Collaborative Disturbance Rejection Control Method among Inverters in Desert Photovoltaic Power Plants

Abstract

This invention discloses a collaborative disturbance rejection control method among inverters in a desert photovoltaic power station, comprising: collecting electrical quantities characterizing the grid operating state and thermal state quantities characterizing the state of each inverter; establishing a dynamic response model for each inverter based on these quantities, generating sensitivity parameters and dynamic capability boundaries; calculating the total regulation budget required at the station level based on the monitored grid disturbance quantities and sensitivity parameters; collaboratively solving the control share to be undertaken by each inverter and the stability control parameters to be configured based on the total regulation budget, sensitivity parameters, and dynamic capability boundaries; issuing and executing control commands to each inverter accordingly; and updating the model parameters and dynamic capability boundaries for the next control cycle based on the system response and thermal state quantities after control execution. This invention, by collecting multi-state quantity modeling and combining three-stage collaborative solving, can complete collaborative disturbance rejection decision-making and execution within hundreds of milliseconds, improving transient support capability and operational reliability under weak grid conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power electronics, and particularly relates to a method for collaborative anti-interference control between desert photovoltaic power station inverters. BACKGROUND

[0002] At present, renewable energy represented by photovoltaic is massively integrated into modern power systems. Due to land resources and sunshine conditions, desert areas become the place for building large-scale centralized photovoltaic power stations. Such power stations are usually located in remote areas and need to be connected to the end of the power grid through long-distance transmission lines, resulting in that the grid connection point presents typical weak grid characteristics, i.e. large equivalent impedance of the power grid, weak voltage support capability and low stability margin. Under this background, the control strategy of the inverter, as the core of photovoltaic power conversion and grid connection, directly determines the response behavior and survival ability of the entire power station in the face of grid disturbances. Therefore, the research on the inverter anti-interference control method which can adapt to the harsh environment of the desert and the challenge of the weak grid and achieve efficient collaboration can not only break through the key technical bottleneck of the performance of large-scale photovoltaic power stations, but also enable the safe and stable operation of a higher proportion of new energy power systems.

[0003] At present, for the collaborative control of multiple inverters in a photovoltaic power station, the industry has proposed various technical solutions. In the early engineering practice, a centralized control architecture is often used, that is, a central controller uniformly collects the information of the entire station, makes calculation decisions, and issues control instructions to each inverter execution unit. Another widely used strategy is the master-slave or peer-to-peer control method based on droop control. In master-slave control, one inverter is designated as the master unit responsible for the key grid connection point voltage and frequency regulation, and other inverters are slave units mainly executing the maximum power point tracking (MPPT) or power following instructions. In the peer-to-peer droop control, all inverters independently adjust their active or reactive power output according to the local measured voltage or frequency deviation based on the preset droop coefficient, and automatically distribute power through implicit interaction based on grid coupling. In recent years, with the development of distributed control theory, multi-agent systems based on consensus algorithm or alternating direction multiplier method (ADMM) have also been introduced into the field of inverter collaborative control, aiming to seek a globally optimal control strategy through neighbor node communication and iterative calculation between inverters.

[0004] Although the existing technology has achieved the collaborative operation of the inverter group to some extent, it still faces problems such as control time lag, dynamic stability vulnerability and individual capability mismatch in the harsh scenario of the desert photovoltaic power station. The mutual interweaving restricts the anti-interference performance of the power station under the weak grid. Therefore, further research is needed to solve the above problems existing in the prior art. SUMMARY

[0005] In view of the above problems of the prior art, the application provides a desert photovoltaic power station inter-inverter collaborative anti-interference control method.

[0006] Technical solution: According to one aspect of the application, the desert photovoltaic power station inter-inverter collaborative anti-interference control method comprises the following steps: collecting electrical quantities representing the operating state of the power grid and thermal state quantities representing the state of each inverter, establishing a dynamic response model for each inverter according to the electrical quantities and the thermal state quantities, and generating sensitivity parameters and dynamic capability boundaries;

[0007] Based on the monitored power grid disturbance quantity and the sensitivity parameters, the total adjustment budget required by the station level is calculated, and according to the total adjustment budget, the sensitivity parameters and the dynamic capability boundaries, the control share required by each inverter and the stability control parameters required to be configured are collaboratively solved;

[0008] According to the solved control share and stability control parameters, control instructions are issued to each inverter and executed;

[0009] According to the system response after control execution and the thermal state quantity, the model parameters and the dynamic capability boundaries of the next control period are updated.

[0010] That is, the electrical quantities representing the operating state of the power grid and the thermal state quantities representing the state of each inverter are collected, a dynamic response model for each inverter is established according to the electrical quantities and the thermal state quantities, and its sensitivity parameters and dynamic capability boundaries are generated; based on the monitored power grid disturbance quantity and the sensitivity parameters, the total adjustment budget required by the station level is calculated; an explicit KKT condition-based closed distribution model is used to map the total adjustment budget to the control share of each inverter at a time within a single control period, and the passive constraints for ensuring stability and the distribution of the control share are solved in the same control period, so as to meet the passivity of the control transition path; according to the solved control share and stability control parameters, control instructions are issued to each inverter and executed; and according to the system response after control execution and the thermal state quantity, the model parameters and the dynamic capability boundaries of the next control period are updated.

[0011] Beneficial effects: The method solves the problem of control time lag by completing the budget distribution and stability constraint linkage solution in a single period; the passivity constraint guarantees the stability of the path, and improves the dynamic stability vulnerability; the capability boundaries are dynamically updated in combination with the thermal state quantity, the individual capability is matched, and the weak grid anti-interference performance is improved. The related technical effects will be described in detail below in combination with specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The flowchart of the desert photovoltaic power station inter-inverter collaborative anti-interference control method provided by the embodiments of the application is shown.

[0013] Figure 2The flowchart provided by the embodiment of the present application for solving the control share borne by each inverter and the stability control parameter to be configured.

[0014] Figure 3 The flowchart provided by the embodiment of the present application for constructing a dynamic passive envelope for the aggregated equivalent output impedance of the inverter group.

[0015] Figure 4 The flowchart provided by the embodiment of the present application for generating a stability control parameter.

[0016] Figure 5 The flowchart provided by the embodiment of the present application for jointly constructing a global optimization objective function with the total regulation budget as a constraint. DETAILED DESCRIPTION

[0017] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0018] It should be noted that the terms first, second, etc. in the specification of the present application and in the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms include and have and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to the clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0019] In order to solve the above problems, the applicant has conducted in-depth retrieval and analysis, and found that:

[0020] Specifically, the existing distributed iterative algorithm is difficult to meet the sub-100ms fast response requirement required by power grid transient stability in response speed; the traditional control method decouples power distribution and stability regulation, and cannot guarantee the stability of the system in the dynamic path of fast power regulation, which is easy to excite subsynchronous or high-frequency oscillation in a weak power grid background; the existing control ignores the real-time and drastic changes of the individual capacity of inverters in the desert environment, resulting in a disconnection between the control instruction and the real physical reachability of the inverter.

[0021] Further, the distributed algorithm based on consensus or ADMM, although has good scalability, but its convergence to the optimal solution process needs multiple rounds of information interaction and iterative calculation between each inverter (or agent). In large-scale photovoltaic power station, it will increase the communication overhead and accumulated calculation delay, so that its total decision time reaches hundreds of milliseconds or even seconds, which cannot meet the suppression demand of rapid dynamic process such as voltage drop and frequency mutation

[0022] Further, the existing control paradigm generally follows the logic of first distribution and then stability maintenance, determines the power distribution target of each inverter according to a certain criterion (such as the vertical coefficient or economic optimization), and then independently adjusts the damping and suppresses the oscillation through virtual impedance and other means. This method ignores the fact that the dynamic path from the current state point to the target state point may have an unstable region. Especially under weak power grid, the rapid and coordinated power regulation behavior of large-scale inverter group will change the aggregated equivalent impedance characteristics of the power station. Without a coordinated impedance shaping strategy, it is easy to trigger a new oscillation mode in the regulation process, leading to system collapse, that is, instability before reaching the destination.

[0023] Further, the existing control model uses static and conservative parameters, such as setting the capacity boundary of the inverter based on the equipment nameplate and fixed derating coefficient. However, in the desert environment, due to high temperature, wind sand, irradiation fluctuation and other factors, the real capacity boundary (such as maximum current, maximum power, power change rate, etc.) of each inverter is changing and has individual differences and nonlinear body. Issuing support instructions that exceed the current real capacity of the inverter that is close to overheating not only cannot achieve the expected support effect, but also may cause control loop saturation, output water, and other problems due to the unattainability of the instruction, which worsens the control quality and even damages the equipment. These three problems are coupled with each other, which together leads to the inadaptation of the existing coordinated control method to the desert photovoltaic power station scene.

[0024] In order to solve these problems, combined with Figures 1 to 5 The application is specifically illustrated by the following embodiments.

[0025] Embodiment one provides the application scene and system architecture of the coordinated anti-interference control method between inverters in desert photovoltaic power station, which serves as the basis for understanding the subsequent embodiments.

[0026] In this embodiment, this method can be applied to a centralized large-scale photovoltaic power station located in a desert area. The power station is connected to the public connection point (PCC) of the regional power grid through a long-distance high-voltage transmission line. Due to the remote geographical location and long transmission distance, the power station exhibits typical weak grid characteristics, that is, the equivalent short-circuit ratio at the PCC point is low (for example, the SCR value is usually lower than 3.0, and even lower than 1.5 in some operating conditions). The power station is deployed with hundreds or even thousands of photovoltaic inverters, and these inverter clusters are connected through the station's power collection line to jointly deliver active power to the power grid.

[0027] The particularity of the desert environment brings challenges to the operation of the inverter. For example, the high temperature and strong radiation during the day cause the junction temperature of the power semiconductor devices (such as IGBT) inside the inverter to approach the safety upper limit; the sandstorm weather not only affects the efficiency of the heat sink, but also causes physical wear to the equipment. The above factors cause the actual maximum power generation capacity, reactive power regulation range, and dynamic response rate of each inverter at any moment to be constantly changing, and there are differences between the states of each inverter.

[0028] In order to cope with the above challenges, the power station deploys a hierarchical collaborative control system. The system includes:

[0029] Inverter local controller: the controller of each inverter, responsible for executing upper-level instructions, controlling the switching action of power semiconductors, collecting its own electrical quantities (voltage, current) and thermal state quantities (such as heat sink temperature T _h , temperature rise rate dT _dt ).

[0030] Station-level coordination controller: usually deployed on the central monitoring system or dedicated edge computing device of the power station; communicates with all inverter local controllers through high-speed industrial Ethernet, periodically (for example, every 100 milliseconds) collects the state data of all inverters and the grid monitoring data at the PCC point; it is responsible for executing the collaborative disturbance control algorithm, calculating the adjustment task share and control parameters that each inverter needs to undertake in the current control period, and issuing these instructions to each inverter local controller for execution.

[0031] In this scenario, when the power grid is disturbed, such as PCC point voltage drop (ΔV<0) or frequency fluctuation (Δf≠0, RoCoF≠0) caused by remote faults, the station-level coordination controller will immediately start the collaborative disturbance control method, dynamically evaluate the available adjustment potential of all station inverters within the time scale of hundreds of milliseconds, and optimize the task allocation, achieving rapid and stable support for the power grid, so that any inverter executes the instructions within its own safe operating boundary.

