Virtual power plant double-layer optimization scheduling method for electricity market transaction

Abstract

The invention discloses an electric power market transaction-oriented virtual power plant double-layer optimization scheduling method, and belongs to the field of electric power market transaction and virtual power plant scheduling optimization. According to the method, aiming at the problem of real-time calculation pressure caused by complexity of nested solution of a double-layer structure and large-scale variables, an upper-layer mixed integer linear programming model taking total bidding profit maximization as a target and a lower-layer mixed integer programming model taking internal scheduling cost minimization as a target are established; and setting a coordination factor including bidding electric quantity deviation penalty and reserve capacity compensation, decomposing a lower-layer model into a plurality of sub-modules according to resource types by adopting a distributed optimization architecture, deploying the sub-modules in an edge controller, performing iterative coordination and parallel solution based on an alternating direction multiplier method, and finally outputting bidding electric quantity and a scheduling instruction. The method is mainly used for formulating a bidding strategy and an internal scheduling scheme when a virtual power plant participates in a power market transaction, so that profit maximization is realized, and the scheduling cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power market transaction and virtual power plant dispatch optimization. More particularly, the present application relates to a virtual power plant bi-level optimization dispatch method for power market transaction. BACKGROUND

[0002] In the power market environment, virtual power plant, as a business model of aggregating multiple distributed energy resources to participate in market transactions, its core function is to form a unified bidding strategy to participate in market bidding through internal resource coordination optimization. The mathematical essence of this process is usually modeled as a bi-level optimization problem, where the upper model is responsible for market bidding decisions to maximize revenue, and the lower model is responsible for the dispatch of internal distributed resources to minimize operating costs. This structure aims to maximize overall benefits.

[0003] However, this bi-level optimization structure faces significant challenges in actual solving. The primary difficulty stems from its inherent nested complexity. The decision of the upper model (bidding power) directly affects the constraint boundary and optimization space of the lower model, while the optimization result of the lower model (actual output and marginal cost) is a key input to the objective function and constraint conditions of the upper model. This close two-way coupling makes the two models unable to be solved independently, and the traditional solution method usually uses an iterative method, taking the solution of one layer model as a fixed parameter into the other layer model and repeatedly operating until the change of the solution of both models meets certain conditions. This nested iteration mechanism leads to a long calculation process, and each overall iteration needs to solve two complex mathematical programming problems (usually mixed integer programming problems) completely, which makes the overall computational complexity rise sharply and makes it difficult to meet the requirements of computational efficiency in practical applications.

[0004] On the other hand, the scale of resources aggregated by virtual power plant is huge and diverse, including distributed photovoltaic, wind turbine, energy storage system, controllable load, and electric vehicle cluster, etc. Each type of resource needs a set of variables and constraints to accurately describe its physical characteristics and operating limits, such as the predicted output curve of photovoltaic and wind power, the state of charge and state of charge continuity constraints of energy storage devices, the start-stop state and minimum operating time constraints of controllable loads, and the charging and discharging power and user demand constraints of electric vehicles. When the dispatching period covers the entire trading day (such as 24 hours, with an interval of one hour or 15 minutes), the number of decision variables and constraints increases linearly or even combinatorially with the number of resources and time periods, forming a super-high-dimensional mathematical programming problem. Large-scale variables and constraints directly lead to heavy computational burden.

[0005] Especially thorny is that the integer decision variables (such as the start-stop state of the equipment, the switching of the load) and the nonlinear constraints (such as the energy state transfer of the energy storage) commonly existing in the optimization problem make the whole problem present non-convexity. Non-convex optimization problems usually have multiple local optimal solutions, and the conventional optimization algorithm is easy to fall into one of the local solutions and cannot get out, and it is difficult to guarantee to obtain the global optimal solution or a high-quality solution. At the same time, solving such a large-scale non-convex mixed integer programming problem itself is very expensive in calculation.

[0006] The above calculation complexity, large-scale variables and constraints, and non-convexity together constitute a huge real-time calculation pressure. The power market, especially the day-ahead market and the intraday market of rolling out, has strict time limit requirements for the submission of the bid, and the time window left for the virtual power plant to carry out optimization calculation is very limited. It is a great test for both the calculation hardware and the algorithm to complete such a complex large-scale optimization solution in a limited time. When dealing with such problems, the traditional centralized optimization method is often unable to meet the real-time requirements of practical applications due to too long calculation time, or has to greatly simplify the model to gain speed, which may sacrifice the optimization accuracy and economy. Therefore, how to efficiently and reliably solve the double-layer optimization scheduling problem of the virtual power plant in a limited time has become a technical difficulty to be solved. SUMMARY

[0007] An object of the present application is to solve at least the above problems and to provide at least the advantages to be described later.

[0008] In order to achieve these objects and other advantages according to the present application, a virtual power plant double-layer optimization scheduling method for power market transactions is provided, comprising: establishing an upper layer model, the upper layer model being a mixed integer linear programming problem with the maximization of the total bid profit of the virtual power plant as the target, and the decision variable being the bid power of each transaction period of the power market; establishing a lower layer model, the lower layer model being a mixed integer programming problem with the minimization of the internal scheduling cost as the target, and the decision variable including the predicted output curve of the distributed photovoltaic, the predicted output curve of the wind turbine, the charge-discharge state and power of the energy storage device, the start-stop state and adjustment amount of the controllable load, and the charge-discharge power of the electric vehicle cluster; setting a coordination factor between the upper layer model and the lower layer model, the coordination factor including a bid power deviation penalty term and a reserve capacity compensation term; adopting a distributed optimization architecture, deploying the upper layer model to the central controller of the virtual power plant, and decomposing the lower layer model into a photovoltaic scheduling submodule, a wind power scheduling submodule, an energy storage scheduling submodule, a load scheduling submodule, and an electric vehicle scheduling submodule according to the resource type; in each iteration of the iterative solution process by the alternating direction multiplier method, performing: The central controller broadcasts the current coordination factor and the bidding power constraint range to all sub-modules; Each sub-module solves the local optimization problem based on the received coordination factor in parallel, and uploads the optimal scheduling scheme and marginal cost information to the central controller; The central controller aggregates the marginal cost information of each sub-module, updates the bidding power and the coordination factor; When the change in bidding power is less than the preset power change threshold and the change in coordination factor is less than the factor change threshold, the iteration is terminated; Output the final bidding power to the power trading platform, and issue resource scheduling instructions to the local controller.

[0009] Preferably, the initial value of the coordination factor is set to a random value in the 80%-120% interval of the power market forecast price; The upper limit of the bidding power constraint range is 95% of the maximum adjustable output of all distributed energy sources; the lower limit is 105% of the maximum adjustable capacity of all controllable loads.

[0010] Preferably, the energy storage scheduling sub-module adopts a time decoupling strategy, divides the 24-hour scheduling period into 4-6 continuous time period groups, solves each time period group using mixed integer programming, connects the time period groups through energy storage state transfer equations, and the energy storage SOC state at the boundary point of the time period group is used as a coupling variable to participate in the alternating direction multiplier method update.

[0011] Preferably, in the electric vehicle scheduling sub-module, a clustering compression algorithm is used, specifically: the electric vehicle cluster is divided into 10-20 equivalent vehicle groups according to the remaining power range, each equivalent vehicle group is characterized by three parameters: the upper limit of the charging and discharging power, the upper limit of the energy capacity, and the charging and discharging efficiency, and when the scheduling instruction is issued, the power is proportionally distributed according to the actual parameters of the vehicles in the equivalent vehicle group.

