A virtual power plant collaborative scheduling system based on distributed resource aggregation

Abstract

The present application relates to the field of power system automation and control technology, disclose a kind of virtual power plant collaborative scheduling system based on distributed resource aggregation, comprising: cloud collaborative scheduling platform and the distributed intelligent gateway connected via communication network, gateway collects resource equipment physical data, using nonlinear observation function mapping is promoted state vector, execute global linear evolution logic, and based on spectral feature drift determination result sends model parameter or state vector to cloud end;Cloud end constructs aggregated linear model, solves convex quadratic programming problem to generate control sequence;Gateway converts sequence into physical instruction to drive equipment by inverse mapping.The present application converts nonlinear control into convex optimization problem by Koopman operator, greatly improves the calculation efficiency, effectively reduces the cloud edge communication load using spectral drift trigger mechanism, and realizes the online model identification of grid-connected power without fluctuation by combining null space excitation injection.

Description

Technical Field

[0001] This invention relates to the field of power system automation and control technology, specifically to a virtual power plant collaborative scheduling system based on distributed resource aggregation. Background Technology

[0002] With the construction of new power systems, distributed energy resources such as distributed photovoltaics, energy storage batteries, and electric vehicle charging piles are being connected to the power grid on a large scale. Virtual power plants, as a key technology for realizing the aggregation management and coordinated optimization of distributed resources, can aggregate massive, heterogeneous, and dispersed resources into a controllable whole to participate in grid dispatch. However, in the actual collaborative dispatch and control process of virtual power plants, multiple technical challenges are faced, including complex equipment physical characteristics, limited communication resources, and difficulties in maintaining online models.

[0003] In terms of equipment modeling and optimization, distributed resource devices typically exhibit nonlinear physical characteristics. For example, the electrochemical reaction process of a battery is affected by the nonlinearity of its state of charge and temperature, and the power electronic switching dynamics of an inverter also exhibit strong nonlinearity. Traditional scheduling methods often use local linearization methods such as Taylor expansion to approximate specific operating points, resulting in insufficient model accuracy during large-scale operation. Alternatively, nonlinear programming algorithms can be directly used for solving the problem, but as the number of aggregated devices increases, the computational complexity of nonlinear optimization rises exponentially, making it difficult to meet the real-time scheduling requirements of virtual power plants at the minute or even second level. How to achieve low-complexity global optimization solutions while preserving the nonlinear dynamic characteristics of the devices is a pressing problem that needs to be solved.

[0004] In terms of cloud-edge collaborative communication, edge devices need to frequently upload status data or model parameters to the cloud to ensure the accuracy of cloud scheduling strategies. In scenarios with massive device access, periodic full data uploads or fixed model synchronization strategies will generate huge data traffic, easily causing network congestion and delays, leading to delayed scheduling commands or distorted status feedback. Existing data transmission mechanisms often lack intelligent perception of model changes, failing to achieve a balance between model accuracy requirements and communication bandwidth resources.

[0005] Furthermore, regarding online maintenance and calibration of the model, the physical parameters of the equipment can drift over time due to changes in ambient temperature or component aging. To maintain control accuracy, online identification or calibration of the system model is necessary, which typically requires injecting excitation signals with rich frequency components into the system to stimulate its dynamic response. However, in grid-connected operation, directly injecting excitation signals into the equipment can cause unexpected fluctuations in the output power of the virtual power plant at the point of common coupling. This not only deviates from the grid's dispatch plan but also affects the power quality and stability of the grid. Therefore, how to achieve seamless online identification of internal equipment without interfering with the normal grid dispatch is another challenge facing the development of virtual power plant technology. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a virtual power plant collaborative scheduling system based on distributed resource aggregation. This system solves the problems of high computational complexity and poor real-time performance in global optimization solutions caused by the nonlinear characteristics of massive heterogeneous resources in existing virtual power plant scheduling, frequent transmission of model parameters during cloud-edge collaboration consuming a large amount of communication bandwidth, and the excitation signals required for online model identification easily causing power fluctuations at the grid connection point, thus affecting the stability of the power grid.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a virtual power plant collaborative scheduling system based on distributed resource aggregation. This system mainly comprises a cloud-based collaborative scheduling platform, a communication network, a distributed intelligent gateway, and distributed resource devices. The cloud-based collaborative scheduling platform and the distributed intelligent gateway are connected via the communication network, while the distributed intelligent gateway is directly connected to the distributed resource devices.

[0008] During system operation, the distributed intelligent gateway is responsible for collecting physical state data from distributed resource devices. This gateway is internally configured with a nonlinear observation function, capable of mapping low-dimensional physical state data into a boosted state vector in a high-dimensional feature space. Based on this boosted state vector, the gateway uses global linear evolution logic to calculate the boosted state vector for the next time step, thereby achieving linearized modeling of the nonlinear physical devices at the edge. Simultaneously, based on the spectral feature drift determination, the gateway decides whether to send the complete model parameters to the cloud-based collaborative scheduling platform or only send the current boosted state vector.