[0032] This specification adopts unified symbols and data item conventions. Among them: V, I respectively represent the three-phase voltage and current phasor of the PCC point; ΔV, Δf, RoCoF respectively represent the voltage deviation, frequency deviation and frequency change rate; T _h , T _j , dT _dt respectively represent the heat sink shell temperature, device junction temperature estimation, temperature rise rate; SCR _proxy represents the equivalent short-circuit ratio proxy quantity; K _i represents the sensitivity vector of the i-th inverter, for example, containing the partial derivative of voltage to reactive power and the partial derivative of frequency to virtual inertia support ; B _i represents the dynamic capability boundary vector of the i-th inverter, for example, containing the maximum active power P _max , the maximum reactive power Q _max , the maximum apparent power S _max , the maximum current I _limit , and the power change rate ramp limit, etc.; H _req , Q _req respectively represent the total budget of virtual inertia and the total budget of reactive power required by the station level to suppress the current disturbance; H _i , Q _i then represent the virtual inertia and reactive power share allocated to the i-th inverter.

[0033] Embodiment two, a method for collaborative anti-disturbance control between desert photovoltaic power station inverters is provided, in a specific implementation, comprising the following steps:

[0034] Step 2.1, collect electrical quantities representing the operating state of the power grid and thermal state quantities representing the state of each inverter, and establish a dynamic response model for each inverter based on this, and generate sensitivity parameters and dynamic capability boundaries.

[0035] Specifically, the station-level coordination controller periodically (for example, with a period of 10-100 milliseconds) collects electrical quantities such as V, I from the power quality monitoring device at the PCC point, and collects thermal state quantities such as T _h , dT _dt from the local controller of each inverter, and optionally, environmental quantities such as wind speed and dust index from the weather station can also be fused.

[0036] On this basis, the controller performs multiple modeling tasks in parallel. For example, by applying the recursive least squares (RLS) algorithm to the short window data of V, I, the equivalent Thevenin impedance of the power grid is identified online, and the SCR _proxy characterizing the strength of the power grid is obtained. At the same time, by analyzing the relationship between the power response of each inverter and the state quantity of the power grid under natural small disturbances of the system, the sensitivity vector K_i . Furthermore, the thermal state quantities are input into a pre-trained, physics-data hybrid dynamic capability boundary model (which will be detailed in subsequent embodiments) to predict and generate the current real, executable dynamic capability boundary B _i .

[0037] It should be understood that, unlike employing fixed, offline-tuned parameters, this step, through online modeling, can capture the dynamic changes of grid operating conditions and individual inverter states, providing inputs for subsequent collaborative decision-making.

[0038] As another implementation, multi-modal measurements and state modeling, including: reading V, I and small-signal perturbation responses, identifying equivalent Thevenin impedance through phasor regression and recursive least squares, obtaining SCR _proxy . Using short-window small-signal method to calculate and , form the sensitivity vector of the i-th inverter, obtaining K _i . Reading T _h , dT _dt , wind speed and dust index (optional), outputting soft boundaries through monotonic thermal reserve network (MTRN), and then using analytical thermal model for upper and lower bound clipping and rate binding, obtaining executable boundary B _i .

[0039] Step 2.2, based on the monitored grid disturbance quantities and sensitivity parameters, calculate the total regulation budget required at the station level, and based on the total regulation budget, sensitivity parameters and dynamic capability boundaries, collaboratively solve the control share each inverter needs to bear and the stability control parameters need to be configured.

[0040] As an example, when a grid disturbance (such as ΔV, Δf, RoCoF exceeding a preset threshold) is monitored, the controller first calculates the total regulation budget H _i required to restore the grid state to the stable interval (for example, suppress RoCoF within 0.5 Hz / s, or control voltage deviation ΔV within ±5%) based on the disturbance quantity and sensitivity aggregation model K _req and Q _req . Further, the controller takes the total budget H _req , Q _req as the optimization objective, inputs the sensitivity K _i of each inverter (as a measure of benefit or cost) and the dynamic capability boundary B _i (as a constraint condition) into a specially designed collaborative optimization algorithm. This algorithm can solve the control share H _i , Q _iand stability control parameters (e.g. virtual impedance parameters Z _out_params ).

[0041] This step optimally matches the adjustment needs of the entire station with the individual capabilities of each inverter, avoiding the problems of traditional one-size-fits-all or simple average allocation, and instead allowing inverter with strong adjustment capability and healthy state to undertake more tasks, and allowing inverter close to derating or with low sensitivity to undertake less or even no tasks, while meeting the needs of the power grid, maximizing the overall benefits and operation safety of the entire station.

[0042] Another example, budget estimation and event triggering, the specific process is: reading ΔV, Δf, RoCoF, K _i , estimating the required budget for disturbance suppression according to the sensitivity condensation model, obtaining H _req , Q _req . According to the SCR _proxy and ΔV, RoCoF threshold, it is judged whether to enter the fast collaborative allocation; if T _h approaches the derating threshold, the hot collaborative flag is set, and σ(t) is generated.

[0043] Another example, self-organizing allocation within 100 milliseconds and passive execution of path, can use the following implementation scheme: reading H _req , Q _req , K _i , the executable boundary B _i , constructing the single mapping of the barrierized water level allocation, directly giving the initial share of each machine H _i , Q _i , avoiding iterative consensus.

[0044] Reading H _i , Q _i (initial value) and the last execution value, performing proximal projection according to the ramp in B _i and the current upper limit to obtain the share H _i , Q _i (after projection) that meets the rate reachability, and suppresses water beating and unreachable overshoot.

[0045] Reading H _i , Q _i (after projection) and B _i , jointly solving the virtual impedance and damping parameters of each machine, so that the aggregate Z _out is kept within the passive envelope, avoiding subsynchronous and high-frequency spikes.

[0046] Step S2.3, according to the control share and stability control parameters solved, issuing and executing control instructions to each inverter.

[0047] Exemplarily, the station-level coordination controller will H_i , Q _i , and Z _out_params are packaged into control instructions and issued to the corresponding ith inverter local controller through a high-speed communication network. After receiving the instructions, the inverter local controller parses and converts them into modifications of the parameters of its underlying control loop (such as the current loop). For example, H _i is converted into the virtual inertia parameter in the virtual synchronous machine control, Q _i is converted into the reactive power limit or response slope in the Q-V droop control, and Z _out_params is configured as a parameter of the virtual impedance link. These modifications will take effect immediately, causing the external characteristics of the inverter to change and collectively form support for the power grid.

[0048] Another example of the implementation of control issuance is also possible: H _i , Q _i (final execution value), and Z _out shaping parameters are issued to the current loop and grid connection controller of each inverter to complete the execution of this control.

[0049] Step 2.4, according to the system response after control execution and the thermal state quantity, feedback updates the model parameters and dynamic capability boundary of the next control period. Among them, according to the system response after control execution, the model parameters of the next control period are updated, including: after the execution of the control instruction, the actual response index of the power grid is collected, and it is compared with the total adjustment budget of this control period to quantify a budget mismatch quantity; based on the budget mismatch quantity, a multiplier fine-tuning term is generated, and the multiplier fine-tuning term is used to modify the initial value of the Lagrange multiplier when solving the initial control share in the next control period.

[0050] Optionally, after the execution of the control instruction, the station-level coordinated controller immediately reads the actual system response (such as whether the change trend of ΔV and Δf is effectively suppressed) of the PCC point, compares it with the expected control target (i.e., the total budget H _req , Q _req ), and quantifies the budget mismatch quantity. This mismatch quantity will be used to modify the internal parameters (such as the Lagrange multiplier) of the collaborative solving algorithm in the next control period, achieving rapid iterative optimization of the control effect. At the same time, the thermal state T _h , dT _dt of each inverter is also continuously monitored. When the temperature of a certain inverter rapidly rises due to the execution of the adjustment task and approaches the derating threshold, this information will trigger the real-time contraction of its dynamic capability boundary B _i . This updated and more conservative B _iThe constraint will be directly as the next control cycle, so that subsequent allocation tasks will not endanger the thermal safety of the inverter. This feedback update step constitutes the time dimension closed loop of the whole control method, so that the system has the ability of self-adaptation and self-learning, which can continuously track the system changes and maintain the long-term optimality of the control strategy.

[0051] Optionally, the thermal derating coordination boundary scheduling (capacity boundary feedforward layer) can also be implemented by the following scheme: reading σ(t), T _h , dT _dt and executable boundary B _i , and before approaching the derating threshold, the boundary is controlled to shrink and redistribute, and the future active / reactive / inertia budget is transferred to the unit with safer temperature, and the updated executable boundary B _i When a certain machine enters the derating area, read H _i , Q _i (final execution value) and executable boundary B _i , and execute fast recovery on the reactive and inertia shares, complete group redistribution in the next beat, and output executable boundary B _i (updated).

[0052] Optionally, fast read and multiplier fine-tuning, specifically including: reading ΔV, Δf and budget mismatch after execution, forming a small multiplier correction term and caching to the next initial value to avoid drift.

[0053] In other words, the desert photovoltaic power station inverter anti-interference control method can also be implemented in the following way:

[0054] Collect electrical quantities representing the state of the power grid and thermal state quantities representing the state of each inverter, and establish a dynamic response model for each inverter based on the electrical quantities and thermal state quantities, and generate its sensitivity parameters and dynamic capability boundary;

[0055] Based on the monitored grid disturbance and sensitivity parameters, calculate the total adjustment budget required at the station level;

[0056] Using a closed distribution model based on the explicit KKT condition, the total adjustment budget is mapped to the control share of each inverter at a time in a single control cycle, and the passive constraint for ensuring stability and the allocation of the control share are solved in the same control cycle to meet the passive control transition path;

[0057] According to the solved control share and stability control parameters, control instructions are issued to each inverter and executed;

[0058] According to the system response and thermal state quantity after control execution, feedback update the model parameters and dynamic capability boundary of the next control cycle.

[0059] Embodiment three provides an optional implementation of the three-stage chain collaborative allocation core architecture, and describes the internal three-stage structure of the collaborative solution link.

[0060] In this embodiment, the control share borne by each inverter and the stability control parameter to be configured are collaboratively solved, which specifically includes three stages that are sequentially cascaded. The three stages are sequentially executed within a control period (for example, within 100 milliseconds), forming a coupled link of allocation-projection-molding, and solving the problem layer by layer.

[0061] In the first stage, a one-time mapping allocation model is used to directly analyze the total regulation budget, the sensitivity parameter and the dynamic capability boundary into the initial control share of each inverter. In other words, after the total regulation budget, the sensitivity parameter and the dynamic capability boundary are directly analyzed by the one-time mapping allocation model, they are used as the initial control share of each inverter.

[0062] An example is that the model is a global optimization problem with constraints. In this example, the objective function comprehensively considers the regulation benefits (determined by the sensitivity K _i of each inverter and the operation cost (such as loss), and the constraint conditions include the total constraint (that is, ΣH _i = H _req , ΣQ _i = Q _req ) and the individual capability constraint (that is, H _i , Q _i must be within the given boundary B _i ). The individual capability boundary B _i is converted into a logarithmic barrier term in the objective function. By deriving the Karush-Kuhn-Tucker (KKT) condition of the optimization problem, an explicit analytical relationship between the share H _i , Q _i and the Lagrange multiplier (which can be understood as the water and electricity price of the whole network) can be obtained. Therefore, once the multiplier is determined through a small amount of calculation, the initial control share of all inverters can be calculated in parallel and directly.

