A method for resiliency scheduling and topology reconfiguration of a multi-energy virtual power plant cluster under non-ideal communication

Abstract

The application discloses a kind of non-ideal communication under the toughness scheduling and topological reconstruction method of multiple energy virtual power plant cluster, it is related to electric power system and its automation field, the digital twin prediction model of mechanism and neural network residual correction fusion is constructed, high fidelity power prediction value is provided when communication is limited;Communication network is abstracted as weighted undirected graph, and the Pareto optimal topological plan is generated offline by multi-objective optimization, and when link interruption or congestion is detected, online matching execution logic link switching is carried out;The prediction value produced by digital twin is mapped as the iteration initial value of distributed consistency algorithm by the derivative of cost function, to realize initial value accurate correction;Adopt cloud edge space-time decoupling architecture, edge layer executes robust consistency control, and cloud layer guides electric carbon collaborative interaction with ladder carbon price.The application guarantees topological survivability under non-ideal communication environment, greatly reduces the number of communications required for convergence, while taking into account physical security, economy and low carbon.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, and in particular to a resilient scheduling and topology reconfiguration method for multi-energy virtual power plant clusters under non-ideal communication conditions. Background Technology

[0002] With the accelerating global energy transition and the digitalization of power systems, the deep integration of multiple energy systems has become an inevitable trend in the power industry. The complementary use of various energy forms, such as wind, solar, natural gas, and energy storage, can effectively improve energy efficiency and reduce carbon emissions. Against this backdrop, multi-energy virtual power plant clusters, as key carriers for aggregating distributed resources and achieving multi-energy complementarity, have demonstrated enormous application potential in the energy internet. Virtual power plants aggregate geographically dispersed and heterogeneous distributed generation, energy storage, and controllable loads into a unified whole, participating in power market dispatch and operation control. However, due to the wide distribution and strong heterogeneity of virtual power plant resources, the control of their physical units is highly dependent on open communication networks, such as 5G communication networks or dedicated power wireless networks. The quality of the communication link directly determines the timeliness and accuracy of control commands, thus affecting the operational safety and economy of the entire cluster.

[0003] For the coordinated control of multi-energy virtual power plant clusters, a common solution in existing technologies is a distributed coordinated control strategy based on multi-agent systems. These schemes typically employ mathematical methods such as consensus algorithms and alternating direction multiplier methods, utilizing information exchange between adjacent agent nodes to achieve optimal power allocation and economical dispatch within the cluster. To address communication uncertainties, some improved schemes attempt to introduce robust control, event-triggered mechanisms, or fault-tolerant optimization frameworks. These frameworks utilize neighborhood feedback and local information to construct fault-tolerant optimization frameworks, or employ adaptive topology adjustments to cope with packet loss and network congestion, thereby maintaining the system's basic control capabilities to a limited extent. These methods exhibit good flexibility and convergence performance under ideal communication conditions, and therefore have been widely studied and initially applied in microgrids and active distribution networks.

[0004] However, the aforementioned existing technologies still have significant shortcomings in practical engineering applications. First, existing distributed consensus algorithms heavily rely on continuous and stable communication interactions. In real industrial environments, non-ideal communication scenarios such as 5G or power wireless private networks often experience high packet loss rates, accompanied by strong latency. This uncertainty can easily lead to algorithmic iterative oscillations, convergence difficulties, and even system instability, failing to meet real-time requirements such as frequency modulation and demand response. Second, existing technologies mostly focus on tolerating uncertainty through control law compensation after communication interference occurs, essentially a passive response paradigm. The system lacks the ability to predict source load fluctuations in advance and cannot proactively perceive the communication network status and make adaptive adjustments. Therefore, there are significant bottlenecks in the control accuracy and efficiency under complex operating conditions. Furthermore, existing research often treats the communication network as a fixed topology or simplified constraint condition, neglecting the dynamic coupling relationship between communication cost, latency tolerance, and control performance. When severe link congestion or network interruption caused by physical failure occurs, there is a lack of effective structured reconstruction mechanisms, making it difficult to guarantee the survivability of the communication topology under extreme conditions. Furthermore, traditional distributed algorithms typically use arbitrary or historical values ​​as initial values ​​for iteration. In non-ideal communication environments, if the initial point is far from the global optimal solution, the system needs to consume a large number of communication time slots to converge, significantly increasing the cumulative risk of packet loss and noise interference during long-term communication. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a resilient scheduling and topology reconfiguration method for multi-energy virtual power plant clusters under non-ideal communication conditions.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: This invention provides a resilient scheduling and topology reconfiguration method for multi-energy virtual power plant clusters under non-ideal communication conditions, comprising the following steps:

[0007] S1. For heterogeneous resources within a virtual power plant cluster, a digital twin prediction model is constructed that integrates a mechanism prediction model and a neural network residual correction model. Under conditions of limited communication or missing data, a high-fidelity power prediction value is generated as a priori estimate of the cluster state.