[0009] After receiving data from the edge, the cloud-based collaborative scheduling platform constructs an aggregated linear model of the entire system and solves a convex quadratic programming problem to generate future boosting control sequences. These boosting control sequences are then transmitted to a distributed intelligent gateway via a communication network. The gateway converts these sequences into physical control commands, which in turn drive distributed resource devices to change their operating states.

[0010] Specifically, the distributed intelligent gateway performs modeling and processing tasks at the edge. The data acquisition module within the gateway collects physical state data at a preset frequency and generates sequences. The Koopman linearization mapping module stores a set of observation functions containing the original physical state variables, combinations of higher-order polynomials, and radial basis functions. This module uses the Koopman state transition matrix to describe the system's free evolution dynamics without control input and the Koopman control matrix to describe the influence of external control inputs on the system state.

[0011] During the calculation, the Koopman linearization mapping module multiplies the current boosted state vector with the Koopman state transition matrix and the control input vector with the Koopman control matrix, then linearly superimposes the products to obtain the boosted state vector for the next time step. Furthermore, based on snapshot matrices and displacement snapshot matrices from historical data, this module solves for and updates the aforementioned state transition and control matrices by minimizing the Frobenius norm of the prediction error.

[0012] To optimize communication resources and ensure model accuracy, the system employs a triggering mechanism based on spectral feature drift. The spectral feature analysis and triggering module performs eigenvalue decomposition on the Koopman state transition matrix to obtain a set of eigenvalues ​​representing the system's dynamic modes. This module continuously monitors the absolute value of the changes in the magnitude and phase of the eigenvalues ​​relative to the reference values ​​at the current moment. Only when the absolute value of the change exceeds a preset threshold, indicating a significant change in the physical and dynamic characteristics of the device, does the gateway generate a model update command and send the updated matrix parameters; otherwise, only state vector data is transmitted.

[0013] To address the inaccuracy issues that may arise during long-term model operation, the system is equipped with an active excitation injection function. The active excitation injection module calculates the instantaneous residual vector between the observed boosted state vector and the predicted state vector, and performs dynamic mode decomposition spectral analysis on the residual sequence within the sliding window. When the analysis identifies a characteristic mode with a stable non-zero argument and a significant modulus, the system locks that specific frequency as the target excitation frequency.

[0014] Subsequently, the active excitation injection module utilizes null-space projection logic to divide the distributed resource devices into complementary first and second device groups. After generating the basic excitation signal, the module superimposes it with positive polarity onto the original control command of the first device group, and simultaneously superimposes it with equal amplitude negative polarity onto the original control command of the second device group. This processing method ensures that the net total power fluctuation of the virtual power plant at the point of common coupling is zero, achieving closed-loop detection of the physical characteristics of the internal devices without affecting grid dispatch.

[0015] At the cloud control level, the cloud-based collaborative scheduling platform includes a linear model aggregation module and a global convex optimization solution module. The linear model aggregation module receives boosted state vectors and Koopman operator matrices from multiple distributed intelligent gateways. Through matrix block diagonalization or linear superposition, it combines the models of each independent subsystem to generate the aggregated state equation for the entire system. The global convex optimization solution module constructs an objective function that includes system operating costs and scheduling command tracking errors, and transforms physical constraints such as battery state of charge limits, inverter power capacity limitations, and temperature safety ranges into linear constraints within the boost space. Based on the aggregated linear model and linearized constraints, this module constructs a convex quadratic programming problem and calculates the optimal boosted control sequence in the future prediction time domain.

[0016] At the execution level, the inverse mapping execution module of the distributed intelligent gateway is responsible for restoring the abstract control signals issued from the cloud into physical actions. This module stores the generalized inverse approximation representation or linear component extraction matrix of the observation function set as the inverse projection matrix. By performing matrix multiplication between the control vector in the boost control sequence and the inverse projection matrix, the module calculates the setpoint with physical units and inputs it into the closed-loop control loop to generate the underlying drive signal. Before the final output, the module also performs a safety boundary check, comparing the drive signal with the maximum allowable current, the upper limit of the DC bus voltage, and the device temperature protection limit to ensure that only signals within the safe and feasible region are loaded into the hardware drive circuit.

[0017] This invention provides a virtual power plant collaborative scheduling system based on distributed resource aggregation. It has the following beneficial effects: 1. By configuring the Koopman linearization mapping module on the distributed intelligent gateway side, the nonlinear physical dynamics of distributed resource devices are mapped into a global linear evolution model within the lift space using a nonlinear observation function. This enables the cloud-based collaborative scheduling platform to construct an aggregated linear model and solve it using a mature convex quadratic programming algorithm. This not only preserves the high-precision description of the nonlinear characteristics of the underlying devices but also significantly reduces the computational complexity of the whole system's collaborative optimization, ensuring the real-time performance and global optimality of scheduling instructions.

[0018] 2. By executing a spectral feature analysis and triggering module based on spectral feature drift, the modulus and phase changes of the Koopman operator eigenvalues ​​are monitored in real time. The model parameters are only updated and uploaded when the physical dynamics characteristics change significantly. At other times, only lightweight boosted state vectors are transmitted. This event-triggered mechanism avoids redundant transmission of high-dimensional model parameters, effectively saving communication bandwidth resources while ensuring the timeliness of the cloud model.