[0063] It should be understood that this step is used to avoid the problem that a plurality of rounds of communication and iteration are required to converge in the prior art (such as the distributed method based on the consistency algorithm), thereby shortening the calculation delay. The output of this stage is the ideal initial control share H _i (initial value) and Q _i (initial value) of each inverter. However, this ideal value does not consider the rate limit of inverter power regulation, which may cause the instruction to be unable to be actually reached within a single control period. This problem will be solved by the next stage.

[0064] In one embodiment, the first stage (closed water level-barrier allocation CBF) includes the following implementation manners:

[0065] read H _req , station-level reactive power budget Q _req , K _i , executable boundary B _i , construct a barrierized water level problem and explicitly derive KKT conditions, solve by two multipliers closed-form (at most two Newton iterations), get H _i , Q _i (initial value).

[0066] In another embodiment, the implementation further includes: read K _i , executable boundary B _i , H _req * , station-level reactive power budget Q _req * , SCR _conf , B _conf , generate cost coefficients and barrier weights according to sensitivity and boundary, get a _i , executable boundary B _i , gamma _i , eta _h , eta _Q . Read a _i , executable boundary B _i , gamma _i , eta _h , eta _Q , H _req * , station-level reactive power budget Q _req * , executable boundary B _i , construct a barrierized water level problem of one mapping and explicitly derive KKT conditions, get two multipliers water level equations F _h (lambda _h )=0, F _Q (lambda _Q )=0 and closed-form relation of each machine. Read the upper shooting multipliers lambda _h_prev , lambda _Q_prev and bus shadow price agent (mapped by delta V, delta f), generate initial value for F _h , F _Q , get lambda _h 0 , lambda _Q 0 . Read F _h , F _Q , lambda _h 0 , lambda _Q 0 , perform at most two Newton iterations, solve initial value share of each machine, get H _i , Q_i (initial value) and compute the total error epsilon _sum Read H _i , Q _i (initial value) and executable boundary B _i Apply barrier clipping and numerical protection to components close to the upper bound, get H _i , Q _i (initial value, clipped).

[0067] The second stage, according to the rate and current limit value contained in the dynamic capability boundary, uses the proximal projection operator to correct the initial control share, to generate the rate corrected share that satisfies the dynamic reachability.

[0068] The share given by the first stage is the state target, but in actual control, the inverter needs time to transfer from the current state to the target state, which is limited by its power change rate (ramp) and current upper limit (I _limit ). The traditional clipping method (peak clipping for the part exceeding) will cause control chattering (water splashing phenomenon) and non-monotonic convergence of the path.

[0069] Specifically, we define the ramp and I _limit given by B _i as the dynamic reachable domain, that is, the set of all points that can be safely reached within the control period from the current share. Next, using the proximal projection operator, the initial control share H _i (initial value) and Q _i (initial value) obtained in the first stage are mapped to the point closest to them within the dynamic reachable domain; mathematically, the change path of the share is always monotonically approaching its allocation target, which is used to suppress chattering and ensure the smoothness and reachability of the control process.

[0070] The output of this stage is the share H _i (after projection) and Q _i (after projection) corrected by the rate and current constraints. This share is physically reachable, but its impact on the equivalent impedance characteristics of the power grid during dynamic execution, as well as the risk of oscillation it may cause, has not been considered.

[0071] According to another aspect of the present application, the second stage (rate conservation projection RGP) can also be: receiving H _i , Q _i (initial value) and the upper limit execution value, according to the ramp and I _limit within the executable boundary B _i , proximal projection is performed to obtain H _i , Q _i (after projection).

[0072] The third stage, the passivity requirement is taken as a hard constraint, the rate-modified share is linked to optimize, and the final executable control share and stability control parameters are generated synchronously.

[0073] Where the passivity is the criterion of system stability, representing that the system cannot generate energy to the grid at any time, but only consume or store. In weak grid environment, large-scale and fast power regulation of inverter group may change the equivalent output impedance of the system, easily exciting subsynchronous or high-frequency oscillation, endangering the system stability. The traditional method allocates power first and then independently adjusts the damping parameters, which cannot guarantee the stability of dynamic path.

[0074] An example is that the real-time evaluated grid strength SCR _proxy is read, and the aggregated equivalent output impedance Z _out of the inverter group is constructed. A dynamic passivity envelope (stability domain) is constructed. Further, the rate-modified share and the stability control parameters Z _out_params (such as virtual resistance, virtual inductance, etc.) to be solved are taken as joint optimization variables, and a lightweight joint optimization is solved under the hard constraint of the passivity envelope. While being as close as possible to the output share of the second stage, a set of Z _out_params is found, so that the equivalent output impedance at any time of the entire dynamic path from the current state point to the target state point always remains within the passivity envelope. The output of this stage is the final executable control share H _i (final execution value), Q _i (final execution value), and the matching stability control parameters Z _out_params . Accordingly, the power allocation and stability control are closely coupled, realizing the leap from end-point stability to path-wide stability, and improving the safety of collaborative control under weak grid.

[0075] Another example is that the third stage (passivity fence PEE and distribution linkage) can be realized by the following steps: reading H _i , Q _i (after projection), and the executable boundary B _i , solving Z _out_params in linkage in the same control period, completing the passivity envelope shaping of the equivalent grid-connected impedance, and obtaining Z _out_params .

[0076] Still another example is that the third stage can also be realized by the following steps: reading H _i , Q _i (after projection, stable), and the control bandwidth parameters, generating the dq small signal equivalent of each machine online to obtain Z _out_i approximation. Reading Z _out_i approximation, SCR _proxy , and σ _grid, set the positive real approximation and incremental passive envelope according to the weak network level, get passivity _envelope and constraint function. Read passivity _envelope , executable boundary B _i , construct D _i , G _V i, L _V i, the feasible region and loss / bandwidth constraint, generate a lightweight convex problem or closed approximation, get Z _out_params candidate. Read H _i , Q _i (Projection, stable) and Z _out_params candidate, passivity _envelope , in the same control cycle, get Z _out_params (final) and stability _evidence (evidence of stability). Read Z _out_params (final) and | |, according to the priority sequence and loss budget review, get Z _out_params (version).

[0077] The traditional method is to allocate first and then adjust the impedance; This step takes passivity as a hard constraint and associates with allocation, meets the transition path at any time within the passive envelope, realizes path passivity, and avoids subsynchronous / high frequency spikes.

[0078] As another optional implementation, the interactive solution includes three stages in turn, and can also include: constructing a one-time mapping allocation model, directly analyzing the total adjustment budget, sensitivity parameters and dynamic capability boundary into the initial control share of each inverter; Use the proximal projection operator to modify the initial control share to solve the dynamic unattainable problem that may exist in the initial control share generated by the first stage, generate the rate corrected share; Take the passivity requirement as a hard constraint, and optimize the rate corrected share to solve the instability problem that may exist in the control path formed by the previous two stages, and synchronously generate the final executable control share and stability control parameters.

[0079] Embodiment four provides a path stability control method based on passivity fence, especially the passivity fence (PEE) mechanism. This mechanism takes the passivity requirement as a hard constraint, optimizes the control share and stability control parameters, so that the inverter group always maintains stability during the whole dynamic adjustment process in response to disturbances, and avoids triggering new oscillations.

[0080] In one specific implementation, the passivity requirement is taken as a hard constraint, and the control shares are jointly optimized with rate corrections to generate the final executable control shares and stability control parameters in synchronization, which includes the following steps: based on the real-time evaluation of the grid strength, the dynamic passivity envelope is constructed for the aggregated equivalent output impedance of the inverter group.

[0081] wherein the aggregated equivalent output impedance refers to the equivalent complex impedance presented by all inverters in the photovoltaic power station in parallel from the grid side, and its expression is Z _out (s)=(Σ _i 1 / Z _out_i (s)) -1 , wherein Z _out_i (s) is the equivalent output impedance approximation of the ith inverter under the dq small signal model, and s is the Laplace operator. The passivity theory points out that if the real part of the impedance of the system is positive at all frequencies, the system is passive, which guarantees the interconnection stability with any passive network; accordingly, the dynamic passivity envelope is introduced, but an adaptive constraint is imposed.

[0082] Specifically, a dynamic passivity envelope is constructed for the aggregated equivalent output impedance of the inverter group, including: on the set of pre-set key frequency points, the passivity envelope is defined as a complex plane conic constraint that requires the real part and the imaginary part of the aggregated equivalent output impedance to satisfy a certain proportional relationship. Wherein the key frequency point set Ω is a frequency band that is prone to instability, which is pre-selected according to historical experience and system analysis, for example, covering the low frequency band of subsynchronous oscillation (such as 10-200Hz), and the high frequency band where LCL filter resonance may exist (such as 2-5kHz). On each frequency point ω in the set, the conic constraint imposed can be expressed as: Re{Z _out (ω)}≥α×|Im{Z _out (ω)}|+ζ. Wherein Re{Z _out (ω)} is the real part of the aggregated equivalent output impedance at the frequency point ω; Im{Z _out (ω)} is the imaginary part; α is the conic angle coefficient of the conic, which determines the allowed impedance phase margin, the larger α is, the more stringent the constraint is, and the more the impedance tends to be pure resistive; ζ is a small positive real offset, which is used to provide additional stability margin. This constraint defines a symmetric conic region centered on the real axis in the complex plane, and the phasor of Z _out (ω) must fall within the region.

[0083] In order to cope with the changes of grid operating conditions, the geometry and safety margin of the conic constraint are further dynamically adjusted according to the real-time evaluation of the grid strength and the severity level of the grid disturbance event, wherein the weaker the grid or the more serious the disturbance, the more stringent the constraint imposed.

[0084] Specifically, the cone angle coefficient a and the offset z are functions of the grid strength proxy SCR _proxy and the grid event flag s _grid . For example, a can be set as a function that monotonically increases as SCR _proxy decreases, such as linearly increasing from 0.1 to 0.6 when SCR _proxy drops from 3.0 to 1.5. Correspondingly, z can be temporarily raised by a fixed value when a severe voltage dip event is detected (s _grid is set to a high level). In this way, the height and thickness of the passive fence can be dynamically adjusted, giving control freedom when the grid is robust and imposing stability constraints when the grid is fragile.

[0085] After constructing the passive envelope, the final executable control share and the stability control parameter are solved as joint optimization variables under the constraint of the passive envelope, so that the dynamic path of the control share in the state transition always meets the stability domain defined by the envelope; or the final executable control share and the stability control parameter are set as joint optimization variables, and the passive envelope is solved as a constraint condition, so that the dynamic change path of the control share in the state transition stage always meets the stability domain defined by the passive envelope, and the control share and the stability control parameter meeting the constraint are solved. Specifically, it includes:

[0086] On the set of critical frequency points, the complex plane cone constraint is linearized to form a set of linear inequality constraints. Since the original cone constraint Re{Z _out (ω)}≥a×|Im{Z _out (ω)}|+z is nonlinear, in order to facilitate fast solving, it can be converted into a linear inequality constraint by first-order Taylor expansion or polyhedron approximation method near its current working point at each frequency point w. The form of these constraints is A x ≥ b, where x is the stability control parameter vector to be optimized, A and b are the coefficient matrix and vector calculated according to the current system state. Correspondingly, based on the linear inequality constraints, and by adding an additional loss upper limit constraint derived from the dynamic capability boundary, a lightweight convex optimization problem with the final executable control share and the stability control parameter as the solving variables is constructed.