[0008] S2. Abstract the communication network of the virtual power plant cluster into a weighted undirected graph. With the optimization objectives of minimizing communication cost, maximizing algebraic connectivity, and satisfying the time delay robustness constraint of the system spectrum radius, solve the multi-objective optimization model and generate and store a set of Pareto optimal topology schemes.

[0009] S3. The edge side continuously monitors the status of the communication link. When an abnormal link is detected, it matches and executes the logical link switching in the Pareto optimal topology plan online according to the current network status to maintain the survivability of the communication topology.

[0010] S4. Before performing distributed consensus control, extract the high-fidelity power prediction value, construct the initial estimate of the consensus marginal cost through the derivative mapping of the cost function of each resource, and use it as the initial value for the iteration of the distributed consensus algorithm.

[0011] S5 adopts a cloud-edge collaborative architecture, deploying physical security control tasks at the edge layer to perform distributed consistency iterations, and deploying economic and low-carbon scheduling tasks at the cloud layer, guiding the electricity-carbon collaborative interaction between clusters through a tiered carbon pricing mechanism.

[0012] In a preferred embodiment of the present invention, in step S1, the method for generating the high-fidelity power prediction value is to superimpose the nominal power value output by the mechanism prediction model with the nonlinear residual value output by the neural network residual correction model.

[0013] In a preferred embodiment of the present invention, in step S1, the mechanism prediction model includes at least a theoretical power model based on irradiance and temperature for photovoltaic power generation units, and a theoretical power model based on wind speed and wind energy utilization coefficient for wind power generation units.

[0014] In a preferred embodiment of the present invention, in step S1, the neural network residual correction model adopts a multilayer perceptron structure, and its input feature vector includes at least the historical power deviation sequence and environmental parameters. The output of the network is the predicted power residual value. The model is trained offline by minimizing the mean square error between the predicted residual and the actual residual.

[0015] In a preferred embodiment of the present invention, in step S2, the objective function of the multi-objective optimization model includes: minimizing the total communication cost, maximizing the algebraic connectivity, and minimizing the node degree variance; the constraints include at least connectivity constraints and time delay robustness constraints, wherein the time delay robustness constraints are characterized by the relationship between the system spectral radius and the maximum expected time delay.

[0016] In a preferred embodiment of the present invention, in step S2, the multi-objective optimization model is solved using a non-dominated sorting genetic algorithm with an elitist strategy to generate a Pareto front as the Pareto optimal topology set, which is then pre-stored in the cloud or on an edge node.

[0017] In a preferred embodiment of the present invention, in step S3, link anomalies include link congestion, packet loss rate exceeding a preset threshold, or physical interruption; the online matching method is as follows: based on the current set of active nodes and the set of available links, select the topology plan that maximizes algebraic connectivity and has the lowest communication cost under similar failure modes from the plan library, and update the routing table or communication matrix of each node to complete the logical link switching.

[0018] In a preferred embodiment of the present invention, in step S4, the initial estimate of the consensus marginal cost is calculated as follows: the high-fidelity power prediction value is substituted into the marginal cost function of each resource, and the marginal cost function is obtained by taking the first derivative of the operating cost function of each resource with respect to the power. The calculated marginal cost value is used as the initial value for iterative calculation of each node in the distributed consensus algorithm.

[0019] In a preferred embodiment of the present invention, in step S5, the distributed consensus iteration formula executed by the edge layer is:

[0020] ;

[0021] in, It is the marginal cost of node i in the kth iteration; It is the marginal cost of neighbor node j in the k-th iteration; It is the set of neighboring nodes that have communication links with node i in the communication topology after the S3 step reconstruction; It is an iterative step-size gain function that decays over time and satisfies the convergence condition in stochastic approximation theory; It is the local power imbalance, which is used as a feedback term to correct the direction of iteration; It is the feedback gain coefficient.

[0022] In a preferred embodiment of the present invention, in step S5, the tiered carbon pricing mechanism is as follows: multiple carbon emission ranges are set, each range corresponds to a different carbon trading price, and the cloud layer broadcasts the tiered carbon pricing signal to the edge nodes to guide low-carbon emission virtual power plants to transfer carbon quotas to high-carbon emission virtual power plants, thereby driving the electricity-carbon collaborative optimization between clusters.