[0019] 3. By applying zero-space projection logic through the active excitation injection module, micro-disturbance excitation signals with equal amplitude and opposite polarity are superimposed between complementary equipment groups, so that the net total power fluctuation of the virtual power plant at the point of common coupling is zero. This allows the system to continuously inject rich frequency domain excitation signals into the system without affecting the stability of external power output and the power quality of the grid, thereby realizing closed-loop detection of equipment physical parameters and dynamic model calibration. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of a virtual power plant collaborative scheduling system based on distributed resource aggregation, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the global linearization modeling process for edge-side nonlinear dynamics in an embodiment of the present invention; Figure 3 This is a schematic diagram of a sparse event-triggered communication process based on spectral feature drift, according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the cloud-based linear model predictive control process based on the lift space in an embodiment of the present invention. Figure 5 This is a schematic diagram of the null space active excitation injection process based on residual DMD spectral analysis according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the inverse mapping and control instruction execution flow according to an embodiment of the present invention.

[0021] Among them, 10. Distributed intelligent gateway; 11. Data acquisition module; 12. Koopman linearization mapping module; 13. Spectral feature analysis and triggering module; 14. Active stimulus injection module; 15. Inverse mapping execution module; 20. Cloud-based collaborative scheduling platform; 21. Linear model aggregation module; 22. Global convex optimization solution module; 30. Communication network; 40. Distributed resource device. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 The present invention provides a virtual power plant collaborative scheduling system based on distributed resource aggregation. The system includes: a distributed resource device 40, a distributed intelligent gateway 10 connected to the distributed resource device 40, a cloud collaborative scheduling platform 20, and a communication network 30 connecting the distributed intelligent gateway 10 and the cloud collaborative scheduling platform 20.

[0024] Distributed resource devices 40 refer to physical entities that constitute a virtual power plant, including but not limited to photovoltaic power generation devices, electrochemical energy storage devices, air conditioning load devices, or combinations thereof. Distributed resource devices 40 have nonlinear physical dynamic characteristics, and their operating state is described by a physical state vector, which includes voltage, current, temperature, state of charge, or power output variables.

[0025] The distributed intelligent gateway 10 is deployed at the edge of the distributed resource device 40 and is equipped with a processor and memory. The distributed intelligent gateway 10 includes a data acquisition module 11, a Koopman linearization mapping module 12, a spectral feature analysis and triggering module 13, an active stimulus injection module 14, and an inverse mapping execution module 15.

[0026] The data acquisition module 11 is connected to the distributed resource device 40 through a sensor interface and is configured to acquire real-time physical state data of the distributed resource device 40 at a preset sampling frequency to generate a physical state vector sequence.

[0027] The Koopman linearization mapping module 12 is connected to the data acquisition module 11. The Koopman linearization mapping module 12 stores a set of predefined nonlinear observation functions, configured to map low-dimensional physical state vectors to a high-dimensional Hilbert space, generating a boosted state vector. The Koopman linearization mapping module 12 is configured to use global linear evolution logic to describe the dynamic process of the distributed resource device 40 within the boosted space. Specifically, the Koopman linearization mapping module 12 calculates the boosted state vector for the next time step by multiplying the current boosted state vector by the Koopman state transition matrix, multiplying the control input vector by the Koopman control matrix, and then linearly superimposing the products. Both the Koopman state transition matrix and the Koopman control matrix are constant matrices describing the dynamic characteristics of the system, realizing a global linearized representation of nonlinear dynamics.

[0028] The spectral feature analysis and triggering module 13 is connected to the Koopman linearization mapping module 12. The spectral feature analysis and triggering module 13 is configured to perform eigenvalue decomposition on the Koopman state transition matrix in the Koopman linearization mapping module 12 to obtain an eigenvalue set. The spectral feature analysis and triggering module 13 stores spectral drift determination logic, configured to compare the magnitude and phase of the eigenvalues ​​at the current time with the deviation from the reference value. When the deviation exceeds a preset threshold, the spectral feature analysis and triggering module 13 generates a model update command and sends the updated Koopman state transition matrix and Koopman control matrix to the cloud-based collaborative scheduling platform 20 via the communication network 30; when the deviation does not exceed the preset threshold, only the current boosted state vector is sent.

[0029] The active excitation injection module 14 is configured to synthesize a micro-perturbation excitation signal based on the spectral analysis results of the prediction residual. The active excitation injection module 14 is also configured with null space projection logic to generate a complementary control signal that satisfies the null space constraint and send the signal to the underlying controller of the distributed resource device 40.

[0030] The inverse mapping execution module 15 is configured to receive linear control commands from the cloud-based collaborative scheduling platform 20 or excitation signals generated by the active excitation injection module 14, and convert the control quantities of the boost space into physical control commands that can be executed by the distributed resource device 40 through the inverse mapping function or pseudo-inverse operation.

