to prevent the subsequent command response from being included in the current event.

[0029] In a preferred embodiment, the edge coordination end forms a group response matrix by combining the power increment symbol sequence of each dynamic fingerprint within the same time interval and the normalized offset trajectory. It extracts a common change basis vector based on the consistent change direction of most nodes in the group response matrix and projects each dynamic fingerprint onto a common perturbation subspace composed of the common change basis vector. The edge coordination end forms a common power change sequence using the projected components, forms a residual envelope according to the distribution of the projection residuals within the long-term reference window, and forms a local difference sequence from the projection residuals exceeding the residual envelope. The edge coordination end compares the start time of the local difference sequence with the response hysteresis segment corresponding to the control command of this node, and identifies nodes whose start time does not fall into the response hysteresis segment and which still retain the local difference sequence after the short-term response window ends as local mismatch candidate nodes.

[0030] In this embodiment, each row of the group response matrix corresponds to a node and each column corresponds to the same aligned sampling position. Common variations and local differences are defined as follows: , , ; Where N is the number of nodes participating in the comparison. Let be the normalized power offset vector of node i. The weights are determined by symbol consistency and trajectory integrity, and their sum is 1. For common changing basis vectors, For projection coefficients, For local difference vectors, This indicates transposition; for example, if 3 out of 4 nodes maintain a downward trend and their recovery trajectories are close, and the other node turns upward in the middle, then the first 3 nodes obtain non-zero weights and form a common change basis vector, and the local difference vector of the other node retains the reverse segment. The edge coordination end forms a residual envelope that varies with the sampling position based on the median absolute deviation of each local difference vector within the long-term reference window. Candidate labels are retained only when the reverse segment crosses the envelope and its starting time does not belong to the control response hysteresis segment of this node. If the local difference returns to the envelope before the end of the short-term response window, the candidate label is cancelled.

[0031] Please refer to Figure 3 , Figure 3 The horizontal axis represents the aligned sampling number in the population response matrix, ranging from 0 to 42, and the vertical axis represents the response component readings. Solid lines represent common variations, dashed lines represent local differences at node I03, thin dotted lines represent the upper envelope, and thin dashed lines represent the lower envelope. The legend indicates the common components, local difference components, and residual envelope boundaries.

[0032] from Figure 3 As can be seen, the common variation curve rises continuously with the sampling sequence number, gradually increasing from approximately 4.2 to over 14, indicating that multiple nodes exhibit unidirectional background changes within the same coordination period. This part is more closely related to irradiation or string-level operational status changes. The I03 local difference curve is negative in the initial segment, reaches a trough near sampling sequence number 16, and then gradually rises before crossing zero in the latter segment. When the local difference approaches or exceeds the upper envelope, it indicates that the node's response cannot be fully explained by the projection of the common variation.

[0033] When establishing the residual envelope at the edge coordination end, the projected residuals of each node are collected for each aligned sampling position within the long-term reference window. The upper and lower boundaries are generated with the median residual value as the center and combined with the median absolute deviation. Only when the sampling positions continuously falling outside the boundary reach a complete response segment do they form a local difference sequence. Individual discrete out-of-bounds points are retained as sampling noise in the common change analysis record. The common change basis vector adopts a direction continuity constraint between adjacent coordination cycles. If the new basis vector is opposite in direction to the previous basis vector and the power increment sign of most nodes has not been reversed, the previous basis vector is maintained and the new data is transferred to the verification area. The start time of the candidate node is also compared with the command generation time, command execution time, and dynamic fingerprint hysteresis segment in sequence. Only when the local difference cannot be explained by the current command or the previous incomplete command, and there are still continuous out-of-bounds segments at the end of the short-term response window, is the candidate identifier submitted to the coding perturbation scheduling process.

[0034] Local mismatch candidate identifiers have a generation time, duration period number, and failure condition. The edge coordination end tracks the sign direction, starting position, and degree of overlap of the out-of-bounds interval of the local difference sequence of the same node within a continuous coordination period. When the direction remains consistent and the out-of-bounds interval persists, the candidate duration period number is accumulated. When the local difference returns to the residual envelope after the common change basis vector changes, the candidate identifier is immediately cleared. After a candidate node enters the coded perturbation batch, it suspends the use of its unattributed response to update the common change basis vector to avoid the probe response changing the candidate judgment in the opposite direction. After the probe ends and the recovered trajectory enters the baseline fluctuation band, the common change update is reopened. When a new control command appears during the waiting period for the same node, the attribution is re-executed with the hysteresis segment corresponding to the new command. The original candidate identifier is only inherited when the new and old local difference sequences are consecutive.

