A vpp privacy protection type energy-data collaborative optimization method and system

Abstract

This invention discloses a privacy-preserving energy-data collaborative optimization method and system for Virtual Power Plants (VPPs), relating to the field of virtual power plant operation optimization technology. The method includes establishing an outer master-slave game model to obtain strategic parameters output by the model; combining these strategic parameters with an inner multi-timescale operation optimization model to obtain operational data; constructing a bidirectional coupling feedback mechanism and a Stackelberg equilibrium mechanism using the strategic parameters and operational data; and employing a distributed collaborative solution method based on the bidirectional coupling feedback mechanism and Stackelberg equilibrium mechanism to obtain the VPP privacy-preserving energy-data collaborative optimization results. This invention constructs a bidirectional closed-loop coupling framework, achieving deep integration and collaborative value-added of data element value and energy physical value.

Description

Technical Field

[0001] This invention relates to the field of virtual power plant (VPP) optimization operation technology, and in particular to a VPP privacy-protected energy-data collaborative optimization method and system. Background Technology

[0002] With the acceleration of the energy transition, Virtual Power Plants (VPPs), as an advanced operating model that aggregates distributed energy resources (DERs) to participate in the electricity market, have become a key technology for improving grid flexibility and promoting the consumption of new energy sources. Through information and communication technologies and optimization algorithms, VPPs integrate resources such as distributed photovoltaics, energy storage systems, and flexible loads to participate in the energy market and ancillary service markets such as peak shaving and frequency regulation, thereby achieving optimal resource allocation and maximizing value.

[0003] However, the optimized operation of VPPs heavily relies on massive amounts of high-quality operational data. This data includes physical parameters such as photovoltaic output and load demand, as well as social information such as user electricity consumption behavior and privacy preferences. While data acquisition and utilization unlock optimization potential, they also present multiple challenges: there is an inherent contradiction between user data privacy protection and the value of data sharing; security issues such as data leakage and tampering pose potential risks to VPP operations; and data quality directly impacts the returns and risks of VPP market decisions. Currently, research in the VPP field is trending from "one-sided optimization of energy flow" to "coordinated optimization of energy flow and information flow," but existing technologies still have significant shortcomings:

[0004] (1) Data value is disconnected from energy optimization: Existing technologies often regard the data required for VPP optimization as a given input or exogenous parameter, ignoring the characteristics of data as a tradable production factor. They fail to build a unified framework to quantify the marginal value of data elements and cannot achieve the optimal trade-off between data investment and energy benefits.

[0005] (2) Static trade-off between privacy protection and operational efficiency: Existing privacy protection methods mostly impose privacy constraints statically in the internal optimization of VPP, without taking the strength of privacy protection as a strategic decision variable linked to external data procurement, making it difficult to dynamically coordinate the configuration of data market procurement and privacy security technology.

[0006] (3) Lack of an economic quantification model for data security risks: Existing research has not established a correlation model between data security investment, leakage probability and economic loss, which makes it impossible for VPP operators to accurately weigh data security costs, procurement costs and benefits.

[0007] (4) The different needs of the ancillary service market for data characteristics are not considered: the existing VPP optimization model does not accurately depict the different requirements of different markets such as energy, peak shaving, and frequency regulation for data accuracy and real-time performance, resulting in low efficiency in data resource allocation.

[0008] (5) The model is isolated and does not form a closed-loop optimization: The data pricing game model and the internal energy scheduling model are independent of each other and cannot capture the core feedback mechanism of "high-quality data drives optimization decision-making to generate high returns, and high returns support higher data procurement budgets", making it difficult to maximize the overall profit of VPP.

[0009] Therefore, there is an urgent need for a VPP privacy-preserving energy-data collaborative optimization method and system to address the shortcomings of existing technologies. Summary of the Invention

[0010] The purpose of this invention is to propose a VPP privacy-protected energy-data collaborative optimization method and system to overcome the limitations of existing VPP optimization technologies in data element utilization, privacy and security collaboration, and multi-market decision-making. It constructs a two-way closed-loop coupling framework that can simultaneously handle external data market games and internal multi-energy flow optimization operations, thereby achieving deep integration and collaborative value-added of data element value and energy physical value.

[0011] On the one hand, to achieve the above objectives, this invention provides a VPP privacy-preserving energy-data collaborative optimization method, comprising:

[0012] S1. Establish an outer-layer master-slave game model and obtain the strategic parameters output by the outer-layer master-slave game model;

[0013] S2. Based on the strategic parameters output by the outer master-slave game model and combined with the inner multi-timescale operation optimization model, obtain the operation data output by the inner multi-timescale operation optimization model.

[0014] S3. Using the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model, a two-way coupling feedback mechanism is constructed, and a Stackelberg equilibrium mechanism is established.

[0015] S4. Based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism, a distributed collaborative solution method is adopted to obtain the VPP privacy-preserving energy-data collaborative optimization results;

[0016] The outer master-slave game model includes a follower model and a leader model, and the inner multi-timescale operation optimization model includes a day-ahead split-bar optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model.

[0017] On the other hand, to achieve the above objectives, the present invention provides a VPP privacy-preserving energy-data collaborative optimization system, comprising: an outer game module, an inner operation optimization module, a coupling feedback module, and a collaborative solution module;

[0018] The outer game module is used to establish an outer master-slave game model and obtain the strategic parameters output by the outer master-slave game model. The outer master-slave game model includes a follower model and a leader model.

[0019] The inner layer operation optimization module is used to obtain the operation data output by the inner layer multi-time scale operation optimization model based on the strategic parameters output by the outer layer master-slave game model and the inner layer multi-time scale operation optimization model. The inner layer multi-time scale operation optimization model includes a day-ahead sub-Blu-ray optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model.

[0020] The coupling feedback module is used to construct a two-way coupling feedback mechanism and establish a Stackelberg equilibrium mechanism by utilizing the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model.

[0021] The collaborative solution module is used to obtain the VPP privacy-preserving energy-data collaborative optimization results by adopting a distributed collaborative solution method based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism.

[0022] Compared with the closest existing technology, the present invention has the following advantages:

[0023] (1) Significantly improved economic benefits: This invention optimizes the allocation of data as an endogenous factor, achieving the best balance between data investment and energy revenue. At the same time, by coordinating and optimizing data and physical resources in a multifunctional market, it fully taps the profit potential of VPPs and can significantly improve the overall net profit of VPP operators.

[0024] (2) Enhanced scientific nature of decision-making: The panoramic model provided by this invention can more accurately reflect the complex interaction between data, security, privacy and energy operation in the real world, so that VPP decision-making is transformed from isolated and static optimization to systematic, dynamic and forward-looking collaborative optimization, resulting in higher decision quality and stronger risk resistance.

[0025] (3) Resource utilization efficiency optimization: Through a refined model, this invention achieves the optimal allocation of data resources, privacy budget, security investment and various physical energy resources in multiple time scales and multiple market environments, avoiding resource waste and improving the operating efficiency of the entire VPP aggregate.

[0026] (4) Controllable privacy and security: While pursuing economic benefits, this invention internalizes privacy protection and data security into controllable cost items through a quantitative model, enabling VPP operators to formulate and implement reasonable privacy and security strategies under clear economic guidance, and achieve a balance between development and security.

[0027] (5) Technical universality and forward-looking: The “energy-data” collaborative optimization framework and method proposed in this invention are not only applicable to the current power market environment, but their core ideas also provide a universal modeling and solution paradigm for the optimization problem of more digital elements and physical systems deeply integrated in the future energy Internet. Attached Figure Description

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

[0029] Figure 1 This is a flowchart of a VPP privacy-preserving energy-data collaborative optimization method according to an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of the structure of a VPP privacy-protected energy-data collaborative optimization system according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] The terminology used in the embodiments section of this invention is for the purpose of explaining specific embodiments of the invention only, and is not intended to limit the invention.

[0033] This invention proposes a privacy-preserving energy-data collaborative optimization method and system for Virtual Power Utility (VPP). The core idea is that the quality of a VPP's operational optimization decisions highly depends on the quality and breadth of the data it can acquire; however, data acquisition is itself a costly and risky market economy process. Therefore, a VPP's globally optimal decision must simultaneously address the strategic problem of "at what cost to acquire what kind of data" and the tactical problem of "how to utilize this data to achieve optimal energy scheduling."

[0034] To address the aforementioned issues, this invention proposes a tightly coupled two-way feedback closed loop. The outer-layer game theory sets the information environment (data level), security tone, and privacy-efficiency trade-offs for the inner-layer optimization. The output (market revenue) of the inner-layer optimization, in turn, serves as the sole benchmark for evaluating the success or failure of the outer-layer game strategy, guiding the Virtual Power Plant Operator (VPPO) to adjust its data strategy. The overall goal is to find a Stackelberg equilibrium at which neither the VPPO nor the distributed resources have an incentive to unilaterally change their strategy.

[0035] like Figure 1 As shown, this embodiment of the invention provides a VPP privacy-preserving energy-data collaborative optimization method, including:

[0036] S1. Establish an outer-layer master-slave game model and obtain the strategic parameters output by the outer-layer master-slave game model. The outer-layer master-slave game model includes a follower model and a leader model.

[0037] S2. Based on the strategic parameters output by the outer master-slave game model and combined with the inner multi-timescale operation optimization model, obtain the operation data output by the inner multi-timescale operation optimization model.

[0038] S3. Using the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model, a two-way coupling feedback mechanism is constructed, and a Stackelberg equilibrium mechanism is established.

[0039] S4. Based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism, a distributed collaborative solution method is adopted to obtain the VPP privacy-preserving energy-data collaborative optimization results;

[0040] The outer master-slave game model includes a follower model and a leader model, and the inner multi-timescale operation optimization model includes a day-ahead split-bar optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model.