[0031] The communication network 30 is the physical link for data transmission. It is configured to support TCP / IP or industrial Ethernet protocols to enable bidirectional data interaction between the distributed intelligent gateway 10 and the cloud-based collaborative scheduling platform 20.

[0032] The cloud-based collaborative scheduling platform 20 is the control center of the virtual power plant, including a linear model aggregation module 21 and a global convex optimization solution module 22.

[0033] The linear model aggregation module 21 is configured to receive boosted state vectors and Koopman operator matrices from multiple distributed smart gateways 10. The linear model aggregation module 21 constructs the aggregated state equations for the entire system through matrix block diagonalization or linear superposition. Since each subsystem model is linear, the aggregated model maintains a linear structure.

[0034] The global convex optimization solution module 22 is connected to the linear model aggregation module 21. The global convex optimization solution module 22 is configured to construct an objective function with system operating cost and scheduling instruction tracking error as variables. Based on the aggregated linear state equations and mapped linear constraints, the global convex optimization solution module 22 constructs the scheduling problem as a convex quadratic programming problem. The global convex optimization solution module 22 is equipped with a numerical solver to calculate the optimal boost control sequence in the future prediction time domain and distribute it to each distributed intelligent gateway 10 via the communication network 30.

[0035] See attached document Figure 2 This invention provides a global linearization modeling method for edge-side nonlinear dynamics. This method is mainly executed by the Koopman linearization mapping module 12 in the distributed intelligent gateway 10, and aims to transform the strong nonlinear physical characteristics of the distributed resource device 40 into a high-dimensional linear model that is easy to compute.

[0036] First, the data acquisition module 11 of the distributed intelligent gateway 10 at discrete time points. Obtain the physical state vector of distributed resource device 40 and control input vector .

[0037] Physical state vector Defined in the original physical observation space, the dimension of this space is denoted as . (i.e., the number of physical variables). Its components correspond one-to-one with the actual physical measurements of the equipment: for distributed resources of electrochemical energy storage type, the physical state vector... Includes variables such as state of charge, terminal voltage, current, and battery module temperature; for distributed resources of photovoltaic power generation type, the physical state vector... It includes variables such as DC-side voltage, DC-side current, and light intensity.

[0038] Control input vector Defined in the external control input space, the dimension of which is denoted as . (i.e., the number of control channels). Its components correspond to the setpoints of the underlying controller, such as active power commands or reactive power commands.

[0039] Subsequently, the Koopman linearization mapping module 12 calls a set of observation functions pre-configured in memory. Observation function set Configured to use the original physical observation space 3D physical state vector The nonlinear mapping is applied to a significantly expanded high-dimensional feature space (i.e., the boosting space) to generate boosting state vectors. The dimension of this lift space is denoted as... And satisfy Much larger The conditions are as follows. The lifting state vector is defined as... .

[0040] To ensure that the enhanced space can fully capture the nonlinear characteristics of the original system and maintain physical interpretability, the observation function set... The constituent elements specifically include three categories: The first category is the original physical state variables themselves, which are used to ensure that physical quantities can be directly extracted during inverse mapping; The second category consists of high-order polynomial combinations of physical state variables, used to capture nonlinear coupling relationships between states. The third type is radial basis functions or Gaussian kernel functions, which are used to approximate highly nonlinear local dynamics.

[0041] By observing the function set The mapping of the distributed resource device 40 in the original low-dimensional state space is embedded into a high-dimensional Hilbert space.

[0042] In generating the boosting state vector Then, the Koopman linearization mapping module 12 establishes the linear evolution relationship of the vector over time steps. This linear evolution relationship is described by the Koopman operator equation, specifically expressed as: ; in, Represents any discrete time. The boosted state vector, Indicates the next moment The boosted state vector, for The dimensional Koopman state transition matrix characterizes the free evolution dynamics of the system in the lift space without control input. for The 3D Koopman control matrix characterizes the linear effect of external control inputs on the system's improved state evolution.

[0043] To obtain the Koopman state transition matrix and Koopman control matrix For the specific numerical values, the Koopman linearization mapping module 12 executes the extended dynamic mode decomposition algorithm. This algorithm first constructs two data matrices based on historically collected time-series data: a snapshot matrix and a snapshot matrix. and displacement snapshot matrix Snapshot matrix Including the front Each sampling time (i.e., time 1 to time 2) The sequence of lifted state vectors; the displacement snapshot matrix. Includes shifting backward by one step A sequence of raised state vectors (i.e., time 2 to time 3) The sequence of boosted state vectors is generated. Simultaneously, the corresponding control input data matrix is ​​constructed. Includes time from time 1 to time 2. The sequence of control input vectors.

[0044] The Koopman linearization mapping module 12 will solve for the Koopman state transition matrix. and Koopman control matrix The process is transformed into a linear least squares optimization problem. The objective of this optimization problem is to minimize the displacement snapshot matrix. The Frobenius norm error between the prediction matrix and the target matrix. The prediction matrix is ​​derived from the Koopman state transition matrix to be solved. Koopman control matrix With snapshot matrix and control input data matrix The calculation yields the Koopman state transition matrix by multiplying the pseudo-inverse of the augmented data matrix with the displacement snapshot matrix. The Koopman linearization mapping module 12 then calculates the Koopman state transition matrix that minimizes the sum of squared prediction errors. and Koopman control matrix The optimal estimate.