[0035] Furthermore, the edge coordination end selects any two sequences from the preset coding set whose cross-correlation peak is less than their respective autocorrelation main peak, and configures staggered start time slots for each local mismatch candidate node according to the order of the start times of the local difference sequences; each coding sequence consists of pairs of control increments with the same amplitude, opposite direction and arranged according to coding symbols, and the edge computing node executes the pairs of control increments within the detection period when the cumulative value of the control increment is 0; the edge coordination end performs correlation decoding on the string current response, the local power response of the candidate node and the power response of the non-candidate node with each coding sequence, and generates a directed control coupling sensitivity matrix indexed by the scrambling node and the scrambling node.

[0036] In this embodiment, each sequence in the encoding set consists of 1s and -1s, and the sum of the sequences is 0. The edge coordination end calculates the cross-correlation value after pairwise cyclic shifts of the sequences and removes sequence combinations whose cross-correlation peaks reach the main autocorrelation peak. The starting time slot is determined by the order of local differences of candidate nodes and the response hysteresis duration, so that when the previous node enters the main response segment, the next node has not yet started detection. The directed control coupling sensitivity is based on: ; in, This represents the relationship between the perturbation of node i and the response of node j, where Q is the number of encoded positions already executed. The sign of node i at encoding position q is either positive 1 or -1. This represents the absolute value of a single control increment. hour This represents the increment of string current after removing common changes. The time represents the power increment of node j; for example, when the encoded symbols are 1, -1, 1, -1 in sequence and the power response of node j flips synchronously with the symbols, the product terms maintain the same sign. Non-zero values ​​are retained. If the change of node j does not change with the encoding, the positive and negative products cancel each other out, and the corresponding matrix element remains 0 or enters the pending confirmation state.

[0037] During the generation of the encoded sequence, the edge coordination end calculates the autocorrelation peak, non-zero shifted autocorrelation value, and pairwise cross-correlation value from the candidate sequence composed of an equal number of positive and negative symbols. Sequences in which both the non-zero shifted autocorrelation value and the cross-correlation peak are lower than the corresponding autocorrelation peak are retained, and different cyclic shift positions are assigned to candidate nodes with similar response hysteresis segments. After each positive control increment, a negative control increment with the same absolute value is configured within the same detection cycle. If the operation is stopped midway due to the switching of the working point, the executed position will not be reissued with the opposite increment. Instead, the edge computing node will restore to the working point before detection and mark the recovery process separately to prevent the recovery response from being regarded as the encoded response. Before correlation decoding, the edge coordination end corrects the original response in sequence according to the common change sequence, the baseline trend before detection, and the response hysteresis time shift, so that the matrix elements correspond to the directional changes caused by the specified scrambling node.

[0038] The encoding positions of the response summary use monotonically increasing serial numbers with batch identifiers. After each encoding position, the edge computing node calculates the local power increment, string current increment, and recovery endpoint between the starting point of that position and the stable point of the response, and writes the start and end indices of the original sampling interval into the summary. During the correlation decoding at the edge coordination end, the response vector is time-shifted according to the hysteresis duration in the dynamic fingerprint of different nodes, so that the response caused by the same encoding symbol falls into the same column position in the correlation calculation. The missing positions generated after the shift do not participate in the accumulation of the numerator and denominator. If the same encoding position is affected by a common change abrupt change, the execution record of that position is retained but it is marked as unusable for numerical decoding. The corresponding row is deleted synchronously in subsequent matrix processing to maintain the consistency between the control action record and the available response data.

[0039] In a preferred embodiment, the edge computing node maintains a baseline fingerprint version and candidate fingerprint versions for each operating point interval. The newly acquired short-term response window is time-scaled and aligned with the baseline fingerprint version within the same operating point interval, and control amplitude is calculated. A version residual envelope is formed based on the residual distribution of the recovery trajectory within the baseline fingerprint version relative to the power evolution baseline. When the aligned response hysteresis segments overlap, the power increment symbol sequence is consistent, and the recovery trajectory returns to the version residual envelope, the newly acquired response is written into the candidate fingerprint version. If a candidate fingerprint version simultaneously contains opposite control directions and the endpoints of the recovery trajectories corresponding to each control direction fall within the version residual envelope, the candidate fingerprint version is merged into the baseline fingerprint version. Response windows that do not meet the aforementioned conditions are written into independent pending return windows. Because the sample area to be attributed is not included in the data writing of the baseline fingerprint version; in specific implementation, the turning point in the baseline fingerprint recovery trajectory is used as the alignment anchor point, and monotonic time scaling is applied to the adjacent sampling segments of the new trajectory, and the offset amplitude is calculated according to the ratio of the absolute values ​​of the new and old control increments. The version residual envelope is composed of the median value and median absolute deviation of the existing recovery trajectory residuals within the same operating point interval. The forward control records and reverse control records in the candidate version are merged only after passing the consistency check. If the hysteresis segment of the new response is separated from the baseline segment, the symbol sequence undergoes an unexplainable flip, or the recovery endpoint is left outside the envelope, it is kept in the sample area to be attributed and accompanied by a common change identifier, command timestamp, and operating point interval identifier. It is only used for re-attribution and does not change the current control basis.