[0041] In summary, steps S1 to S4, by constructing a closed-loop system encompassing "strategic game theory, operational optimization, two-way feedback, and collaborative problem-solving," achieved multiple technological breakthroughs: Firstly, by dynamically linking data quality with market prediction errors, risk costs, and returns, a unified quantitative framework was built, effectively resolving the disconnect between data element value assessment and energy optimization decisions, achieving the optimal balance between data investment and energy returns. Secondly, by including privacy budget allocation, data security investment, and data procurement prices as strategic decision variables, and through dynamic collaborative configuration of data market procurement and privacy security technologies, dynamic synergy between information security, user privacy protection, and operational efficiency was achieved. This approach significantly reduces the total system cost; it establishes a full-time-scale model covering day-ahead, intraday, and real-time periods, and finely allocates data resources and privacy budgets based on the revenue characteristics of different ancillary services. This enables global coordination of bidding and scheduling strategies in the external data market and internal multi-markets, solving the challenge of coordinated allocation of data and physical resources across multiple time scales and ancillary service markets. Through the bidirectional coupling of outer-layer master-slave game theory and inner-layer multi-stage stochastic optimization, a closed-loop adaptive optimization of "strategy formulation - operational implementation - revenue feedback - strategic calibration" is formed, addressing the shortcomings of isolated models and broken decision-making chains in existing models. This effectively supports VPP operators in maximizing long-term comprehensive profits.

[0042] As one possible implementation, in the above embodiments, step S1 may specifically include the following steps:

[0043] Acquire virtual power plant operators and distributed resources in the outer master-slave game;

[0044] A follower model is constructed, the level of willingness to share data is defined as a decision variable for distributed resources, and the optimal response function of distributed resources is obtained based on privacy costs and data purchase prices.

[0045] Based on the optimal response function of the distributed resources, the aggregated data level of the virtual power plant operator is obtained by using the relative value of the distributed resource data as weights.

[0046] A leader model is constructed, defining data procurement price, data security protection investment level, and privacy budget as decision variables for virtual power plant operators, and constructing the objective function of virtual power plant operators in combination with the aggregated data level of the virtual power plant operators;

[0047] Based on the decision variables and objective function of the virtual power plant operator, set the leader constraint conditions;

[0048] By combining the optimal reaction function of the distributed resources with the aggregated data level of the virtual power plant operator and the leader constraint, the strategic parameters output by the outer master-slave game model are obtained.

[0049] The outer model describes the strategic interaction between the VPPO and distributed resources surrounding data sharing. The VPPO is the leader, setting the data price; the resources are the followers, determining their willingness to share. The specific steps for establishing the outer master-slave game model are as follows:

[0050] Step 1.1: Establish the follower model: rational decision-making for distributed resources;

[0051] This step aims to construct a behavioral model of distributed resources in the data market. Its core logic is that each resource owner is a rational decision-making entity. Based on the data prices published by VPPO, they weigh the economic benefits of data sales against the privacy losses associated with data sharing, thereby determining their optimal level of data sharing. This decision-making process constitutes the lower-level problem in the outer master-slave game.

[0052] (1) Decision variables and objective function:

[0053] For any distributed resource i∈I (e.g., a household user with a smart meter, electric vehicle, or rooftop solar panels), the decision to be made is to select a data sharing willingness level, denoted as θ. i This variable is a continuous variable, with values ​​ranging from 0 to 1 (data sharing willingness level, 0 ≤ θ). i ≤1). Where I represents the overall set of all distributed resources participating in data sharing and VPP collaborative optimization.

[0054] When θ i When =0, it means that distributed resource i refuses to share any data.

[0055] When θ i When =1, it means that the distributed resource i is willing to share all its available data, such as accurate load curves, real-time power generation, and the state of charge (SOC) of electric vehicles.

[0056] When θ i Taking the middle value indicates partial sharing, such as sharing only aggregated data with a coarser time scale, or data with added noise.

[0057] The decision objective of distributed resource i is to maximize its own utility, and its utility function U is defined. i for:

[0058] (1)

[0059] The utility of distributed resource i comes from the economic compensation obtained from selling data, minus the subjective costs caused by privacy leaks. This function reflects the user's privacy trade-off behavior when faced with economic incentives.

[0060] π dataIndicates the data purchase price, π data ∙θ i This refers to data revenue, which is a price (e.g., yuan / unit data level) uniformly set by the VPPO for each unit of data sharing level. The more data distributed resource i shares (θ... i The larger the size of the property, the greater the economic compensation received.

[0061] The privacy cost is a non-linear cost representing the subjective negative effects of shared data on distributed resource i, such as privacy loss and decreased security. The more data shared, the higher the privacy cost. To quantify the privacy cost, a convex quadratic function is used, as shown in formula (2):

[0062] (2)

[0063] in, This is the privacy sensitivity coefficient; the higher the value, the more sensitive the user is to privacy when sharing the same data (θ). i The stronger the marginal privacy aversion caused by this, the greater the sensitivity factor (κ). For example, a user who is extremely sensitive to personal information will have a high sensitivity coefficient (κ). i .

[0064] ν i For linear coefficients, this coefficient captures the "initial attitude" of distributed resource i towards data sharing. Its sign and magnitude have practical significance:

[0065] If ν i A value less than 0 indicates that the user has an inherent willingness to share. For example, a user may want to obtain better energy management services from the VPPO by sharing data, in which case sharing the data itself can bring certain positive utility (negative cost).

[0066] If ν i A value greater than 0 indicates that even sharing a very small amount of data will incur a fixed startup cost, such as the inconvenience of needing to install additional equipment or changing usage habits.

[0067] The significance of the convexity of the privacy cost function: the first derivative of the cost function It is increasing, among which, This represents the current stage of privacy operation costs. This means that marginal privacy costs are increasing: when sharing the first unit of data, users may not feel much impact, but as more and more data is shared, the user's sense of "unease" or "loss" increases with each additional unit shared. This convexity assumption aligns with the loss aversion principle prevalent in behavioral economics, ensuring that the optimization problem has a unique solution.

[0068] (2) Derivation of the optimal reaction function for resource i:

[0069] Given the purchase price π published by VPPO data Distributed resource i determines its optimal data sharing willingness level by solving the above utility maximization problem. This is a convex optimization problem with simple boundary constraints.

[0070] First, ignoring the boundary constraints [0,1], we solve the first-order conditions (i.e., the utility function with respect to θ). i To find the internal extreme points, we can use the derivative of the derivative to be zero.

[0071] (3)

[0072] Solving the above equation yields the candidate optimal solution. :

[0073] (4)

[0074] Then, the boundary constraint 0≤θ needs to be considered. i ≤1. Therefore, the optimal response function of distributed resource i is the projection of its candidate solution onto the interval [0,1]:

[0075] (5)

[0076] in, Let π be the optimal response function for distributed resource i. data ≤ν i When the price offered by VPPO is very low, even insufficient to cover the initial cost (or inherent negative utility) of users sharing data, the rational choice for users is not to share data at all; i ≤π data ≤2κ i +ν i The higher the price, the more data users are willing to share, and the slope... Influenced by the privacy sensitivity coefficient, users with stronger privacy awareness (κ) i The larger the π value, the less sensitive its sharing behavior is to price; data ≥2κ i +ν i Users will choose to fully share their data when the marginal benefit is still greater than or equal to the marginal privacy cost.

[0077] Formula (5) clearly depicts the market supply curve. When the price falls below a certain threshold (ν... i At a price that is not high enough, users perceive the benefits as insufficient and refuse to share. As the price increases, the willingness to share increases linearly. When the price is high enough, users are willing to share all the data. This represents the optimal response of all followers. Together, they constitute the Nash equilibrium at a given price.

[0078] (3) Horizontal modeling of VPPO aggregated data:

[0079] VPPO is not concerned with decisions made for individual resources, but rather with the collective behavioral outcomes of all resources, i.e., the level of aggregated data available to the entire VPP, which is defined as the weighted sum of the optimal sharing levels of all resources:

[0080] (6)

[0081] Where, Θ * (π data ) represents the aggregated data level available across the entire VPP, with weights ω. i This represents the relative value of data provided by different distributed resources, and the weights can be set based on a variety of factors:

[0082] Resource type: Charging data for electric vehicles may be more valuable for scheduling than load data for a typical household, and therefore should be given higher weight.

[0083] Data quality: Users with high historical data accuracy and reliability may have greater weight in newly shared data.

[0084] Data dimension: Resources that provide multiple information such as power, state of charge (SOC), and controllability have higher data weight.

[0085] In formula (6), the aggregated data level Θ * Its core function: This aggregate variable serves as the central bridge connecting the outer game theory and the inner optimization. VPPO adjusts the data purchase price π. data To influence Θ * , and Θ * The value of this value will directly determine the information quality of the inner-layer energy optimization model, thereby affecting the market revenue of VPPO. Thus, the rational decision-making model of the followers (distributed resources) constitutes a key constraint on the VPPO (leader) profit maximization problem.

[0086] Step 1.2: Establish a leader model to maximize the strategic profits of VPPO;

[0087] As a leader, a VPPO needs to rationally anticipate the collective reaction of their followers. * Based on this, an optimal strategy is formulated to maximize its long-term overall net profit.

[0088] (1) Define VPPO decision variables:

[0089] π dataData purchase price (currency / unit of data level).

[0090] S: Level of investment in data security protection (dimensionless, representing the overall security level).

[0091] ε=(ε day ,ε intra ,ε real ): Privacy budget allocated to the day-ahead, intraday, and real-time phases. The privacy budget is inversely proportional to the injected noise, reflecting the strength of privacy protection. Among them, ε day For the privacy budget allocated to the day-ahead phase, ε intra For the privacy budget allocated to the intraday phase, ε real The privacy budget allocated to the real-time phase.

[0092] (2) VPPO objective function (maximizing overall net profit):

[0093] (7)

[0094] Among them, Π VPPO To maximize the overall profit of VPPO. Detailed explanation of each cost and benefit in formula (7):

[0095] 1) E p [R market* The expected market revenue is the most complex part of the entire model, representing the output of the inner multi-stage stochastic optimization. It indicates the expected market revenue under a given data environment (S, Θ). * Under the conditions of privacy protection, this refers to the expected market revenue that VPPO can obtain by optimally scheduling its internal energy assets to participate in the energy market, peak-shaving market, and frequency regulation market. This part represents the ultimate value of all VPPO's operational activities and is the sole criterion for evaluating the success or failure of its data strategy.

[0096] 2) π data ∙Θ * Data acquisition cost is the direct variable cost paid by the VPPO to acquire data resources, determined by the VPPO's decision variable π. data And the equilibrium response of resources Θ * The joint decision-making process reflects the interactive nature of the game. VPPO needs to weigh the benefits of acquiring high-quality data against the costs incurred.

[0097] 3) C risk (Θ * Market risk costs are economic losses caused by inaccurate forecasts, including penalties for power deviation assessments, penalties for insufficient reserve capacity, and missed trading opportunities due to forecasting errors. The accuracy of forecasting is positively correlated with the level of data.