[0087] Wherein, the stability control parameter vector x _i may include the virtual resistance D _i , the virtual conductance G _ V i, the virtual inductance L _V i, etc. of the i-th inverter. The additional loss upper limit constraint P _loss (x _i )≤P _cap，i, to limit the extra power loss due to the introduction of these virtual impedance parameters, with an upper bound P _cap，i determined dynamically by the thermal margin of the dynamic capability boundary B _i of the inverter. The resulting lightweight convex optimization problem, e.g., a quadratic programming (QP) or a second-order cone programming (SOCP) problem, can have an objective function set as minimizing the regulation cost, e.g., min∑ _i w _i ×||x _i || _2 , i.e., minimizing the two-norm of all inverter stability parameters, to seek the minimum control cost. On this basis, by solving the convex optimization problem, the control share and stability control parameters that meet the stability and loss boundary are obtained. Among them, the station-level coordination controller uses an embedded convex optimization solver to solve the problem within milliseconds to obtain the stability control parameters x _i , i.e., Z _out_params , which each inverter needs to configure finally.

[0088] As a preferred embodiment, when generating the stability control parameters, differentiated configuration of the regulation task is also included: based on the sensitivity parameters of each inverter, the influence effectiveness of each inverter on the key indicators of the power grid is evaluated, and the priority of stability regulation for each inverter is set accordingly. Or, based on the sensitivity parameters, the priority of stability regulation for each inverter is set, especially the degree of influence on the key indicators of the power grid (such as voltage-reactive power sensitivity | _i ). For example, the inverters can be sorted according to the size of the voltage-reactive power sensitivity | _i . The higher the sensitivity of the inverter, the greater the influence of its regulation action on the PCC point voltage, and therefore the higher the priority.

[0089] According to the set priority, the stability control parameters are configured differently, and the main damping regulation task is allocated to the inverters with high priority, so as to maximize the group regulation effect while meeting the total loss budget of the station.

[0090] Specifically, when solving the above convex optimization problem, different weights w _i may be set for inverters of different priorities, or post-processing allocation based on priority is performed after solving. High-priority inverters are allowed to configure larger virtual resistance D _i (i.e., to undertake the main damping task), while low-priority inverters are kept as original as possible and only a small amount of adjustment is made. In this way, the total introduced loss Σ_i P _loss (x _i ) not exceeding the budget P _loss_budget , realize the group damping effect.

[0091] In some optional embodiments, if the above convex optimization problem has no solution under certain extreme working conditions, i.e. no set of parameters can satisfy all the constraints at the same time, the system will start a fallback adjustment mechanism. For example, the controller will send a feedback signal to the second stage (RGP) to request it to reduce the step size of the share correction; or send a signal to the first stage (CBF) to request it to increase the barrier weight to generate a more conservative initial share. This cross-stage feedback adjustment mechanism improves the robustness and adaptability under complex working conditions.

[0092] The present application takes passive requirements as a hard constraint, and optimizes the control share and stability parameters in linkage, solving the problem of missing dynamic path stability caused by decoupling of power distribution and stability control. The method closely combines abstract algorithms (lightweight convex optimization) and specific power grid technical fields (subsynchronous oscillation suppression), wherein the input parameters are the SCR _proxy and the rate-corrected share representing the control target, and the output result is a specific virtual impedance Z _out_params . At preset key frequency points, a complex plane conic constraint (passive envelope) dynamically adjusted according to the strength of the power grid is constructed, and a joint optimization problem with the envelope and the upper limit of the loss as constraints is solved, so that the aggregated equivalent output impedance of the entire dynamic adjustment trajectory of the inverter group from the current working point to the target working point is always locked within the preset stability domain. In a weak power grid environment, such path full-range stable design eliminates the problem of exciting subsynchronous or high-frequency oscillation due to large-scale power rapid adjustment, solves the problem of instability before reaching the destination in the prior art, and improves the survival ability and power grid friendliness of the power station under extreme working conditions.

[0093] Embodiment five provides an exemplary scheme of a one-mapping type control share allocation method based on KKT conditions, which sets forth a closed-form water level-barrier allocation algorithm for realizing fast and accurate allocation of control share within hundreds of milliseconds.

[0094] In the present embodiment, a one-mapping allocation model is constructed, and the total adjustment budget, the sensitivity parameter and the dynamic capability boundary are directly analyzed as the initial control share of each inverter, which specifically includes: the dynamic capability boundary is converted into a logarithmic barrier term, and a cost term related to the sensitivity parameter is combined to construct a global optimization objective function with the total adjustment budget as a constraint. The global optimization objective function is used to complete the allocation of the total adjustment budget with the minimum comprehensive cost.

[0095] Exemplarily, the objective function J can be constructed as follows: J = {H i , Q i}min∑ i (C i (H i , Q i )- PH(H i , B i )- PQ(Q i , B i )) ; s.t.∑ _i H _i = H _req ,∑ _i Q _i = Q _req ; wherein, C _i (H _i , Q _i ) is the cost term of the i-th inverter, P _h (H _i , B _i ) and P _Q (Q _i , B _i ) are the logarithmic barrier terms corresponding to virtual inertia and reactive power share, respectively.

[0096] Specifically, a global optimization objective function is jointly constructed with total regulation budget as a constraint, including: generating a cost term corresponding to control effectiveness and operating efficiency based on sensitivity parameters and inverter estimated loss. For example, the cost term C _i may be specifically embodied as a _i / H _i + b _i / Q _i + γ _i × H _i / Q _i . Wherein, the coefficient a _i is related to frequency regulation effectiveness and active side loss, a _i is proportional to 1 / | | _i , that is, the higher the frequency sensitivity of the inverter, the lower the cost of the allocated unit H _i . Similarly, b _i is related to voltage regulation effectiveness and reactive side loss, and is proportional to 1 / | | _i . γ _i represents the cross-coupling effect and corresponding loss between active-reactive regulation.

[0097] Further, based on the credibility of the dynamic capability boundary and the confidence of the grid strength evaluation, the weight of the logarithmic barrier term is dynamically set, wherein the lower the credibility or confidence, the greater the barrier weight applied, thereby enhancing the robustness of the allocation. In other words, the weight of the logarithmic barrier term is determined based on two factors, namely the credibility of the dynamic capability boundary and the confidence of the grid strength evaluation; the corresponding influence relationship is: the lower the credibility or confidence, the greater the barrier weight applied. The form of the logarithmic barrier term is P _h = η _h × log(H _max，i - H _i ) and P _Q = η _Q × log(Q _max，i - Q _i ). When the allocated share H _i or Q _i approaches its boundary H _max，i or Q _max，i (sourced from B _i ), the term tends to negative infinity, forming a penalty in the objective function and preventing the share from crossing the boundary. The weights η _h and η _Q are dynamically set, for example, η _h = η _0 × (1 / B _conf_i ) × κ(SCR _conf ). Wherein η _0 is the basic weight; B _conf_i is the credibility of the dynamic capability boundary B _i , if the boundary prediction model has high prediction uncertainty for the current state, B _conf_i is low, and the weight η _h increases, making the allocation result more conservative; κ(SCR _conf ) is a function of the credibility of the grid strength SCR _proxy , if the grid strength recognition result is unstable, κ increases, and a stronger barrier is also applied.

[0098] Further, the Karush-Kuhn-Tucker (KKT) condition of the global optimization objective function is derived to obtain an explicit analytical relationship between the initial control share and the Lagrange multiplier, and the initial control share is directly solved based on the relationship; or, the dynamic capability boundary is converted into a logarithmic barrier term, combined with a cost term related to the sensitivity parameter, to jointly construct a global optimization objective function with the total adjustment budget as a constraint; by deriving the Karush-Kuhn-Tucker condition of the global optimization objective function, the original multi-agent consensus problem is converted into a water level equation about two core Lagrange multipliers for solving, and the initial control share is obtained.

[0099] By taking the partial derivatives of the Lagrangian function with respect to H _i , Q _i and the Lagrange multiplier λ _h , λ _Q and setting them to zero, a set of algebraic equations can be obtained. This set of equations directly reveals the optimal shares H _i , Q _i of each inverter as an explicit function of its own parameters (a _i , b _i , B _i etc.) and the global water level signal λ _h , λ _Q . In other words, once λ _h and λ _Q are determined, H _i and Q _i can be directly calculated by simple algebraic expressions without iteration. While solving λ _h and λ _Q , one can solve the total quantity constraints Σ _i h _i (λ _h )=H _req and Σ _i q _i (λ _Q )=Q _req , which can be quickly converged by one or two iterations of Newton's method.

[0100] To further accelerate the solution, the process of directly solving the initial control shares based on the relationship also includes: setting a dynamic initial value for the solution of the Lagrange multiplier, which is composed of the final value of the multiplier in the last control period and a feedback correction term. Specifically, the initial value λ _h 0 of the Newton iteration in the current period is set as λ _h_prev + Δλ _h , where λ _h_prev is the final multiplier value obtained by solving in the last control period (e.g. 100 milliseconds ago). Due to the continuity of the system state, this value is a very good estimate of the current optimal solution.

[0101] where the feedback correction term is generated according to the mismatch between the total adjustment budget and the actual system response after the execution of the control command in the last period. That is, the feedback correction term is obtained based on the mismatch between the total adjustment budget and the actual system response after the execution of the control command in the last period. Specifically, the station-level controller compares the budget H _req_prev of the last period with the actual frequency response improvement to obtain the budget mismatch epsilon _h; the multiplier fine-tuning term Δλ is generated by a proportional-integral (PI) controller _h ; wherein the fine-tuning term is used to compensate for control errors caused by model inaccuracy or time delay, etc.

[0102] Further, starting from the dynamic initial value, the explicit analytical relationship is used for solving, and the initial control share is quickly converged. Since the initial value is close to the optimal solution, the Newton iteration of λ _h , λ _Q can be converged to a high enough precision within two times, and the initial shares H _i , Q _i of all inverters can be calculated in parallel through the explicit relationship. The time complexity of the whole process is approximately O(n), which is much lower than the O(n 2 ×k) of the traditional distributed iterative algorithm, and meets the control requirement of hundreds of milliseconds.

[0103] As an optional implementation, after obtaining the initial share, a feasible region clipping step can be added. For example, if the calculated H _i is very close to H _max，i , it is forced to be set to H _max，i - δ _h , wherein δ _h is a small safety margin (such as 1%-5% of the boundary), to avoid numerical overflow or excessive sensitivity to the barrier function.

[0104] The present application solves the problem of time lag in control decision caused by the distributed iterative algorithm in the background technology by constructing a one-time mapping barrier water level allocation model and using the explicit analytical relationship of KKT condition for solving. Specifically, the dynamic capability boundary B _i of each inverter is converted into a logarithmic barrier term in the objective function, and the global adjustment budget H _req , Q _req is taken as a total constraint, and by solving the Lagrange multiplier, the optimal initial share of all inverters can be directly and analytically obtained in parallel. Further, the final value of the multiplier of the last period is adopted and combined with the budget mismatch feedback to dynamically set the initial value of the current period, so that the solving process can be converged within two Newton iterations. The communication and calculation complexity of the traditional distributed algorithm O(n 2 ×k) is reduced to the local calculation complexity of approximately O(n), so that the optimal coordinated decision time of thousands of inverters in the whole station can be controlled within hundreds of milliseconds. In the scenario of desert photovoltaic power station responding to weak grid transient disturbance, the super-high decision-making capability enables the support response of the power station to timely suppress the first drop of voltage or frequency, and provides transient stability margin for the power grid.

