Transient voltage collaborative support method and system based on hybrid topology

Abstract

The application specifically provides a transient voltage collaborative support method and system based on a hybrid topology, comprising: modeling a new energy station hybrid topology as a graph structure, encoding the topology graph by using a graph convolution network to generate a topology embedding vector, outputting a reactive power distribution coefficient increment by a local intelligent agent corresponding to each reactive power source according to a local state and the topology embedding vector, and determining a given value of reactive power of each source after normalization and in combination with a total reactive power demand. The application adopts a centralized training and decentralized execution framework, and online reasoning does not depend on real-time communication between intelligent agents; a hybrid coefficient weighting mechanism is used to realize smooth switching of AI distribution and fixed droop distribution, and the application automatically degrades to droop autonomy when communication is interrupted. The application realizes adaptive adjustment of a reactive power distribution coefficient with a station topology, and takes into account optimization performance and engineering reliability.

Description

Technical Field

[0001] This invention belongs to the field of power quality management technology, specifically relating to a transient voltage collaborative support method and system based on hybrid topology. Background Technology

[0002] During a grid short-circuit fault, the voltage drop at the grid connection point of new energy sources and the continuous increase in active current driven by unbalanced power may cause the system to exceed the voltage stability boundary, triggering transient voltage instability. Transient voltage instability will cause new energy units to disconnect from the grid due to their weak withstand capability. After disconnection, the reduced short-circuit capacity will further deteriorate the transient voltage, potentially triggering a chain reaction of grid disconnection accidents. At the same time, severe voltage fluctuations can also cause power quality problems such as equipment damage and power outages.

[0003] In traditional control strategies, both SVG and SVC switch to voltage closed-loop control during transient periods, with multiple devices controlling the same grid connection point voltage in parallel. Furthermore, this relies on centralized coordination of various reactive power sources at the station control level, issuing unified commands via communication.

[0004] The hybrid topology of renewable energy power plants is a non-standard topology composed of multiple voltage levels, multiple branches of collector lines and transformers. This leads to significant differences in the support effect of the same reactive power increment on the grid connection voltage of various reactive power sources. However, the traditional fixed-ratio allocation strategy only allocates reactive power according to the rated capacity of the equipment. This results in sources with close electrical distances being overloaded due to excessive reactive power, while sources with far electrical distances are idle, affecting the overall efficiency of transient voltage support. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a transient voltage collaborative support method and system based on hybrid topology to solve the aforementioned technical problems.

[0006] In a first aspect, the present invention provides a transient voltage collaborative support method based on hybrid topology, comprising: The hybrid topology is modeled as a topology graph. The set of nodes in the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point nodes. The set of edges in the topology graph includes collector line branches and transformer branches. The topology graph is encoded using a graph convolutional network to generate a topology embedding vector; Each reactive power source's corresponding local agent outputs the local reactive power allocation coefficient increment based on the local state acquired in the current cycle and the topology embedding vector. The allocation coefficient of each reactive power source is calculated based on the incremental reactive power allocation coefficient of each reactive power source. The total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point is obtained, and each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

[0007] In an optional implementation, the characteristics of each node in the topology graph include: the current reactive power output of the reactive power source corresponding to the node, the upper limit of the reactive power capacity of the reactive power source, the current voltage of the node, and the device type identifier of the reactive power source. Different types of reactive power sources correspond to different device type identifiers, so that the graph convolutional network can distinguish between static var generators, static var compensators and new energy unit groups based on the device type identifier. The characteristics of each side include: the impedance and admittance values ​​of the corresponding collector line or transformer branch, the DC side voltage of the static var generator, the current active power output of the new energy unit group, and the voltage change rate at the grid connection point, among at least one of these.

[0008] In one optional implementation, the local agent is trained and deployed using a centralized training, distributed execution framework: During the offline training phase, the critic network of the station control layer acquires the global state and global actions for centralized training, and the Actor network of each local agent updates its parameters based on the global reward signal. The global state includes the local state of all reactive power sources, and the global actions include the allocation coefficient increment of all reactive power sources. During the online execution phase, only the Actor network of each local agent performs inference based on the local state obtained in the current period and the topology embedding vector and outputs the incremental source reactive power allocation coefficient. The local agents do not rely on real-time communication to coordinate.

[0009] In an optional implementation, the global reward signal includes a combination of the following evaluation items: grid connection point voltage tracking error, sum of squares of the capacity utilization of each reactive power source, change in the allocation coefficient of each reactive power source, and grid connection point voltage over-limit penalty item.

[0010] In an optional implementation, the allocation coefficient is smoothly switched with the traditional drooping allocation coefficient through a hybrid mechanism: Using the mixing coefficient α as the weight, the allocation coefficient output by the local agent is weighted and summed with the pre-set original drooping allocation coefficient to obtain the actual reactive power allocation coefficient. When communication is interrupted or the confidence level of the local agent output is lower than a preset threshold, the mixing coefficient α is set to 0, and the system switches to a fixed droop mode that allocates power according to the rated capacity of each reactive power source.