[0098] (8)

[0099] Where, γ imb γ is the penalty coefficient for power deviation. res The penalty coefficient for insufficient reserves. The variance of power prediction error. This represents the variance of the load forecasting error. The variance of the forecasting error is a decreasing function of the data levels, and can be expressed in, for example, as follows:

[0100] , (9)

[0101] Where, α P β is the baseline coefficient for the power prediction error variance. P α is the attenuation coefficient of the power prediction error variance. L β is the baseline coefficient for the variance of load forecasting error. L This is the attenuation coefficient of the load forecast error variance.

[0102] Data is considered a factor of production here, and its input (Θ) * Increasing the variance (σ) can reduce uncertainty in the production process. 2 (Reduce), thereby reducing risk losses.

[0103] 4) C security (S,Θ * The total cost of data security (TWC) captures the "liability" of data, highlighting its potential for leakage and loss. VPPO can manage this risk through proactive data security protection investment level (S).

[0104] C security =C sec (S)+E[loss](10)

[0105] C sec (S) = c1S + c2S 2 (11)

[0106] (12)

[0107] Among them, C security Let c1 be the coefficient of the first term of the active defense cost, and c2 be the coefficient of the second term of the active defense cost.

[0108] Active defense cost C sec (S): This is a convex function of the data security protection investment level S, representing the increasing marginal cost of security investment. For example, the cost increase of purchasing a more advanced firewall or hiring more experienced security experts is non-linear.

[0109] Expected leakage loss E[loss]: Consists of leakage probability and leakage loss, where leakage probability p(S) is modeled as an exponential decay function of data security protection investment level S, p(S) = p0∙e -λS The risk of leakage decreases exponentially with increased security investment, which aligns with common understanding in the security field; leakage losses include compensation to resource providers. (Due to the leakage of user data) and VPPO's own losses V VPPO (Such as VPPO's own goodwill, compliance fines, intellectual property losses, etc.). Here, p0 is the baseline leakage probability (when S=0), and λ is the leakage attenuation coefficient of security investment.

[0110] (3) Constructing leader constraints:

[0111] 1) Privacy budget allocation constraints:

[0112] ε day +ε intra +ε real ≤ε total (13)

[0113] VPPO has a fixed total privacy budget ε total This necessitates a cross-period allocation between decision quality (requiring low noise, i.e., high ε) and the strength of user privacy protection (requiring high noise, i.e., low ε). This constitutes a profound cross-period trade-off. Allocating more budget to the day-ahead phase can improve the accuracy of long-term planning; allocating it to the real-time phase can improve control performance, but at the expense of privacy protection or decision quality in other phases.

[0114] 2) Boundary constraints of strategic variables

[0115] 0≤π data ≤π max (14)

[0116] S min ≤S≤S max (15)

[0117] ε day ≥0,ε intra ≥0,ε real ≥0 (16)

[0118] Where, π max S is the upper limit for data procurement price. min To minimize the investment in data security protection, S max For the maximum level of data security protection investment. Formula (14) prevents market manipulation and overcompensation, Formula (15) is the technical and budgetary limit for security investment, and Formula (16) ensures that the privacy budget is non-negative.

[0119] As one possible implementation, in the above embodiments, step S2 may specifically include the following steps:

[0120] Based on the strategic parameters output by the outer master-slave game model, the aggregation data level, data security protection investment level, and privacy budget are obtained. The privacy budget includes the day-ahead privacy budget, the intraday privacy budget, and the real-time privacy budget.

[0121] By leveraging the aggregated data level and the level of investment in data security protection, a set of uncertainties in data-energy coupling is constructed;

[0122] Based on the uncertainty set of the data-energy coupling and the day-ahead privacy budget combined with physical system and market constraints, a day-ahead sub-Bruker optimization model is constructed;

[0123] The basic scheduling and bidding plan are obtained using the aforementioned day-ahead Brussels bar optimization model;

[0124] Federated learning is used to update the global prediction model, and the optimized prediction model is obtained by combining the intraday privacy budget.

[0125] Based on the aforementioned optimized prediction model, a model prediction and control optimization problem is constructed, and combined with intraday stage-specific constraints, an intraday federated learning-model prediction and control model is constructed.

[0126] Based on the intraday federated learning-model prediction control model, intraday adjustment instructions are obtained;

[0127] Based on the data security protection investment level, a communication encryption scheme is determined, and a linearized system state space model is established in conjunction with the real-time privacy budget.

[0128] Based on the linearized system state-space model, a privacy-preserving output feedback mechanism is established.

[0129] Based on the privacy protection output feedback mechanism, a control objective and privacy cost function are established, and a real-time differential privacy control model is constructed by combining real-time frequency modulation resource dynamics and allocation constraints.

[0130] Real-time control commands are obtained using the aforementioned real-time differential privacy control model;

[0131] Based on the basic scheduling and bidding plan, the intraday adjustment instructions, and the real-time control instructions, the operational data output by the inner multi-timescale operation optimization model is obtained.

[0132] The inner model describes the given information and security environment (Θ) formed in the outer game. *Under the conditions of S, ε), how does the VPPO execute its core energy asset dispatching and market bidding tasks? This is a typical multi-stage, data-driven stochastic optimization problem, deeply integrating peak shaving and frequency regulation ancillary services. The specific steps are as follows:

[0133] Step 2.1: Establish the day-ahead blue bar optimization model (DRO);

[0134] Before the market closed, VPPO, based on limited current information, developed a robust scheduling and bidding plan for the next 24 hours. This step employed a Distributed Robust Optimization (DRO) framework, which, unlike stochastic programming, does not rely on elusive, precise probability distributions, nor is it overly conservative like traditional robust optimization. Instead, it assumes the true distribution exists within a "fuzzy set" and optimizes the expected return in the worst-case scenario. This approach is particularly well-suited for handling data-driven uncertainties with incomplete distribution information.

[0135] (1) Construct the objective function:

[0136] (17)

[0137] Among them, R day* For the objective function of the current Bruker bar optimization model, u day Let P be the vector of all decision variables in the day-ahead phase, containing all day-ahead decision variables (such as planned output of each unit, energy storage charging and discharging plans, market bidding volume, etc.), where P is the probability distribution, u is the set of uncertainties in data-energy coupling, and E is the vector of all decision variables in the day-ahead phase. P [∙] represents the mathematical expectation of the probability distribution P, and T represents the total number of day-ahead scheduling periods. The day-ahead energy market electricity price for time period t. Bidding for energy market The peak-shaving service price for time period t. This represents the total peak-shaving capacity declared by VPP during time period t. The price for peak-shaving services during time period t. This represents the total peak reduction capacity declared by VPP during time period t. The price for frequency adjustment services during time period t. This represents the total peak-shaving capacity undertaken by VPP during time period t. The price for down-frequency regulation services during time period t. C represents the total downshaving capacity undertaken by VPP during time period t. operation Total operating costs, including fuel costs Equipment degradation costs Demand response compensation costs etc., that is C priv,dayThis refers to current-day privacy operating costs. Current-day privacy operating costs are expressed as follows:

[0138] (18)

[0139] Where β is the current-day privacy cost discount factor, and ∆f is the sensitivity of the query function.

[0140] To protect trade secrets (such as bidding strategies), VPPO injects differential privacy noise into its day-ahead optimization, causing the final decision to deviate from the theoretically optimal solution. This cost quantifies the economic loss resulting from this decision bias. The variance of the noise and This is directly proportional, clearly demonstrating the fundamental trade-off between privacy protection and decision-making accuracy.

[0141] The set of uncertainties u in data-energy coupling is represented as:

[0142] (19)

[0143] Where M(Ξ) is the sub-Bruker optimization objective function mapping operator with the uncertainty set Ξ as input, which is essentially "the optimal value of the objective function in the worst-case scenario ξ∈Ξ", and E P [ξ] is the mathematical expectation vector of the random prediction error vector ξ with respect to the probability distribution P. ξ includes the prediction errors of all random variables such as photovoltaic output, load, and electricity price. Cov is the expected vector of prediction error (correlated with the aggregated data level). P [ξ] is the covariance matrix of the random vector ξ with respect to the distribution P, ∑(Θ * ) represents the covariance matrix of the prediction error (related to the aggregated data level), and W1(p1,p0) is the 1-Wasserstein distance between the true distribution p1 and the reference distribution p0. The set defined by formula (19) is one of the core technologies of the model, which quantifies the data (Θ) * How do security (S) influence the decision-making environment, where:

[0144] Data accuracy effect. Nominal forecast. The deviation varies with Θ * It increases and decreases. For example, photovoltaic prediction error can be modeled as... ,in, To consider the aggregated data level Θ * The nominal expected error of photovoltaic power output prediction The baseline nominal expectation (Θ) for photovoltaic power output prediction error * →0 (the limit value), ∆ is the reference error.

[0145] : Data volume effect. The more data (Θ)* The larger the eigenvalues ​​(∑), the more accurate the understanding of uncertainty, the smaller the eigenvalues ​​of the covariance matrix, the more compact the uncertainty set, and the more aggressive rather than conservative decisions VPPO can make.

[0146] W1[p1,p0]≤ρ(S): Security Trust Effect. W1 is the 1-Wasserstein distance, measuring the deviation between the true distribution p1 and a reference empirical distribution p0. The higher the level of data security protection investment S, the lower the risk of data tampering or model poisoning due to network attacks, and the higher the VPPO's trust in the reference distribution p0, thus the smaller the allowable distribution deviation ρ(S). For example, ρ0 is the baseline upper limit of the 1-Wasserstein distance (the limit when the data security protection investment S=0), and s is the season.

[0147] (2) Constructing the physical system and market constraints:

[0148] 1) Power balance constraints (considering uncertainties):

[0149] (20)

[0150] in, The base state planned output of unit i in time period t. To predict the output of photovoltaic power in time period t, Let be the power output prediction error of photovoltaics in time period t, and ES be the set of energy storage resources. Let i be the planned discharge power of energy storage device i during time period t. The charging power of energy storage device i during time period t. The total upmodulation capacity declared by VPP in time period t. The total down-regulation capacity declared by VPP in time period t. For fixed load, Let HP be the load forecast error for time period t, HP be the heat pump resource set, EB be the electric boiler resource set, and P be the load forecast error for time period t. i,t Let ξ be the net output of energy storage device i during time period t, and ξ be the prediction error.