[0023] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0024] (1) This invention provides a resilient scheduling and topology reconstruction method for multi-energy virtual power plant clusters under non-ideal communication conditions. By constructing a digital twin prediction model that integrates a mechanism model and neural network residual correction in the cloud, it can output high-fidelity power prediction values ​​by relying on the twin's advanced extrapolation when the communication link is lost or interrupted, replacing the missing real-time measurement data. The prediction model uses historical data to train the residual network, so that the prediction results closely match the actual operating state. Even during the short period of complete communication interruption, the system can still obtain reliable resource state estimates, thereby avoiding the control blind spot caused by data loss. Compared with the passive fault-tolerant strategy of the prior art that relies on real-time feedback, this invention changes the control logic from waiting for communication to resume to active calculation and extrapolation, reducing the rigid demand for continuous communication interaction from the source. Thus, in non-ideal communication environments, it can maintain the continuous generation and issuance of control commands, effectively preventing control oscillations or malfunctions caused by data interruption.

[0025] (2) This invention abstracts the communication network into a weighted undirected graph, and obtains a series of Pareto optimal topology schemes in advance through multi-objective optimization. When the edge node detects link congestion or physical interruption in real time, it matches and executes logical link switching online. It can comprehensively consider communication cost, algebraic connectivity and system spectrum radius constraints, so that the reconstructed topology can simultaneously meet the requirements of economy and dynamic stability. When a local link fails due to interference or fault, it can quickly switch to the alternative logical path to maintain the connectivity of the overall network and prevent the cluster from splitting into multiple islands. Compared with the existing technology that regards the communication network as a fixed topology or simple constraint, this invention actively gives the communication topology structure elasticity, so that the communication network still maintains effective algebraic connectivity under continuous high packet loss rate or strong delay interference, providing a reliable information interaction channel for the upper-layer consensus algorithm, thereby avoiding convergence failure caused by topology breakage.

[0026] (3) This invention constructs the initial value of the distributed robust consensus algorithm by mapping the high-fidelity power prediction value output by the digital twin through the derivative of each resource cost function. Since the digital twin model has high prediction accuracy after residual correction, the initial value is extremely close to the global optimal marginal cost. As a result, the initial error norm of the algorithm is greatly reduced, and the number of communication iterations required for convergence decreases logarithmically. Compared with the existing technology that uses random or historical values ​​as initial values, this invention avoids consuming a large number of communication time slots to gradually approach the optimal solution due to the initial point deviating too far from the optimal solution. In a non-ideal communication environment, each communication interaction may encounter packet loss or noise interference. The very few iterations mean that the system can complete convergence in a short time, thereby avoiding the risk of cumulative failure caused by long-term continuous communication and effectively ensuring the speed of frequency response and power allocation.

[0027] (4) The present invention adopts a cloud-edge collaborative architecture, which deploys physical security control tasks with fast response speed and high real-time requirements, such as frequency recovery, to be executed entirely at the edge layer, while deploying economic scheduling and carbon trading guidance tasks with longer time scales and higher computational complexity at the cloud layer. The invention also drives the electric carbon collaborative interaction between clusters through a tiered carbon pricing mechanism. Compared with the existing technology that couples security control and economic scheduling to be executed at the same layer, the present invention can eliminate the mutual constraints between fast dynamics and slow optimization. When non-ideal communication causes cloud-edge interaction delays or interruptions, the edge layer can still independently complete physical quantity control and will not delay the security response due to waiting for cloud instructions. At the same time, the cloud layer can still re-optimize and issue scheduling strategies after communication is restored, thereby achieving a multi-dimensional unity of system security, economy and low carbon under harsh communication conditions. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of the MEVPP cluster cloud-edge collaborative resilience regulation architecture under non-ideal communication conditions of the present invention;

[0030] Figure 2 This is a flowchart illustrating the digital twin modeling process of the present invention;

[0031] Figure 3 This is a schematic diagram of the multi-energy virtual power plant cluster architecture based on IEEE 33 nodes of the present invention. Detailed Implementation

[0032] The technical solutions of 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.

[0033] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0034] Example 1:

[0035] A resilient scheduling and topology reconfiguration method for multi-energy virtual power plant clusters under non-ideal communication conditions includes the following steps:

[0036] S1. For heterogeneous resources within a virtual power plant cluster, a digital twin prediction model is constructed that integrates a mechanism prediction model and a neural network residual correction model. Under conditions of limited communication or missing data, a high-fidelity power prediction value is generated as a priori estimate of the cluster state.