[0011] In an optional implementation, after each reactive power source determines its own reactive power setpoint based on its respective allocation coefficient and the total reactive power demand, and before issuing execution instructions to each reactive power source, the following further step is included: Verify that the sum of the allocation coefficients of all reactive power sources is 1; Verify that the reactive power setpoint of each reactive power source does not exceed the preset percentage of the upper limit of the reactive power capacity of the source; Verify that the change in the allocation coefficient of each reactive power source relative to the previous cycle does not exceed the preset step size limit. If any of the above checks fails, the allocation coefficient of the previous period will be used to replace the allocation coefficient of the current period.

[0012] In an optional implementation, it further includes: The low voltage cross-pass curve, overcurrent protection logic, and overvoltage protection logic have higher priority than the reactive power distribution instructions output by the local intelligent agent. When any hard protection condition is triggered, the reactive power setpoint of the corresponding reactive power source is directly taken over by the hard protection logic, and the output of the local agent is shielded.

[0013] In an optional implementation, the converter of the new energy unit does not participate in the reactive power distribution under transient operating conditions: During the transient period, the active current of the new energy unit converter is limited to 0.3 times the rated value, and the released capacity is used for reactive power support; the reactive current of the new energy unit converter is output autonomously according to the low voltage crossing national standard curve and the degree of voltage drop at the grid connection point, and does not receive the given reactive power value.

[0014] In an optional implementation, it further includes: Real-time monitoring of grid connection point voltage and grid connection point current; when the grid connection point voltage exceeds the preset voltage threshold range or the grid connection point current exceeds the preset current threshold, it is determined that a transient operating condition has been entered, and the transient voltage collaborative support is triggered. When the grid connection point voltage recovers to the preset hysteresis range and remains within the preset time, it is determined that the transient operating condition has been exited, and the allocation coefficient is gradually returned to the steady-state preset ratio through the ramp function.

[0015] Secondly, the present invention provides a transient voltage cooperative support system based on a hybrid topology, comprising: The topology modeling module is used to model the hybrid topology as a topology graph. The node set of the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point nodes; the edge set of the topology graph includes collector line branches and transformer branches. The feature encoding module is used to encode the topology graph using a graph convolutional network to generate a topology embedding vector; The incremental output module is used to output the incremental reactive power allocation coefficient of each reactive power source based on the local intelligent agent obtained in the current cycle and the topology embedding vector. The coefficient calculation module is used to calculate the allocation coefficient of each reactive power source based on the reactive power allocation coefficient increment of each reactive power source. The power determination module is used to obtain the total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point. Each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

[0016] The beneficial effects of this invention are as follows: The transient voltage collaborative support method and system based on hybrid topology provided by this invention, while retaining the hierarchical autonomous physical framework, only improves the multi-source reactive power distribution link, and has the following beneficial effects: This invention uses a graph convolutional network to encode the hybrid topology of the new energy power station, so that the distribution coefficient reflects the electrical distance from each reactive power source to the grid connection point, realizing topology adaptive distribution. At the same time, it suppresses single-source overload through the capacity utilization evaluation item and extends the equipment life. It adopts a centralized training and distributed execution framework, and online inference does not rely on real-time communication between agents. With the hybrid smooth switching mechanism, it automatically degenerates to fixed droop distribution when communication is interrupted, taking into account both optimization performance and engineering reliability. Moreover, this invention only replaces the distribution coefficient generation link, and the physical control framework such as local PI closed loop and LVRT hard protection is completely retained, which is compatible with existing equipment and has low transformation cost. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic block diagram of a system according to an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0022] The transient voltage collaborative support method based on hybrid topology provided in this embodiment of the invention is executed by a computer device, and correspondingly, the transient voltage collaborative support system based on hybrid topology runs in the computer device.

[0023] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a transient voltage collaborative support system based on a hybrid topology. Depending on different requirements, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0024] like Figure 1 As shown, the method includes: S1. Model the hybrid topology as a topology graph. The node set of the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point nodes; the edge set of the topology graph includes collector line branches and transformer branches. S2. Encode the topology graph using a graph convolutional network to generate a topology embedding vector; S3. The local agent corresponding to each reactive power source outputs the incremental reactive power allocation coefficient based on the local state obtained in the current cycle and the topology embedding vector. S4. Calculate the allocation coefficient of each reactive power source based on the incremental reactive power allocation coefficient of each reactive power source; S5. Obtain the total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point, and each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

[0025] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0026] The hybrid topology of the new energy power station is modeled as a graph structure G=(V,E), where V is the set of nodes and E is the set of edges. The specific modeling method is as follows: Each static var generator (SVM), static var compensator (SVC), renewable energy unit group, and grid connection point in the renewable energy power station is defined as a node in the diagram. For a renewable energy unit group, multiple unit converters on the same collector line are aggregated into an equivalent unit group node, whose aggregated capacity is the sum of the capacities of each unit, and whose aggregated reactive power output is the sum of the reactive power outputs of each unit. Assuming there are N reactive power source nodes (including SVG, SVC, and unit group nodes) plus 1 grid connection point node in the entire power station, the total number of nodes |V| = N + 1.