[0105] Example 6 provides a specific implementation scheme for the rate conservation and chatter suppression method based on near-end projection, and elaborates on the rate conservation projection (RGP) algorithm; the initial control share is a static target, and does not consider the strict limitation of the current and power change rate of the inverter's internal physical components (such as power modules, inductors, etc.) when the inverter moves from the current operating point to the target point; this algorithm can make the generated commands smooth, reachable and chatter-free.

[0106] In one specific implementation, based on the rate and current limits contained in the dynamic capability boundary, the initial control share is corrected using a near-end projection operator to generate a rate-corrected share that satisfies dynamic reachability. The specific implementation includes: defining the rate and current limits as a dynamic reachability domain, and using a near-end projection operator to map the initial control share to the point closest to it within the dynamic reachability domain to form the rate-corrected share.

[0107] The dynamic reachability domain refers to the region X, which is the actual execution share of the inverter in the previous control cycle. _i_prev (where X represents H or Q) as the starting point, in the control period T _s Within this range, considering both the maximum allowable rate of change (ramp) and the maximum allowable current (I), _limit After that, all share points X that can be safely reached. _i A set; it will change with X _i_prev and B _i The near-end projection operator Prox(·) is a mathematical tool that updates in real time based on changes in the point (here, the initial control share X). _i_Target Find the nearest point in the specified set (in this case, the dynamically reachable domain). Make the corrected share X _i_new Is it towards target X? _i_Target A forward step, the largest possible step, but without crossing the line.

[0108] Specifically, this process can be decomposed into separate or combined projections of rate and current. The reactive power component Q... _i Taking the correction as an example, from the dynamic capability boundary B _i Read the reactive power change rate limit ramp _Q (The unit is usually Mvar / s); based on this, the calculation is performed in a single control cycle T. _s Maximum permissible variation ΔQ _max =κ _Q ×ramp _Q ×T _s ; where κ _Q As a safety factor, its value is usually between [0.6, 0.9], to reserve a certain margin.

[0109] For the initial reactive power component Q _i_TargetThe calculation process for velocity projection can be expressed as: ΔQ = clip(Q _i_Target -Q _i_prev -ΔQ _max ΔQ _max ), to obtain the share Q after velocity projection _rate =Q _i_prev +ΔQ. The clip(value, min, max) function restricts value to between min and max; this step ensures that the change in share does not exceed ΔQ. _max .

[0110] B _i Maximum current limit I _limit Based on the current PCC voltage V and active power P _i This is converted into an instantaneous limit Q on the reactive power component. _imax For example, Q _imax =sqrt((S _max ) 2 -(P _i ) 2 ), where S _max =V×I _limit ; For Q _rate Perform another amplitude limiting operation: Q _i_new =clip(Q _rate -Q _imax Q _imax ).

[0111] As described above, this mapping process ensures that the control share of each inverter monotonically approaches its allocation target along the time axis, suppressing control chattering caused by direct limiting. Near-end projection searches for the optimal approximation throughout the feasible region, guaranteeing the monotonicity of the convergence path, avoiding the phenomenon of the share hovering near the limit, and improving the stability of the control process.

[0112] To further address potential control chattering caused by multi-machine interactions or model mismatch, this invention also provides a preferred implementation, namely, generating rate-corrected shares, which further includes: monitoring the time series of each inverter's control shares, detecting whether the sign of its first-order difference has flipped, and if so, generating a chattering trend marker. Accordingly, the system records the share execution value Q for each control cycle. _i (k). And calculate its first-order difference sequence ΔQ. _i (k)=Q _i (k)-Q _i (k-1). Check sign(ΔQ) _i (k)) and sign(ΔQ) _i(k-1)) are of different signs. For example, if the last cycle share is increasing (the difference is positive) and this cycle becomes decreasing (the difference is negative), it is considered that the chattering trend may occur, and the chattering trend flag flag_damp is set for the inverter.

[0113] In some embodiments, according to the chattering trend flag, an adaptive step factor less than 1 is used to reduce the share increment generated by the proximal projection operator, and the control share of the last cycle is updated based on the reduced increment to obtain the final rate-corrected share. In other words, based on the chattering trend flag, an adaptive step factor is used to reduce the share increment, and the control share of the last cycle is updated to obtain the final rate-corrected share; wherein the adaptive step factor is set to be less than 1; and the reduction of the share increment is generated by the proximal projection operator.

[0114] When flag_damp is set, the system will activate the active damping mechanism. At this time, the final share update rule will be adjusted from Q _i_new = Q _i_prev + ΔQ to Q _i_new = Q _i_prev + μ _i × ΔQ. Wherein μ _i is an adaptive step factor, whose initial value is 1, and will be dynamically reduced after detecting the chattering trend, for example, μ _i ← max(0.3, μ _i × 0.8); the active damping mechanism suppresses the oscillation trend. When no sign flip is detected for consecutive multiple cycles, μ _i will gradually recover to 1.

[0115] As an optional solution, the detection of chattering can not only be based on the first-order difference, but also be based on higher-order difference or frequency domain analysis method. For example, by performing short-time Fourier transform on the share sequence, if there is an abnormal energy concentration in a certain frequency band, it is also considered that the chattering trend occurs. The adjustment strategy of the step factor can also use a more complex control law, for example, it is linked with the amplitude and frequency of the chattering. Each of the above variants is within the protection concept of the present application.

[0116] This invention employs a rate-conserving near-end projection operator to address the common problems of control chattering (splashing) and non-monotonic convergence paths in traditional control methods, which are often caused by simple amplitude limiting. Specifically, this method transforms the rate and current limits provided by a high-fidelity dynamic capability boundary model into a dynamically reachable domain. When the ideal command generated by cooperative allocation exceeds this domain, the near-end projection operator, by solving a micro-optimization problem, pulls the command back to the point within the feasible domain closest to the ideal command. This correction method, based on minimum distance mapping, ensures that the time series of control shares monotonically approximates the target, avoiding chattering caused by repeated jumps of commands near constraint boundaries. Furthermore, by monitoring the first-order difference sign of the share series and introducing an adaptive step factor, oscillations caused by complex factors such as multi-machine coupling can be actively suppressed. In the actual operation of desert photovoltaic power plants, this chatter-free control execution not only improves the system's response quality to grid disturbances but also reduces the switching stress on the power semiconductor devices inside the inverter, extending equipment lifespan.

[0117] Example 7: Optional technical solutions for providing a hybrid modeling and forward-looking thermal management method for dynamic capability boundaries, especially a physical-data hybrid modeling method, and a cooperative scheduling strategy based on this model that can prevent thermal runaway.

[0118] In one specific implementation, a dynamic response model is established for each inverter to generate its dynamic capability boundary. This specifically includes: inputting thermal state variables into a neural network with monotonic structure constraints to infer a soft boundary set. The monotonic structure constraints are manifested in that the weights of the neural network are constrained to be monotonically non-increasing with respect to the input representing temperature and monotonically non-decreasing with respect to the input representing wind speed.

[0119] Among them, the neural network is preferably a monotonic heat reserve network, which receives data from the radiator shell temperature T. _h Temperature rise rate dT _dt The thermal state vector θ is composed of optional environmental quantities (such as wind speed and dust index). _T As input; its internal weights are subject to specific structural constraints, for example, with respect to temperature T. _h The relevant weights are forced to be non-positive, and the weights related to wind speed are forced to be non-negative; this ensures that the network output, i.e., the available capacity of the inverter, monotonically remains unchanged with increasing temperature and monotonically remains unchanged with increasing wind speed. This step incorporates prior physical knowledge into the data-driven model, avoiding black-box prediction and enhancing the model's generalization ability and interpretability. The network outputs a soft boundary set B. _soft , including S _max_soft (apparent power soft boundary), I _limit_soft (Soft boundary of current), etc.

[0120] The soft boundary set is input into the analytical thermal model, which is then used to trim the physical upper and lower bounds to obtain the final dynamic capability boundary. The analytical thermal model is based on the R-axis of the semiconductor device and heat dissipation system. _Th -C _Th (Thermal resistance-thermal capacity) equivalent circuit model; soft boundary set B receiving the MTRN output. _soft The model undergoes physical verification and tailoring based on first principles (energy conservation, Fourier's law of heat conduction, etc.). For example, the model will be adjusted based on the current junction temperature T. _j and the maximum allowable junction temperature T _j_max And the thermal time constant of the device, to calculate the time required to satisfy T within a short time window in the future (e.g., 1 second). _j The absolute upper limit of the inverter's output current, without exceeding the limit. If B... _soft The given I _limit_soft If the limit is exceeded, it is clipped to that limit value. The dynamic capability boundary B is generated through the concatenation and fusion of the data model and the physical model. _i It not only utilizes the data model's ability to learn the effects of complex environments, but also ensures the physical rigidity of its output results, thereby improving its accuracy and reliability.

[0121] Based on the above boundary model, this invention further provides a thermal management method that proactively performs boundary contraction and capacity margin transfer across units before the inverter actually derating.

[0122] As an example, the method further includes: calculating the time to reach the threshold based on the thermal state parameters of each inverter and a preset derating temperature threshold; specifically, this includes: subtracting the current temperature given by the thermal state parameters from the preset derating temperature threshold to obtain a temperature margin; and dividing the temperature margin by the temperature rise rate given by the thermal state parameters to determine the time to reach the threshold. The calculation formula is: t _Thr =(T _Thr -T _h ) / max(dT _dt , ε). Where, t _Thr That is, the time to reach the threshold; T _Thr The heatsink temperature threshold set by inverter manufacturers that requires them to begin derating; T _h dT represents the current real-time monitored radiator temperature. _dt ε represents the current rate of temperature rise; ε is a very small positive number used to prevent the denominator from being zero.

[0123] Next, if the time to reach the threshold is less than a preset time, a boundary contraction ratio is determined based on this threshold time through a preset mapping relationship. This ratio is then used to contract the dynamic capability boundary of the inverter, forming a capacity margin to be transferred. The preset time is a warning threshold, for example, 300 seconds. When t..._Thr Less than 300 seconds, the system considers that the inverter exists the risk of thermal runaway. At this time, the boundary pre-contract mechanism will be started. The contraction ratio β is a function of t _Thr , for example, it can be a linear or nonlinear mapping relationship, t _Thr is shorter, β is larger, and the contraction is more severe. For example, Q _i in B _max will be updated to β × Q _max . The capacity (1-β) × Q _max contracted forms the capacity margin to be transferred.

[0124] According to the thermal state quantity of other inverters, the capacity margin to be transferred is allocated to the dynamic capacity boundary of other inverters to generate an updated dynamic capacity boundary. The station-level coordinated controller will evaluate the thermal state of all other inverters in the power station, i.e. their t _Thr (or simply temperature margin). Accordingly, according to the on-demand allocation and capacity priority principle, the above capacity margin to be transferred is allocated to those inverters with healthy thermal state (such as t _Thr is the longest) to temporarily expand their dynamic capacity boundary B _j . This active cross-unit capacity transfer can reduce the burden on the inverter that is about to overheat without affecting (even improving) the overall regulation capacity of the station, avoiding unplanned derating events and improving the overall power generation and operation stability of the station. In some embodiments, the allocation of capacity margin can not only be based on the thermal state, but also on the sensitivity K _j , and be preferentially allocated to those inverters with higher regulation benefits to achieve global optimization.