[0045] After completing the above calculations, the distributed intelligent gateway 10 will obtain a constant Koopman state transition matrix. and Koopman control matrix The current linear proxy model of the distributed resource device 40 is stored locally and prepared for subsequent state prediction and data reporting to the cloud. This approach transforms the nonlinear differential-algebraic equations that originally required iterative solutions into matrix-vector multiplication operations, thereby reducing computational complexity while preserving the nonlinear characteristics of the device across its entire working domain.

[0046] See attached document Figure 3 This invention provides a sparse event-triggered communication mechanism based on spectral feature drift. This mechanism is executed by the spectral feature analysis and triggering module 13 in the distributed intelligent gateway 10. It aims to dynamically adjust the data transmission strategy by monitoring changes in the essential dynamic characteristics of the physical system, thereby reducing the communication bandwidth usage while ensuring the accuracy of the cloud model.

[0047] The spectral feature analysis and triggering module 13 first reads the current time-time Koopman state transition matrix identified by the Koopman linearization mapping module 12. To extract the physical and dynamic characteristics contained in this matrix, the spectral feature analysis and triggering module 13 performs analysis on the Koopman state transition matrix. Perform eigenvalue decomposition. This operation produces a set of complex eigenvalues, denoted as . In this set, each eigenvalue This corresponds to a specific dynamic mode of the system within its lift space. Specifically, it refers to the magnitude of the eigenvalue. In discrete-time systems, the energy decay or growth rate of this mode is characterized when... The time indicates that the mode is stable; the argument of the eigenvalues It characterizes the inherent oscillation frequency of this mode, corresponding to the periodic behavior in the physical system.

[0048] After acquiring the current eigenvalue set, the spectral feature analysis and triggering module 13 retrieves the reference eigenvalue set stored in the memory. The set of benchmark eigenvalues This refers to the set of feature values ​​corresponding to the model parameters that were successfully triggered and uploaded to the cloud-based collaborative scheduling platform 20 last time. The spectral feature analysis and triggering module 13 calculates the spectral drift of the current feature value set relative to the baseline feature value set. The calculation of this drift includes two dimensions of judgment logic: One method is modal damping drift determination, which calculates the absolute value of the change in modulus of the dominant eigenvalue (i.e., the eigenvalue whose modulus is closest to 1) and determines whether it exceeds a preset damping drift threshold. ; The second is modal frequency drift determination, which calculates the absolute value of the change in the amplitude of the dominant eigenvalue and determines whether it exceeds a preset frequency drift threshold. .

[0049] When any of the above conditions is met, it indicates that the physical characteristics of the distributed resource device 40 have changed significantly. For example, the battery's internal resistance may increase due to aging, causing a change in the time constant, or the control parameters of the photovoltaic inverter may change, causing a shift in the output resonant frequency. At this time, the spectral feature analysis and triggering module 13 generates a model update trigger signal. In response to this trigger signal, the distributed intelligent gateway 10 updates the latest Koopman state transition matrix. and Koopman control matrix The entire package is then sent to the cloud-based collaborative scheduling platform 20 via communication network 30, while simultaneously updating the current feature value set to the new baseline feature value set. .

[0050] Conversely, if neither of the above two conditions is met, it indicates that the current spectral drift is within the allowable error tolerance, and the physical dynamics of the distributed resource device 40 remain stable relative to the model copy held in the cloud. At this time, the distributed intelligent gateway 10 suspends the uploading of model parameters and only sends the current boost state vector through the communication network 30. Due to the boosting of the state vector The amount of data is much smaller than the Koopman state transition matrix. and Koopman control matrix With such a large amount of data, this sparse triggering mechanism enables the system to maintain a low-bandwidth communication mode for most of the runtime, and only occupies high bandwidth for model synchronization at critical moments when physical characteristics drift, thus achieving an adaptive balance between communication resources and model accuracy.

[0051] See attached document Figure 4This invention provides a cloud-based linear model predictive control method based on improvement space. The method is executed collaboratively by the linear model aggregation module 21 and the global convex optimization solution module 22 in the cloud-based collaborative scheduling platform 20. It aims to achieve global optimization scheduling of heterogeneous resources of the entire system by using the linear proxy model uploaded from the edge side.

[0052] The linear model aggregation module 21 first receives real-time boosted state vectors and Koopman model parameters from each distributed intelligent gateway 10. The linear model aggregation module 21 is configured to construct a system-wide aggregated state vector in cloud memory. This vector is logically composed of boosted state vectors from all distributed resource subsystems arranged in a predetermined order. Correspondingly, the linear model aggregation module 21 constructs the state transition logic and control response logic of the aggregated system. Specifically, the linear model aggregation module 21 combines the Koopman state transition matrices and Koopman control matrices of each independent subsystem according to the order corresponding to the state vectors to generate the system-wide aggregated Koopman state transition matrix and aggregated Koopman control matrix. Through this block diagonalization or linear superposition construction method, the cloud integrates multiple heterogeneous subsystem models that were originally independent into a unified high-dimensional global linear dynamic system model.