[0040] The baseline fingerprint version is stored using a read-only main body and incremental record separation method. The candidate fingerprint version only records the hysteresis segment, slope, symbol sequence and recovery trajectory that have changed relative to the baseline version. The time scaling alignment restricts the mapping of adjacent sampling positions to remain monotonic and must not cross the turning point of the recovery trajectory. The control amplitude reduction only acts on the voltage, current and power offset without changing the hysteresis duration. When merging candidate versions, the distance from the recovery endpoint to the power evolution baseline is calculated for records with opposite control directions. Merging is terminated if either direction does not return to the version residual envelope. When multiple samples to be attributed are continuously formed in the same operating point interval, the edge computing nodes are clustered according to the common change identifier and command timestamp. Only samples that can be explained by the same external change are used to update the baseline power evolution baseline, and not used to update the control response slope.

[0041] Both the baseline fingerprint version and the candidate fingerprint version have a parent version identifier. The candidate version merging adopts three consecutive processing states: temporary writing, integrity verification, and version switching. The integrity verification checks the working point interval, control direction, sampling position, and recovery endpoint. Before the verification is completed, the edge computing node still uses the original baseline version to form a dynamic fingerprint. After the version switch, the original baseline version is retained until the new version completes the verification of at least one positive control event and one negative control event. If the new version generates continuous samples to be attributed during the verification period, the original baseline version is restored and the new version is put into a frozen state. The sample area to be attributed maintains the original order according to the event index. The mean merging of samples is not performed. When re-attributing, the time relationship between common changes, command responses, and recovery trajectories can be restored in sequence.

[0042] Furthermore, the edge coordination end constructs a directed adjacency graph according to the node sequence of the photovoltaic optimizer in the string and the response propagation order of the previous control cycle, decomposing the group response matrix into a graph-smooth common response component and a graph-sparse differential response component. The graph-smooth common response component satisfies the minimum total response difference between adjacent nodes, while the graph-sparse differential response component satisfies the minimum set of non-zero supporting nodes when reconstructing the group response matrix. The edge coordination end updates the common change basis vector based on the directional consistency of the graph-smooth common response component within consecutive control cycles, and matches the propagation starting point of the graph-sparse differential response component with the control command timestamp chain. Non-zero supporting nodes whose propagation starting point does not fall into any control command's corresponding response hysteresis segment are identified as local mismatch candidate nodes. The decomposition is achieved through: ; Where Y is the group response matrix, G is the graph smoothing common response component, S is the graph sparse differential response component, and L is the graph Laplacian matrix obtained by symmetrization of the directed adjacency graph. and The non-negative constraint coefficients are... For the Frobenius norm, It is the sum of the 2norms of each row; for example, when three adjacent nodes decrease synchronously while the fourth node increases independently, the changes of the first three nodes are preferentially entered into G, and the row where the fourth node is located is preferentially entered into S. If the propagation starting point of the fourth node does not fall into any command lag section, then the local mismatch candidate mark is retained.

[0043] Please refer to Figure 4 , Figure 4 The horizontal axis represents the aligned sampling number of the population response matrix, ranging from 0 to 42, and the vertical axis represents the decomposed component readings. Solid lines represent the original response, dashed lines represent the smoothed components, and dotted lines represent the sparse components. The legend indicates the response source and decomposition result corresponding to the three types of curves.

[0044] from Figure 4 As can be seen, the smooth component of the graph remains positive overall and gradually increases with the sampling sequence number, indicating that there is a continuous, unidirectional cascade-level response background between adjacent optimizers. The original response curve is negative in the early part, gradually rises in the middle part, and turns positive in the later part, reflecting that individual nodes still retain local reverse changes outside the common background. The sparse component curve is mainly in the negative range, forming a deep trough in the middle part, indicating that local differences are concentrated in the finite sampling segment and the finite number of node rows.