[0151] Formula (20) is the most basic equilibrium constraint, which must hold under all possible scenarios within the uncertainty set u. It requires that at any time t, the sum of the power of all power sources (controllable generators G, photovoltaic PV, and energy storage discharge dis) must equal the sum of the power of all loads (energy market bids). Fixed load (Power consumption of heat pumps (HP) and electric boilers (EB)) and the capacity of ancillary services provided (peak shaving) ,FM The sum of the prediction error (ξ) and the prediction error (ξ). This is a robust constraint that ensures the scheduling scheme remains feasible even when faced with the worst-case prediction error.

[0152] 2) Heat / Cold Power Balance Constraints (by Season): Winter Heating: In winter, the heat generated by all heat sources (CHP, heat pumps, electric boilers, and thermal storage devices) must meet the building's heat load. The heat required to charge the thermal storage device The uncertainty of heat load was also taken into account. .

[0153] (twenty one)

[0154] Wherein, CHP represents the resource set of the combined heat and power unit, TS_dis represents the subset of discharge resources for the thermal storage device, and H... i,t Let be the thermal power output of thermal device i during time period t.

[0155] Summer cooling: In summer, the cooling load is met by the heat pump in cooling mode.

[0156] (twenty two)

[0157] in, Let i be the cooling power generated by the heat pump through the consumption of electrical energy during time period t. The cold load forecast error for time period t is... The building's cooling load power during time period t.

[0158] 3) CHP Unit Operation Constraints (Polygonal Feasible Region): The core characteristic of CHP units is the strong coupling between their electrical and thermal outputs. Their feasible region is a convex polygon, bounded by the maximum / minimum electrical output, maximum / minimum thermal output, and the upper and lower boundaries of the electro-thermal conversion (the "thermal-determined-electricity" line). This reflects the differences in energy quality and the rigid constraints of thermo-electric coupling, which is crucial for integrated energy system modeling.

[0159] Electrical power output (23)

[0160] Thermal power output (24)

[0161] Minimum electrical output (25)

[0162] Maximum electrical output (26)

[0163] The upper boundary of electricity is determined by heat (27)

[0164] The lower boundary of the electric field is determined by heat (28)

[0165] Minimum thermal output (29)

[0166] Minimum thermal output (30)

[0167] in, It is the electrical power output of CHP unit i during time period t; This refers to the power generation efficiency of CHP unit i; It is the fuel input power consumed by CHP unit i during time period t; It is the thermal power output of CHP unit i during time period t; This refers to the heating efficiency of the CHP unit i. This is the lower limit of the minimum electrical output of the CHP unit i; This is the upper limit of the maximum electrical output of CHP unit i; It is the upper limit slope of the "heat-driven power generation" operating zone of CHP unit i; b i It is the intercept of the upper limit line of the "heat-driven power" operating zone of CHP unit i on the thermal power axis; It is the lower limit slope of the "heat-driven power generation" operating zone of CHP unit i; a i It is the intercept of the lower limit line of the "heat-driven power generation" operating zone of CHP unit i on the thermal power axis; It is the lower limit of the minimum thermal output of CHP unit i; It is the upper limit of the maximum thermal output of CHP unit i.

[0168] 4) Energy storage dynamics and operational constraints (general form, applicable to electric energy storage ES, electric vehicle EV, thermal storage tank TS): Energy storage constraints ensure the spatiotemporal translation of energy. They include non-convex integer variables to prevent simultaneous charging and discharging, and consider physical limitations such as charging and discharging efficiency and ramp rate, making the problem a mixed integer programming problem.

[0169] Energy state evolution (31)

[0170] Capacity limitations (32)

[0171] Charging power limit (33)

[0172] Lower limit of charging power (34)

[0173] Charging and discharging are mutually exclusive, so binary variables (35) are introduced.

[0174] State of charge constraint (36)

[0175] Climbing rate constraint (37)

[0176] Electric vehicle availability constraints (38)

[0177] Electric vehicle availability constraints (39)

[0178] Among them, SOC i,t It is an energy storage device during the time period The state of charge (SOC) represents the percentage of energy currently stored relative to the rated capacity. i,t-1 It is the state of charge of energy storage device i in time period t-1 (i.e., the previous time period); It is the charging efficiency of energy storage device i, which represents the overall efficiency of energy conversion and storage during the charging process; ∆t is the charging power of energy storage device i during time period t; ∆t is the duration of each scheduling period. It is the discharge power of energy storage device i during time period t; It is an energy storage device The discharge efficiency represents the overall efficiency of energy release and output during the discharge process; It is the limit of the permissible state of charge of the energy storage device, which is usually set to prevent over-discharge; This is the upper limit of the state of charge allowed for the energy storage device, which is usually set to prevent overcharging; This is the maximum allowable charging power of energy storage device i; It is a binary variable used to indicate whether the energy storage device i is in a charging state during time period t (1 for yes, 0 for no). This is the maximum allowable discharge power of energy storage device i; It is a binary variable used to indicate whether energy storage device i is in a discharging state during time period t (1 for yes, 0 for no); SOC i,T It is the state of charge of energy storage device i at the end of the scheduling cycle (time period T); RD is the target state of charge (SOC) that energy storage device i is expected to achieve or maintain at the end of the dispatch cycle; i P is the maximum allowable discharge ramp rate of energy storage device i, i.e., the maximum reduction in discharge power between adjacent time periods; i,t It is the net output of energy storage device i during time period t, usually defined as ;P i,t-1 It is the net output of energy storage device i in time period t-1; RU i It is the maximum allowable charging ramp rate of energy storage device i, that is, the maximum increase in charging power between adjacent time periods; For the VPP access period of electric vehicle i; For the period when electric vehicle i leaves VPP; For electric vehicles i during the departure period The actual state of charge; Let i be the target state of charge for electric vehicle i when it leaves the field.

[0179] 5) Coupling Constraints between Heat Pumps and Electric Boilers: Heat pumps (HP) and electric boilers (EB) are key coupling components connecting the electrical and heating / cooling networks. Their constraints reflect energy conversion efficiency, with coefficients of performance (COPs) typically much greater than or close to 1, and are crucial for improving overall system energy efficiency. The correct coupling equations need to be selected based on the operating mode (winter / summer).

[0180] Heat pump heating (40)

[0181] Heat pump refrigeration (41)

[0182] Heat pump power limit (42)

[0183] Electric boiler heating (43)

[0184] Electric power limits for electric boilers (44)

[0185] Total power limit, if sharing a line (45)

[0186] in, It is the heat power generated by heat pump i through the consumption of electrical energy during time period t (output in heating mode). It is the coefficient of performance of heat pump i when it is running in heating mode in season s (such as winter), which represents the ratio of its output heat power to input electrical power; It is the electrical power consumed by heat pump i during time period t; It is the cooling power generated by heat pump i through the consumption of electrical energy during time period t (output in cooling mode). It is the coefficient of performance of heat pump i when it is running in cooling mode in season s (such as summer), which represents the ratio of its output cooling power to its input electrical power; This is the maximum electrical power limit that the heat pump i is allowed to consume; It is the heat power generated by the electric boiler i through the consumption of electrical energy during time period t; It is the electro-thermal conversion efficiency of electric boiler i; It is the electrical power consumed by electric boiler i during time period t; This is the maximum allowable electrical power consumption limit for electric boiler i; This is the upper limit of the total power of the power supply line when the heat pump i and the electric boiler i share the same power supply line.

[0187] 6) Demand Response Constraints: Demand response must not only limit the total reduction amount and instantaneous reduction rate, but also consider the physical characteristics of the load and user comfort, such as minimum sustained reduction time and minimum sustained recovery time, to avoid user discomfort and frequent equipment start-ups and shutdowns. This introduces complex integer logic constraints, significantly increasing the difficulty of solving the problem.

[0188] Total reduction limit (46)

[0189] Instantaneous reduction rate limit (47)

[0190] Minimum duration of reduction (48)

[0191] Minimum duration of recovery (49)

[0192] Power reduction and state relationship (50)

[0193] in, It is the actual load reduction power of demand response resource i in time period t; It is a demand response resource The maximum total load reduction allowed throughout the entire scheduling cycle; It is the maximum instantaneous load reduction ratio of demand response resource i at any given time period; It is the baseline load power of demand response resource i in time period t (i.e., the natural load before any reduction). It is the shortest number of periods that each load reduction must last when demand response resource i is invoked; It is a binary variable that indicates whether the demand response resource i is in a load reduction state during time period k (1 for yes, 0 for no). It is a binary variable that indicates whether the demand response resource i is in a load reduction state during time period t; It is a binary variable representing whether the demand response resource i is in a load reduction state during time period t-1; It is the shortest recovery period that demand response resource i must wait between two load shedding events; It is the maximum allowable load reduction power for demand response resource i during time period t.

[0194] 7) Ancillary service capacity feasibility constraints: Peak shaving capacity constraints: Upward peak shaving capacity is limited by the remaining upward adjustment margin of all resources that can increase power generation (generator units, energy storage discharge) or reduce load on the basis of the current plan. Downward peak shaving capacity is limited by the remaining downward adjustment margin of all resources that can reduce power generation or increase load (load reduction increase, energy storage charging) on ​​the basis of the current plan.

[0195] Increase the upper limit of peak capacity (51)

[0196] Lower the peak capacity limit (52)

[0197] in, It is the total peak-shaving capacity declared by VPP during time period t; G up It is a collection of dispatchable generator sets (such as CHP and gas turbines) that can provide peak shaving services. This is the maximum technical output of unit i; It is the base-state planned output of unit i in time period t; It is the upward rotation of unit i as a backup, which can increase the output capacity; ES is a collection of electrical energy storage resources; This is the maximum allowable discharge power of energy storage device i; It is the planned discharge power of energy storage device i during time period t; i represents the additional discharge capacity that the energy storage device i can provide; DR represents the demand response resource set. It is the maximum allowable power reduction of demand response resource i; It is the planned power reduction of demand response resource i in time period t; It is the additional scalable capacity of demand response resource i; It is the total peak-shaving capacity declared by VPP in time period t; G dn It is a set of dispatchable generator sets that can provide peak shaving services; This is the minimum technical output of unit i; The downward rotation of unit i for standby reduces the required output capacity. This is the maximum allowable charging power of energy storage device i; It is the charging power of energy storage device i during time period t; It is the additional charging capacity that energy storage devices can provide; It is the maximum allowable load increase power of demand response resource i (such as power restoration after load interruption).