[0125] The present application solves the problem of the decline of the overall regulation capacity or the loss of power generation of the entire power station caused by the approach of individual inverters to the thermal limit in the desert high-temperature environment through the cross-unit thermal standby capacity transfer method. The method does not wait for the passive derating of the inverter to actually occur, but calculates the threshold time t _Thr to predict the risk in advance. Once it is predicted that an inverter exists the risk of overheating, the system does not limit the inverter, but actively and optimally transfers the part of the capacity margin released by the boundary pre-contract of the inverter to the healthy inverters that still have the capacity according to the real-time thermal state and regulation benefit (sensitivity K _j ) of other inverters. This collaborative scheduling converts the loss originally belonging to a single device into available standby capacity shared by the entire station. Under the premise of ensuring the thermal safety of the entire station, the total available regulation capacity and the total power generation of the entire station at any time are maintained, the power generation loss caused by the harsh desert environment is reduced, and the asset utilization rate and economic benefit of the power station are improved.

[0126] On this basis, the application also adopts the combination of high-fidelity dynamic capability boundary modeling and rate conservation proximal projection correction to solve the control mismatch and equipment safety problems caused by the disconnection between control instructions and the real physical capability of the inverter. Further, by fusing a neural network with monotonic physical constraints and an analytical thermal model, the system can depict the real and dynamic capability boundary B _i of each inverter in a harsh desert environment (such as high temperature and sandstorm) in real time, and solve the problems of waste of healthy inverter capability and overloading of endangered inverters caused by the use of static and conservative parameters. On this basis, when the initial instruction of coordinated allocation is issued, the rate conservation projection algorithm converts the rate and current limits defined in B _i into a dynamic reachable domain, and uses a proximal projection operator to make the final issued instruction increment located within this domain; this correction method mathematically guarantees monotonic convergence, reduces control chattering, and improves control quality. The two work together to not only make each instruction of coordinated control have a clear target, but also make each instruction be within the capability, maximize the available regulation potential of the whole station, smooth the control process, ensure the safety of equipment, and improve the overall power generation efficiency and asset health level of the power station.

[0127] Embodiment eight provides an optional solution of real-time power grid strength evaluation and its adaptive application in coordinated control, which describes the strength and weakness of the perceived power grid, and dynamically adjusts the strength and conservativeness of the core coordinated control strategy based on the perception results, for realizing high adaptability to different power grid conditions.

[0128] In a specific embodiment, the method further comprises applying recursive least squares estimation to the electrical quantities to generate an equivalent short circuit ratio proxy quantity representing the strength of the power grid in real time. Accordingly, the traditional power station control strategy is based on an offline set fixed short circuit ratio (SCR) value, which cannot reflect the real-time dynamic changes of the strength of the power grid caused by factors such as power grid topology changes and generator start-stop. Therefore, the application introduces online real-time estimation, and the specific implementation is as follows: the station-level coordination controller collects the three-phase voltage v(t) and three-phase current i(t) waveforms of the PCC point at a high sampling rate (for example, 10 kHz). After obtaining the power grid fundamental frequency and phase through a phase-locked loop, the short window Fourier transform or oversampling phasor measurement unit algorithm is used to extract the time series of the fundamental voltage phasor V and current phasor I.

[0129] The system uses an equivalent Thevenin circuit to simulate the characteristics of the external power grid, that is, V = E _Th -Z _Th ×I. Wherein, V and I are measurable PCC point phasors, E _Th (equivalent Thevenin potential) and Z _Th(Equivalent Thevenin impedance) is the unknown to be identified. By performing a difference operation on the equation, we can obtain ΔV = -Z. _Th ×ΔI, eliminating the effect on E _Th The system continuously monitors minute natural fluctuations in V and I, generating a series of ΔV and ΔI data pairs.

[0130] The above difference equation is written in the standard form of the least squares method and solved using the Recursive Least Squares (RLS) algorithm. The RLS algorithm is an online, iterative parameter estimation algorithm that updates the previous parameter Z using the new data pair (ΔV, ΔI) at each new sampling point. _Th This algorithm provides an estimate without storing all historical data. It is computationally inefficient, converges quickly, and can track time-varying parameters, making it suitable for this online application scenario.

[0131] In obtaining Z _Th After obtaining the estimated value, calculate the short-circuit capacity of the power grid, SCC=V. _n 2 / |Z _Th |, where V _n The rated voltage at point PCC is used; the equivalent short-circuit ratio surcharge (SCR), characterizing the grid strength, is calculated. _proxy =SCC / S _n S _n Rated capacity of the photovoltaic power plant; SCR _proxy The value is updated every second or every few hundred milliseconds, reflecting the strength of the power grid in real time.

[0132] As a preferred implementation, to improve identification accuracy when grid background disturbances are low, the station-level coordination controller can proactively issue safe, weak disturbance injection commands to the inverter group. For example, it can instruct all inverters to superimpose a pseudo-random binary sequence (PRBS) signal with an amplitude not exceeding 1% of the rated power onto their reactive power commands. Since the statistical characteristics of this injected signal are known, the RLS algorithm can extract Z from the background noise. _Th Information with higher confidence level (SCR) _conf (Higher) SCR _proxy .

[0133] Furthermore, based on this equivalent short-circuit ratio surrogate quantity, the weight of the logarithmic barrier term used in the collaborative solution and the constraint strength of the dynamic passive envelope are adjusted synchronously. This is specifically reflected in two aspects: adjustments to the collaborative allocation strategy; in the CBF allocation algorithm, the weight η of the logarithmic barrier term is set to SCR. _proxy The function. When SCR _proxyWhen the grid is weak, the value of η will automatically increase. This will make the barrier function stronger, resulting in more conservative allocation results, and make the operating point of each inverter away from its capacity boundary, leaving more stability margin to deal with the risk of interaction oscillation under weak grid. The adjustment of the path stability constraint, in the PEE mechanism, the geometry of the passive envelope (the cone angle coefficient α of the conic constraint) is directly determined by SCR _proxy . When SCR _proxy is low, the value of α will automatically increase, and the phase constraint of the aggregated output impedance of the inverter group will become more stringent, requiring it to be closer to pure resistive; it can suppress the subsynchronous oscillation that is prone to occur under weak grid, providing adaptive stability guardrails for large-scale power regulation. From the above, the present application builds a complete closed loop from real-time perception to adaptive decision-making, making the entire collaborative anti-disturbance control method dynamically adjust its behavior mode.

[0134] By introducing the online evaluation and application of real-time grid strength (SCR _proxy ), the present application solves the problem of poor scene adaptability caused by the offline and fixed grid parameter setting of existing control strategies. Specifically, the method continuously identifies the equivalent Thevenin impedance of the grid through the recursive least squares (RLS) algorithm, and quantifies the strength of the grid in real time. This SCR _proxy , as a key adaptive adjustment factor, is input to both the collaborative allocation (CBF) and the path stability (PEE) core modules. When the grid becomes weaker (SCR _proxy decreases), the system automatically and synchronously performs two adjustments: first, the weight of the logarithmic barrier term in the CBF model is increased, making the share allocation more conservative and reserving more stability margin for the inverter; second, the passive envelope constraint in the PEE model is tightened, requiring the aggregated impedance of the inverter group to exhibit stronger damping characteristics. This design enables the entire collaborative control system to have the ability to balance with the situation, automatically relaxing the constraints to pursue dynamic performance and economic benefits when the grid is strong, and automatically tightening the constraints to prioritize system stability when the grid is weak. This not only enables the present application to run safely and stably under various unknown or changing grid conditions, but also avoids the engineering cost of extensive field debugging and parameter re-tuning required by traditional methods.

[0135] Embodiment Nine, a numerical calculation case of collaborative anti-disturbance control is provided, which is specific and can be reproduced step by step, as follows:

[0136] Assume that a photovoltaic power station is composed of 3 same type inverters (INV1, INV2, INV3), and the rated capacity of the power station S _n = 30 MVA. Currently, a grid disturbance occurs, and a voltage drop of -0.15 p.u. is monitored at the PCC point, and the frequency change rate RoCoF reaches 0.5 Hz / s.

[0137] Step S9.1, input generation.

[0138] Assume measured SCR _proxy =1.5 (typical weak grid). The station-level controller obtains the status of each inverter as follows: {inverter, Q _i_prev (Mvar), _i , Q _max，i (Mvar), Q _min，i (Mvar), ramp _Q，i (Mvar / s)}={ (INV1, 0, 0.008, 10, -10, 5), (INV2, 0, 0.008, 5, -5, 4), (INV3, 0, 0.007, 10, -10, 5)}; where INV2 has its dynamic capability bound B _2 limited due to poor heat dissipation.

[0139] Step S9.2, budget calculation.

[0140] To compensate for the voltage dip of -0.15 p.u., consider the weak grid correction factor ψ(1.5)=1.2, calculate the total reactive power budget Q _req ≈0.15 / (avg(|B |))x1.2≈22.5Mvar. For simplicity, take Q _req =20Mvar. (calculation process omitted), assume that H _req =30MWs is calculated.

[0141] Step S9.3, collaborative solution.

[0142] Take reactive power as an example, the cost coefficient b _i is inversely proportional to | | _i , and the barrier weight η _Q is set to a higher value due to SCR _proxy =1.5. The boundary Q _max，2 =5 of INV2 will function earlier than other inverters. Solve Σ _i q _i (λ _Q )=20. Assume that through two Newton iterations, the global reactive power level λ _Q =-1.5 is obtained. According to the explicit relationship, substitute λ _Q to obtain: Q _1_Target =9.2Mvar; Q _2_Target =4.5Mvar (as the boundary is limited, its allocation share is naturally smaller); Q _3_Target= 6.3 Mvar; (9.2 + 4.5 + 6.3 = 20 Mvar, meets total constraint). Calculate achievable: control period T _s = 0.1 s. AQ _max，1 = 50.1 = 0.5 Mvar; AQ _max，2 = 40.1 = 0.4 Mvar; AQ _max，3 = 5 x 0.1 = 0.5 Mvar. Perform projection, Q _1_new = Q _1_prev + clip(Q _1_Target - Q _1_prev , -0.5, 0.5) = 0 + clip(9.2, -0.5, 0.5) = 0.5 Mvar; Q _2_new = 0 + clip(4.5, -0.4, 0.4) = 0.4 Mvar; Q _3_new = 0 + clip(6.3, -0.5, 0.5) = 0.5 Mvar; shares after rate correction are {0.5, 0.4, 0.5} Mvar. It can be seen that, despite the large target, in a single period, each inverter can only emit the maximum increment allowed by its rate.

[0143] According to SCR _proxy = 1.5, set passive envelope as Re{Z _out} > 0.5 x |Im{Z _out}|. At the same time, set loss upper limit according to each inverter thermal margin. Based on {0.5, 0.4, 0.5} Mvar, solve lightweight convex optimization problem. INV1 is given the highest priority due to the highest sensitivity. Obtain final instructions: the result given by the solver can be a small adjustment to the shares, and specific virtual resistance values are generated.