[0037] S2. Abstract the communication network of the virtual power plant cluster into a weighted undirected graph. With the optimization objectives of minimizing communication cost, maximizing algebraic connectivity, and satisfying the time delay robustness constraint of the system spectrum radius, solve the multi-objective optimization model and generate and store a set of Pareto optimal topology schemes.

[0038] S3. The edge side continuously monitors the status of the communication link. When an abnormal link is detected, it matches and executes the logical link switching in the Pareto optimal topology plan online according to the current network status to maintain the survivability of the communication topology.

[0039] S4. Before performing distributed consensus control, extract the high-fidelity power prediction value, construct the initial estimate of the consensus marginal cost through the derivative mapping of the cost function of each resource, and use it as the initial value for the iteration of the distributed consensus algorithm.

[0040] S5 adopts a cloud-edge collaborative architecture, deploying physical security control tasks at the edge layer to perform distributed consistency iterations, and deploying economic and low-carbon scheduling tasks at the cloud layer, guiding the electricity-carbon collaborative interaction between clusters through a tiered carbon pricing mechanism.

[0041] The cloud-edge collaborative resilience regulation architecture upon which this embodiment is based is as follows: Figure 1 As shown, the architecture is logically divided into a cloud collaboration layer and an edge control layer. The cloud collaboration layer is responsible for performing computationally complex and time-scaled global optimization tasks, including building and updating digital twin models, solving for Pareto optimal communication topology plans, and formulating electricity-carbon coordinated scheduling strategies. The edge control layer is deployed at each virtual power plant substation and is responsible for performing physical control tasks with high real-time requirements and short time scales. Specifically, this includes collecting local measurement data, monitoring communication link status, executing distributed robust consensus algorithms, and autonomously switching communication logical links according to cloud-based plans. The cloud and edge interact to a limited extent through a dedicated power wireless network and upload runtime status data aggregated by the edge layer. Even in extreme cases where the cloud-edge communication link experiences a brief interruption, the edge layer can still independently and stably complete the coordinated control of its own virtual power plant and adjacent nodes, relying on locally cached digital twin models and topology plans.

[0042] The concept behind this invention is to realize a paradigm shift from passive fault tolerance to active resilience. Specifically, instead of passively waiting for communication interference to occur and then compensating through control laws, it uses a digital twin model that integrates mechanisms and data to proactively and forward-lookingly deduce the system state under conditions of limited communication or missing data. This provides the control algorithm with an initial value close to the optimal solution. At the same time, through a topology reconstruction mechanism that combines offline optimization and online matching, a survivable communication base map is prepared in advance for the upper-level control algorithm. Thus, the strategy of exchanging computation for communication can effectively reduce the dependence on continuous, high-quality communication interaction, thereby improving the operational resilience and scheduling efficiency of the entire cyber-physical system in non-ideal communication environments.

[0043] In this embodiment, step S1 is used to address the decision-making blind spot problem caused by data loss when communication is limited in traditional control methods. Existing technologies typically rely on real-time measurement data for feedback control. Once data packet loss or transmission delay occurs, the controller cannot obtain an accurate system state, leading to inaccurate control commands or system oscillations. To overcome this deficiency, such as... Figure 2 As shown, this embodiment introduces digital twin technology to construct a hybrid driving model that can actively generate high-fidelity power prediction values.

[0044] Specifically, for the heterogeneous distributed resources within the virtual power plant cluster, including photovoltaic power generation units, wind power generation units, and energy storage units, mechanism models based on their nominal physical characteristics are established respectively.

[0045] Taking a photovoltaic power generation unit as an example, the theoretical output power mechanism formula is as follows:

[0046] .

[0047] in, It is the theoretical output power calculated by the mechanistic model; The photovoltaic modules are tested under standard conditions (irradiance 1000 W / m). 2 Rated power at battery temperature 25 ℃; γ is the actual total irradiance received by the tilted surface of the photovoltaic module; γ is the power temperature coefficient of the photovoltaic module. This is the current operating temperature of the photovoltaic cell.

[0048] Taking a wind power generation unit as an example, the theoretical output power mechanism formula is as follows:

[0049] .

[0050] in, It is the theoretical output power calculated by the wind turbine mechanism model; A is the air density; A is the swept area of ​​the fan impeller. It is the wind energy utilization coefficient, which is a nonlinear function of the tip speed ratio λ and the pitch angle β; This is the real-time wind speed.