[0012] Preferably, a relaxation mechanism is also included in the iterative solving process, specifically: when the improvement amount of the objective function for the last 3 iterations is less than the improvement threshold, the integer variables in the mixed integer programming are fixed as the current optimal solution, and the remaining continuous variable optimization problem is solved using the interior point method until the termination condition is met.

[0013] Preferably, a real-time calculation monitor is also included, when the number of iterations reaches 5 times and still does not converge: Start the reduced order mode, fix the predicted output curve of the photovoltaic scheduling sub-module and the wind power scheduling sub-module as the day-ahead prediction value, relax the charging and discharging state variables of the energy storage scheduling sub-module to continuous variables, and keep the integer constraints of the controllable load scheduling sub-module and the electric vehicle scheduling sub-module to continue solving.

[0014] Preferably, in the updating rule of the coordination factor, a dynamic step adjustment coefficient is set, which is adaptively adjusted according to the historical variation trend of the deviation between the bidding power and the actual output in the last three iterations: when the deviation shows an expanding trend, the step is reduced to stabilize the convergence; and when the deviation shows a shrinking trend, the step is increased to accelerate the convergence.

[0015] Preferably, a communication interruption fault tolerance mechanism is further included, when the central controller does not receive feedback information of any one or more of the sub-modules within a preset timeout, the local cache scheduling scheme of the corresponding sub-module is automatically enabled to participate in the current round of aggregated updating, and the local cache scheduling scheme is the optimal scheduling scheme successfully received from the corresponding sub-module last time.

[0016] Preferably, an economic safety check step is set after the aggregated updating of the central controller and before the convergence determination, whether the estimated total profit of the current iteration is lower than a preset break-even threshold is checked, if yes, a conservative bidding strategy is triggered, the upper limit of the bidding power range is lowered to 80% of the maximum adjustable output of the distributed energy, and the optimization of each sub-module and the central aggregation are re-performed.

[0017] The present application at least includes the following beneficial effects: Firstly, the present application fundamentally changes the solution paradigm of the problem by constructing a distributed optimization architecture, decoupling the traditional centralized double-layer coupled problem into bidding decision of the upper layer and multi-resource parallel scheduling of the lower layer. The upper layer central controller focuses on market game, and the lower layer sub-modules optimize locally by using edge computing capacity, and interact with each other through coordination factors and marginal cost information, replacing the traditional full-variable nested iteration. This architecture greatly reduces the single-point computing load and communication overhead, enables the optimization problem of large-scale resource aggregation to be solved efficiently within the limited time window specified by the power market, and solves the core contradiction between the complexity of double-layer structure and the real-time calculation, providing key technical support for virtual power plants participating in real-time market.

[0018] Secondly, the present application provides intelligent starting guidance for the iteration process by designing a random initialization strategy of the coordination factor and a compact constraint range of the bidding power. The initial value of the coordination factor is randomly generated based on the predicted price, so that it can fall into the effective neighborhood with a high probability, avoiding the pre-oscillation caused by the initial value deviating too much, and significantly improving the convergence stability of the algorithm. The upper and lower limits of the bidding power constraint range are set considering the uncertainty of distributed energy output and the adjustment potential of controllable load, which ensures the coverage of the search space to feasible solutions while avoiding redundant exploration in invalid areas. The two work together to effectively overcome the initial parameter sensitivity problem, ensuring that the algorithm can quickly and stably enter the convergence track, and improving the overall optimization efficiency.

[0019] Third, the present application is aimed at the model dimension disaster problem caused by the energy coupling constraint of the energy storage device in the whole time period. A time decoupling strategy is creatively proposed. By dividing the long-period scheduling into several continuous short-period groups, the originally high-dimensional and coupled mixed integer programming problem is decomposed into multiple low-dimensional and parallel processing sub-problems, which greatly reduces the solving complexity of a single problem. At the same time, the state of charge (SOC) of the boundary points of the time period group is ingeniously set as a coupling variable, and is coordinated and updated in the framework of the alternating direction multiplier method, ensuring the continuity and consistency of the segmented optimization results on the global time scale. This method realizes the order of magnitude reduction of the computational complexity on the premise of maintaining the optimization accuracy, and solves the dimension explosion problem in the energy storage scheduling.

[0020] Fourth, the present application abstracts a large number of electric vehicles with different characteristics into a small number of representative equivalent vehicle groups by using a clustering compression algorithm, and realizes the extreme simplification of the optimization model. The creativity of this method lies in abandoning the traditional idea of accurate modeling of each individual, and instead starting from the macro characteristics of the group, using a small integer variable to represent the scheduling potential of the entire cluster, thereby compressing the variable and constraint scale by several orders of magnitude. This not only solves the "constraint explosion" problem caused by individual differences, making it possible to solve complex integer programming problems in real time, but also ensures the executability of the final scheduling instruction at the physical level through the equal proportion allocation mechanism, and ingeniously balances the calculation efficiency and scheduling accuracy.

[0021] Fifth, the present application effectively deals with the local optimal trap problem caused by the non-convexity of mixed integer programming by introducing an intelligent relaxation mechanism. This mechanism can dynamically monitor the optimization process, and when it detects that the improvement of the objective function has stagnated, it automatically fixes the integer variables at the current optimal solution, thereby converting the non-convex mixed integer programming problem into a convex continuous optimization problem that is easy to solve. This adaptive problem conversion capability enables the algorithm to escape from the constraint of local optimal points and continue to search for better solutions in a larger solution space, significantly improving the quality of the final solution and the global convergence ability of the algorithm. This method avoids the risk of getting stuck in computational difficulties of traditional branch and bound methods, and enhances the reliability and economy of the optimization results without significantly increasing the computational burden.

[0022] Sixth, the application provides a reliable guarantee mechanism for coping with market bidding deadlines by setting a real-time calculation monitor and introducing a reduced-order mode. When the number of iterations reaches a safety threshold and the model still does not converge, the mechanism automatically triggers the model reduction process: fixes the predicted output of uncontrollable resources, relaxes the integer constraint of energy storage, while retaining the integer characteristics of key loads and electric vehicles for accurate optimization. This differentiated and targeted simplification strategy does not simply abandon optimization, but sacrifices part of the secondary accuracy to achieve a sharp increase in calculation speed based on a deep understanding of the influence weight of each subsystem on the final decision, ensuring that a feasible and more economical suboptimal solution is generated within a strict time limit, effectively avoiding the risk of transaction interruption and zero-revenue situation due to overtime.

[0023] Seventh, the application successfully solves the inherent problem of fixed step size in complex working conditions by designing a dynamic step size adjustment mechanism based on historical deviation trends, making the update process of the coordination factor intelligent and adaptive. The mechanism can dynamically adjust the step size according to the recent optimization situation: reduce the step size to enhance stability in oscillation trends, and increase the step size to speed up the process in convergence trends. This feedback adjustment mechanism makes the algorithm adaptive to different initial conditions, resource states, and market price fluctuations, showing stronger robustness and better overall convergence performance, avoiding the problem of sustained oscillation or slow convergence caused by fixed step size, and improving the response speed and quality of virtual power plant bidding decisions.