[0045] The basic edges of the directed adjacency graph are established according to the connection order of the photovoltaic optimizer in the string. If the response start point of a certain node continuously leads another node in the previous control cycle and there is a non-zero coupling sensitivity between the two, a propagation edge pointing from the leading node to the lagging node is added to the basic edge. The propagation edge is deleted after the time order is lost in the continuous cycle. The update of the graph smooth common response component is weighted by row vectors with the same direction in the continuous cycle. The graph sparse differential response component retains the node position and start time according to the non-zero support row. The edge coordination end traces back the earliest non-zero support node along the propagation edge and matches the command timestamp on the tracing path with the lag segment of each node one by one. If the path start point can correspond to a certain control command, the subsequent non-zero support row is interpreted as a propagation response. Only the path start point without a corresponding command is regarded as a local mismatch candidate.

[0046] In a preferred embodiment, the edge coordination end establishes a detection conflict map based on the directed control coupling sensitivity matrix formed in the previous detection cycle. Local mismatch candidate nodes with bidirectional non-zero coupling or non-zero matrix elements pointing to the same disturbed node are connected as conflict edges. The edge coordination end then divides the coding perturbation batches according to the independent node set of the detection conflict map. The coding sequences within the same batch have different cyclic shift positions and satisfy cross-correlation peak constraints. Isolation time slots covering the string current recovery section are set between different batches. The edge computing node executes the coding sequence of the current batch only when the operating point interval corresponding to its dynamic fingerprint has not switched. When the operating point interval switches, the remaining coding symbols are terminated, and the edge coordination end sends the code to the edge coordination end. The coordination end submits the executed encoding positions; in specific implementation, the edge coordination end arranges candidate nodes from most to least according to the number of conflicting edges, and puts each candidate node into a batch that does not have conflicting edges with the nodes already in the batch. The encoding sequences in the same batch form different correlation peak positions through cyclic shifting. The end position of the isolation time slot is taken as the latest position of the recovery trajectory of each node in the previous batch entering the baseline fluctuation zone. Before executing each encoding position, the edge computing node checks the current working point interval identifier with the interval identifier when the batch is issued. If the identifier is consistent, it continues to execute. If the identifier changes, it retains the completed encoding positions, actual control increment and response time range and terminates the unexecuted part, so that subsequent decoding only uses the actually completed encoding content.

[0047] When updating the collision map, two nodes corresponding to bidirectional non-zero coupling directly establish a collision edge. When multiple scrambling nodes point to the same scrambled node, paired collision edges are established between these scrambling nodes. Coupled elements in the pending confirmation state are treated as having collision edges. Batch division adopts a node-by-node insertion method. Each candidate node can only enter a set that has no collision edges with other nodes in the batch. If it cannot be inserted into an existing set, a new batch is established. The cyclic shift position of the same batch is allocated according to the relevant peak spacing. No new perturbations are issued within the isolation time slot. Only the recovery trajectory and string current are continuously collected. If the recovery trajectory has not entered the reference fluctuation band at the end of the predetermined isolation interval, the current isolation time slot is extended and the next batch is frozen to avoid cross-batch responses from superimposing into the same encoding position.

[0048] When the edge coordinator issues a batch, it simultaneously sends the batch version, the binding relationship between nodes and coding sequences, the cyclic shift position, the allowed range of working points, and the working point before probing. Edge computing nodes only accept instructions with versions higher than those of locally completed batches. After a batch is partially executed, unexecuted nodes do not use their original cyclic shift positions to enter the next batch. Instead, they are reassigned by the edge coordinator based on the reserved coding columns. The recovery records of executed nodes are saved independently of the next batch. If a new local mismatch candidate node appears during the isolation slot extension, it is only added to the scheduling set and not inserted into the current batch or isolation slot. The conflict graph is reconstructed only after the current batch has been fully recovered and the response digest verification is completed. This ensures that the coding matrix of each batch is not dynamically changed after issuance.