[0198] 8) Frequency regulation capacity constraints: Frequency regulation capacity is not only limited by the instantaneous power ramp-up capability of resources, but also more strictly limited by their energy endurance. For example, for energy storage, its up-regulation capacity is limited by the energy that its current SOC can discharge at rated power without falling below the minimum SOC during the frequency regulation duration, ensuring that the total frequency regulation capacity provided by the VPP complies with market rules.

[0199] Upper limit of frequency modulation capacity (53)

[0200] Lower frequency modulation capacity limit (54)

[0201] Total power constraint of FM signal (55)

[0202] in, It is the total frequency regulation capacity declared by VPP in time period t; FR is the set of fast regulation resources that can provide frequency regulation services (such as electric energy storage and some gas turbines). It is the maximum technical contribution of resource i; It is the base state planned output of resource i in time period t; It is the upward adjustment margin of resource i based on power ramping capability; It is the frequency modulation energy endurance safety factor of resource i ( ), used to reserve a safety margin; SOC i,t It is the real-time state of charge of resource i (mainly energy storage) in time period t; This is the limit of the allowed charge state for resource i; It is the rated energy capacity of resource i; This is a typical duration of frequency modulation service (e.g., 15 minutes). It is the upper limit of the frequency modulation power of resource i based on energy endurance capability; It is the total down-regulation capacity declared by VPP in time period t; It is the minimum technical output of resource i; It is the downward adjustment margin of resource i based on power ramping capability; This is the upper limit of the charge states allowed for resource i; It is the discharge efficiency of resource i; It is the upper limit of the downward frequency modulation power of resource i based on its energy absorption capacity (considering charging efficiency). It is the total power of the frequency modulation signal provided by VPP, i.e., the average frequency modulation capacity; It is the upper limit of the total power of a single frequency regulation signal as stipulated by the power system, used to ensure the safe and stable operation of the system.

[0203] 9) Resource operating domain coupling constraints (ensuring ground-state plan compatibility with ancillary services): Resource operating domain coupling constraints ensure that when resources provide frequency regulation and peak shaving services, their ground-state plan ( Sufficient, non-overlapping adjustment margins must be reserved. For example, the planned output of an energy storage system must be set at a level that can meet both the upward and downward demands for frequency regulation and peak shaving, while its State of Charge (SOC) must be maintained within a "safe window" that can support the frequency regulation process. This profoundly reflects the resource synergy and competition under multi-market participation.

[0204] Consider the lower limit of output power for frequency modulation (56)

[0205] Considering the upper limit of output for frequency modulation (57)

[0206] Energy storage frequency regulation capacity is strongly coupled with SOC (58)

[0207] Overall, the lower limit is adjusted downwards (59).

[0208] Overall, the upper limit is adjusted upwards (60).

[0209] in, This is the limit of the permissible state of charge for energy storage devices. This is the upper limit of the allowed state of charge for energy storage device i; It refers to the charging efficiency of the energy storage device i; SOC i,t It is the real-time state of charge of energy storage device i in time period t; This represents the total downscaling capacity declared by VPP during time period t. It is the total downshaving capacity undertaken by VPP in time period t; This represents the total peak-shaving capacity declared by VPP during time period t. It is the total peak-shaving capacity undertaken by VPP during time period t.

[0210] Step 2.2: Establish intraday federated learning-model predictive control (FL-MPC);

[0211] Once in real-time operation, VPPO utilizes incoming, rolling data to revise the daily plan at a higher time resolution (e.g., 15 minutes). To protect user privacy and reduce communication overhead, Federated Learning (FL) is used for distributed, privacy-preserving prediction model updates, and Model Predictive Control (MPC) is employed for closed-loop optimization. The focus of this phase is on executing the peak shaving plan and maintaining readiness for frequency regulation services.

[0212] (1) Establishing a federated learning process:

[0213] 1) Initialize the global prediction model on the VPPO server. (Used for forecasting photovoltaic power, load, prices, etc.);

[0214] 2) For communication rounds n=1 to N FL :

[0215] a. VPPO broadcasts the current global model To a subset of the selected resource nodes;

[0216] b. Each selected node i utilizes its latest local data D. i,τ Calculate local model update ;

[0217] c. Nodes add differential privacy noise to local updates: ,in, Let be the amount of local model update for distributed node i in the nth round of federated learning. Let n be the update amount of the privacy-preserving local model after adding noise to node i in the nth round. i Differential privacy noise added for local model updates For n i Follows a normal distribution. The standard deviation of the differential privacy noise in federated learning. ;

[0218] d. The node will update after adding noise. Send back to the VPPO server;

[0219] e.VPPO aggregates all received updates: , where η FL The global learning rate for federated learning;

[0220] 3) Output the final model Used for subsequent MPC prediction, where These are the final global model parameters after federated learning training is completed.

[0221] (2) Establish the MPC optimization problem:

[0222] At the starting point τ0 of each rolling window, solve the following finite-time optimization problem:

[0223] (61)

[0224] C comm =c comm ∙Θ *(62)

[0225] (63)

[0226] Among them, u intra It is the decision variable vector that needs to be solved in intraday rolling optimization, which usually contains the power adjustment instructions of each resource within the rolling window; τ0 is the start time of the current rolling optimization window; H is the length of the rolling optimization window (the number of time periods covered). E[∆P] represents the real-time market electricity price for the time period τ. T ] is the expected value of the power exchange deviation, which, when multiplied by the electricity price, constitutes the expected settlement cost or revenue in the real-time market; ∆P T It is the power exchange deviation between VPP during time period τ and the real-time market plan; ρ imb It is the penalty coefficient for power imbalance, reflecting the severity of the deviation assessment; This is the expected variance of the power imbalance, which is used to minimize operational risk. This is the net power imbalance within the VPP caused by factors such as prediction errors; ρ PR It is the penalty coefficient for deviations in the peak shaving plan tracking; It is the planned peak-shaving power formulated in the day-ahead phase and for the time period τ; It is the actual executable peak-shaving power during the time period τ, based on the latest forecasts and conditions; λ comfort It is the penalty coefficient for thermal comfort deviation, reflecting the degree of importance attached to user comfort; ∆T τ Indoor temperature T in,τ Its set value T set Deviation; This is the expected variance of the temperature deviation, which is used to maintain indoor thermal comfort; C comm The total communication cost incurred during the federated learning process; C priv,intra These are the privacy operation costs incurred during the daytime implementation of privacy protection measures; c comm It is the cost coefficient per unit of communication rounds and per unit of data level; Θ * The aggregated data level is determined by the outer game theory, representing the data dimension and scale; η is the privacy cost coefficient, used to quantify performance loss into economic cost; ∆f FL It refers to the sensitivity of the query function in the federated learning model; ε intra The privacy budget allocated to the intraday phase determines the intensity of the noise addition; N FL It represents the total number of communication rounds performed during the federated learning process.

[0227] (3) Establish intraday phase-specific constraints:

[0228] 1) Scrolling state update constraint: At the beginning of each scrolling window, MPC uses the latest actual measurement value. This is used to initialize the system state (such as energy storage SOC, indoor temperature, etc.) to achieve closed-loop feedback.

[0229] (64)

[0230] in, The initial state value for MPC rolling optimization is assigned from the measured value, with no prediction error;

[0231] 2) Frequency regulation readiness status maintenance constraints: During the daytime phase, even if frequency regulation services are not actively provided, the VPPO must ensure, through rolling optimization, that the SOC of its frequency regulation resources (especially energy storage) is maintained within a "ready state" or "adjustable bandwidth" that can respond to frequency regulation commands in the next time period at any time. This is crucial to ensuring the reliability of frequency regulation services.

[0232] (65)

[0233] in, This refers to the upper limit of frequency modulation capacity declared by VPP during time period t. This represents the upper limit of the down-regulation capacity declared by the VPP during time period t.

[0234] 3) Dynamic Indoor Temperature Constraints: This constraint is a mathematical description of thermal inertia, directly linking power dispatching with user thermal comfort. The heat capacity C of the building envelope... th and heat loss coefficient K loss This results in slow temperature changes, providing VPPOs with the flexibility to utilize buildings as "virtual energy storage." MPCs optimize electrical and thermal loads to achieve time-shifted energy while maintaining comfort.

[0235] (66)

[0236] (67)

[0237] Among them, T in,τ+1 The predicted indoor temperature for the next optimization period τ+1; T in,τ is the indoor temperature during the current optimization period τ (the initial value is usually a measured value); ∆t is the time step of the optimization model (e.g., 15 minutes); C th It is the equivalent heat capacity of the building envelope, reflecting the building's heat storage capacity; It is the heating power generated by the heat pump during time period τ; It is the heating power generated by the electric boiler during time period τ; It is the heat power generated by the combined heat and power unit during time period τ; It is the heat release power of the thermal storage device during time period τ; K is the charging power of the thermal storage device during time period τ (considered as heat load); loss It is the building's overall heat loss coefficient, reflecting the rate of heat dissipation through walls, doors, windows, etc.; T out,τ It is the outdoor ambient temperature during the time period τ; It is the lower limit of the indoor temperature allowed by the user, i.e., the minimum comfortable temperature; It is the upper limit of the indoor temperature allowed by the user, i.e. the highest comfortable temperature.

[0238] Step 2.3: Establish a real-time differential privacy control model;

[0239] In the real-time operation phase, ranging from seconds to minutes, the VPPO needs to respond quickly to Automatic Generation Control (AGC) commands issued by the grid dispatch center to provide high-quality frequency regulation ancillary services. The core characteristics of this phase are extremely high timeliness requirements and stringent tracking accuracy. However, to ensure the privacy and security of real-time operational data on the user side (such as the actual charging power of electric vehicles, indoor temperature of buildings, and the precise state of charge of energy storage), and to prevent the leakage of sensitive information during transmission and processing, it is necessary to systematically inject differential privacy noise that meets strict mathematical definitions into the control loop.