[0024] Eighth, the application significantly enhances the resilience and availability of distributed optimization systems in unreliable industrial network environments by introducing a communication interruption fault-tolerant mechanism. The mechanism breaks the traditional dependence of distributed algorithms on strong synchronization communication of all subsystems by presetting timeout judgments and enabling local cache history solutions, allowing the global iterative optimization process to continue when individual subnodes fail to communicate, rather than being completely interrupted. It allows the system to output a feasible scheduling scheme based on part of the latest data and part of the reliable historical data, which is not the absolute optimal solution at the moment, but ensures the feasibility of scheduling instructions and the timeliness of bidding, enabling the virtual power plant to calmly cope with network fluctuations and avoid the loss of market opportunities due to local communication problems, improving the engineering practical value of the system.

[0025] Ninth, the application deeply fuses the optimization process in the pure mathematical sense and the economic profit and loss bottom line of business operation by embedding the economic security check step, so that the algorithm has online economic risk perception and active avoidance ability. The mechanism estimates the profit of the current scheme after each iteration, and triggers a conservative strategy once a possible loss is found, which forcibly guides the optimization direction to the profit interval by tightening the upper limit of the bid. This is equivalent to setting an "economic sentinel" for the algorithm, effectively preventing the algorithm from converging to a mathematically feasible but economically loss-making solution in scenarios such as market price fluctuations, ensuring the basic economy of the virtual power plant operation, and improving the optimization goal from pure mathematical optimization to business feasibility, greatly enhancing the robustness of the decision and the predictability of the economic benefit.

[0026] Other advantages, objects, and features of the application will be apparent from the following specification, and will be appreciated by those skilled in the art. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The flowchart of one of the technical solutions of the application. DETAILED DESCRIPTION

[0028] The application will be further described in detail below with examples, so that those skilled in the art can implement it according to the description.

[0029] It should be noted that the experimental methods described in the following embodiments are conventional methods, and the reagents and materials, unless otherwise specified, can be obtained commercially.

[0030] As shown in the following formula: Figure 1 The application provides a virtual power plant double-layer optimization scheduling method for power market transactions, which decouples the double-layer structure by distributed optimization architecture and alternating direction multiplier method, and the specific steps are as follows: Step 1, model decomposition and distributed deployment: A virtual power plant double-layer optimization scheduling architecture is established. The upper model is deployed in the virtual power plant central controller, and the objective function is to maximize the total bidding profit of the power market, which is formalized as a mixed integer linear programming problem. The decision variable is the bidding power of each trading period (unit: MW), and the objective function is expressed as: Where is the market price of period t (unit: yuan / MWh), is the actual output (unit: MW), and p is the coordination factor (unit: yuan / MW). The constraint condition limits the range of the bidding power: P min ≤ ≤Pmax .

[0031] The lower layer model is decomposed into five sub-modules according to resource types: Photovoltaic dispatching submodule: output photovoltaic predicted power curve (unit: MW), solve linear programming problem based on irradiance prediction.

[0032] Wind power dispatching submodule: output wind turbine predicted power curve (unit: MW), solve linear programming problem based on wind speed prediction.

[0033] Energy storage dispatching submodule: decision charging and discharging power and SOC state (unit: MW), solve mixed integer programming problem.

[0034] Load dispatching submodule: decision controllable load start-stop state and adjustment amount (unit: MW), solve mixed integer programming problem.

[0035] Electric vehicle dispatching submodule: output cluster charging and discharging power (unit: MW), solve integer programming problem.

[0036] Each submodule is deployed in the local controller of the corresponding resource (such as new energy station, energy storage EMS, load PLC, and charging pile controller) to realize edge computing.

[0037] Step two, iterative solution by alternating direction multiplier method: The central controller initializes the coordination factor p and the bidding power constraint range (P min , P max ), and executes the following loop: 1. Parameter broadcast: the central controller sends the current coordination factor p k and the bidding power constraint range ([P min , P max ]) to all submodules.

[0038] 2. Submodule parallel optimization: each submodule independently solves the local optimization problem based on the received p k and the bidding power constraint. The objective function is to minimize the sum of resource dispatching cost and deviation penalty: The photovoltaic dispatching submodule and the wind power dispatching submodule both dispatch the predicted power curve; the energy storage dispatching submodule outputs the charging and discharging power and the SOC state; the load dispatching submodule outputs the start-stop state and the adjustment amount; and the electric vehicle dispatching submodule outputs the cluster charging and discharging power. After synchronous calculation of each submodule, the optimal dispatching scheme and marginal cost information (unit: yuan / MW) are uploaded to the central controller.

[0039] 3. Central aggregation update: Update the bidding power based on the marginal cost sum and the actual output sum uploaded by the sub-modules , update as follows: ; Update the coordination factor, .

[0040] Step three, convergence judgment and result output: Check the termination condition after each iteration: The change in bidding power, ; the change in coordination factor, .

[0041] If both conditions are met, terminate the iteration. The central controller submits the final bidding power to the power trading platform and issues dispatching instructions to each sub-module to perform actual output control.

[0042] This embodiment decouples the upper and lower layer models through a distributed architecture. The central controller only processes bidding power optimization, and resource scheduling is calculated by edge sub-modules in parallel, avoiding nested solution. Sub-module parallel optimization reduces single-point computing load; ADMM algorithm realizes fast convergence through coordination factor and marginal cost feedback, meeting the time limit requirements of the power market.

[0043] The hardware support of the distributed optimization architecture includes: a central controller deployed in the virtual power plant master control center, selected from an industrial server (such as a server equipped with an Intel Xeon processor and 64 GB random access memory (RAM)), running the upper layer model optimization algorithm. Photovoltaic scheduling sub-module and wind power scheduling sub-module, deployed in the local controller of the new energy station (such as ARM Cortex-A72 edge computing device). The energy storage scheduling sub-module is deployed in the energy storage power station energy management system (EMS). The load scheduling sub-module is deployed in the demand response terminal (such as SIEMENS S7-1500 programmable logic controller (PLC)). The electric vehicle scheduling sub-module is deployed in the charging pile centralized controller (such as Raspberry Pi 4B). The functions performed by each sub-module are shown in Table 1, real-time data is transmitted using IEEE C37.118.2 protocol, communication delay is not more than 50 ms, and time window for each iteration is set as: 300 s for day-ahead market reservation, 60 s for intra-day market reservation. According to the physical characteristics and constraint homogeneity, photovoltaic and wind power output are uncontrollable but predictable, energy storage has time-coupled state variables (state of charge (SOC)), load has start-stop integer constraints, and electric vehicle has high-dimensional discrete variable characteristics. Parallel computing is realized based on the calling of Gurobi / CPLEX solver by each sub-module for independent calculation, and the central controller aggregates the calculation results through message passing interface (MPI).

[0044] Table 1 Sub-module table The implementation of the present embodiment includes the following and working mechanism: 1. Distributed ADMM framework, according to the double-layer optimization problem can be split into: upper problem (central controller), P min ≤ ≤P max . Objective: trade-off market revenue and output deviation penalty (regulated by coordination factor ρ). Lower problem (distributed resource submodule): . Objective: minimize scheduling cost while tracking bid power . The double-layer problem is decoupled into independent sub-problems by coordination factor ρ, avoiding direct nested solution.

[0045] 2. ADMM iteration mechanism: Parameter broadcast, send the current coordination factor ρ k , bid constraints ([P min , P max ]). According to ρ k As a Lagrange multiplier, it forces the lower output to approximate the upper bid quantity.