[0049] In this embodiment, the edge coordination end arranges the coded sequences within the same coded perturbation batch according to their coded positions to form a coding matrix, arranges the corresponding string current responses and node power responses to form a response matrix, and synchronously deletes the corresponding rows in the coding matrix and response matrix based on the executed coded positions. The direct response coefficients corresponding to each perturbed node are obtained using the column space decomposition of the retained coding matrix. The edge coordination end eliminates common power change sequences from the power responses of non-candidate nodes, performs time-series layering according to the response start time of each non-candidate node relative to the perturbed node, and obtains the indirect response coefficients after deducting the determined direct response components layer by layer. The direct and indirect response coefficients are written into the directed control coupling sensitivity matrix. Matrix elements corresponding to coding sequences that cannot maintain linear independence from other retained coding sequences are marked as pending confirmation. The retained coding matrix is ​​denoted as... The response matrix is ​​denoted as The direct response coefficient matrix is ​​based on: ; Where R represents the set of lines to retain after deleting the unexecuted positions. This represents the Morpenrose generalized inverse; for example, when the preserved coded sequences of two scrambling nodes are still linearly independent. The direct coefficients of each response column are given separately. If the two columns overlap after deleting the unexecuted position, the corresponding coefficients are not directly written into the coupling sensitivity matrix but are marked as pending confirmation. The remaining responses of non-candidate nodes are then subtracted from the known direct components according to the starting time from early to late to obtain the indirect response coefficients formed along the propagation chain.

[0050] Row deletion in the coding matrix and response matrix strictly uses the same set of retained rows. Before column space decomposition, the rank of the retained coding columns is calculated. The direct response coefficients are obtained when the rank equals the number of scrambling nodes. If the rank is insufficient, all elements corresponding to linearly correlated columns will be set to pending confirmation. After deducting common changes in the power response of non-candidate nodes, it is divided into a direct layer and several propagation layers according to the response start time relative to the scrambling node. The direct layer is obtained from the coding matrix. The remaining coefficients of the subsequent propagation layer are calculated after subtracting the reconstructed responses of all earlier layers. If the remaining response of a certain layer does not remain correlated with any coding column, it is not written into the indirect response coefficients. The direct coefficients and indirect coefficients are stored in the source batch, the set of retained rows, and the operating point interval, respectively. Subsequent supplementary probes only replace matrix elements with the same source and the same operating point interval.

[0051] The indirect response coefficients are obtained layer by layer, with the latest start time of the direct response layer as the boundary of the next layer. The response of each node in the next layer is first subtracted from the product of all confirmed previous layer coefficients and the corresponding coding column, and then the relevant projection is performed on the remaining response. If the sign of the projection result is inconsistent with the direction of the propagation edge, it is not written into the matrix. The rank of the retained coding matrix is ​​determined by column elimination. During the elimination process, the coding column with missing principal components and its linear combination source are marked as pending confirmation, rather than arbitrarily retaining one column. The coupling sensitivity matrix stores the power response coefficient and the string current response coefficient respectively. The feasible upward direction of the group uses the power response coefficient to form the target term, and the string current constraint uses the current response coefficient to form the boundary term. The two do not replace each other.

[0052] Furthermore, the edge coordination end maps the matrix elements in the directed control coupling sensitivity matrix that are in a state of unconfirmation to candidate coupling edges. It then sets the corresponding matrix elements of each candidate coupling edge to zero and re-solves for the cooperative target operating point. The difference vector between the re-solved result and the original cooperative target operating point represents the sensitive direction of the candidate coupling edge. The edge coordination end incorporates locally mismatched candidate nodes whose difference vectors are linearly correlated into the same supplementary detection set. It then reselects encoding sequences that keep the encoding columns of locally mismatched candidate nodes linearly independent, and restricts the supplementary detection to the operating point interval and control direction corresponding to the dynamic fingerprint before the supplementary detection. After the supplementary response is decoded, the edge coordination end replaces the unconfirmed state with a newly added direct response coefficient, and assigns the variable that cannot be determined by the same... A set of matrix elements is used to reconstruct the pre-probe response matrix and the supplementary response matrix. After the matrix elements are restored to the pending confirmation state, the coding perturbation batch is redistributed. In specific implementation, only one candidate coupling edge is set to zero each time, while the remaining matrix elements remain unchanged. The difference vector is formed by the difference between the two cooperative target operating points. Candidate nodes whose corners of the difference vectors fall into the same directional interval are included in the same supplementary set. The edge coordination end reselects column vectors from unused cyclic shift positions or non-conflicting coding sequences and applies paired perturbations along the original control direction within the original operating point interval. The response matrices before and after the supplementary probe are reconstructed by the updated matrix. Only the newly added direct response coefficients that simultaneously satisfy the two sets of reconstruction relationships are released from the pending confirmation state.