[0027] The collector lines and transformer branches connecting the nodes within a new energy power station are defined as edges in the graph. For any electrical connection branch between any two nodes i and j, an undirected edge is established. The edge attribute contains the electrical parameters of the branch, including resistance. Reactance X ij and susceptance (B) ij For transformer branches, their impedance must be reduced to the same voltage reference.

[0028] Feature vector of each node The structure is as follows:

[0029] in: This represents the current reactive power output (per unit value) of the reactive power source corresponding to this node. This is the upper limit of the reactive power capacity of the reactive power source (per unit value). For grid connection nodes, this value is 0. This represents the current voltage amplitude (per unit) at this node. This is the DC-side voltage (per unit) of the static var generator. For non-SVG nodes, this value is taken as a default constant. This represents the current active power output (per unit value) of the reactive power source corresponding to this node. For SVG nodes and devices with negligible active power output, this value is 0. Encode the device type for this node: SVG = 1, SVC = 2, unit group = 3, and grid connection point = 0. This is the voltage phase angle at that node.

[0030] Each edge eigenvectors The structure is as follows:

[0031] in: Branch resistance (per unit value); Branch reactance (per unit value); Branch susceptance (per unit value); This is used to identify the branch type; 0 is used for collector lines and 1 for transformers.

[0032] For two nodes that are not directly electrically connected, there is no edge and no edge feature.

[0033] Special handling for grid connection nodes: The grid connection point node serves as the connection point between the entire power station and the external power grid, and its voltage... As the control objective for reactive power support across the entire field, this node does not allocate reactive power output, but participates in the feature aggregation of each reactive power source node as an important neighbor node in the information propagation of graph convolution.

[0034] Detailed rules for unit group aggregation: For the m-th collector line Taiwanese generator converters are aggregated into generator group nodes. :

[0035] The equivalent impedance between the aggregation node and the PCC is taken as the cumulative impedance from the beginning of the collector line to the PCC.

[0036] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0037] A Graph Convolutional Network (GCN) is used to encode the topological graph G=(V,E) constructed in step S1, along with its node feature matrix X and edge feature matrix E, to generate a topological embedding vector h that reflects the electrical coupling relationships of the entire hybrid topology. G .

[0038] The node feature matrix output in step S1 is X∈R |V|×8 (The feature dimension d of each node) x =8). The edge feature matrix is ​​E∈R |E|×2 (The feature dimension d of each edge) e =2). Since subsequent graph convolution operations require adding node features and edge features in the same space, the edge features are first mapped to the node feature space through a linear projection layer:

[0039] in ∈R {8×2} The edge feature projection weight matrix is... ∈R 8 The bias term, after projection ∈R 8 The dimensions are consistent with the node features.

[0040] This implementation employs a two-layer graph convolutional structure. The first graph convolutional layer maps the input node features to the hidden layer space (64 dimensions), and the second graph convolutional layer keeps the hidden layer features in the same space, ultimately outputting the embedded representation of each node. Let the adjacency relationship be: for node i, its set of neighboring nodes is... (Including self-loops, i ∈ ).

[0041] The aggregation and update rules for the first layer of graph convolution are as follows:

[0042]

[0043] in The 8-dimensional input features of neighbor node j Let be the 8-dimensional projected edge features between node i and its neighbor node j, and MEAN be the aggregation function that averages the elements. ∈R 64×8 This is the first layer weight matrix. ∈R 64 The first layer bias is used, and ReLU is the activation function. ∈R 64 This is the hidden embedding of node i in the first layer output.

[0044] The aggregation and update rules for the second-layer graph convolution are as follows:

[0045]

[0046] in The first-level hidden embedding for neighbor node j, ∈R 64×64 This is the weight matrix for the second layer. ∈R 64 For the second layer bias, ∈R 64 The final node embedding is the output of node i after two layers of graph convolution.

[0047] The aggregation function uses mean aggregation (MEAN), which takes the arithmetic mean of the features of neighboring nodes (including self-loops). The inclusion of self-loops ensures that the node's own features are preserved in each layer of the graph convolution. If a node has no actual physical neighbors (an isolated node), its neighbor set only contains itself, and its own features are still preserved after aggregation.

[0048] The final embeddings of all nodes are aggregated into a fixed-length topological embedding vector using global average pooling. :

[0049] Where |V| represents the total number of nodes. ∈R 64 This is a 64-dimensional global topology embedding vector. This vector encodes the electrical coupling information of the entire new energy power station's hybrid topology, including the equipment type, capacity, operating status of each node, impedance information of each branch, and the overall topology connection relationships.