[0046] Submodule parallel optimization: photovoltaic scheduling submodule and wind power scheduling submodule, using linear programming, constraint conditions are 、 , where, PV scheduling submodule actual output (MW) at time t; PV output and irradiance prediction function relationship (such as linear mapping based on PV panel efficiency, installation capacity); Irradiance prediction value at time t (such as W / m 2 ); Wind power scheduling submodule actual output (MW) at time t; Wind power output and wind speed prediction function relationship (such as linear piecewise function consistent with wind turbine power curve); Wind speed prediction value at time t (such as m / s). Output predicted power curve (uncontrollable resources follow passively).

[0047] Energy storage scheduling submodule, load scheduling submodule, electric vehicle scheduling submodule, using mixed integer programming: Energy storage scheduling submodule, constraint condition is resource physical constraint, charge and discharge power limit, , where, > 0 indicates charging, < 0 indicates discharging; and are the maximum charge power and the maximum discharge power, respectively.

[0048] SOC (State of Charge) constraints: when charging, when discharging, SOC min ≤ SOC t ≤ SOC max , where SOC t is the state of charge at time period t; η ch , η dis are the charging and discharging efficiencies, respectively; is the length of the time period; is the rated capacity; SOC min , SOC max are the upper and lower limits of SOC, respectively.

[0049] Charge-discharge state exclusivity (integer variable constraints): , , QUOTE are the charging state indicator variable and the discharging state indicator variable, respectively.

[0050] Load scheduling sub-module, constraint condition is resource physical constraint: start-stop state constraint: , where = 1 indicates that the load is running, = 0 indicates that it is shut down; regulation range: , where is the actual load power at time period t; P load,min , P load,max are the minimum and maximum powers when the load is running, respectively; start-stop time constraint (if applicable): , where τ on is the minimum continuous running time to prevent frequent start-stop.

[0051] Electric vehicle scheduling sub-module, constraint condition is resource physical constraint: cluster charge-discharge power limit: , where is the charging power of the electric vehicle cluster at time period t; is the number of electric vehicles connected at time period t; is the maximum charging power of a single vehicle; user demand constraint: , where is the charging power of the i-th vehicle; T i is the available charging time period of the i-th vehicle; is the demand power of the i-th vehicle; integer variable constraint: QUOTE , is a non-negative integer, representing the number of access vehicles.

[0052] 3. Centralized aggregation update: update the bidding power, The principle is to determine the next round of bidding power by minimizing the objective function, considering three factors: market revenue term (P ), which is directly related to market price and bidding power; deviation penalty term (ρ ), which controls the deviation between bidding and actual output through coordination factor ρ k ; and lower-level scheduling cost gradient (∑ marginal cost), which integrates the marginal value of operation cost of each sub-module. The function is piecewise linear, and the closed-form solution is obtained by solving the extreme point to ensure the economic and feasible of the bidding power.

[0053] Update the coordination factor ρ , using the dual ascent method principle, dynamically adjust the penalty strength through the deviation between bidding power and actual output (P ): if the bidding power is greater than the sum of actual output, ρ increases, strengthening the punishment for "overbidding"; if the bidding power is less than the sum of actual output, ρ decreases, weakening the punishment to encourage bidding. The coefficient 0.5 (this step value can be adjusted according to the actual system characteristics) is used to control the adjustment step, ensuring the convergence stability.

[0054] 4. Convergence criterion: the iteration must meet the following conditions simultaneously: , , according to the bidding power change threshold (0.5 MW), to ensure the stability of the market declaration result and avoid frequent small adjustments to interfere with transaction decisions; the coordination factor change threshold (10 yuan / MW) ensures that the deviation penalty term converges to a steady state, making the upper and lower optimization objectives consistent. The two conditions work together to ensure optimization accuracy and avoid invalid iterations.

[0055] The traditional bi-level problem uses a nested iteration mode, which needs to handle the interaction of the full amount of variables simultaneously, with high computational complexity, which increases exponentially with the growth of resource size and time period. This method breaks through the bottleneck by two key optimizations: distributed architecture dimension reduction, which decomposes the overall problem into N sub-modules for parallel computing, each sub-module only handles the optimization problem of local resources, and the computational complexity is significantly reduced, with a predicted one order of magnitude improvement in computing efficiency; lightweight communication mechanism, based on the marginal cost feedback mechanism of ADMM, only passing scalar information (such as actual output , marginal cost), replacing the traditional full variable interaction mode, with significantly reduced communication overhead, only a small amount of aggregated information needs to be transmitted, reducing data transmission pressure.

[0056] For the large-scale variable scheduling pressure of multiple types of resources (photovoltaic, wind power, energy storage, etc.), the following strategies are used to realize real-time response: parallel computing acceleration, different types of sub-modules (linear programming of photovoltaic / wind power, mixed integer programming of energy storage / load, integer programming of electric vehicles) are independently solved synchronously, combined with the localized processing capability of edge computing hardware, the single-point computing time delay is greatly shortened, and the resource contention of centralized computing is avoided; The update step (0.5) of the coordination factor p is strictly guaranteed, which is verified by theory to meet the Lyapunov stability condition, ensuring linear convergence in the iteration process and avoiding oscillation; The search space is strictly limited by the constraints [P min 、P max ] of the bidding power, which prevents the optimization result from diverging from the boundary and ensures the effectiveness of the iteration direction. Therefore, by using distributed decoupling, parallel computing and stable convergence mechanism, the method can reduce the computing and communication burden and meet the strict real-time requirement of the power market when dealing with large-scale energy resource scheduling. The solving method of the application is essentially a distributed optimization framework based on ADMM, which decouples the complex bi-level problem into sub-modules that can be solved in parallel through the coordination factor p and marginal cost feedback mechanism, significantly improves the real-time computing efficiency while ensuring convergence. The mathematical rigor relies on the convex optimization theory and the dual ascent method, and the engineering feasibility is realized through edge computing and resource-specific optimizers.

[0057] In another embodiment of the application, the virtual power plant central controller initializes the coordination factor before iteration. The initial value of the coordination factor is set to a random value in the range of 80%-120% of the predicted electricity price of the electricity market. The central controller automatically selects the specific value through the built-in random number generator, covering the typical electricity price fluctuation range. This design is based on the continuity of the electricity market, so that the initial value falls into the reasonable neighborhood with a high probability, significantly reducing the oscillation in the early stage of iteration.

[0058] The upper limit of the bidding power constraint range is set to 95% of the maximum adjustable output of all distributed energy sources. Distributed energy sources include photovoltaic units, wind power units and energy storage devices, and their maximum adjustable output is determined by the device rated power and real-time working condition. The lower limit of the constraint range is set to 105% of the maximum adjustable capacity of all controllable loads. Controllable load refers to a cluster of industrial or commercial loads with start-stop adjustment capability, and the maximum adjustable capacity is calibrated according to historical operation data. The central controller automatically aggregates resource parameters to calculate the constraint range, leaving a 5% safety margin to avoid touching the physical boundary, while preventing search space redundancy.

[0059] The coordination factor is randomly initialized and combined with the bidding constraint for an iterative process. The random initial value provides a diversified starting point for the bias penalty term, and the compact constraint range guides the first iteration into the effective search domain. After receiving the parameters, each submodule synchronously performs local optimization: the photovoltaic scheduling submodule and the wind power scheduling submodule quickly respond to the prediction curve in the edge device, the energy storage scheduling submodule adjusts the charging and discharging strategy according to the 95% output upper limit, the load scheduling submodule optimizes the start-stop plan according to the 105% capacity lower limit, and the electric vehicle scheduling submodule processes discrete variables. The central controller transmits the actual output and marginal cost through a lightweight communication protocol, dynamically monitors the bidding power change (ΔP ≤0.5MW) and the coordination factor change (Δρ ≤10 yuan / MW), and realizes efficient convergence.