[0240] (1) Establish the state-space model of the linearized system:

[0241] To design the controller, a mathematical model describing the key dynamic processes within the VPP is first required. Using a state-space model and linearizing it around typical operating points, we obtain:

[0242] 1) Equations of state:

[0243] (68)

[0244] x=[P ES SOC ES H TS SOC TS ,P ES ,T in ,...] T (69)

[0245] in, Let A be the system state change rate vector at time t, and let A be the state matrix, describing the system state variables themselves and their dynamic coupling relationships; x(t) is the system state vector, which contains all the key variables describing the internal dynamics of the VPP; B is the input matrix, describing how the control input affects the changes in the state variables; u(t) is the control input vector, which is the instruction that the controller needs to calculate and issue; w(t) is the process noise, which follows a Gaussian distribution with zero mean and covariance matrix Q, i.e., w(t) ~ N(0,Q), used to model unconsidered modeling errors and external disturbances; N(0,Q) is a multivariate Gaussian distribution with zero mean and covariance matrix Q; x is the core state vector of the VPP linearized system; P ES It is the actual output of electrical energy storage; SOC ES It is the state of charge of electrical energy storage; H TS It refers to the heat release / absorption power of the thermal storage tank; SOC TS It refers to the heat storage state of the thermal storage tank; T in It refers to the indoor temperature of the building.

[0246] 2) Output equation:

[0247] y(t)=Cx(t)+v(t)(70)

[0248] w=[P ES,ref ,P HP,ref H EB,ref ,...] T (71)

[0249] Where y(t) is the ideal measurement output vector, which is a physical quantity that can be obtained directly or indirectly through a sensor; C is the output matrix, describing how the state variables are mapped to the measurement output; v(t) is the measurement noise, which follows a Gaussian distribution with zero mean and covariance matrix R, i.e., v(t) ~ N(0,R), used to model the sensor's measurement error; w is the core reference command vector for VPP real-time control; P ES,ref It is a power reference command issued to the energy storage; P HP,ref This is a power reference command sent to the heat pump; H EB,ref It is a thermal power reference instruction sent to the electric boiler.

[0250] (2) Establish a privacy protection output feedback mechanism:

[0251] To protect data privacy, the VPP's central controller does not receive the raw measurement value y(t), but rather a carefully processed signal containing additional privacy noise:

[0252] (72)

[0253] in, This is the privacy-preserving measurement signal actually received by the controller; n(t) is the injected privacy noise. This noise follows a Laplace distribution, and its scale parameter is determined by the privacy budget and the function sensitivity, i.e. ; Laplance(∙) is the Laplace noise distribution operator; ε real This is the privacy budget allocated to the real-time control phase. A smaller value means more noise is added, resulting in stronger privacy protection but poorer control performance; ∆f comtral It is the sensitivity of the query function, that is, the maximum change in output that can be caused by changing a user's input data in the worst case, which is the lower limit of the noise required to protect privacy.

[0254] (3) Establish the control objective and privacy cost function:

[0255] The goal of a real-time controller is to find an optimal control law u(t) that minimizes the following overall performance metrics:

[0256] (73)

[0257] (74)

[0258] (75)

[0259] Where E[∙] is the mathematical expectation operator, used to solve the statistical expectation of the objective function under the combined effects of process noise, measurement noise and differential privacy noise; r(t) is the reference trajectory, which may include the set value of indoor temperature, the charging and discharging power of the energy storage plan, etc. It is the weighted sum of squares of the output tracking error; the weight matrix Q determines the importance of the tracking accuracy of different output variables; It is the weighted sum of squares of the control energy; the weight matrix R is used to penalize excessive control actions to ensure smooth system operation; ϕ is the penalty coefficient for frequency modulation tracking deviation, which is usually set very large to ensure a fast and accurate response to AGC commands, which is a core requirement of frequency modulation service; P GAC (t) is the standardized AGC command signal issued by the power grid, which is usually a time-varying signal between -1 and 1, where -1 represents a reduction in full power and +1 represents an increase in full power; P GAC ,actual (t) is the normalized frequency modulation power of the actual output of VPP; It is the actual output of frequency modulation resource i at time t; It is the ground-state planned output of frequency modulation resource i; P FR,total This is the total FM capacity promised by VPP; C priv,real is the real-time privacy operating cost; μ is the privacy cost discount factor, used to quantify the loss of control performance into economic cost.

[0260] (4) Establish real-time frequency regulation resource dynamics and allocation constraints:

[0261] 1) Dynamic constraints of frequency regulation energy storage: Formula (76) describes the dynamic change of the state of charge (SOC) of an energy storage device specifically designed to provide frequency regulation services, indicating that the rate of change of the SOC of the energy storage is related to its reference discharge power for tracking AGC commands. It is directly proportional to its discharge efficiency. and rated energy capacity Inversely proportional. This dynamic must be monitored in real time to ensure that the SOC of the energy storage remains within a safe operating range during the response to frequency regulation commands, preventing over-discharge or over-charging, and ensuring the continuity of frequency regulation services and equipment safety.

[0262] During discharge (76)

[0263] in, Let i be the rate of change of the state of charge of the electrical energy storage. The reference discharge power for tracking AGC commands for energy storage i.

[0264] 2) Frequency regulation power allocation constraint: Equation (77) requires that the sum of the deviations of the actual output of all resources participating in frequency regulation within the VPP from the ground state plan must be exactly equal to the total frequency regulation power (P) required by the power grid. FR,total •P GAC This requires the VPP to have a fast and reliable power allocation algorithm that can decompose the total demand into various frequency-regulated resources (such as energy storage, controllable loads, etc.) on a second-scale timescale. The allocation strategy can be based on a pre-agreed proportional coefficient or embed a fast online optimization module that considers the adjustment capability and cost of each resource.

[0265] (77)

[0266] 3) Frequency modulation ramp rate constraint: Formula (78) requires that each frequency modulation resource Actual output change rate It cannot exceed its physical maximum gradeability. Frequency modulation (FM) services, especially secondary FM, require resources to be able to follow rapidly changing AGC signals, thus placing demands on ramp rate. This constraint is typically much higher than the requirements for participating only in energy dispatch or peak shaving. It ensures the executability of frequency regulation commands and avoids tracking errors and service quality degradation caused by equipment inability to keep up with response times.

[0267] (78)

[0268] 4) Frequency regulation control end-to-end time delay constraint: Formula (79) requires that the total end-to-end time delay from the start of sensor measurement to the completion of control command calculation and issuance to the actuator (such as the energy storage converter) must be less than the upper limit δ specified by the power grid. max This constraint forces VPPO to make a difficult trade-off between the security of the control system and the real-time response. In the real-time phase, in order to meet the millisecond-level response requirements, it is often necessary to sacrifice some security strength, such as using lightweight cryptographic algorithms or symmetric encryption algorithms with lower computational overhead, to ensure the real-time performance of the frequency modulation service.

[0269] (79)

[0270] in, The computational latency is primarily consumed by running the control algorithm and performing necessary encryption and decryption operations. This latency is highly dependent on the cryptographic algorithm Alg used. i The complexity, while the algorithm strength Alg i This is also related to the level of security investment (S) determined by VPPO at the strategic level. Higher-strength algorithms are more secure, but also have higher computational overhead and longer latency. δ comm It refers to communication latency, which includes the time for data transmission and network queuing.

[0271] As one possible implementation, in the above embodiments, step S3 may specifically include the following steps:

[0272] Based on the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model, the bidirectional coupling variables of the virtual power plant master-slave participants are obtained.

[0273] A bidirectional coupling feedback channel is established based on the bidirectional coupling variables of the virtual power plant master-slave participants.

[0274] A bidirectional coupling feedback mechanism is constructed using the aforementioned bidirectional coupling feedback channel;

[0275] The Stackelberg master-slave game subject boundaries are defined based on the bidirectional coupling feedback mechanism, and a Stackelberg equilibrium mechanism is established.

[0276] The specific steps for designing the coupling mechanism between the inner and outer layer models and the Stackelberg equilibrium mechanism are as follows:

[0277] Step 3.1: Design the coupling mechanism between the inner and outer layer models;

[0278] The "energy-data" collaborative optimization model constructed in this invention forms a complete two-way closed-loop feedback system, and its coupling relationship is reflected in the following two levels:

[0279] (1) Outer layer → Inner layer (strategy determines the operating environment):

[0280] The strategic decision-making of the outer master-slave game sets the specific operating environment for the inner energy optimization:

[0281] 1) Data quality determines information quality: Aggregated data level Θ * This directly determines the accuracy of the prediction model and the size of the uncertainty set in the inner optimization. Higher Θ * This means more accurate photovoltaic and load forecasting, enabling VPPO to make more forward-looking and accurate scheduling decisions, reducing conservatism and thus obtaining higher returns in the market.

[0282] 2) Security investment determines trust level: The level of investment in data security, S, shapes the risk-taking nature of operational decisions by influencing the conservatism of the uncertainty set. A higher S reduces the risk of data tampering, enhances the VPPO's trust in the data, and enables them to adopt more proactive market strategies.

[0283] 3) Privacy budget determines efficiency trade-offs: The allocation of the privacy budget ε directly translates into the operating costs of the three inner phases. In the day-to-day phase, ε day Influences decision-making bias; during the intraday phase, ε intra Impact on federated learning efficiency; in the real-time phase, ε real This impacts control performance. This forces operations teams to strike a balance between privacy protection and operational efficiency.

[0284] (2) Inner layer → Outer layer (Operational feedback strategic value):

[0285] The operational results of multi-stage optimization at the inner layer provide value assessment for strategic decision-making at the outer layer:

[0286] 1) Market revenue is the core value benchmark: Expected market revenue E[R] market* [This is an assessment of outer-layer strategies (π)] data The sole criterion for success or failure is the objective function Π(S,ε). It quantifies the core profitability of VPPO under a specific data environment and directly determines the outer objective function Π. VPPO Size.

[0287] 2) Privacy costs impact budget allocation: Privacy operation costs C at each stage of the inner layer priv Privacy constitutes a significant portion of the total cost of VPPO, which necessitates that external decision-makers allocate privacy budgets prudently to avoid sacrificing too much operational efficiency for over-protection and to achieve synergistic optimization between privacy protection and economic benefits.

[0288] Step 3.2: Design the Stackelberg load balancing mechanism;

[0289] The solution to the model of this invention is a Stackelberg equilibrium, and the strategy combination is as follows. An equilibrium is established if and only if the following conditions are met:

[0290] (1) Follower optimality: at the equilibrium price π given by VPPO data* Under these conditions, no distributed resource has an incentive to unilaterally change its data sharing volume. Each resource maximizes its own utility at the current price. For each distributed resource i, the following holds true:

[0291] (80)

[0292] in, Maximize the utility of each resource at the current price.