[0053] The sensitive direction of the candidate coupling edge is determined by the difference vector between the coordinating target working points before and after zeroing. The edge coordination end normalizes the difference vector and calculates the pairwise inner product. Nodes with an absolute value close to 1 and the same sign are grouped into the same supplementary probe set, while nodes with opposite signs are arranged into adjacent batches. The supplementary encoding selection is constrained by preserving the column rank under the row set and prioritizes avoiding the cyclic shift positions that cause column correlation in the original probe. The control increment direction of the supplementary probe must not cross the boundary of the original working point interval. If a boundary switch occurs during execution, the supplementary result will not be used to remove the pending confirmation state. The updated coupling matrix reconstructs the original response matrix and the supplementary response matrix respectively. Only when both sets of reconstructed residuals fall into their respective baseline residual envelopes are the newly added direct response coefficients written. Otherwise, the pending confirmation state is restored and the conflict graph is reconstructed.

[0054] After the supplementary probe set is formed, the edge coordinator checks whether the current working point interval of each node in the set is still consistent with the original probe. When any node switches intervals, it is removed from the set and kept in the pending confirmation state. The remaining nodes recalculate the coded column rank. The response matrix before the supplementary probe and the supplementary response matrix adopt the same common change removal method and the same response hysteresis alignment method. The reconstructed residuals are compared with their respective residual envelopes before the probe. The original residual envelopes are not modified in reverse using the supplementary probe data. After the new direct response coefficients are added to remove the pending confirmation state, the edge coordinator resolves the cooperative target working point and compares the target difference vector. If the difference vector still contradicts the zeroing test results, the corresponding coupling edge is set to pending confirmation again and its scheduling priority in the next conflict graph is increased.

[0055] In a preferred embodiment, the edge coordination end establishes a string power increment prediction relationship based on the directed control coupling sensitivity matrix after completing supplementary detection, and defines the cooperative target operating point by the adjustable range of each photovoltaic optimizer, the string current change boundary, and the cumulative change boundary of the control quantity; each edge computing node applies the corresponding control increment according to the staggered execution order, and transmits back the actual string power increment and the dynamic fingerprint of the node at each execution stage; the edge coordination end forms a stage tolerance range based on the historical prediction residual of the dynamic fingerprint within the same operating point range, matches the direction and value of the actual string power increment with the predicted power increment stage by stage, and sends a recovery instruction carrying the previous target operating point to the corresponding edge computing node for execution stages with inconsistent directions or exceeding the stage tolerance range, freezes the corresponding matrix elements and suspends writing the response of the corresponding execution stage into the dynamic fingerprint, and uses the recovered operating point as the starting state of the next round of encoded perturbation sequence; the prediction increment and judgment residual of stage m are respectively: , ; in, To control the increment in stages, To measure the increase in string power, For stage prediction residuals; for example, if the predicted increment is positive while the measured increment is negative, the direction is directly determined to be inconsistent, even if the prediction and measurement are in the same direction but... When the values ​​are found to be inconsistent when the tolerance range formed by the median value and median absolute deviation of the historical residuals in the same working point interval is exceeded, the edge coordination end restores the previous target working point and freezes the coupling matrix elements involved in this stage. After the restored trajectory enters the reference fluctuation zone, the encoded perturbation sequence is redistributed with the dynamic fingerprint corresponding to the restored point and the working point interval.

[0056] The staggered execution adopts a phased state recording method. Each phase records the target node, starting operating point, target operating point, control increment, predicted string power increment, measured string power increment, and corresponding matrix element version. The next phase is released only after the recovery trajectory of the previous phase enters the baseline fluctuation zone. The phase tolerance interval is established separately according to the historical prediction residuals under the same operating point interval and the same control direction. When the historical samples are insufficient, only the direction consistency is determined without updating the dynamic fingerprint. The recovery command carries the previous target operating point and its command version number. The edge computing node only performs recovery when the version number is consistent with the current phase. The data collected during the recovery period is entered into an independent recovery record. The frozen matrix elements do not participate in the solution of the feasible upward direction of the population before the subsequent coding perturbation is reconfirmed. The first complete dynamic fingerprint after recovery serves as the starting basis for the next round of common change extraction and coding allocation.

[0057] Before each execution phase begins, the edge coordinator re-verifies the target node's current operating point, dynamic fingerprint version, and coupling matrix version. If any version is inconsistent with the phase state record, the current phase and its subsequent phases are canceled and the solution is recalculated. The actual string power increment is calculated with reference to the power evolution baseline of the stable interval before the phase begins. The phase end position is taken as the earlier position where the target node's recovery trajectory enters the reference fluctuation zone or the next phase is allowed to start. The prediction residual only uses the data within the complete interval of this phase. After the recovery command is executed, the edge computing node maintains the previous target operating point until the power symbol sequence continuously returns to zero symbols and the string current enters the recovery section. The recovery record is not incorporated into the normal control event. The frozen matrix elements can only be released after obtaining identifiable direct response coefficients in the new coded perturbation.