[0050] Step S2, GCN forward inference, is executed in conjunction with the feature import in step S1: The station control layer collects dynamic quantities of each node every 100ms, combines them with static quantities to construct complete X and E, inputs them into the GCN for one forward propagation, and outputs... .Should This process remains unchanged within the 100ms period, allowing the Actor networks of each local agent to reuse the data multiple times during this period (once every 10ms), avoiding high-frequency repetitive computations in the GCN and reducing the computational burden on the station control layer. When the topology changes (e.g., branch disconnection, equipment commissioning / decommissioning), it is automatically reflected in E and X during the next feature import, and the GCN re-encodes and generates new features. .

[0051] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0052] S301. Local Intelligent Agent Architecture In this implementation, each controllable reactive power source (SVG, SVC, each generator group) is configured with a local intelligent agent. i Let i = 1, 2, ..., N, where N is the total number of controllable reactive power sources. Each local agent contains an Actor network to implement an end-to-end mapping between the system's real-time state and the reactive power distribution control strategy. Agents do not rely on real-time communication for cooperation; they make independent decisions based solely on their local state and shared topology embedding vectors.

[0053] S302. Construction and Acquisition of Local State Within each control cycle (typically 10ms), the local agent Agent_i acquires the local state vector. The structure is as follows:

[0054] The meaning, source, and acquisition method of each component are as follows: Grid connection point voltage The current per-unit voltage at the grid connection point is measured by a voltage transformer at the grid connection point and digitized after A / D sampling. Grid connection point current. The current per-unit value at the grid connection point is obtained by measuring the current transformer at the grid connection point.

[0055] The source is effortless and generates no effort. This represents the per-unit value of the reactive power currently output by the reactive power source. SVG or SVC is measured by the local controller, while for a group of units, it is obtained by summing the reactive power outputs of all its subordinate units. Reactive power capacity upper limit. This is the per-unit value representing the upper limit of the reactive power capacity of the reactive power source. It is a factory parameter of the equipment, stored offline in the local controller, and is only updated when the equipment is put into operation, deactivated, or when the parameters are changed.

[0056] SVG DC side voltage This is the per-unit value of the DC-side voltage of the SVG converter, measured internally by the SVG local controller and acquired via fiber optic communication; for non-SVG agents, this component is set to a default constant of 1.0. (Intrinsic active power output) This is the per-unit value of the current active power output of the reactive power source. SVG or SVC is obtained from local measurement, and the value of the generator group is obtained by summing the active power output of each generator.

[0057] Voltage change rate The rate of change of the grid connection point voltage is calculated from the difference in voltage values ​​over two consecutive sampling periods: Power grid impedance estimation The estimated equivalent impedance of the external power grid is obtained through the station control layer impedance online identification module, with a typical update cycle of 100ms.

[0058] The above dynamic quantities The data is collected via the station control layer communication network with a sampling period of 100 μs. At the beginning of each control cycle (10 ms), each local agent reads the latest value to construct a complete local state vector. Static quantity Update only when equipment is put into service, deactivated, or when parameters are changed.

[0059] Local state vector With the global topology embedding vector generated in step S2 By combining these elements, the complete input to the Actor network is formed:

[0060] in =8, therefore the input dimension is 72.

[0061] S303. Actor Network Structure and Inference Process Each local agent i The Actor network uses a fully connected deep neural network, with the following structure: Input layer: 72 dimensions, receives the concatenated local state and topological embedding vector; First hidden layer: 128 neurons, activation function is ReLU; The second hidden layer consists of 64 neurons with the ReLU activation function. Output layer: 1 neuron, activation function is tanh, output range is (-1, 1).

[0062] The forward propagation process of an Actor network can be represented as follows:

[0063] in Agent i Actor network, Its network parameters, ∈(-1,1) represents the original output of the network.

[0064] S304. Output and Limiting of Allocation Coefficient Increment The raw output of the Actor network The increment of the original reactive power distribution coefficient is obtained after scaling. :

[0065] in Set the maximum adjustment range for a single step (0.05~0.1 is recommended) to ensure that each adjustment does not exceed the preset step size limit and prevent sudden changes in the allocation coefficient from causing system shock.

[0066] The allocation coefficient for the current cycle is obtained by superimposing the increment of the allocation coefficient from the previous cycle and then limiting it:

[0067]

[0068] The `clip(·, 0, 1)` method restricts the value to the range [0, 1]. The outputs of each agent... Then, global normalization is performed to ensure Σλ i =1.

[0069] S305. Coordination of Inference Frequency and Timing The Actor network inference of each local agent coordinates with the feature acquisition in step S1 and the GCN inference in step S2 on the time axis as follows: Feature acquisition (period 100μs): Each local controller acquires dynamic values ​​and sends them to the station control layer; GCN inference (period 100ms): The station control layer performs GCN forward propagation every 100ms to update h. G ; Actor inference (10ms cycle): Each local agent performs Actor network forward inference once every 10ms. Within each 100ms GCN cycle, h G Keeping it constant, each agent uses the same h G and constantly updated local state i (k) Perform 10 inferences (once every 10ms) and output Δλ.i (k).

[0070] Actor inference is performed on the local controller (without relying on the station control layer), and the time for a single inference is less than 2ms, which is much less than the 10ms control cycle, thus meeting the real-time requirements.