[0293] (2) Leader optimality: when it is anticipated that the resource provider will follow its optimal reaction function Under the premise of taking action, VPPO's chosen data pricing, security investment, and privacy budget strategy represents the optimal choice for maximizing its long-term overall profit. Among these, E[R] market* This is obtained by solving the complete internal multi-stage energy optimization problem. For VPPO, the following conditions are met:

[0294] (81)

[0295] Among them, S * For VPPO to achieve the optimal level of data security protection investment, ε * The optimal differential privacy budget allocation strategy for VPPO.

[0296] (3) Consistency: Ensure that the data level expected by the strategic layer is consistent with the data level actually available to the operational layer, forming a self-consistent closed-loop system. Aggregated data level satisfies:

[0297] (82)

[0298] At the Stackelberg equilibrium point, the system reaches an optimal state in multiple senses:

[0299] Data market clearing: Data supply (willingness to share resources) and data demand (VPPO's willingness to purchase) at equilibrium price π data* A balance is reached, and data, as a factor of production, achieves optimal allocation.

[0300] Risk-reward balance: The marginal investment of VPPO in data security equals the marginal risk reduction benefit it brings, and the marginal cost of privacy protection equals the marginal operational efficiency improvement it brings, thus forming the best trade-off between security, privacy and efficiency.

[0301] Strategic and operational synergy: The energy system achieves optimal operation under given information constraints. The resource allocation at the strategic level and the scheduling decisions at the operational level are highly coordinated, and the entire VPP system reaches a globally optimal state.

[0302] Incentive compatibility: The equilibrium outcome ensures that all participants (VPPO and distributed resources) spontaneously improve the overall efficiency of the system while pursuing their own interests to maximize, thus forming a good incentive mechanism.

[0303] As one possible implementation, in the above embodiments, step S4 may specifically include the following steps:

[0304] S4-1. Based on the bidirectional coupling feedback mechanism, obtain a multimodal intelligent agent, which includes a data market intelligent agent, a distributed resource intelligent agent, a day-ahead scheduling intelligent agent, an intraday rolling optimization intelligent agent, and a real-time control intelligent agent;

[0305] S4-2. Using the multimodal agent, initialize the multimodal agent strategy, which includes a data market agent strategy, a distributed resource agent strategy, and operation layer agent decision variables.

[0306] S4-3. Based on the aforementioned multimodal agent strategy, a distributed collaborative solution method is adopted to obtain a VPP privacy-preserving energy-data collaborative optimization scheme, including:

[0307] S4-3-1. Update the outer game agent of the distributed resource agent according to the multimodal agent strategy to obtain the distributed consensus result;

[0308] S4-3-2. Based on the distributed consensus result and combined with the multimodal agent strategy, perform inner layer operation agent update on the day-ahead scheduling agent, the intraday rolling optimization agent and the real-time control agent to obtain the inner layer operation agent optimization result;

[0309] S4-3-3. Use the optimization results of the inner layer operating agent to perform Monte Carlo simulation and obtain global gradient estimation;

[0310] S4-3-4. Based on the global gradient estimation and the multimodal agent policy, the Actor-Critic method is used to perform leader policy update on the data market agent to obtain a VPP privacy-preserving energy-data collaborative optimization scheme.

[0311] S4-4. Based on the VPP privacy-preserving energy-data collaborative optimization scheme and the Stackelberg equilibrium mechanism, a convergence judgment is performed to obtain the VPP privacy-preserving energy-data collaborative optimization result, including:

[0312] The convergence residual is obtained using the aforementioned VPP privacy-preserving energy-data collaborative optimization scheme;

[0313] Determine whether the convergence residual satisfies the convergence judgment condition. If it does, obtain the VPP privacy-preserving energy-data collaborative optimization result based on the VPP privacy-preserving energy-data collaborative optimization scheme combined with the Stackelberg equilibrium mechanism. Otherwise, return to step S4-3-1.

[0314] Traditional methods decompose complex bi-level optimization problems and typically employ centralized iteration or simple distributed computation. The core innovations of this algorithm are: 1) modeling all participants in the outer-level game and the different timescales of the inner-level optimization problem as independent "agents"; 2) designing a "virtual coordinator" that does not perform centralized computation but guides all agents to collaboratively converge to Stackelberg equilibrium and the globally optimal scheduling scheme through a carefully designed "consensus-innovation" information exchange protocol; 3) embedding the dynamic constraints of the physical system (such as the dynamics of energy storage SOCs and thermal dynamics) as prior knowledge into the agents' policy updates or objective functions, greatly accelerating convergence and ensuring the feasibility of the solution. The specific steps of the algorithm are as follows:

[0315] Step 3.1: Defining the agent and refactoring the problem;

[0316] The original problem is decomposed into the following five types of agents, each responsible for solving a subset of the original problem:

[0317] 1) Designing data market intelligent agents:

[0318] Type: Leader agent; Decision variable: Data purchase price π data Data security protection investment level S, privacy budget allocation ε; Objective: Maximize Π defined by formula (7) VPPO Key information: Θ needs to be obtained from other intelligent agents. * (Aggregate data level) and E[R] market* The estimated value of (expected market revenue).

[0319] 2) Design a distributed resource intelligent agent:

[0320] Type: Follower agent (one per resource); Decision variable: its own data sharing level θ i Objective: Maximize the self-utility function U defined by formula (1) i Key information: Receive the current π from the data market agent. data .

[0321] 3) Design the day-ahead scheduling agent:

[0322] Type: Operational layer intelligent agent; Decision variables: All bids and scheduling plans in the day-to-day market. day Objective: To solve the bibliophilic bar optimization problem defined by formula (17) and maximize R. day* Key information: Receives Θ from the data market agent. * , S, ε day To construct an uncertainty set u.

[0323] 4) Design an intraday rolling optimization agent:

[0324] Type: Operational layer intelligent agent; Decision variable: Intraday adjustment instruction u intra Objective: To solve the MPC problem defined by formula (61); Key information: Receive Θ from the data market agent * and ε intra It also receives updated prediction models from the federated learning process.

[0325] 5) Design a real-time control agent:

[0326] Type: Operational layer intelligent agent; Decision variable: Real-time control command u real Objective: To solve the stochastic optimal control problem defined by formula (73), minimizing frequency modulation tracking error and privacy cost; Key information: Receive ε from the data market agent real And S (the latter affects the delay caused by the choice of communication encryption algorithm).

[0327] Step 3.2: Initialize the data market agent strategy, the distributed resource agent strategy, and the operation layer agent decision variables;

[0328] Let the iteration count k=0, and initialize the policies of all agents:

[0329] Data Market Intelligent Agent Strategy ψ 0 =(π data,0 ,S 0 ,ε 0 );

[0330] Distributed resource intelligent agent strategy ;

[0331] Operational layer intelligent agent decision variable u day,0 u intra,0 u real,0 , where u day,0 For the initial scheduling / control decision variables in the day-ahead phase, u intra,0 For the initial scheduling / control decision variables during the intraday phase, u real,0 These are the initial scheduling / control decision variables for the real-time phase.

[0332] Step 3.3: Distributed collaborative solution main loop;

[0333] Step 3.3.1: Outer game agent update (consensus phase);

[0334] 1) For k=0,1,2,...,K max Execution Policy Broadcast: The data market agent broadcasts its current policy ψ k ;

[0335] 2) Distributed resource agent update: Each resource agent i computes its optimal response in parallel:

[0336] (83)

[0337] in, π represents the optimal data sharing amount for distributed resource agent i in the (k+1)th iteration; data,k In the k-th iteration, the data market agent broadcasts the data sharing equilibrium price to distributed resource i, representing the revenue compensation per unit of data sharing. This is the core incentive signal driving resource i to determine the sharing amount; v i κ is the linear coefficient in the privacy cost function of resource i, reflecting its initial attitude towards data sharing; i The privacy sensitivity coefficient of resource i (>0) is the value of the resource i. The larger the value, the more sensitive it is to privacy.

[0338] Then, to protect privacy, Laplace noise was added:

[0339] (84)

[0340] in, To determine the level of privacy in data sharing after adding noise; ε is the Laplace noise generated for resource i in the k-th iteration; Laplace(0,b) is a Laplace distribution with mean 0 and scale parameter b; res For user privacy parameters; ε res A differential privacy budget is set for the resource side; the smaller the value, the stronger the privacy protection.

[0341] 3) Distributed consensus computing aggregated data level:

[0342] Each resource agent i initializes a consensus state. ;

[0343] For consensus iterations l=0,1,2,...,L-1:

[0344] (85)

[0345] Once consensus is reached, all agents obtain the global aggregated data level Θ. k+1 :

[0346] (86)

[0347] in, Let ω be the state vector of resource agent i at the start of the iteration; i The weight of the data provided to resource i reflects its relative value; L is the total number of iterations of the consensus algorithm; Let N be the state vector of resource agent i during the (l+1)th consensus iteration; i ω is the set of neighbors of resource agent i (communication topology definition); i,j These are the elements of the consensus weight matrix W; Let be the state vector of resource agent i during the l-th consensus iteration; and These are the two components of the state vector after consensus is reached.

[0348] Step 3.3.2: Inner layer operational agent update (physical information-guided innovation stage);

[0349] 1) Parallel optimization of operational layer intelligent agents:

[0350] ① A scheduling agent recently solved the Distributed Brussels Optimization (DRO) problem. Through dual transformation, the following deterministic second-order cone programming (SOCP) problem was solved:

[0351] (87)

[0352] Among them, R day,* The optimal expected return in the day-ahead energy market is obtained by solving the inner-layer sub-Bruker bar optimization model, representing the maximum market return of the VPPO under all operational constraints and uncertainty boundaries; w is the vector of decision variables in the day-ahead phase (e.g., unit output, market bid volume); λ, γ, and μ are dual variables; f is the coefficient vector of the decision variables in the objective function; ρ(S) is the Wasserstein sphere radius related to the safety input S, and ρ(S) is usually a decreasing function of S; Ξ(Θ) is the uncertainty covariance matrix related to the aggregated data level Θ, and Ξ is a decreasing function of Θ; Ξ(Θ) k+1 ) represents the level Θ of the global aggregated data from the (k+1)th round. k+1 The updated uncertainty covariance matrix is ​​given by Θ k+1 The decision reflects the effect of improved data quality on narrowing the uncertainty boundary; b is the constant term vector in the dual problem; A i d i c i e iG and h are the coefficient matrix and vector for defining the i-th second-order cone constraint; G and h are the coefficient matrix and vector for defining the linear physical constraint.