[0060] In another embodiment of the present application, to further solve the model complexity problem caused by time coupling in energy storage scheduling, a time decoupling strategy is introduced in the energy storage scheduling submodule. The specific implementation is as follows: Intra-group application of energy storage state transition equation: within each time period group, the energy storage scheduling submodule updates the SOC state according to the energy storage state transition equation. The equation is specifically expressed as: Charging: Discharging: This equation ensures the physical consistency and continuity of the SOC state in each time period within the group, and is the basic constraint of energy storage operation.

[0061] Time period group boundary SOC as a coupling variable: the SOC value at the end time of each time period group (such as the SOC of the last time period of the gth group ) is defined as a coupling variable, denoted as . This variable is used to connect adjacent time period groups and ensure the continuity of the full-time SOC trajectory.

[0062] Mechanism of coupling variable participating in ADMM update: in the ADMM iteration process, the central controller is responsible for coordinating the boundary SOC values of each time period group. The specific steps are as follows: The energy storage scheduling submodule uploads the termination SOC values of each group to the central controller; The central controller calculates the boundary SOC deviation according to the global consistency principle: Wherein, the deviation is included in the update term of the coordination factor ρ, realizing the punishment adjustment of the boundary SOC; the central controller broadcasts the corrected boundary SOC value to the energy storage scheduling submodule as the initial SOC value of the next time period group; the energy storage scheduling submodule re-solves the intra-group optimization problem based on the updated boundary SOC value.

[0063] Through the above mechanism, the originally full-period coupled energy storage scheduling problem is decoupled into multiple period group independent optimization problems, greatly reducing the model complexity, while the global optimality and convergence are guaranteed through the coordination mechanism of ADMM.

[0064] The technical solution comprises a storage scheduling period grouping strategy, and a storage scheduling submodule divides a 24-hour scheduling period into 4-6 continuous period groups. Each group comprises adjacent trading periods, and the period group length is uniformly distributed. The storage scheduling submodule is deployed in an energy management system of the energy storage power station, and receives a period grouping instruction issued by the central controller. The period group division is based on the scheduling period integrity principle, and ensures that the periods in the group are continuous and cover the complete operation period.

[0065] Intra-group mixed integer programming solving, each period group is taken as an independent optimization unit. The storage scheduling submodule solves the mixed integer programming problem based on the initial state of charge value of the current period group. The decision variables include the charging and discharging power values in each period of the period group, and charging and discharging state indicator variables. The constraint conditions include the upper and lower limits of the charging and discharging power, the charging and discharging efficiency, the state of charge safety boundary and the state exclusion. The solving process calls a Gurobi or CPLEX solver (not limited to Gurobi / CPLEX commercial solvers, other similar or self-developed solvers can also be used), and outputs the charging and discharging power sequence in the period group and the state of charge value at the termination time.

[0066] Period group state transfer coordination, the state of charge value at the boundary point of the period group is taken as a coupling variable. The storage scheduling submodule uploads the state of charge value at the termination time to the central controller. The central controller updates the boundary point state of charge value through the alternating direction multiplier method: The termination state of charge value of the previous period group is taken as the initial value of the next period group; The boundary state of charge deviation of adjacent groups is included in the coordination factor update term; The central controller broadcasts the corrected boundary state of charge value to the storage scheduling submodule, triggering the next round of intra-group optimization.

[0067] Alternating direction multiplier method coordination mechanism, the storage scheduling submodule participates in the global iteration and performs hierarchical coordination: 1. receiving the coordination factor, the bidding power constraint range and the boundary state of charge correction value issued by the central controller; 2. solving the intra-group optimization problem in sequence according to the period group sequence; 3. uploading the charging and discharging scheme of each period group and the boundary state of charge value; 4. receiving the updated boundary state of charge value for the next round of intra-group optimization; This mechanism decouples the full-period coupling constraint into a group coordination problem, and realizes time dimension decomposition through the boundary state of charge variable.

[0068] In another embodiment of the present application, in the process of double-layer optimization scheduling of virtual power plant facing power market transaction, the electric vehicle scheduling submodule is responsible for processing the charging and discharging scheduling problem of large-scale electric vehicle cluster. There are significant differences among individual electric vehicles in residual power, battery capacity, maximum charging and discharging power, and user travel demand. If each electric vehicle is independently modeled and optimized, the number of decision variables and constraint conditions will increase linearly with the number of electric vehicles, resulting in a sharp expansion of the optimization model size, too high computational complexity, and inability to meet the real-time requirements of power market bidding.

[0069] To solve this technical problem, the clustering compression algorithm is used in the electric vehicle scheduling submodule. The algorithm first sorts all electric vehicle clusters within the aggregation range of the virtual power plant according to their current residual power values, and divides them into a fixed number of equivalent vehicle groups within the entire residual power range. The number of equivalent vehicle groups is usually set to ten to twenty. Each equivalent vehicle group no longer represents a single vehicle, but an equivalent scheduling unit composed of multiple vehicles with similar residual power.

[0070] Each equivalent vehicle group is characterized by three core parameters: a unified upper limit of charging and discharging power, unit MW; a unified upper limit of total energy capacity, unit MWh; and an average charging and discharging efficiency. These parameters are calculated by the maximum value or weighted average value of the corresponding parameters of all vehicles in the group. When performing optimization calculations, the decision variable of the electric vehicle scheduling submodule is no longer the power of each vehicle, but the charging and discharging power of each equivalent vehicle group. This greatly reduces the number of integer variables and related constraints in the optimization problem, compressing a problem that may contain thousands of variables into a problem that contains only dozens of variables.

[0071] When the central controller finally issues the scheduling instruction, the scheduling instruction is for the total power value of each equivalent vehicle group. The total power needs to be further decomposed and allocated to each real electric vehicle in the group. The power allocation adopts the principle of equal proportion allocation, that is, according to the proportion of the actual maximum acceptable power of each electric vehicle in the total maximum acceptable power of all vehicles in the group, to calculate the actual executed power value, so as to ensure the physical executability of the scheduling instruction and meet the constraints of individual vehicles.

[0072] The existing method for handling the scheduling of a cluster of electric vehicles is to model and optimize each electric vehicle independently. This method attempts to accurately consider the individual state and constraints of each vehicle in order to obtain the most accurate optimization result in theory. However, the size of the optimization model is proportional to the number of electric vehicles. When a virtual power plant needs to manage thousands of electric vehicles, the number of integer variables and constraints in the model will reach a huge scale, forming a problem of "constraint explosion". This results in a very long optimization calculation time, often taking hours or even longer to complete an optimization solution, which completely fails to meet the calculation time limit of minutes usually required in the day-ahead market or intraday market, and lacks feasibility in practical applications.

[0073] Compared with the prior art described above, the clustering compression algorithm used in the embodiment has fundamental differences and significant progress. Instead of pursuing accurate independent depiction of each individual, the method abstracts a large number of individuals into a small number of equivalent units through clustering. This processing method essentially compresses the dimension of the optimization problem from a very high dimension related to the number of vehicles to a fixed, lower dimension. Whether the cluster contains one thousand or ten thousand electric vehicles, the complexity and computational burden of the optimization model are basically constant and controllable after clustering and compression. This makes real-time solution of complex optimization problems possible, successfully solving the computational bottleneck problem caused by large-scale access of electric vehicles, and is a key technical improvement that can be applied to actual power market trading scenarios.