[0071] S306. Decoupling of Training and Execution Actor network parameters θ for each local agent i The training process is already implemented offline. The training phase employs a centralized training, distributed execution (CTDE) framework: the Critic network in the station control layer is trained centrally using global state and global actions, while each Actor network updates its parameters based on the global reward signal. After training, only the parameters θ of each Actor network are processed. i Deployed to the corresponding local controller, the Critic network is deprecated during the online execution phase. During online execution, each local agent relies only on its local state s. i (k) and shared topological embedding h G Forward reasoning is performed without relying on real-time communication between agents or real-time computation at the station control layer, thus ensuring system reliability in scenarios where communication is interrupted.

[0072] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0073] S401. Cumulative Calculation of Distribution Coefficients Allocation coefficients of each reactive power source in the current cycle Based on the distribution coefficient of the previous period The increment output of step S3 is superimposed get:

[0074] Among them, the previous period allocation coefficient The values ​​are stored in the registers of each local controller, and the initial values ​​(at k=0) are set according to the ratio of the rated capacity of each reactive power source.

[0075] The above formula ensures that each source fairly shares the total reactive power demand according to its capacity at the initial moment.

[0076] S402. Single-source limiting To prevent individual reactive power source allocation coefficients from exceeding the physically feasible range, single-source limiting is applied to the accumulated results:

[0077] This limiting operation ensures the allocation factor for each reactive power source. All values ​​are within the closed interval [0,1]. Since each agent outputs independently... Before the limit is reached, some sources may appear. For cases where the value is >1 or <0, this operation can be used to bring it back to a reasonable range.

[0078] After completing global normalization, the normalized allocation coefficients are obtained. Subsequently, in order to balance the performance of AI optimization with the security of system operation, this invention does not directly... Instead of issuing commands to individual reactive power sources for execution, an α-hybrid smooth switching mechanism is introduced to weight and fuse the AI ​​output with a pre-set fixed droop distribution coefficient to obtain the final reactive power distribution coefficient for actual execution. .

[0079] Preset initial droop distribution coefficients: A fixed set of initial droop distribution coefficients is pre-set for each reactive power source. This coefficient is determined solely based on the rated capacity ratio of each reactive power source and does not rely on any communication or AI calculations.

[0080] in For the first The rated reactive power capacity of each reactive power source. This initial droop factor is pre-stored in the read-only memory of each local controller, serving as the final physical safety net baseline for reliable system operation. In a specific example of this invention (a 200MW photovoltaic power plant), the initial droop allocation factors for SVG, SVC, and the unit group are respectively... , , .

[0081] Real-time determination of the mixing coefficient α: Defining the mixing coefficient The AI ​​confidence weight for the current control cycle is continuously adjustable between 0 and 1, and is specifically determined by the health of the communication link and the confidence of the AI ​​model. Ideal operating state: When the station control layer communication link is normal (heartbeat signal is continuously online, data delay is less than 5ms) and the confidence level of each local agent's network output is... All are not lower than the preset threshold. At that time, the system determines that the current AI decision is reliable. The value is 1.0.

[0082] Fully degraded state: When the communication link at the station control layer is completely interrupted (no heartbeat signal is received for more than 50ms), or the confidence level of any agent is... consistently below After three confirmations, the system determined that the AI ​​had malfunctioned. Forced to be set to 0.

[0083] Smooth transition state: When occasional packet loss occurs in the communication link or the confidence level is low... When fluctuations occur, to avoid jumps in allocation coefficients caused by frequent system switching, Perform first-order low-pass filtering smoothing using the following formula:

[0084] in The instantaneous target weight is obtained by directly looking up a table based on the current communication status. The smoothing time constant (a value of 0.1 to 0.5 seconds is recommended). To control the cycle. This filtering mechanism ensures The changes are made slowly in the form of a ramp to avoid impacting the system due to sudden changes in execution coefficients caused by instantaneous communication jitter.

[0085] Weighted fusion calculation of actual execution allocation coefficients: This determines the fusion coefficients for the current period. Then, the actual allocation coefficients of each reactive power source. The following weighted summation formula is used to obtain:

[0086] The physical meaning of this formula is as follows: when (When AI is working normally) The system fully executes the allocation strategy optimized by GCN-DDPG topology awareness, demonstrating the advanced performance of this invention; when (In case of communication interruption or AI malfunction) The system seamlessly degrades to the original fixed droop distribution that does not rely on any intelligent computing, ensuring that the basic functions of transient support are not lost; when When the transition process or confidence level is questionable, the system output is a linear interpolation of the AI ​​optimization scheme and the fixed fallback scheme, which retains some AI gain and constrains the extreme output that the AI ​​may produce through the physical baseline.

[0087] Weighted normalization check: To prevent slight deviations in the sum due to rounding errors after weighted summation, the normalization is checked again before performing the safety constraint check. Perform normalization correction:

[0088] Strictly meets the requirements after correction and This ensures the physical integrity of the allocation ratio of each source.

[0089] In one embodiment of the present invention, based on step S5, a possible embodiment will be given below, and its specific implementation will be described in a non-limiting manner.