[0051] Furthermore, while the aforementioned mechanism model can reflect the basic physical laws of resources, its accuracy is limited by factors such as model simplification, parameter aging, and environmental noise, resulting in a nonlinear residual between it and the actual output. To correct this deviation, this embodiment constructs a multilayer perceptron neural network as a residual correction model. The input feature vector of this network... This includes, but is not limited to: the deviation sequence between historical measured power values ​​and mechanistic model predictions, and environmental parameters such as current and historical temperature, humidity, and irradiance. The network output is the predicted power residual value. .

[0052] During the offline training phase, the multilayer perceptron network is trained using a historical dataset collected from actual operation. This dataset contains the input parameters and output values ​​of the mechanistic model at the same time point. and the corresponding measured power values The goal of training is to minimize the loss function, which is defined as the mean square error between the residuals predicted by the neural network and the theoretical residuals.

[0053] Furthermore, after the model training is completed, it is deployed in the cloud. During real-time operation, the cloud-based digital twin performs the following calculations: based on real-time collected meteorological and other feedforward data, it calls the mechanistic model to calculate the theoretical power. ; the feature vector of the current state The input is fed into a pre-trained multilayer perceptron network to obtain the predicted residual. High-fidelity power prediction It is generated by the following formula:

[0054] .

[0055] Understandably, this predicted value This serves as a high-fidelity prior estimate of the cluster state under non-ideal communication environments. For example, when a virtual power plant experiences packet loss in its communication link with the cloud or other nodes, the node can directly use [the prior estimate] during the several control cycles it waits for data recovery. This data is used as the actual value of its current output power in subsequent collaborative calculations, thereby avoiding control blind spots caused by missing data.

[0056] In this embodiment, step S2 is used to address the problem that static communication topologies in the prior art cannot adapt to link congestion or failure, resulting in poor system survivability; this embodiment proposes an adaptive reconfiguration mechanism that combines offline optimization and online matching.

[0057] Specifically, the communication network of the virtual power plant cluster is abstracted as a weighted undirected graph G=(V,E,W). Here, V is the set of nodes (each virtual power plant agent); E is the set of all possible logical communication links between nodes; and W is the set of link weights, where the weight values ​​are... It can comprehensively characterize the communication cost between node i and node j.

[0058] The core of this step is to solve a multi-objective optimization model to generate a set of Pareto optimal topology schemes. The objective function and constraints of the optimization model are as follows:

[0059] The first objective function is to minimize the total communication cost, and its expression is:

[0060] .

[0061] in, It is a binary decision variable; a value of 1 indicates a link. Selected into the current communication topology; a value of 0 indicates that it has not been selected.

[0062] The second objective function is to maximize the survivability of the communication topology by maximizing the algebraic connectivity of the graph. To achieve this, L is the Laplacian matrix of graph G. This is its second smallest eigenvalue, and its magnitude directly reflects the full connectivity of the graph; the larger the value, the stronger the connectivity. In this embodiment, for ease of calculation, this objective is transformed into minimizing its reciprocal or negative value, such as... .

[0063] Objective function three aims to minimize the variance of node degree to avoid monopolistic key nodes and balance communication load. The expression is:

[0064] .

[0065] in, It represents the degree of node i in the currently selected topology, i.e., the number of links connected to this node. It is the average degree of all nodes.

[0066] Furthermore, the optimized model must satisfy the following constraints:

[0067] Connectivity constraints: ,in It is a preset minimum algebraic connectivity threshold used to ensure the basic connectivity of the topology.

[0068] Delay robustness constraints: This constraint ensures that the upper-level distributed control algorithm remains stable under the maximum expected time delay τ; where ρ(·) represents the spectral radius of the matrix, A and B are the parameter matrices in the system state-space model, and K is the control gain matrix. This constraint ensures that the dynamic error of the control system converges even under the worst-case time delay.

[0069] Furthermore, the multi-objective optimization model is solved using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with an elitist strategy. This algorithm, by simulating the natural evolutionary process, can efficiently search for a set of non-dominated solutions that satisfy the constraints, i.e., the Pareto front. Each solution on the Pareto front corresponds to a specific communication topology. .

[0070] Optimal topology plan The corresponding key parameters are pre-stored in the cloud database and can be further distributed and cached in the local storage units of each edge node so that they can be quickly accessed in the event of network anomalies.