[0074] In another embodiment of the application, in the iterative solution process of the double-layer optimization scheduling of the virtual power plant, the optimization problems of the load scheduling submodule and the electric vehicle scheduling submodule include integer decision variables such as start-stop state and access quantity. The existence of these integer variables makes the entire optimization problem a non-convex mixed integer programming problem. Non-convexity causes the objective function to have multiple local optimal points, and the conventional iterative algorithm is easily trapped in a local optimal solution and cannot jump out, cannot continue to converge to the global optimal solution, and forms a local optimal trap. This will cause the optimization to terminate prematurely, and the final bidding strategy and economy are not optimal.

[0075] To solve this technical problem, the embodiment introduces a relaxation mechanism in the iteration process. The mechanism continuously monitors the substantial progress of the optimization process. After each iteration is completed, the system calculates the improvement of the current objective function value compared with the previous iteration. When the system detects that the improvement of the objective function of three consecutive iterations is less than the preset improvement threshold, it is determined that the optimization process may have fallen into a local optimal region and the progress has stagnated.

[0076] At this time, the relaxation mechanism is triggered. The mechanism first fixes all integer variables in the mixed integer programming problem to their optimal solution values in the current iteration. The physical states represented by these integer variables, such as the start-stop of the load and the access of the electric vehicle, are temporarily determined. After fixing the integer variables, the original mixed integer programming problem is converted into a pure continuous variable optimization problem.

[0077] For this transformed continuous variable optimization problem, the present embodiment uses the interior point method for solving. The interior point method is a high-efficiency algorithm specially used for solving large-scale continuous convex optimization problems, and its characteristics are polynomial time complexity and good convergence. The system calls the interior point method solver to continuously optimize the problem until the original convergence termination condition is met, i.e., the bid power change and the coordination factor change both reach their threshold requirements. By transforming the non-convex problem into a convex problem for solving, the local optimal trap is effectively jumped out, and the solution is guided to continue searching in a better direction.

[0078] The existing technology usually uses classical algorithms such as branch and bound or cutting plane method to solve the mixed integer programming problem in virtual power plant dispatch, and these algorithms are always used throughout the solving process. Such methods can theoretically find the global optimal solution, but the calculation time increases exponentially with the problem size, and the calculation cost is huge. More importantly, in complex non-convex problems, the search tree of these methods may be pruned too early, or trapped in a local optimum of a certain branch and cannot be pulled out, and in fact it is difficult to guarantee to obtain a high-quality feasible solution within a limited time. For real-time power market transactions with extremely high real-time requirements, the practicability is poor.

[0079] Compared with the above existing technology, the relaxation mechanism adopted by the present embodiment represents a different way of solving problems. Instead of pursuing to solve all problems in a single algorithm framework, it adopts a hierarchical and adaptive strategy. The core advantage of this method is its dynamic nature. It does not assume the nature of the problem in advance, but dynamically determines whether it is trapped in a local optimum by monitoring the performance of the optimization process in real time, and accordingly triggers different solving strategies. This mechanism cleverly changes the nature of the problem when the algorithm is stuck, and converts the difficult non-convex integer programming problem into a convex continuous optimization problem that is easy to solve efficiently, thereby significantly improving the convergence reliability and calculation efficiency of the algorithm under the premise of ensuring the solving accuracy, effectively avoiding the local optimal trap caused by the non-convexity of the problem, and more suitable for practical engineering applications.

[0080] In another embodiment of the present application, the power market transaction, especially the intraday rolling market, has extremely strict real-time requirements for the submission of bidding strategies. Although the virtual power plant bi-level optimization scheduling method significantly improves efficiency through distributed architecture and parallel computing, its iterative solution process may still fail to naturally converge to the preset threshold condition before the bidding deadline. If waiting for complete convergence, there will be a risk of overtime, resulting in the inability to submit any bidding strategy, thereby missing market opportunities and causing economic losses.

[0081] To solve the optimization interruption risk caused by the market bidding deadline, the present embodiment introduces a real-time computing monitor. The monitor is independent of the optimization algorithm and continuously tracks the core state (iteration number) of the entire iterative solution process. It sets a clear safety boundary, i.e., five iterations. This number is determined based on a large number of simulation experiments, which can achieve an engineering balance point between calculation time and market returns.

[0082] Once the real-time computing monitor detects that the iteration number has reached five and the convergence determination condition is still not met, it will immediately send an overtime warning signal to the central controller. At this time, the system will not continue to attempt the sixth iteration that may not be completed in time, but will automatically start a pre-designed reduced-order mode. The core idea of this mode is to sacrifice part of the accuracy and complexity of the model to achieve a sharp improvement in calculation speed, ensuring the generation of a feasible suboptimal solution and successful submission before the final deadline, while ensuring the decision-making function.

[0083] The reduced-order mode achieves acceleration through a series of targeted model simplification operations. For the photovoltaic scheduling submodule and the wind power scheduling submodule, since their output mainly depends on external environment prediction and is uncontrollable, the reduced-order mode directly fixes their predicted output curve to the latest day-ahead prediction value, thereby completely bypassing the optimization problem of these two submodules and saving a lot of calculation time. For the energy storage scheduling submodule, the reduced-order mode relaxes its charge and discharge state variable from an integer variable that must take 0 or 1 to a continuous variable that can take a continuous value between 0 and 1. This operation removes the integer constraint in the problem and converts the mixed integer programming into nonlinear programming, significantly reducing the difficulty and time consumption of solving. For the load scheduling submodule and the electric vehicle scheduling submodule, since their start-stop and access decisions have a more critical impact on the feasibility and economy of the final scheduling scheme, the reduced-order mode chooses to retain their integer constraints and continue to optimize them accurately to ensure the effectiveness of the core scheduling instructions.

[0084] Through the above combination strategy, the reduced mode quickly converts an originally complex and time-consuming full version optimization problem into a simplified version problem. The central controller quickly completes the last iteration calculation based on the simplified model, and immediately outputs the obtained bidding power and scheduling instruction, thereby ensuring that the virtual power plant can successfully submit the bid before the market closes, effectively avoiding the business risk of transaction interruption due to optimization calculation timeout.

[0085] The existing technology usually adopts two strategies when dealing with the risk of calculation timeout. One is a rigid strategy, that is, always insist on running the complete optimization model until complete convergence or absolute timeout. This strategy often leads to submission failure due to insufficient estimation of calculation time consumption. The other is a simple and rough conservative strategy, that is, terminate the iteration a long time in advance, and directly use an early intermediate result which may be of poor quality as the final output for submission. Although this strategy guarantees submission, it sacrifices too much economy, because the unoptimized bidding strategy may lack competitiveness in the market or lead to high operating costs.

[0086] Compared with the above existing technology, the adaptive reduced mode based on real-time calculation monitor adopted by the embodiment provides a more intelligent and reliable solution. Instead of passively waiting for timeout or actively giving up optimization, the method sets a key risk intervention point. Instead of adopting a single simplification strategy, it carries out differentiated and refined model reduction processing according to the physical characteristics of different sub-modules and the influence degree on the overall optimization target. This processing mode is a trade-off based on a deep understanding of the problem structure, which maximizes the calculation accuracy of the parts that have the greatest impact on the decision quality, while decisively simplifying the parts that have relatively small impact on the results but high calculation cost. Thus, a bidding scheme that is superior to the conservative strategy in feasibility and economy is submitted under strict time constraints, fundamentally solving the contradiction between optimization interruption and bidding opportunity loss.