[0090] Real-time monitoring of grid connection point voltage U t (k) and grid connection point current I s (k). When any of the following conditions are met, the system is determined to be in a transient operating condition, triggering the transient voltage collaborative support process: Condition 1: The voltage at the grid connection point exceeds the preset voltage threshold range.

[0091] Condition 2: The grid connection point current exceeds the preset current threshold. The above determination is executed once in each control cycle (10ms), and the complete control process of steps S1 to S5 is started immediately after the determination is triggered.

[0092] In addition to pure threshold determination, this implementation also deploys a lightweight LSTM prediction network as an optional auxiliary enhancement.

[0093] S501. Mechanism for generating total reactive power demand Total reactive power demand The voltage is generated independently by the outer loop of the Static Var Generator (SVG). As the fastest continuous reactive power regulation device in the entire field, the SVG's outer loop monitors the grid connection point voltage U in real time. t (k) calculates the total reactive power injection required to maintain the voltage.

[0094] The SVG voltage outer loop employs a PI control structure combined with a self-decrease additional term, and its voltage setpoint... It consists of two parts: a steady-state voltage reference and a transient additional quantity.

[0095] in This is the steady-state voltage reference value (typical value 1.0 pu). The transient additional control quantity is calculated from the current output of the SVG using an adaptive attenuation coefficient:

[0096] in This is the adaptive attenuation factor for SVG (recommended value range: 0.005~0.02 pu / Mvar). The current output reactive power of SVG is given by the saturation function sat(), which restricts the independent variable to the interval [-1, 1].

[0097] The PI regulator in the outer voltage loop calculates the total reactive power demand based on the deviation between the voltage setpoint and the actual voltage value. :

[0098] voltage deviation , and These are the outer loop PI parameters of the voltage. To control the cycle. The meaning of the positive and negative signs: a positive value indicates that the system needs capacitive reactive power (the voltage is too low and needs to be increased), and a negative value indicates that the system needs inductive reactive power (the voltage is too high and needs to be decreased).

[0099] S502. Complete Steps for Generating Total Reactive Power Demand SVG generation The complete calculation process is as follows: The SVG local controller collects the grid connection point voltage. And SVG's current reactive power output ; according to and Calculate transient additional quantity ; Calculate the voltage setpoint ; Calculate voltage deviation Calculated by PI controller ; right Limit the amplitude, with the upper limit being the total reactive power demand limit for the entire site. (Determined by the sum of the capacities of all sources in the field), the lower limit is ; Will The broadcast is transmitted to the local controllers of each reactive power source (SVG itself, SVC controller, and controllers of each unit group) through the station control layer communication network.

[0100] S503.SVC and unit converter do not participate. generate It should be noted that static var compensators (SVCs) and converters for new energy units are not included. The generation of SVC. During transients, SVC maintains closed-loop power control, and its reactive power output is based solely on the allocation factor it receives. and Given a specific value, the total reactive power demand is not actively calculated. Furthermore, the converters of the new energy generating units do not participate in station-level collaborative calculations during transient periods, relying entirely on local detection to autonomously output the total reactive power demand according to the national standard LVRT curve. This design ensures that only the SVG, with the fastest response time, is responsible for sensing voltage deviations and generating total demand commands, avoiding target conflicts caused by multiple sources simultaneously calculating total demand.

[0101] S504. Determination of Reactive Power Setpoints for Each Reactive Source Each reactive power source local controller receives the broadcast from the station control layer. Then, the allocation coefficients calculated in step S4 and stored in the local register are combined. The final reactive power setpoint of the source is calculated using the following formula:

[0102] After the reactive power setpoints for each source are determined, it is necessary to verify whether the capacity constraints are met.

[0103] If a certain source calculates... If it exceeds 95% of its rated capacity, its setpoint is clamped to... and the excess reactive power The remaining sources are redistributed proportionally according to their normalized allocation coefficients. The given values ​​for each source must also satisfy the following: for SVG, its maximum transient reactive power output current can reach 1.2 to 1.5 times the rated current, i.e. For SVC, its output is limited by the TSC group switching step size, and the actual output is closest to... The discrete range value; for the unit converter, its reactive power output upper limit is 0.3 pu (30% of the rated capacity), that is... .

[0104] After each source determines its initial setpoint and before issuing the final execution command to each reactive power source, the following three safety constraint checks are forcibly executed to ensure that the allocation scheme always operates within the physical limits and safety boundaries of the equipment. In this embodiment, the verification process is centrally executed in the station control layer coordination controller, and the final command is issued to each local controller after the verification is passed.

[0105] First verification: Allocation coefficient and normalization verification Verify that the sum of the actual allocation coefficients for all reactive power sources is 1. Rounding errors in numerical calculations or slight offsets after α-weighting may cause Σ to be slightly different. Theoretically, it is 1, but in computer floating-point representation, it is 10. -6 The magnitude of the deviation. The calibration tolerance set in this implementation method. =10 -4 :

[0106] If the verification passes, maintain the current allocation coefficients; if the verification fails, force renormalization until the verification tolerance is met.