[0071] In this embodiment, during step S3, as the system operates in real time, the edge controllers deployed at each virtual power plant substation continuously monitor the quality of their communication links. The monitored metrics include: packet loss rate, round-trip time, and signal-to-noise ratio of the received signal. The packet loss rate is calculated based on the proportion of data packets for which no acknowledgment is received per unit time, and the round-trip time is measured by sending heartbeat packets.

[0072] When the edge controller detects that the packet loss rate of a critical link exceeds a preset threshold, such as a packet loss rate higher than 20% within 5 consecutive seconds, or a round-trip delay exceeding the maximum value allowed by the system stability margin, or even a physical interruption occurs, such as multiple consecutive heartbeat requests without response, a topology reconfiguration event is triggered.

[0073] In this embodiment, the node that triggers the reconstruction event will match the current network state, such as the set of currently activated nodes and the set of links that still have communication capabilities, in the local topology plan library. The matching strategy can be based on the lookup table method, that is, according to the preset current state-optimal topology mapping relationship, the plan that can maximize algebraic connectivity and has low communication cost under similar failure modes is selected.

[0074] Furthermore, once a suitable contingency plan is matched... The node or the designated first node will be responsible for initiating the logical link switching process. The switching process is achieved by updating the routing tables or communication matrices of each node, informing all relevant nodes to activate the contingency plan. The system defines logical links and simultaneously shuts down or puts to sleep existing links that are faulty or too costly. The entire process requires no physical line modifications and only completes network reconstruction at the logical level of the communication protocol stack. This restores the connectivity of the entire cluster communication network in a very short time, providing the upper-layer control algorithm with a robust information interaction base map that is robust even under packet loss and latency interference.

[0075] In this embodiment, step S4 is used to eliminate the shortcomings of traditional distributed consensus algorithms, such as slow convergence and susceptibility to interference in non-ideal communication environments due to improper selection of initial values. In the prior art, the initial values ​​of the algorithm are usually arbitrarily set, leading to an initial error norm. The problem is very large, requiring a significant amount of communication time slots for iterative correction. In non-ideal communication environments, each communication interaction carries the risk of packet loss and latency.

[0076] To address this issue, this embodiment directly utilizes the high-fidelity power prediction value output by the digital twin model in step S1 before performing distributed robust consistency control at the edge layer. To construct an initial estimate of the consistent marginal cost.

[0077] Specifically, assume that the power generation cost of each virtual power plant unit can be approximated by a quadratic function:

[0078] .

[0079] in, Is node i at the output power Operating costs at that time; , , These are the coefficients of the quadratic, linear, and constant terms of the cost function. They are dimensionless or constants with corresponding dimensions, and their specific values ​​are determined by the economic characteristics of each resource. Marginal cost is the first derivative of the cost function with respect to power.

[0080] .

[0081] The core of this algorithm lies in using the power value predicted by the digital twin. Substituting directly into the marginal cost function above, we calculate the corresponding marginal cost and use it as the initial value for the consensus algorithm iteration:

[0082] .

[0083] Furthermore, since the digital twin model in step S1 undergoes residual correction, its output power prediction value... Compared with actual operating point Very close, thus the initial value of the marginal cost calculated from this is It will also inevitably approach the globally optimal marginal cost. This means the initial error norm It has been greatly reduced in size.

[0084] Based on the convergence analysis of spectral graph theory and distributed consensus algorithms, the number of iterations required for convergence is... It is proportional to the natural logarithm of the initial error norm, and the relationship can be approximated by the following formula:

[0085] .

[0086] in, ρ is the preset convergence accuracy threshold; ρ is the convergence spectrum radius of the algorithm, and its value is between 0 and 1.

[0087] It is understandable that, due to the initial value correction strategy of this invention, The value is extremely small, making the molecule The values ​​decrease logarithmically. For example, when the initial error of the traditional algorithm is 10, while the initial error of the modified algorithm in this invention is 0.1, the values ​​of ρ and... Under fixed conditions, The reduction will be very significant, resulting in a substantial decrease in the number of rounds of communication and interaction between the algorithm and the physical world. This fundamentally avoids the systemic risks of packet loss and noise accumulation during long-term, multi-round communication, and enables fast and stable control.

[0088] In this embodiment, step S5 achieves the unification of multiple objectives—physical security control and economical low-carbon scheduling—through a cloud-edge spatiotemporal decoupling architecture. Physical security control tasks with extremely high real-time requirements, such as rapid frequency recovery, are entirely deployed and executed at the edge layer. When a system frequency deviation is detected, each edge controller executes a robust consistency algorithm, the core iterative formula of which is:

[0089] .