[0087] In another embodiment of the application, in the double-layer optimization scheduling process of the virtual power plant participating in the electricity market transaction, the update strategy of the coordination factor is crucial to the convergence performance of the algorithm. The fixed step size is used to update the coordination factor. This fixed step size method keeps the step size parameter constant throughout the iteration process. When the market electricity price fluctuates sharply or the distributed energy and load conditions within the virtual power plant are complex, the fixed step size is difficult to adapt to the changing optimization conditions. If the step size is set too large, the deviation between the bidding power and the actual output will continue to oscillate in the iteration, resulting in unstable convergence; if the step size is set too small, the deviation reduction speed is too slow, and a large number of iterations are required to reach the convergence condition, which is difficult to meet the urgent time limit requirement of the electricity market bidding. This contradiction between convergence efficiency and stability is an inherent defect of the fixed step size method.

[0088] To solve the above technical problems, the updating rule of the coordination factor is improved in the embodiment, and a dynamic step adjustment mechanism is introduced. The core of the mechanism is a dynamic step adjustment coefficient, which is not fixed but is adaptively adjusted according to real-time feedback of the optimization process. Specifically, in each iteration, the central controller records the deviation value between the bidding power and the actual total output. The system maintains a history record containing the deviation values of the last three iterations, and analyzes the change trend presented by the sequence. The calculation of the dynamic step adjustment coefficient is based on the judgment of the short-term historical trend. If the analysis finds that the deviation values of the last three iterations show a continuously expanding trend, it is determined that the current iteration may be in an unstable or oscillating state, and the dynamic step adjustment coefficient will automatically decrease, thereby reducing the actual adjustment step of the coordination factor, stabilizing the iteration process and suppressing oscillation. On the contrary, if the analysis finds that the deviation values show a continuously decreasing trend, it is determined that the current iteration is effectively progressing towards convergence, and the dynamic step adjustment coefficient will automatically increase, thereby expanding the actual adjustment step of the coordination factor, accelerating the convergence process and faster meeting the termination condition.

[0089] The dynamic step adjustment mechanism adopted in the embodiment is in sharp contrast to the above-mentioned fixed step technology. The adjustment strategy of the existing technology is static and passive, and cannot respond to the dynamic changes in the optimization process. The technical solution is dynamic and active, which makes the updating process of the coordination factor adaptive, and can adjust the control parameters in real time according to the current working condition like an intelligent controller. This adaptability makes the whole double-layer optimization algorithm exhibit more robust and efficient performance when facing different initial conditions, such as different initial values of the coordination factor or different bidding power constraint ranges, and variable market environment, such as rapid fluctuations in electricity prices. It effectively bridges the gap between convergence efficiency and stability, avoids the performance defects of the fixed step method in complex application scenarios, and improves the overall quality and reliability of the virtual power plant bidding decision.

[0090] In another embodiment of the present application, in the distributed architecture of the virtual power plant double-layer optimization scheduling system, reliable communication between the central controller and multiple sub-modules is the basis for the continuous process of iterative solution. Existing technical solutions generally lack effective communication fault handling mechanisms. Under this existing architecture, the central controller waits for all sub-modules, including photovoltaic scheduling sub-modules, wind power scheduling sub-modules, energy storage scheduling sub-modules, load scheduling sub-modules, and electric vehicle scheduling sub-modules, to return their optimal scheduling schemes and marginal cost information before proceeding with aggregation updates and starting the next iteration. However, the actual industrial network environment is not ideal, and communication delays or even interruptions may occur due to hardware failures, signal interference, or network congestion. If one or more local controllers of the sub-modules, such as photovoltaic controllers or load PLCs deployed in remote stations, fail to respond in a timely manner due to network problems, the central controller will remain in a waiting state. This dependence on strong synchronization of the entire subsystem makes a single node communication failure sufficient to cause the entire optimization iteration process to stall, making it impossible to continue or output results, ultimately missing the bidding window of the electricity market and causing direct economic losses.

[0091] To solve the above technical problems, the present embodiment introduces a set of communication interruption fault tolerance mechanism. The core of this mechanism is to set a preset timeout time for the central controller, which is determined based on the average delay of the communication network and the time budget of the iteration calculation. After initiating parameter broadcast for each iteration, the central controller starts a timer. If the central controller detects that any one or more sub-modules have not returned feedback information when the preset timeout time is exhausted, it will no longer wait unnecessarily. The system will automatically trigger the fault tolerance process and enable the corresponding local cache scheduling scheme for these lost sub-modules. This local cache scheduling scheme is not randomly generated, but a copy of the optimal scheduling scheme successfully received and verified from the sub-module in the last iteration, which is stored in the cache area of the central controller. Subsequently, the central controller will use these cache schemes, together with the latest schemes returned by other normally online sub-modules, to perform aggregation calculation and coordination factor update for this round, so that the iteration process can continue to proceed.

[0092] The communication interruption fault-tolerant mechanism adopted by the embodiment is essentially improved over the method relying on perfect communication. The existing technology is a "all or nothing" rigid strategy, whose normal operation is completely based on the ideal assumption that all communication links are always smooth, and the system resilience is insufficient. The technical solution is an intelligent degradation operation strategy, which acknowledges and accepts the possibility of partial communication failure in complex industrial environments, and replaces the missing real-time data by using historical data (i.e. the effective scheme of the last round). This method ensures that when non-global and local communication problems occur, the entire optimization scheduling process will not be completely interrupted. The system can output a scheduling result based on part of the latest data and part of the historical data. Although this result may not be the optimal solution because it uses cached data, it slightly compromises on economy, but it guarantees the feasibility of the scheduling scheme and the availability of the system, so that the virtual power plant can submit a bid on time and avoid greater business risks caused by the interruption of the optimization process, significantly enhancing the robustness of the entire system in the face of uncertainty.

[0093] In another embodiment of the application, in the double-layer optimization scheduling process of the virtual power plant participating in the electricity market transaction, the core goal of the algorithm is to seek an optimal bidding and internal scheduling scheme to maximize profits. However, the existing technical solution usually regards the optimization process as a purely mathematical problem, and its termination condition only depends on whether the change in the bidding power and the coordination factor reaches the preset mathematical threshold. This method has a fundamental flaw: it may converge to a mathematically completely feasible (i.e. meeting all constraint conditions) but economically loss-making solution. This situation is particularly prominent when market electricity prices fluctuate sharply. For example, when market electricity prices plummet due to unexpected events, the algorithm may still mechanically pursue accurate tracking of the upper-level bidding power and the lower-level actual output, and may be forced to call high-cost distributed energy (such as starting expensive oil-fired units) or adjust controllable loads at a high cost in order to fulfill a bidding plan made during a high-price period. As a result, the electricity sales revenue cannot cover the high internal scheduling cost, ultimately leading to overall operational losses of the virtual power plant. The existing technology lacks the ability to identify and avoid such potential economic risks online.