[0107] The second verification: reactive power capacity constraint verification Verify the initial setpoints of each reactive power source. Does it exceed the preset percentage η of the maximum reactive power capacity (in this implementation, η = 95%)?

[0108] If all sources are satisfied, maintain the current given value. If there is a source j that exceeds the constraint, perform the following iterative correction: Step 1: Clamp the given value of source j to Its allocation coefficient is adjusted accordingly. ; Step 2: Allocate the reactive power deficit that source j could not distribute due to clamping. The remaining sources (those not exceeding the limit) will be redistributed proportionally according to their allocation coefficients:

[0109] After correction, the given values ​​for each source are recalculated and the capacity constraints are checked again. If other sources still exceed the limits after the initial allocation, the above steps are repeated iteratively until all sources meet the capacity constraints, or the rollback mechanism is triggered when the number of iterations exceeds the preset upper limit (10 times in this implementation).

[0110] The third verification: Constraint verification of the rate of change of allocation coefficients. Verify whether the change in each source allocation coefficient relative to the previous period exceeds the preset step size limit. (This implementation method takes) =0.1):

[0111] If the change in a certain source j exceeds Then the source allocation coefficient will be clamped to Then, normalization is re-performed on all sources to ensure that the sum is 1.

[0112] rollback mechanism triggering conditions and execution logic If any of the following situations occur in the above three checks, the security check is deemed to have failed: Even after exceeding the preset limit (10 times) of iteration clamping, it is still impossible to make all sources meet the capacity constraint; After normalization Σ The deviation from 1 exceeds the tolerance ε sum =10 -4 Furthermore, forced normalization still fails to achieve convergence; Detect any It is an abnormal value such as a non-numerical value or infinity.

[0113] If any check fails, the system does not use the given value calculated in the current cycle, but instead uses the final execution allocation coefficient from the previous cycle. Replace the current period's total reactive power demand Q_total(k-1) with the total reactive power demand of the previous period:

[0114] Simultaneously, the mixing coefficient α(k) is forcibly set to 0, and the system automatically degenerates into a fixed drooping allocation mode. This ensures that if the AI ​​module fails to converge due to numerical anomalies, the system can still continue to operate using the previously valid allocation scheme. The system also sends a safety verification alarm signal to the station control layer, prompting maintenance personnel to intervene and investigate.

[0115] S505. Coordination Logic at the Execution Level Determine the reactive power setpoint for each source. Then, their local controllers execute according to their respective physical control laws: SVG will The current is converted into a q-axis current command in the inner current loop, which is then output as a modulated signal by the PI regulator in the inner current loop to drive the converter. SVC will... Compared to the current actual output, the thyristor conduction angle is adjusted via a PI regulator, and TSC capacitor banks are sequentially switched on or off according to the group switching logic. The unit converter does not execute [operations] during transient periods. Instead, it outputs independently according to the LVRT national standard curve, that is:

[0116] and The active current is simultaneously limited to 0.3 pu, freeing up capacity for reactive power support.

[0117] S506. Switching between transient and steady states When the grid connection point voltage recovers to the hysteresis range [0.95pu, 1.05pu] and remains there for a preset time (typically 1 second), the transient operating condition is exited. The generation mechanism gradually returns to steady state: the transient additional quantity ΔU of SVG attach (k) at a ramp rate R ramp =0.05 pu / s decays to zero, voltage setpoint Gradually returning to U base The allocation coefficient λ of each source i (k) Synchronous ramp transition back to steady-state preset ratio (according to rated capacity ratio). The voltage transition is smooth throughout the process, avoiding secondary impact.

[0118] In some embodiments, the hybrid topology-based transient voltage collaborative support system may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the hybrid topology-based transient voltage collaborative support system may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functionality of transient voltage co-support based on hybrid topology.

[0119] In this embodiment, the hybrid topology-based transient voltage collaborative support system can be divided into multiple functional modules according to its functions, such as... Figure 2 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0120] The topology modeling module is used to model the hybrid topology as a topology graph. The node set of the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point nodes; the edge set of the topology graph includes collector line branches and transformer branches. The feature encoding module is used to encode the topology graph using a graph convolutional network to generate a topology embedding vector; The incremental output module is used to output the incremental reactive power allocation coefficient of each reactive power source based on the local intelligent agent obtained in the current cycle and the topology embedding vector. The coefficient calculation module is used to calculate the allocation coefficient of each reactive power source based on the reactive power allocation coefficient increment of each reactive power source. The power determination module is used to obtain the total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point. Each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

[0121] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0122] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0123] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0125] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A transient voltage collaborative support method based on hybrid topology, characterized in that, include: The hybrid topology of the new energy power station is modeled as a topology graph. The node set of the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point node; the edge set of the topology graph includes collector line branches and transformer branches. The topology graph is encoded using a graph convolutional network to generate a topology embedding vector; Each reactive power source's corresponding local agent outputs the local reactive power allocation coefficient increment based on the local state acquired in the current cycle and the topology embedding vector. The allocation coefficient of each reactive power source is calculated based on the incremental reactive power allocation coefficient of each reactive power source. The total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point is obtained, and each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