[0053] Subsequently, the global convex optimization solution module 22 constructs a model predictive control problem based on the aggregated linear model. The global convex optimization solution module 22 first defines a rolling prediction time domain. In each scheduling cycle, the global convex optimization solution module 22 constructs a comprehensive optimization objective function. This objective function is configured to quantitatively evaluate performance indicators in two main dimensions: one is the tracking deviation between the actual output power of the virtual power plant at the point of common coupling and the grid dispatch command, aiming to maximize the response accuracy to the grid command; the other is the amplitude or rate of change of the control actions of each distributed resource device 40, aiming to minimize the mechanical wear or adjustment costs of the equipment. Since the improved state vectors of each subsystem already include physical quantities such as power output through linear mapping, the aforementioned tracking deviation indicators are constructed as quadratic functions of the aggregated state vector and the aggregated control input vector.

[0054] While constructing the objective function, the global convex optimization solution module 22 performs a linearization mapping operation of physical constraints. The global convex optimization solution module 22 transforms the physical-level operational constraints of each distributed resource device 40 (including but not limited to battery state of charge limits, inverter power capacity limits, and temperature safety ranges) into linear constraints within the boost space. Specifically, the global convex optimization solution module 22 identifies the components in the boost state vector corresponding to the original physical variables and sets corresponding linear inequality boundaries for these components. This ensures that any solution generated during the mathematical optimization process, when restored to the physical space, strictly meets the safe operation requirements of the devices.

[0055] Combining the aforementioned quadratic objective function, the linear state evolution equation based on the aggregation matrix, and the mapped linear inequality constraints, the global convex optimization solution module 22 mathematically transforms this complex scheduling optimization problem into a standard convex quadratic programming problem. Utilizing the mathematical property of convex optimization problems—that a local optimum is equivalent to a global optimum—the global convex optimization solution module 22 calls a numerical solver to calculate the optimal aggregated control input sequence in the future rolling prediction time domain within polynomial time. Finally, the global convex optimization solution module 22 extracts the control command for the current moment from this optimal sequence and, through the communication network 30, splits it according to the device index and distributes it to the corresponding distributed intelligent gateways 10.

[0056] See attached document Figure 5 This invention provides a null space active excitation injection mechanism based on residual DMD spectrum analysis. This mechanism is specifically executed by the active excitation injection module 14 in the distributed intelligent gateway 10. It aims to solve the technical problem that the linear agent model gradually becomes inaccurate due to the drift of equipment physical characteristics during long-term operation. It achieves online closed-loop correction of the model through an implicit detection method that is transparent to power grid scheduling.

[0057] The active stimulus injection module 14 first performs real-time generation and spectral characteristic monitoring of the model prediction residuals. The active stimulus injection module 14 is configured to synchronously receive the actual observation boosted state vector from the data acquisition module 11 and the single-step predicted state vector from the Koopman linearization mapping module 12. The active stimulus injection module 14 calculates the vector difference between the two, generates an instantaneous residual vector, and stores the residual vectors from multiple consecutive time points in a sliding observation window in chronological order. To extract valuable physical information from the unordered error data, this module performs dynamic mode decomposition spectral analysis on the residual sequence within the sliding window.

[0058] Specifically, the active excitation injection module 14 constructs a time snapshot matrix based on the residual sequence and extracts the dominant eigenvalues ​​of the matrix through singular value decomposition or similarity transformation. The active excitation injection module 14 identifies the error nature based on the magnitude and argument distribution of the eigenvalues: if the eigenvalue spectrum shows a uniform distribution and no obvious dominant mode, it is determined that the current residual is Gaussian white noise and the model accuracy meets the requirements; if the analysis results separate eigenmodes with stable non-zero arguments and obvious magnitudes, it is determined that the current model has a dynamic missing point at a specific frequency, and this specific frequency is locked as the target excitation frequency.

[0059] For the locked target excitation frequency, the active excitation injection module 14 enters the stage of synthesizing and injecting micro-perturbation excitation signals. The active excitation injection module 14 first uses signal generation logic to generate a sine wave or pseudo-random sequence containing the target frequency component as the basic excitation signal. To ensure that the injection process of this excitation signal does not disrupt the overall tracking performance of the virtual power plant to grid dispatch commands, i.e., to achieve no net impact on grid-side power exchange, the active excitation injection module 14 adopts a zero-space projection control strategy.

[0060] The zero-space projection control strategy, based on the redundant degrees of freedom of the multi-micro-source aggregation system, first divides the distributed resource devices 40 within the system, which have similar adjustment capabilities and whose current power operating points are all within a preset safety margin from their rated capacity upper limits, into a complementary first device group and a second device group. Subsequently, the active excitation injection module 14 constructs control allocation logic that satisfies the zero-space constraints, superimposing the synthesized basic excitation signal with positive polarity into the original control command of the first device group, and simultaneously superimposing it with equal amplitude negative polarity into the original control command of the second device group.