[0094] To solve the above technical problems, the embodiment embeds an economic safety check step in the optimization iteration process. The execution timing of this step is carefully set after the central controller completes the current round of aggregation update, but before the formal convergence judgment. The core operation is to calculate the estimated total profit corresponding to the scheduling scheme generated by the current iteration round. When calculating the estimated total profit, the central controller multiplies the updated bid power after this round with the currently known market forecast price sequence to obtain the estimated electricity sales revenue; at the same time, it aggregates the marginal cost information uploaded by each sub-module and estimates the total internal scheduling cost combined with the scheduling scheme; the total revenue minus the total cost is the estimated total profit. Subsequently, the system compares this estimated total profit with a preset break-even threshold. The break-even threshold is the economic bottom line of virtual power plant operation, usually set to zero or a small positive number, to ensure at least no loss. If the current estimated total profit is found to be lower than the break-even threshold, it is determined that the current optimization direction has economic risks, and the conservative bidding strategy is triggered immediately.

[0095] Once the conservative bidding strategy is triggered, the system does not directly output the current possible loss scheme, but actively takes intervention measures. The core of the intervention is to tighten the search space of the optimization problem, and the upper limit of the bid power constraint range is significantly reduced from 95% of the maximum adjustable output of all distributed energy to 80%. This operation forcibly limits the maximum capacity of the virtual power plant to sell electricity in the market, and essentially changes from selling more electricity to selling appropriate and profitable electricity. Subsequently, the system restarts a complete sub-module optimization and central aggregation process based on this new and stricter bid power constraint range. This is equivalent to imposing an economic-oriented hard constraint on the optimization algorithm, guiding it to find the optimal solution in a safer and more conservative feasible region, thereby actively avoiding the loss area.

[0096] The economic safety check and conservative strategy intervention mechanism adopted by the embodiment is in sharp contrast to the above-mentioned technical solution that simply pursues mathematical convergence. The existing technology is an economic blind optimization, and the economic benefits of the final output scheme have a great contingency. However, the present technical solution introduces a key economic sentinel that actively evaluates and intervenes the economic feasibility of the scheme at the end of each iteration. This method deeply embeds economic judgment into the core loop of the optimization process, enabling the algorithm to have economic risk perception and avoidance ability. It ensures that the optimization and scheduling decisions of the virtual power plant are not only mathematically feasible, but more importantly, economically reasonable, firmly holding the bottom line of profit and loss operation. Especially in the case of complex and variable market environment, it greatly enhances the robustness and predictability of the economic benefits of virtual power plant operation.

[0097] While embodiments of the application have been disclosed in connection with the above specification, it will be evident to those skilled in the art that many modifications, substitutions, and alterations to the embodiments of the application can be made and that many specifically protected details shown can be substituted by other specifically protected details. Accordingly, it is intended that the application not be limited to the specific details shown and described above but that it be given broadest scope indicated by the following claims and their equivalents.

Claims

1. A method for double-layer optimization scheduling of a virtual power plant for power market transaction, characterized in that, The method comprises the following steps: establishing an upper layer model, the upper layer model being a mixed integer linear programming problem with the maximum total bidding profit of the virtual power plant as the target, and the decision variable being the bidding power of each trading period in the electricity market; establishing a lower layer model, the lower layer model being a mixed integer programming problem with the minimum internal scheduling cost as the target, and the decision variable including the predicted output curve of the distributed photovoltaic, the predicted output curve of the wind turbine generator, the charging and discharging state and power of the energy storage device, the start and stop state and the adjustment amount of the controllable load, and the charging and discharging power of the electric vehicle cluster; setting a coordination factor between the upper layer model and the lower layer model, the coordination factor including a bidding power deviation penalty term and a reserve capacity compensation term; adopting a distributed optimization architecture, deploying the upper layer model to the central controller of the virtual power plant, and decomposing the lower layer model into a photovoltaic scheduling submodule, a wind power scheduling submodule, an energy storage scheduling submodule, a load scheduling submodule and an electric vehicle scheduling submodule according to the resource type; iteratively solving the process by using the alternating direction multiplier method, and performing the following steps in each iteration: the central controller broadcasts the current coordination factor and the bidding power constraint range to all submodules; each submodule solves a local optimization problem based on the received coordination factor and uploads the optimal scheduling scheme and marginal cost information to the central controller in parallel; the central controller aggregates the marginal cost information of each submodule, updates the bidding power and the coordination factor; the iteration is terminated when the change amount of the bidding power is less than a preset power change threshold and the change amount of the coordination factor is less than a factor change threshold; the final bidding power is output to the electricity trading platform, and the resource scheduling instructions are issued to the local controller.

2. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 1, characterized in that, The initial value of the coordination factor is set to a random value in the 80%-120% interval of the predicted electricity price of the electricity market; The upper limit of the bidding power constraint range is 95% of the maximum adjustable output of all distributed energy sources; and the lower limit is 105% of the maximum adjustable capacity of all controllable loads.

3. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 1, characterized in that, The energy storage scheduling submodule adopts a time decoupling strategy, divides the 24-hour scheduling period into 4-6 continuous time period groups, solves a mixed integer programming problem in each time period group, connects the time period groups through the energy storage state transfer equation, and uses the energy storage SOC state at the boundary point of the time period group as a coupling variable to participate in the alternating direction multiplier method update.

4. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 1, characterized in that, In the electric vehicle scheduling submodule, a clustering compression algorithm is adopted, specifically: the electric vehicle cluster is divided into 10-20 equivalent vehicle groups according to the remaining power range, each equivalent vehicle group is represented by three parameters, namely the upper limit of the charging and discharging power, the upper limit of the energy capacity and the charging and discharging efficiency, and when the scheduling instruction is issued, the power is proportionally distributed according to the actual parameters of the vehicles in the equivalent vehicle group.

5. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 1, characterized in that, In addition, a relaxation mechanism is set in the iterative solving process, specifically: when the objective function improvement amount of the last three iterations is less than an improvement threshold, the integer variable in the mixed integer programming is fixed as the current optimal solution, and the remaining continuous variable optimization problem is solved by using the interior point method until the termination condition is met.

6. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 1, characterized in that, In addition, a real-time calculation monitor is set, and when the iteration number reaches 5 times and the convergence is not achieved: The reduced order mode is started, the predicted output curves of the photovoltaic scheduling submodule and the wind power scheduling submodule are fixed as the day-ahead prediction values, the charge and discharge state variables of the energy storage scheduling submodule are relaxed as continuous variables, and the integer constraints of the controllable load scheduling submodule and the electric vehicle scheduling submodule are retained to continue solving.

7. The virtual power plant bi-level optimal dispatch method for power market transaction according to claim 6, wherein, In the updating rule of the coordination factor, a dynamic step adjustment coefficient is set, which is adaptively adjusted according to the historical variation trend of the deviation between the bidding power and the actual output in the last three iterations: when the deviation shows an expanding trend, the step is reduced to stabilize the convergence; when the deviation shows a shrinking trend, the step is increased to accelerate the convergence. 8.The virtual power plant bi-level optimization scheduling method for power market transaction of claim 1, wherein, Further comprising a communication interruption fault tolerance mechanism, when the central controller does not receive feedback information from any one or more of the submodules within a preset timeout, the local cache scheduling scheme of the corresponding submodule is automatically enabled to participate in the current round of aggregated updating, and the local cache scheduling scheme is the optimal scheduling scheme successfully received from the corresponding submodule last time. 9.The virtual power plant bi-level optimization scheduling method for power market transaction of claim 1, wherein, After the central controller aggregates and updates, before convergence determination, an economic safety check step is set to check whether the estimated total profit of the current iteration is lower than the preset break-even threshold, if it is lower, a conservative bidding strategy is triggered, the upper limit of the bidding power range is lowered to 80% of the maximum adjustable output of the distributed energy, and each submodule optimization and central aggregation are performed again.

Images