2. The method according to claim 1, characterized in that, In the topology graph, the characteristics of each node include: the current reactive power output of the reactive power source corresponding to the node, the upper limit of the reactive power capacity of the reactive power source, the current voltage of the node, and the equipment type identifier of the reactive power source. Different types of reactive power sources correspond to different equipment type identifiers, so that the graph convolutional network can distinguish static var generators, static var compensators and new energy unit groups according to the equipment type identifier. The characteristics of each side include: the impedance and admittance values ​​of the corresponding collector line or transformer branch, the DC side voltage of the static var generator, the current active power output of the new energy unit group, and the voltage change rate at the grid connection point, at least one of the following:

3. The method according to claim 1, characterized in that, The local intelligent agent is trained and deployed using a centralized training and distributed execution framework. During the offline training phase, the critic network of the station control layer acquires the global state and global actions for centralized training, and the Actor network of each local agent updates its parameters based on the global reward signal. The global state includes the local state of all reactive power sources, and the global actions include the allocation coefficient increment of all reactive power sources. During the online execution phase, only the Actor network of each local agent performs inference based on the local state obtained in the current period and the topology embedding vector, and outputs the incremental source reactive power allocation coefficient. The local agents do not rely on real-time communication to coordinate.

4. The method according to claim 3, characterized in that, The global reward signal includes a combination of the following evaluation items: grid connection point voltage tracking error, sum of squares of the capacity utilization rate of each reactive power source, change in the allocation coefficient of each reactive power source, and grid connection point voltage over-limit penalty item.

5. The method according to claim 1, characterized in that, The allocation coefficient is smoothly switched from the traditional drooping allocation coefficient through a hybrid mechanism: Using the mixing coefficient α as the weight, the allocation coefficient output by the local agent is weighted and summed with the pre-set original drooping allocation coefficient to obtain the actual reactive power allocation coefficient. When communication is interrupted or the confidence level of the local agent output is lower than a preset threshold, the mixing coefficient α is set to 0, and the system switches to a fixed droop mode that allocates power according to the rated capacity of each reactive power source.

6. The method according to claim 1, characterized in that, After each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand, and before issuing execution instructions to each reactive power source, the process also includes: Verify that the sum of the allocation coefficients of all reactive power sources is 1; Verify that the reactive power setpoint of each reactive power source does not exceed the preset percentage of the upper limit of the reactive power capacity of the source; Verify that the change in the allocation coefficient of each reactive power source relative to the previous cycle does not exceed the preset step size limit. If any of the above checks fails, the allocation coefficient of the previous period will be used to replace the allocation coefficient of the current period.

7. The method according to claim 1, characterized in that, Also includes: The low voltage cross-pass curve, overcurrent protection logic, and overvoltage protection logic have higher priority than the reactive power distribution instructions output by the local intelligent agent. When any hard protection condition is triggered, the reactive power setpoint of the corresponding reactive power source is directly taken over by the hard protection logic, and the output of the local agent is shielded.

8. The method according to claim 1, characterized in that, The converter of the new energy unit does not participate in the reactive power distribution under transient operating conditions: During the transient period, the active current of the new energy unit converter is limited to 0.3 times the rated value, and the released capacity is used for reactive power support; the reactive current of the new energy unit converter is output autonomously according to the low voltage crossing national standard curve and the degree of voltage drop at the grid connection point, and does not receive the given reactive power value.

9. The method according to claim 1, characterized in that, Also includes: Real-time monitoring of grid connection point voltage and grid connection point current; when the grid connection point voltage exceeds the preset voltage threshold range or the grid connection point current exceeds the preset current threshold, it is determined that a transient operating condition has been entered, and the transient voltage collaborative support is triggered. When the grid connection point voltage recovers to the preset hysteresis range and remains within the preset time, it is determined that the transient operating condition has been exited, and the allocation coefficient is gradually returned to the steady-state preset ratio through the ramp function.

10. A transient voltage collaborative support system based on hybrid topology, characterized in that, include: The topology modeling module is used to model the hybrid topology as a topology graph. The node set of the topology graph includes static var generator nodes, static var compensator nodes, at least one new energy unit group node, and grid connection point nodes; the edge set of the topology graph includes collector line branches and transformer branches. The feature encoding module is used to encode the topology graph using a graph convolutional network to generate a topology embedding vector; The incremental output module is used to output the incremental reactive power allocation coefficient of each reactive power source based on the local intelligent agent obtained in the current cycle and the topology embedding vector. The coefficient calculation module is used to calculate the allocation coefficient of each reactive power source based on the reactive power allocation coefficient increment of each reactive power source. The power determination module is used to obtain the total reactive power demand generated by the static var generator based on the voltage deviation at the grid connection point. Each reactive power source determines its own reactive power setpoint based on its own allocation coefficient and the total reactive power demand.

Images