[0061] Through the aforementioned dual injection method, at any sampling moment, the positive power deviation generated by the excitation of the first device group and the negative power deviation generated by the second device group precisely cancel each other out algebraically, resulting in a net total power fluctuation of the entire resource aggregate at the point of common coupling being zero. This mechanism ensures that the active excitation signal circulates only within the system and remains invisible to the external power grid. However, at the microscopic device level, each injected physical device actually experiences the input signal containing rich target frequency components, thus forcing the device to elicit its true physical response at that frequency.

[0062] During the execution of null space excitation, the active excitation injection module 14 synchronously triggers the high-frequency data recording function to collect input-output data pairs containing system dynamic information, including the specific frequency excitation signal and its corresponding system response. Using this new data containing system dynamic information, the module executes a directional correction algorithm for the model parameters. Specifically, it employs recursive least squares or Kalman filtering algorithms, iteratively updating only the feature subspace parameters in the Koopman state transition matrix corresponding to the target frequency, until the prediction residual at that frequency converges to a preset noise level. After the update is complete, the active excitation injection module 14 automatically cancels the excitation signal, and the system returns to the normal scheduling mode. This process achieves adaptive model repair without disconnecting the device from the network or interrupting normal business operations.

[0063] See attached document Figure 6This invention provides a method for inverse mapping and control instruction execution. This method is specifically implemented by the inverse mapping execution module 15 in the distributed intelligent gateway 10. As an execution link connecting the cloud abstract computing space and the underlying physical entity device, it ensures that the optimal strategy obtained from the high-dimensional linear space can be accurately transformed into the actual action of the physical device.

[0064] The inverse mapping execution module 15 first performs instruction reception and source identification operations. This module is configured to simultaneously listen for the globally optimal control instruction sequence issued from the cloud-based collaborative scheduling platform 20, as well as the zero-space excitation signal generated by the local active excitation injection module 14. The inverse mapping execution module 15 integrates instruction arbitration logic. When it receives a superimposed instruction containing zero-space excitation, it prioritizes processing the synthesized instruction to support online model correction; in normal mode, it processes the global scheduling instructions from the cloud. Since the instructions issued from the cloud are usually based on calculations in an enhanced high-dimensional linear space, or data that has been normalized for numerical stability, these instructions cannot be directly recognized by the underlying power electronic controller.

[0065] Therefore, the inverse mapping execution module 15 performs the inverse mapping transformation operation. This operation aims to reduce the abstract computational control quantity to an engineering control quantity with a definite physical unit. For the boosted control input form used in the Koopman operator system, the inverse mapping execution module 15 calls the pre-stored inverse projection matrix in memory. This inverse projection matrix is ​​an approximate representation of the generalized inverse of the Koopman observation function set, or it is constructed from the linear component extraction matrix of the observation function. The inverse mapping execution module 15 projects the received high-dimensional or normalized control vector back from the high-dimensional feature space to the low-dimensional physical control space by performing matrix multiplication with this inverse projection matrix, thereby solving for the corresponding physical setpoint, such as the active power setpoint of a photovoltaic inverter or the charging and discharging current setpoint of an energy storage system.

[0066] After acquiring the physical setpoints, the inverse mapping execution module 15 enters the low-level drive signal generation stage. This module inputs the aforementioned physical setpoints into the local closed-loop control loop, such as a proportional-integral-derivative (PID) controller or a current inner-loop controller. For power electronic interfaces such as voltage source converters, the inverse mapping execution module 15 further converts the power or current command into a pulse-width modulation (PWM) signal. Specifically, the inverse mapping execution module 15 calculates the duty cycle sequence of the switching transistors required for the target power output, generating high-frequency on / off signals to drive the IGBT or MOSFET power semiconductor devices.

[0067] Before final execution, the inverse mapping execution module 15 performs a safety boundary check function in parallel. Although the cloud scheduling has considered linearization constraints, to prevent equipment damage caused by communication errors or instantaneous model deviations, the inverse mapping execution module 15 compares the generated physical drive signal with the absolute safety thresholds of the device hardware on a millisecond-level timescale. These thresholds include the maximum allowable current, the upper limit of the DC bus voltage, and the IGBT junction temperature protection limit. Only when the drive signal is within the safe and feasible region will the inverse mapping execution module 15 load the signal to the hardware drive circuit, driving the distributed resource device 40 to change its operating state, thereby completing a complete closed loop from cloud algorithm to edge physical action. At the same time, the module feeds back the actual state after execution to the data acquisition module 11, providing initial conditions for the next cycle.