[0144] Step S9.4, output final control instructions The station-level coordinated controller issues the following instructions to each inverter (only reactive power and virtual resistance are listed): {inverter, Q _i_final (Mvar), D _i (p.u.)} = { (INV1, 0.48, 0.05), (INV2, 0.40, 0.02), (INV3, 0.49, 0.03)};

[0145] In the weak grid environment, facing serious voltage drop, the method can allocate tasks according to the boundary and sensitivity of each inverter, let INV2 undertake less share consistent with its ability; through RGP projection, the target is decomposed into physically reachable incremental instructions; through PEE optimization, while issuing reactive support, a larger virtual resistance is configured for high sensitivity INV1, actively injecting damping, ensuring the stability of the regulation process. The whole process is completed in a single control cycle, showing its real-time, adaptability and safety.

[0146] Embodiment ten provides a robust system state awareness and high-confidence boundary modeling method for obtaining high-quality and high-confidence grid strength and inverter capability boundary information; providing a basis for collaborative control.

[0147] As a preferred implementation scheme of passive (only relying on natural disturbance) grid strength evaluation, the embodiment provides an active detection type evaluation method, which can still obtain high-precision SCR _proxy when the grid background noise is weak.

[0148] Specifically, the station-level coordination controller can periodically (e.g., every 5 minutes) initiate active detection. Within the detection window, the controller will design a weak energy, good spectrum characteristic and harmless to the grid disturbance injection signal. Preferably, the signal is a pseudo-random binary sequence (PRBS), whose amplitude is strictly limited within 1% of the inverter rated current. The controller will synchronize the signal command to some or all of the inverter in the power station, instructing them to superimpose the disturbance signal in their own reactive or active control loop.

[0149] Since the injected disturbance signal ΔI _inject is known to the system, in the ΔV=-Z _Th ×ΔI relationship of embodiment eight, the ΔI part contains a high signal-to-noise ratio known component ΔI _inject . Accordingly, the voltage response ΔV measured at the PCC point also contains a specific response component to the known injected signal. Recursive least squares (RLS) algorithm can use cross-correlation and other techniques to extract this injection-response relationship from the background noise, and identify a more accurate Thevenin impedance Z _Th .

[0150] On this basis, the system will also generate a confidence index SCR _proxy accompanied by SCR _conf . The calculation of this confidence can integrate multiple factors: the residual size of the identification algorithm, that is, ΔV+Z _ThThe norm of ΔI, the smaller the residual, the higher the confidence; the correlation coefficient between the injected signal and the voltage response, the stronger the correlation, the higher the confidence; the convergence of the covariance matrix of the RLS algorithm. SCR _conf The indicators will serve as the basis for making weighted decisions, such as adjusting the barrier weight.

[0151] Further, the mixed boundary modeling method is refined, and the thermal state vector θ _T Before inputting into the MTRN network, the system will perform a series of fine data preprocessing. For example, for the collected temperature T _h , wind speed and other time series data, use median filter or sliding window average method to remove instantaneous outliers or sensor noise; normalize the input of different physical dimensions (such as temperature in Celsius, wind speed in meters / second) (for example, maximum-minimum normalization), eliminate the influence of different characteristic scale differences on neural network training.

[0152] In order to make the prediction behavior of the MTRN network conform to the physical law, in addition to monotonic constraints, more prior knowledge can be introduced in the network structure design and training stage. For example, the activation function of the network is preferably a piecewise linear unit (such as ReLU and its variants), which makes the model locally linear, facilitating analysis and understanding. In the loss function of training, in addition to the traditional prediction error term, a physically infeasible penalty term can also be added. This penalty term will impose a larger penalty when the network output violates the known, simplified physical relationship (for example, the approximate positive correlation between temperature rise rate and output power), guiding the network to learn the intrinsic relationship in physics.

[0153] After obtaining the final dynamic capability boundary B _i , the system will calculate the accompanying boundary confidence B _conf_i for it. B _conf_i is a value between 0 and 1, and its calculation can be a comprehensive score, and its influencing factors include: the stability of the input sensor reading, the larger the reading variance, the lower the confidence; the amplitude of the MTRN model output being trimmed by the analytical thermal model, if the physical model has a large correction to the data model, it means that the data model's prediction at this operating point may not be accurate, and the confidence should be reduced; the time since the last device maintenance or parameter calibration, the longer the time, the more likely the model is mismatched, and the lower the confidence.

[0154] In addition, the system will also generate a boundary failure flag boundary_fault. When a key sensor (such as a temperature sensor) reading is lost, exceeds a reasonable physical range, or fails to pass a consistency check for consecutive multiple periods, the flag is set. B _conf_i and boundary_fault are key to improving the robustness of the entire collaborative control system. In subsequent allocation decisions, B _conf_iLow inverters will be assigned more conservative tasks (such as by increasing their barrier weight), while inverters marked boundary_fault can be temporarily isolated from the cooperative control system, running only in the most conservative local mode until the fault is recovered.

[0155] As another possible implementation, further comprising: reading T _h , dT _dt , wind speed (optional), dust index (optional), through median filtering and unit normalization, obtaining θ _T (hot state vector). Wherein, the window length is 1-5s; the outlier rejection threshold is 3σ; the temperature unit is unified as Celsius, and the time reference is synchronized to the control cycle. Reading θ _T , through MTRN with monotonic structure, the soft boundary is inferred to obtain the soft boundary set B _soft ={S _max_soft , I _limit_soft , ramp _soft}. Reading the soft boundary set B _soft , θ _T , through R _Th -C _Th thermal model and device loss model, the upper and lower bounds are clipped to obtain B _clip ={S _max_clip , I _limit_clip , ramp _clip}. According to the current T _h and the allowed ΔT, the S and I changes in the future τ seconds are limited: S _max_clip ≤f1(T _h , ΔT, R _Th , C _Th ); ramp _clip ≤f2(C _Th , heat dissipation conditions).

[0156] Reading B _clip , combining manufacturer limits and on-site strategies, forming executable boundaries, obtaining executable boundaries B _i ={P _max , Q _max , S _max，i_limit , ramp} and B _conf . Wherein, P _max , Q _max are projected to the power plane by S _max_clip ; I _limit comes from device and inductance constraints; ramp is the minimum value of ramp _clip and grid connection strategy; B _conf is based on sensor reliability, model residual and recent stability record comprehensive calculation.

[0157] Embodiment 11 provides an optional implementation of the fine-grained disturbance budgeting and multi-event coordinated triggering method, which describes the transition of the system from the original disturbance monitoring to the adjustment task setting, and makes orderly and conflict-free response decisions under complex grid and device states.

[0158] Generate the total adjustment budget H _req from the monitored grid disturbance _req It is not a simple linear mapping, but a fine-grained calculation process considering multiple factors. Take the calculation of reactive power budget Q _req as an example. The preliminary calculation is Q _req_initial =ΔV / K _V_avg , where K _V_avg is the average voltage-reactive power sensitivity of the entire station; in a weak grid, the grid impedance is dominant, and the response of voltage to reactive power will become sluggish and nonlinear. To compensate for this effect, the invention introduces a weak grid correction function ψ(SCR _proxy ). The final budget is corrected to Q _req =Q _req_initial ×ψ(SCR _proxy ). The function ψ(x) is a monotone non-increasing function, ψ(x)≥1. For example, it can be designed as ψ(x)=1 when x (i.e. SCR _proxy ) is greater than 3.0, and ψ(x) increases linearly from 1.0 to 1.3 when x decreases from 3.0 to 1.5; the weaker the grid, the more the system will automatically and nonlinearly expand the budget of reactive power support to achieve the same voltage support effect.

[0159] Based on the consistency and clipping calculation of multi-source confidence, Q _req and H _req are the ideal values of the demand side, which also need to be consistent with the actual ability of the supply side. The effective budget H _req * , Q _req * issued to the distribution algorithm is the result of integrating demand confidence and supply confidence. For example, Q _req * =clip(w _Q ×Q _req ,0,ΣQ _max_i ). Where the weight w _Q is a function of disturbance measurement confidence Q _conf and the average boundary confidence avg(B _conf_i ) of the entire station, for example, w _Q =min(Q _conf ,avg(B _conf_i )). If the disturbance signal itself has a lot of noise (Q _confor most of the inverter's capability boundary model is in low confidence state, w _Q will be less than 1, and the budget will be discounted. The final clip function makes sure the total budget will not exceed the sum of all the inverter capability boundaries in the current state. This process avoids the system chasing an inaccurate or unachievable target in a state of uncertainty.

[0160] To make the system response both timely and robust, avoiding overreaction to small perturbations or transient signals, the invention designs a sophisticated and clear event triggering and arbitration mechanism, which is as follows: multi-level thresholds and time window de-bounce system sets multi-level thresholds for key perturbation quantities (such as ΔV, RoCoF) and internal state quantities (such as t _Thr ). For example, for ΔV, set -10% as the first-level warning threshold and -20% as the second-level serious threshold. The triggering of an event not only requires the perturbation quantity to exceed the threshold, but also requires it to be outside the threshold for a preset time window (for example, 3-5 sampling periods, about 30-50 milliseconds). This amplitude + duration double criterion filters transient peak noise and constitutes the triggering de-bounce logic.

[0161] When multiple events of different types (such as grid event σ _grid and thermal warning event σ _Thermal ) occur at the same time, a clear priority arbitration rule table will determine the final response behavior of the system; for example:

[0162] Highest priority: second-level (serious) grid stability event, such as SCR _proxy sharp drop to below 1.2 or voltage drop exceeding -20%. At this time, the system will unconditionally execute the coordinated control that prioritizes grid stability, such as raising passive constraints to the strongest level.

[0163] Second highest priority: first-level (general) grid stability event, such as voltage drop -10%.

[0164] Medium priority: device-level emergency warning, such as t _Thr of a certain inverter is less than the warning threshold (such as 300 seconds).

[0165] Arbitration logic example: when a first-level voltage drop event and a thermal warning event occur at the same time, the system will adopt a coordinated disposal strategy. On the one hand, it will start coordinated reactive power support to deal with voltage drop; on the other hand, when allocating reactive power shares, it will give priority to the B _iThe system can also reduce the total reactive power budget appropriately to balance the grid support and the thermal safety of the equipment. If the secondary voltage drop and the thermal warning occur simultaneously, the arbitration logic will determine that the grid safety is the priority, the system will temporarily ignore the thermal warning, and instruct all inverters with the ability to output reactive power, including the one that is about to be derated, to output reactive power at full capacity, and then deal with the thermal problem after the voltage stabilizes.

[0166] Through the budget generation and event arbitration mechanism, the rationality, orderliness and robustness of the collaborative control behavior are realized, and intelligent decisions can be made in complex and variable real environments.

[0167] The application solves the problems of decision confusion, response disorder or overreaction of the existing control system when facing multiple, concurrent or even conflicting events by designing a refined event triggering mechanism including multiple thresholds, time window debouncing and priority arbitration. The mechanism deeply integrates abstract business rules (expert decision logic) and specific technical fields (power station operation control). For example, when a primary voltage drop event and a device thermal warning occur simultaneously, the internal arbitration rule table will cooperatively call multiple technical modules for disposal: on the one hand, reactive power support response is started, and on the other hand, when the share allocation is performed, the dynamic capability boundary B _i of the thermal warning inverter will be taken into account, and the inverter will be contracted to achieve load reduction. The design makes the response behavior of the system always orderly and oriented to the global optimum; avoids single and rigid triggering mode, and gives the entire collaborative control system comprehensive decision-making ability, which can accurately identify the main contradiction in complex working conditions, make responses beneficial to the overall interests of the system, and improve the intelligent level, robustness and reliability of the entire power station in extreme scenarios.