[0090] in, It is the marginal cost of node i in the kth iteration; It is the marginal cost of neighbor node j in the k-th iteration; It is the set of neighboring nodes that have communication links with node i in the communication topology after the S3 step reconstruction; It is an iterative step-size gain function that decays over time and satisfies the convergence condition in stochastic approximation theory; It is the local power imbalance, which is used as a feedback term to correct the direction of iteration; It is the feedback gain coefficient.

[0091] Furthermore, as the iteration progresses, the marginal costs of each node will gradually converge, and this will be further determined by the mapping relationship between marginal cost and power. It can adjust the output power of each unit, and ultimately share the power disturbance, restoring the system frequency from the deviation state to its rated value.

[0092] Meanwhile, economic activities and carbon trading tasks with longer time scales are deployed in the cloud layer; after receiving the aggregated status information sent by each edge node, the cloud collaboration layer performs electricity-carbon collaborative optimization. In this embodiment, a tiered carbon price model is introduced as an economic incentive mechanism. This model sets multiple carbon emission ranges, each range corresponding to a different carbon trading price. When the carbon emissions of a virtual power plant exceed its free allowance and enter a higher range, the unit cost of purchasing carbon allowances will increase in a tiered manner.

[0093] By broadcasting this tiered carbon pricing signal to each edge node, the cloud guides the optimized redistribution of carbon emission allowances within the cluster. Virtual power plants with a high proportion of clean energy and low carbon emissions can profit by selling excess carbon allowances, while virtual power plants with high carbon emissions face higher compliance costs. Driven by this economic signal, each virtual power plant naturally adjusts its power generation plan and electricity-energy interaction strategy in the process of maximizing its own interests, thereby achieving Pareto improvement in system economy and low carbon emissions on a global scale.

[0094] To verify the effectiveness of the method proposed in this invention, this embodiment is implemented in an IEEE 33-node system (refer to) consisting of three heterogeneous multi-energy virtual power plants (MEVPP1, MEVPP2, and MEVPP3). Figure 3 The system was simulated and verified using a 5G wireless communication environment. The baseline packet loss rate was set at 15%, the signal-to-noise ratio (SNR) was 20 dB, and random delays of 200 ms to 500 ms were introduced to simulate harsh industrial communication conditions. The verification results demonstrate the effectiveness of the method in the following three aspects.

[0095] Noise Resistance and Convergence Verification: In the simulated scenario, the proposed method uses digital twin predictions to correct the initial iteration values. Even with a 15% packet loss rate in the link, the initial marginal cost resultant point of the algorithm is already within the neighborhood of the optimal solution after startup. Data shows that the algorithm only needs about 50 iterations for the marginal costs of each virtual power plant to converge to the global Nash equilibrium point, and the total power generation cost quickly reaches the minimum. This proves that the proposed method significantly reduces the dependence on continuous, lossless communication by minimizing the initial error norm, and achieves fast convergence under non-ideal communication conditions.

[0096] Frequency security verification: The simulation simulated a 10% instantaneous power deficit disturbance within MEVPP2. Following this disturbance, the distributed frequency consensus protocol deployed at the edge layer using the proposed method was activated. Due to the communication topology reconstruction mechanism ensuring the connectivity of critical links, each edge node was able to complete the necessary information exchange and state synchronization. During this process, the system frequency dropped to a minimum of 49.75 Hz, and after approximately 200 control cycles (about 2 seconds), the frequency was smoothly and stably restored to its rated value of 50 Hz. Throughout the entire dynamic adjustment process, the system did not experience secondary frequency drops or oscillations due to communication delays or packet loss, demonstrating the independent physical security control capability of the edge layer under non-ideal communication conditions.

[0097] Economic and Carbon Reduction Effect Verification: This embodiment compares the cloud-edge collaborative method of the present invention with the traditional independent operation mode of each virtual power plant. In a typical 24-hour operation scenario, the independent operation mode, due to the lack of inter-cluster electricity-carbon coordination, allows each virtual power plant to rely solely on its own resources for scheduling, resulting in high operating costs. However, with the solution of the present invention, the cloud layer guides electricity-carbon interaction between MEVPP1 (mainly photovoltaic and energy storage, with surplus carbon allowances) and MEVPP2 (mainly gas turbines, with high carbon emissions) through a tiered carbon pricing mechanism. MEVPP1 sells its surplus photovoltaic power and carbon allowances to MEVPP2, while MEVPP2 reduces its high-cost, high-emission gas turbine power generation. Simulation results show that the total cluster operating cost of the proposed method is reduced by 89.9% compared to the independent operation mode, and the total cluster carbon emissions are reduced by 35%.