Claims

1. A virtual power plant collaborative scheduling system based on distributed resource aggregation, characterized in that, include: A cloud-based collaborative scheduling platform (20), and multiple distributed intelligent gateways (10) connected to the cloud-based collaborative scheduling platform (20) via a communication network (30), each of the distributed intelligent gateways (10) being connected to a distributed resource device (40); The distributed intelligent gateway (10) is used to collect physical state data, construct a local linear agent model through nonlinear mapping and global linear evolution logic, and send model parameters or only the current boosting state vector to the cloud collaborative scheduling platform (20) based on the spectral feature drift judgment result. The cloud-based collaborative scheduling platform (20) is used to receive multi-source data and construct a system-wide aggregated linear model, perform global collaborative optimization based on the aggregated linear model, solve the convex quadratic programming problem to generate and distribute the boosting control sequence; The distributed intelligent gateway (10) is also used to receive the boost control sequence, convert it into physical control instructions through inverse mapping, and drive the distributed resource device (40) to change its operating state.

2. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 1, characterized in that, The distributed intelligent gateway (10) includes: The data acquisition module (11) is used to acquire the physical state data and generate a physical state vector sequence; Koopman linearization mapping module (12) is used to perform the mapping and the global linear evolution logic; The spectral feature analysis and triggering module (13) is used to perform the analysis of the spectral feature drift determination results and the data transmission strategy; An active excitation injection module (14) is used to synthesize micro-perturbation excitation signals based on predicted residuals and generate complementary control signals that satisfy null space constraints; The inverse mapping execution module (15) is used to perform the conversion and driving operations of the physical control instructions.

3. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 2, characterized in that, The Koopman linearization mapping module (12) stores a set of observation functions containing the original physical state variables, combinations of higher-order polynomials, and radial basis functions; The Koopman linearization mapping module (12) is used to multiply the current boosted state vector with the Koopman state transition matrix, multiply the control input vector with the Koopman control matrix, and linearly superimpose the product results to calculate the boosted state vector at the next moment. The Koopman linearization mapping module (12) is also used to solve the Koopman state transition matrix and the Koopman control matrix by minimizing the Frobenius norm error based on the snapshot matrix and the displacement snapshot matrix.

4. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 2, characterized in that, The spectral feature analysis and triggering module (13) is used to perform eigenvalue decomposition on the Koopman state transition matrix to obtain an eigenvalue set; The spectral feature analysis and triggering module (13) is used to calculate the absolute value of the change in the magnitude of the feature value at the current time relative to the reference value, and the absolute value of the change in the phase of the feature value at the current time relative to the reference value; When the absolute value of the change exceeds a preset threshold, a model update instruction is generated, and the updated Koopman state transition matrix and Koopman control matrix are sent through the communication network (30).

5. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 2, characterized in that, The active excitation injection module (14) is used to calculate the instantaneous residual vector between the actual observed lifted state vector and the predicted state vector; The active excitation injection module (14) is used to perform dynamic mode decomposition spectrum analysis on the residual sequence within the sliding observation window. When the characteristic modes with stable non-zero amplitude and obvious modulus are separated, the target excitation frequency is locked and a basic excitation signal containing the target excitation frequency is synthesized.

6. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 5, characterized in that, The active incentive injection module (14) is configured with zero-space projection logic, which divides the distributed resource device (40) into a complementary first device group and a second device group. The active excitation injection module (14) is used to superimpose the basic excitation signal with positive polarity onto the original control command of the first device group according to the zero-space projection logic, and simultaneously superimpose it with negative polarity of equal amplitude onto the original control command of the second device group.

7. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 1, characterized in that, The cloud-based collaborative scheduling platform (20) includes a linear model aggregation module (21) and a global convex optimization solution module (22). The linear model aggregation module (21) is used to receive the boosted state vector, Koopman state transition matrix and Koopman control matrix from multiple distributed smart gateways (10); The linear model aggregation module (21) is used to combine the lifting state vectors of each subsystem to generate the aggregated state equation of the whole system by matrix block diagonalization or linear superposition. The global convex optimization solution module (22) is used to construct an objective function that includes system operating costs and scheduling instruction tracking errors, and to calculate the optimal boost control sequence in the future prediction time domain based on the aggregated state equation.

8. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 7, characterized in that, The global convex optimization solution module (22) is used to convert the upper and lower limits of the battery state of charge, the power capacity limit of the inverter and the temperature safety range of the distributed resource device (40) into linear constraints within the improvement space. The global convex optimization solution module (22) is used to combine the objective function, the aggregate state equation and the linear constraints to construct and solve the convex quadratic programming problem.

9. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 2, characterized in that, The inverse mapping execution module (15) is used to store an inverse projection matrix, which is a generalized inverse approximation of the observation function set or a linear component extraction matrix; The inverse mapping execution module (15) is used to perform matrix multiplication between the control vector in the boost control sequence and the inverse projection matrix to calculate the physical setpoint. The inverse mapping execution module (15) is used to input the physical set value into the closed-loop control loop to generate the underlying drive signal.

10. The virtual power plant collaborative scheduling system based on distributed resource aggregation according to claim 9, characterized in that, The inverse mapping execution module (15) is used to perform a security boundary check before executing the underlying drive signal; The inverse mapping execution module (15) is used to compare the underlying drive signal with the maximum allowable current, the upper limit of DC bus voltage and the IGBT junction temperature protection limit, and load the underlying drive signal into the hardware drive circuit only when the underlying drive signal is within the safe and feasible domain.

Images