[0168] Embodiment twelve provides an optional implementation of a safe execution and hierarchical degradation collaborative control strategy, which is an important supplement to all the aforementioned algorithms and strategies at the physical execution level, and is used to describe how the application makes control instructions safely land, how to cope with sudden changes in the ability of the device, and how to degrade when the system encounters a fault, so that the entire photovoltaic power station and the grid connection point can operate safely and stably.

[0169] According to one aspect of the application, when the current proactive strategy cannot be avoided or a sudden reason causes a certain inverter to actually enter the hardware derating state, the pre-derating pre-contracting proactive strategy will be activated. Specifically, the station-level coordination controller monitors the derating state of the inverter in real time through two ways: one is to directly read the state flag σ_derate uploaded by the local controller of the inverter, which is set when the inverter itself protection logic triggers derating; the other is to compare the issued instruction with the actual output, and when it is found that the actual output power (active or reactive) of a certain inverter is continuously lower than the control share H _i , Q_i At this moment, the controller also determines that it has entered a de-rating state in fact.

[0170] Correspondingly, once the i-th inverter is detected to enter de-rating, the controller immediately reads the current maximum feasible output upper limit of the inverter (for example, its Q _derated_max after de-rating) and calculates the share of the control task that is suspended and not executed on the inverter, i.e., Q _Tmp_i = Q _i_final - Q _derated_max . This Q _Tmp_i (and similarly H _Tmp_i ) is immediately recovered and merged into the temporary emergency budget pool Q _Tmp =∑Q _Tmp_i at the station level.

[0171] On this basis, the temporary budget Q _Tmp will be directly injected into the allocation algorithm of the CBF in the next control cycle (i.e., within 100 milliseconds) as a high-priority bias item. Specifically, the total regulation budget of the next cycle will be revised as Q _req_new = Q _req_orig + Q _Tmp . The CBF algorithm will automatically redistribute the regulation gap caused by the single-inverter de-rating according to the current capability boundary B _j and the sensitivity K _j of all other healthy inverters. From the above, the present application makes the sudden de-rating of a single inverter not cause the failure of the entire station to support the power grid, and the system can transfer the function of the damaged part to the healthy part, maintaining the continuity and stability of the overall function.

[0172] According to another aspect of the present application, the abstract parameters (H _i , Q _i , Z _out_params ) output by the optimization algorithm are converted into specific settings that the inverter can understand and execute. The present application designs a mapping and delivery mechanism at this link, including: H _i , which represents the physical quantity of inertia size, is mapped to the rotational inertia constant J in the virtual synchronous generator (VSG) control algorithm at the inverter control level, or to the dynamic gain coefficient of the improved frequency-active droop control. Q _i is mapped to the dynamic operating boundary of the inverter reactive power controller. For example, in Q-V (reactive power-voltage) droop control, Q _i is set as the upper limit Q _max and the lower limit Q _min of the reactive power output of the inverter in the control cycle. The virtual resistance D _i , virtual inductance L _Vi, are directly mapped to the corresponding gain coefficients of the internal virtual impedance loop of the inverter current control loop.

[0173] For such sensitive parameters as virtual impedance, the station-level coordination controller does not directly issue the final value, but decomposes it into a series of small incremental steps, and issues them step by step within several control cycles (for example, 200-500 milliseconds) to achieve smooth ramping of the parameter, that is, ramping down.

[0174] Before issuing any instructions, the controller first queries the health status of the target inverter and whether there is a local alarm. Only when the inverter returns a normal state and allows receiving new instructions, the instructions are issued. After the instructions are issued, the controller requires the inverter to read back and confirm that the parameters are correctly set. If the - read-back fails continuously for multiple times, the inverter will be marked as communication abnormal.

[0175] According to another aspect of the present application, in order to cope with extreme but possible failures such as communication interruption and controller downtime, the present application designs a hierarchical system-level degradation strategy, so that the operation of the power station is predictable and safe in any case. The triggering condition system continuously monitors the key events that may trigger degradation, including:

[0176] If the communication interruption time with more than 10% of the inverters exceeds the preset threshold (such as 5 seconds). If there is a large and non-converging mismatch between the actual response of the system and the control budget for multiple consecutive periods. If the station-level coordination controller main algorithm process is unresponsive, or the key input (such as PCC power meter) data is lost.

[0177] Once the triggering condition of the degradation level and strategy is met, the system will enter different levels of degradation mode according to the severity and range of the fault: primary degradation (individual isolation), when a single or a small number of inverters are communication abnormal or have their own faults, the station-level controller will isolate them from the object list of collaborative control, and send instructions to them to restore local control. The isolated inverters will return to the most conservative local droop control mode preset in their firmware; while the remaining healthy inverters will continue to be collaboratively controlled under the command of the station-level coordinator.

[0178] Secondary degradation (full-station collaborative degradation), when the optimization solving algorithm of the station-level coordination controller fails continuously, or most of the inverter boundary models are in a low confidence state, the controller will suspend the complex optimization algorithm and execute a simplified but still collaborative backup control strategy. For example, according to the total budget and inverter rated capacity, a simple proportional distribution is performed.

[0179] Third level degradation (full station localization) is triggered when the station level coordination controller itself is down or communication with the full station is interrupted. At this time, the heartbeat monitoring mechanism of each inverter will activate its final, independent local survival mode due to timeout. In this mode, all inverters will run in the factory default, extremely conservative fixed parameter droop control mode without any communication and coordination.

[0180] The recovery mechanism system will continue to try to recover from the degraded mode. For example, a background process will periodically detect the communication link and algorithm health. Once the conditions that triggered the degradation disappear, the system will perform a soft start procedure after a stable period of time (e.g. 1 minute) to gradually reintegrate the inverters from the local mode into the coordinated control system, achieving automatic recovery of functionality.

[0181] From the above, the performance of the application under normal operating conditions, more in abnormal and extreme operating conditions, shows its toughness and robustness, and constitutes a complete, industrial-grade control solution that can be applied in actual engineering.

[0182] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details of the above-described embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A method for inter-inverter coordinated anti-disturbance control of a desert photovoltaic power station, characterized in that, The method comprises the following steps: Collecting electrical quantities representing the operating state of the power grid and thermal state quantities representing the state of each inverter, and establishing a dynamic response model for each inverter based on the electrical quantities and the thermal state quantities, to generate sensitivity parameters and a dynamic capability boundary; Based on the monitored grid disturbance quantity and the sensitivity parameters, calculating the total adjustment budget required by the station level, and based on the total adjustment budget, the sensitivity parameters and the dynamic capability boundary, cooperatively solving the control share required by each inverter and the stability control parameters required to be configured; According to the solved control share and stability control parameters, control instructions are issued to each inverter and executed; According to the system response and the thermal state quantity after control execution, the model parameters and the dynamic capability boundary of the next control cycle are fed back and updated; The dynamic capability boundary is input into a neural network with a monotonic structure constraint to obtain a soft boundary set, which is input into an analytical thermal model to obtain the dynamic capability boundary after clipping by the physical upper and lower bounds; The cooperatively solved control share required by each inverter and the stability control parameters required to be configured comprise three stages in sequence, specifically: using a one-time mapping distribution model to directly analyze the total adjustment budget, the sensitivity parameters and the dynamic capability boundary into the initial control share of each inverter; based on the rate and current limit value contained in the dynamic capability boundary, the initial control share is corrected using a proximal projection operator to generate a rate-corrected share that satisfies dynamic reachability; the passive requirement is taken as a hard constraint to optimize the rate-corrected share, and the final executable control share and stability control parameters are generated synchronously; The passive requirement is taken as a hard constraint to optimize the rate-corrected share, and the final executable control share and stability control parameters are generated synchronously, including: based on the real-time evaluation of the grid strength, the dynamic passive envelope is constructed for the aggregated equivalent output impedance of the inverter group; the final executable control share and stability control parameters are taken as joint optimization variables, and the passive envelope is solved under the constraint of the passive envelope, wherein the dynamic path of the control share in the state transition always satisfies the stability domain defined by the envelope limit; The dynamic passive envelope is constructed for the aggregated equivalent output impedance of the inverter group, including: on the set of preset key frequency points, the passive envelope is defined as a complex plane conic constraint that requires the real part and the imaginary part of the aggregated equivalent output impedance to satisfy a specific proportional relationship; the geometric shape and safety margin of the conic constraint are dynamically adjusted according to the real-time evaluation of the grid strength and the severity level of the grid disturbance event determined based on the grid disturbance quantity, wherein the weaker the grid or the more severe the disturbance, the stricter the constraint applied; The one-time mapping distribution model is used to directly analyze the total adjustment budget, the sensitivity parameters and the dynamic capability boundary into the initial control share of each inverter, including: the dynamic capability boundary is converted into a logarithmic barrier term, and a cost term related to the sensitivity parameters is combined to construct a global optimization objective function with the total adjustment budget as the constraint; the Karush-Kuhn-Tucker condition of the global optimization objective function is derived to obtain an explicit analytical relationship between the initial control share and the Lagrange multiplier, and the initial control share is directly solved based on the relationship.

2. The method of claim 1, wherein, The final executable control share and the stability control parameter are solved as joint optimization variables under the constraints of the passive envelope, including: On the set of critical frequency points, the complex plane conic constraint is linearized to form a linear inequality constraint; Based on the linear inequality constraint, an additional loss upper limit constraint derived from the dynamic capability boundary is added to jointly construct a lightweight convex optimization problem with the final executable control share and the stability control parameter as the solving variables; By solving the convex optimization problem, the control share and the stability control parameter that meet the stability and loss boundary are obtained.

3. The method according to any one of claims 1, 2, characterized in that, The stability control parameter is generated, and the method further includes: Based on the sensitivity parameter of each inverter, the influence effectiveness of the inverter on the key indicators of the power grid is evaluated, and the priority of the stability adjustment of each inverter is set according to the evaluation result; According to the set priority, the stability control parameter is configured differently, and the main damping adjustment task is allocated to the inverter with high priority.

4. The method of claim 1, wherein, A global optimization objective function with the total adjustment budget as a constraint is jointly constructed, including: Based on the sensitivity parameter and the estimated loss of the inverter according to the control share, a cost item corresponding to the control effectiveness and operating efficiency is generated; Based on the reliability of the dynamic capability boundary and the confidence of the grid strength evaluation, the weight of the logarithmic barrier item is dynamically set, and the lower the reliability or confidence, the greater the barrier weight.

5. The method of claim 1, wherein, The initial control share is directly solved based on the relationship, including: A dynamic initial value for the solution of the Lagrange multiplier is set, and the initial value is composed of the multiplier final value of the last control period and a feedback correction term; The feedback correction term is generated according to the mismatch between the total adjustment budget and the actual system response after the execution of the control instruction in the last period; Using the dynamic initial value as the starting point, the initial control share is solved using the explicit analytical relationship to quickly converge.

6. The method of claim 1, wherein, According to the rate and current limit values contained in the dynamic capability boundary, the initial control share is corrected using a proximal projection operator to generate a rate-corrected share that meets the dynamic reachability, including: The rate and current limit values are defined as a dynamic reachable domain, and the proximal projection operator is used to map the initial control share to the nearest point in the dynamic reachable domain to form the rate-corrected share; In the mapping process, the control share of each inverter monotonically approaches its allocation target along the time axis, and the control chattering caused by direct clipping is suppressed.

Images