Claims

1. A photovoltaic optimizer control system based on edge computing, characterized in that, This includes edge computing nodes that establish control data channels with each photovoltaic optimizer within the same photovoltaic string, and edge coordination terminals that aggregate state summaries of each edge computing node; The edge computing nodes align the voltage, current, power, and duty cycle sequences according to the timing of control quantity changes, forming a dynamic fingerprint characterizing the control response. The edge coordination terminal extracts cross-node in-direction response components from each dynamic fingerprint as common power changes, and determines local mismatch candidate nodes based on the remaining local difference components, and assigns distinguishable coded perturbation sequences to the local mismatch candidate nodes. The edge computing node executes the encoded perturbation sequence and returns a response summary containing the encoded position and power response. The edge coordinator decodes to obtain the inter-node control coupling sensitivity, constructs the group feasible upward direction of string power based on the inter-node control coupling sensitivity, and generates a cooperative target operating point and staggered execution order within the string current constraint and the adjustable range of each photovoltaic optimizer.

2. The photovoltaic optimizer control system based on edge computing according to claim 1, characterized in that, The dynamic fingerprint establishes a response start point according to the command timestamp corresponding to each duty cycle change, and a long-term baseline window is captured before the response start point and a short-term response window is captured after the response start point. The edge computing node obtains the power evolution baseline when no new control quantity is applied from the long-term reference window, and obtains the voltage offset trajectory, current offset trajectory and power offset trajectory after the control quantity is changed from the short-term response window. The power evolution baseline and each offset trajectory are normalized to the same scale according to the control increment magnitude and the reference power. The dynamic fingerprint includes a control direction identifier, a response hysteresis segment, an initial response slope, a power increment symbol sequence, a recovery trajectory, and a corresponding operating point interval, and is stored separately according to the operating point interval and the control direction.

3. The photovoltaic optimizer control system based on edge computing according to claim 2, characterized in that, The edge coordination terminal forms a group response matrix by combining the power increment symbol sequence of each dynamic fingerprint in the same time interval and the normalized offset trajectory. It extracts a common change basis vector based on the consistent change direction of most nodes in the group response matrix, and projects each dynamic fingerprint onto a common perturbation subspace composed of the common change basis vector. The edge coordination end forms a common power change sequence with projection components, forms a residual envelope according to the distribution of projection residuals within the long-term reference window, and forms a local difference sequence with projection residuals that exceed the residual envelope. The edge coordination terminal compares the start time of the local difference sequence with the response hysteresis segment corresponding to the control command of this node, and determines the node whose start time does not fall into the response hysteresis segment and still retains the local difference sequence after the short-term response window ends as a local mismatch candidate node.

4. The photovoltaic optimizer control system based on edge computing according to claim 3, characterized in that, The edge coordination end selects any two sequences from the preset coding set whose cross-correlation peak is smaller than their respective autocorrelation main peak, and configures staggered start time slots for each local mismatch candidate node according to the order of the start times of the local difference sequences. Each coded sequence consists of pairs of control increments with the same amplitude, opposite direction, and arranged according to the coded symbols. The edge computing node executes the pairs of control increments during the detection period when the cumulative value of the control increment is 0. The edge coordination terminal performs correlation decoding on the string current response, the local power response of the candidate node, and the power response of the non-candidate node with each coded sequence to generate a directed control coupling sensitivity matrix indexed by the scrambling node and the scrambling node.

5. The photovoltaic optimizer control system based on edge computing according to claim 2, characterized in that, The edge computing node maintains a baseline fingerprint version and a candidate fingerprint version for each working point interval. It performs time scaling alignment and control amplitude conversion on the newly acquired short-term response window and the baseline fingerprint version in the same working point interval, and forms a version residual envelope based on the residual distribution of the recovery trajectory in the baseline fingerprint version relative to the power evolution baseline. When the aligned response hysteresis segments overlap, the power increment symbol sequences are consistent, and the recovery trajectory returns to the version residual envelope, the newly acquired response is written into the candidate fingerprint version. After the candidate fingerprint versions simultaneously contain opposite control directions and the endpoints of the recovery trajectories corresponding to each control direction fall into the residual envelope of the version, the candidate fingerprint versions are merged into the baseline fingerprint version. Response windows that do not meet the conditions are written to a separate attribution sample area, which is not involved in the data writing of the baseline fingerprint version.