[0098] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0099] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A resilient scheduling and topology reconfiguration method for multi-energy virtual power plant clusters under non-ideal communication conditions, characterized in that, Includes the following steps: S1. For heterogeneous resources within a virtual power plant cluster, a digital twin prediction model is constructed that integrates a mechanism prediction model and a neural network residual correction model. Under conditions of limited communication or missing data, a high-fidelity power prediction value is generated as a priori estimate of the cluster state. The method for generating the high-fidelity power prediction value is as follows: the nominal power value output by the mechanism prediction model is superimposed with the nonlinear residual value output by the neural network residual correction model. S2. Abstract the communication network of the virtual power plant cluster into a weighted undirected graph. With the optimization objectives of minimizing communication cost, maximizing algebraic connectivity, and satisfying the time delay robustness constraint of the system spectrum radius, solve the multi-objective optimization model and generate and store a set of Pareto optimal topology schemes. S3. The edge side continuously monitors the status of the communication link. When an abnormal link is detected, it matches and executes the logical link switching in the Pareto optimal topology plan online according to the current network status to maintain the survivability of the communication topology. S4. Before performing distributed consensus control, extract the high-fidelity power prediction value, construct the initial estimate of the consensus marginal cost through the derivative mapping of the cost function of each resource, and use it as the initial value for the iteration of the distributed consensus algorithm. S5 adopts a cloud-edge collaborative architecture, deploying physical security control tasks at the edge layer to perform distributed consistency iteration, and deploying economic and low-carbon scheduling tasks at the cloud layer, guiding the electricity-carbon collaborative interaction between clusters through a tiered carbon pricing mechanism; The distributed consensus iteration formula executed by the edge layer is as follows: ； in, It is the marginal cost of node i in the kth iteration; It is the marginal cost of neighbor node j in the k-th iteration; It is the set of neighboring nodes that have communication links with node i in the communication topology after the S3 step reconstruction; It is an iterative step-size gain function that decays over time and satisfies the convergence condition in stochastic approximation theory; It is the local power imbalance, which is used as a feedback term to correct the direction of iteration; It is the feedback gain coefficient.

2. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S1, the mechanism prediction model includes at least a theoretical power model based on irradiance and temperature for photovoltaic power generation units, and a theoretical power model based on wind speed and wind energy utilization coefficient for wind power generation units.

3. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S1, the neural network residual correction model adopts a multilayer perceptron structure, and its input feature vector includes at least the historical power deviation sequence and environmental parameters. The output of the network is the predicted power residual value. The model is trained offline by minimizing the mean square error between the predicted residual and the actual residual.

4. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S2, the objective function of the multi-objective optimization model includes: minimizing the total communication cost, maximizing the algebraic connectivity, and minimizing the node degree variance; the constraints include at least connectivity constraints and time delay robustness constraints, wherein the time delay robustness constraints are characterized by the relationship between the system spectral radius and the maximum expected time delay.

5. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S2, the multi-objective optimization model is solved using a non-dominated sorting genetic algorithm with an elitist strategy to generate a Pareto front as the Pareto optimal topology set, which is pre-stored in the cloud or on an edge node.

6. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S3, link anomalies include link congestion, packet loss rate exceeding a preset threshold, or physical interruption. The online matching method is as follows: based on the currently active node set and the available link set, select the topology plan that maximizes algebraic connectivity and has the lowest communication cost under similar failure modes from the plan library, and update the routing table or communication matrix of each node to complete the logical link switching.

7. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S4, the initial estimate of the consensus marginal cost is calculated as follows: the high-fidelity power prediction value is substituted into the marginal cost function of each resource. The marginal cost function is obtained by taking the first derivative of the operating cost function of each resource with respect to the power. The calculated marginal cost value is used as the initial value for iterative calculation by each node in the distributed consensus algorithm.

8. The resilient scheduling and topology reconfiguration method for a multi-energy virtual power plant cluster under non-ideal communication conditions according to claim 1, characterized in that, In step S5, the tiered carbon pricing mechanism is as follows: multiple carbon emission ranges are set, each range corresponds to a different carbon trading price, and the cloud layer broadcasts the tiered carbon pricing signal to the edge nodes to guide low-carbon emission virtual power plants to transfer carbon quotas to high-carbon emission virtual power plants, thereby driving the electricity-carbon collaborative optimization between clusters.

Images