6. The photovoltaic optimizer control system based on edge computing according to claim 3, characterized in that, The edge coordination terminal constructs a directed adjacency graph according to the node sequence of the photovoltaic optimizer in the string and the response propagation order of the previous control cycle, and decomposes the group response matrix into a graph smooth common response component and a graph sparse differential response component. The graph smooth common response component satisfies the minimum total response difference between adjacent nodes, and the graph sparse differential response component satisfies the minimum set of non-zero support nodes when reconstructing the group response matrix. The edge coordination terminal updates the common change basis vector based on the directional consistency of the graph smooth common response components within a continuous control cycle, and matches the propagation starting point of the graph sparse differential response components with the control command timestamp chain, and determines the non-zero support node whose propagation starting point does not fall into any control command corresponding to the response hysteresis segment as the local mismatch candidate node.

7. The photovoltaic optimizer control system based on edge computing according to claim 4, characterized in that, The edge coordination end establishes a detection conflict map based on the directed control coupling sensitivity matrix formed in the previous detection cycle, connects local mismatch candidate nodes that have bidirectional non-zero coupling or have non-zero matrix elements pointing to the same disturbed node as conflict edges, and divides the coded perturbation batches according to the independent node set of the detection conflict map. The coded sequences within the same batch cover the cyclic shift position of the string and satisfy the cross-correlation peak constraint. Isolation time slots covering the current recovery section of the string are set between different batches. The edge computing node executes the encoding sequence of the current batch only when the working point interval corresponding to the dynamic fingerprint in this section has not changed. When the working point interval changes, the remaining encoding symbols are terminated and the executed encoding position is submitted to the edge coordination terminal.

8. The photovoltaic optimizer control system based on edge computing according to claim 7, characterized in that, The edge coordination end arranges the coding sequences within the same coding perturbation batch according to the coding position to form a coding matrix, arranges the corresponding string current response and node power response to form a response matrix, and synchronously deletes the corresponding rows in the coding matrix and the response matrix according to the executed coding position, and uses the column space decomposition of the retained coding matrix to obtain the direct response coefficients corresponding to each perturbation node. The edge coordination terminal eliminates the common power change sequence from the power response of non-candidate nodes, performs time-series layering according to the response start time of each non-candidate node relative to the harassing node, and calculates the indirect response coefficient after deducting the determined direct response component layer by layer. The direct response coefficient and the indirect response coefficient are written into the directed control coupling sensitivity matrix. Matrix elements corresponding to coding columns that cannot maintain linear independence from other reserved coding columns are marked as pending confirmation.

9. The photovoltaic optimizer control system based on edge computing according to claim 8, characterized in that, The edge coordination end maps the matrix elements in the directed control coupling sensitivity matrix that are in the unconfirmed state to candidate coupling edges, sets the matrix elements corresponding to the candidate coupling edges to zero one by one, and re-solves the cooperative target working point. The difference vector between the re-solved result and the original cooperative target working point represents the sensitive direction of the candidate coupling edge. The edge coordination end includes the local mismatch candidate nodes that are linearly correlated by the difference vector into the same supplementary detection set, reselects the encoding sequence that keeps the encoding column of the local mismatch candidate node linearly independent, and restricts the supplementary detection to the working point interval and control direction corresponding to the dynamic fingerprint before the supplementary detection; After the supplementary response is decoded, the edge coordination end replaces the pending state with the newly added direct response coefficient, and restores the matrix elements that cannot be reconstructed from the same set of matrix elements to the pending state before redistributing the encoded perturbation batches.

10. The photovoltaic optimizer control system based on edge computing according to claim 1, characterized in that, The edge coordination terminal establishes a string power increment prediction relationship based on the directional control coupling sensitivity matrix after completing supplementary detection, and defines the cooperative target operating point by the adjustable range of each photovoltaic optimizer, the current change boundary of each string, and the cumulative change boundary of the control quantity. Each edge computing node applies corresponding control increments according to the staggered execution order, and sends back the actual string power increment and the dynamic fingerprint of this section at each execution stage; The edge coordination terminal forms a stage tolerance range based on the historical prediction residuals of the dynamic fingerprint within the same working point interval. It temporarily places the corresponding matrix elements of the execution stage that are inconsistent with the predicted trend or exceed the stage tolerance range into a state awaiting recovery. For the direction of the recovery instruction of the previous target working point, it freezes the corresponding matrix elements and suspends the writing of the response of the corresponding execution stage into the dynamic fingerprint. The recovered working point is used as the starting state of the next round of the encoded perturbation sequence.

Images