[0353] ② Intraday rolling optimization agent solves finite-time domain optimization problem within time window τ0:

[0354] (88)

[0355] Among them, u intra Let be the decision variable vector for the intraday phase; τ be the time step index in the intraday rolling optimization; τ0 be the start time of the current rolling optimization window; H be the prediction time domain (window length); y τ Let r be the system's output vector during time interval τ; τ To output y τ The reference trajectory; Q and R are the weight matrices for the output tracking error and control action; g τ η is the control input decision vector for time period τ in intraday rolling optimization; η is the privacy cost discount factor; ∆f FL Sensitivity of the query function in the federated learning model; x τ+1 Let x be the state vector of the system in time interval τ+1; τ Let A be the state vector of the system during time interval τ; A and B are the system matrices of the state-space model; g(∙) is the function that defines the operational constraints.

[0356] 2) Federal assessment of expected market revenue:

[0357] Each operational agent m∈{DA,INTRA,REAL} performs N locally s Sub-Monte Carlo simulation, calculating the stochastic gradient of its local value function:

[0358] (89)

[0359] The secure aggregator computes the global gradient estimate:

[0360] (90)

[0361] Where m is the operational agent type index (DA day-ahead, INTRA day-ahead, REAL real-time). For the local stochastic gradient estimation of the operating agent m in the k-th main iteration; N s The number of Monte Carlo simulations performed locally for each operational agent; The portion of market revenue contributed by the operating intelligent agent m For income The gradient of the comprehensive strategy ψ; k g is the global synthesis strategy vector for VPPO in the k-th main iteration; kThis is the global gradient estimate for the k-th main iteration; This is the local stochastic gradient estimate for the operating agent m in the kth main iteration.

[0362] Step 3.3.3: Leader Strategy Update (Innovation Phase Guided by Reinforcement Learning);

[0363] Leader strategy update, also known as data market agent strategy update (Actor-Critic), includes Critic update (evaluating the value function) and Actor update (improving the strategy).

[0364] 1) Critic Update (Evaluation Value Function):

[0365] Update the value network parameter ϕ using the temporal difference error:

[0366] (91)

[0367] Among them, the time-series difference target value y k for:

[0368] (92)

[0369] Among them, ϕ k+1 V represents the parameter set updated by the Critic neural network after the (k+1)th iteration; ϕ (ψ k ) is a Critic neural network with parameter ϕ, used to estimate the expected total return under policy ψ; Π VPPO (ψ k ) represents the true profit of VPPO under strategy ψ, defined by formula (7); γ is the discount factor (0≤γ≤1), used to measure the current value of future returns; Let ϕ be a Critic neural network used to estimate the expected total return under policy ψ.

[0370] 2) Actor Update (Improvement Strategy):

[0371] Update the policy network parameters ω using policy gradient ascent:

[0372] (93)

[0373] Advantage function A(ψ) k ,∆ψ k+1 Approximately:

[0374] (94)

[0375] Action ∆ψ k+1 Generated by the Actor network: The new strategy is ψ k+1 =ψ k +∆ψ k+1 ;

[0376] Where, ω k+1 ω represents the parameter set updated by the Actor neural network after the (k+1)th iteration. k Let α be the current parameter set of the Actor neural network at the k-th iteration; ω The learning rate (step size) of the Actor neural network; ∇ ω logπ ω (a|ψ) is the logarithmic gradient of the policy with respect to its parameter ω (score function); a is the action output by the Actor network, which in this context is the policy update; ∆ψ is the learning rate (step size) of the Actor network; ψ k+1 Let A(ψ) be the global synthesis policy vector updated by VPPO after the (k+1)th iteration. k ,∆ψ k+1 Let Δψ be the advantage function, representing the action ∆ψ to be taken in state ψ. k+1 Advantages relative to the average level; This is the value estimate of Critic after the (k+1)th iteration; This is the value estimate of Critic after the k-th iteration; Let be an Actor neural network with parameter ω, representing the strategy of taking action a in state ψ.

[0377] Step 3.4: Design convergence criteria;

[0378] Calculate the convergence residual:

[0379] (95)

[0380] Convergence criterion: If ∆ k ≤ε tol or k≤K max The algorithm terminates and outputs the equilibrium solution ψ. * And each operational strategy; otherwise, let k=k+1 and return to step 3.3.

[0381] Where, ∆ k π represents the convergence residual of the k-th iteration, measuring the change of key variables between successive iterations; data,k+1 The data sharing equilibrium price broadcast by the data market agent to distributed resource i in the (k+1)th iteration; Π VPPO (ψ k+1 ) represents the comprehensive strategy ψ of VPPO in round k+1. k+1 The true total profit; ε is the sum of iterative changes in the amount of data shared across all distributed resources; tolK is the preset convergence tolerance. max ψ is the maximum allowed number of iterations. * This is the (approximate) Stackelberg equilibrium strategy that the algorithm ultimately outputs.

[0382] This algorithm constructs a hierarchical distributed computing framework. Its core employs multi-agent deep reinforcement learning as the strategic decision-making engine, where the data market agent learns optimal pricing and security strategies through the Actor-Critic method. At the collaboration level, a distributed consensus algorithm achieves privacy-preserving global information synchronization, and a federated learning mechanism completes distributed prediction model updates. At the optimization level, for multi-timescale operational problems, distributed Brussels bar optimization is used to handle day-ahead uncertainties, model predictive control executes intraday rolling optimization, and differential privacy technology is innovatively embedded into the real-time control loop. Finally, confidential calculation of the entire system's revenue is completed through federated evaluation. This organic integration of algorithms forms the first complete solution capable of simultaneously handling strategic game theory, multi-stage optimization, and privacy protection.

[0383] Further reference Figure 2 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a VPP privacy-preserving energy-data collaborative optimization system. This device embodiment is similar to... Figure 1 The method embodiments shown correspond to those described.

[0384] like Figure 2 As shown, a VPP privacy-preserving energy-data collaborative optimization system in this embodiment includes: an outer game theory module, an inner operation optimization module, a coupling feedback module, and a collaborative solution module;

[0385] The outer game module is used to establish an outer master-slave game model and obtain the strategic parameters output by the outer master-slave game model. The outer master-slave game model includes a follower model and a leader model.

[0386] The core function of this module is to construct an outer master-slave game theory system that includes follower and leader models, clarifying the interaction logic and decision-making objectives between VPP operators and distributed resource owners. The follower model, with the distributed resource owner as the rational decision-making subject, will fully weigh the economic benefits and non-linear privacy costs of data sharing based on the data procurement price set by the VPP operator, determining the optimal data sharing level through utility maximization. The leader model, with the VPP operator as the decision-making core, is guided by maximizing long-term comprehensive net profit, rationally anticipating the decision-making reactions of the distributed resource owner, and jointly optimizing the data procurement price, data security protection investment level, and privacy budget allocation scheme across the day-to-day and real-time stages. Finally, through game equilibrium analysis, it outputs core strategic parameters such as aggregated data level, security investment intensity, and privacy budget allocation ratio, providing basic constraints and key information input for inner-layer operational optimization.

[0387] The inner layer operation optimization module is used to obtain the operation data output by the inner layer multi-time scale operation optimization model based on the strategic parameters output by the outer layer master-slave game model and the inner layer multi-time scale operation optimization model. The inner layer multi-time scale operation optimization model includes a day-ahead sub-Blu-ray optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model.

[0388] This module uses the strategic parameters output by the outer master-slave game model as its core basis, integrating the day-ahead sub-Blu-ray optimization model, the intraday federated learning-model predictive control model, and the real-time differential privacy control model to construct a multi-timescale closed-loop operation optimization system. In the day-ahead phase, relying on the sub-Blu-ray optimization method, combined with the aggregated data level and security investment dynamically adjusted uncertainty set from the outer output, it optimizes market bidding and ground-state scheduling plans, while quantifying the decision-making bias costs brought by the privacy budget. In the intraday phase, the distributed privacy-preserving predictive model is updated and iterated through a federated learning mechanism, and rolling optimization is carried out in conjunction with model predictive control to correct the day-ahead plan and maintain the frequency modulation resource readiness state. In the real-time phase, based on a linearized state-space model, differential privacy noise matching the privacy budget is injected into the control loop to quickly respond to AGC commands and provide high-quality frequency modulation services. By solving the sub-models at each stage, key operational data such as market revenue, privacy operation costs, and resource scheduling commands for each time period are output, accurately reflecting the VPP operational effectiveness under the strategic parameters.

[0389] The coupling feedback module is used to construct a two-way coupling feedback mechanism and establish a Stackelberg equilibrium mechanism by utilizing the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model.

[0390] This module focuses on the deep integration of outer-layer strategic parameters and inner-layer operational data. Its core is the construction of a two-way coupled feedback mechanism and a Stackelberg equilibrium mechanism, forming a closed-loop system of "strategy-operation-feedback-optimization." In the two-way coupled feedback mechanism, the strategic parameters output from the outer layer define the core boundaries of the inner-layer operations—the level of aggregated data determines prediction accuracy, security investment affects decision-making trust, and privacy budget allocation is transformed into operational costs. Meanwhile, the operational data output from the inner layer provides a value benchmark for outer-layer strategic adjustments—market revenue is used to evaluate strategic effectiveness, and privacy costs constrain the direction of strategic optimization. The Stackelberg equilibrium mechanism ensures the optimality and strategy consistency of both sides in the game, meaning that distributed resource owners maximize utility at the equilibrium price, VPP operators maximize profits after anticipating resource party reactions, and the expected data level at the strategic layer remains consistent with the actual data level at the operational layer, thereby ensuring overall system synergy and optimality.

[0391] The collaborative solution module is used to obtain the VPP privacy-preserving energy-data collaborative optimization results by adopting a distributed collaborative solution method based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism.

[0392] This module follows a bidirectional coupled feedback mechanism as its core principle and employs a distributed collaborative solution method guided by physical information to achieve efficient global optimization. It decomposes the original problem into five types of agents: data market, distributed resources, day-ahead scheduling, intraday rolling optimization, and real-time control. Each agent solves sub-problems in parallel. A distributed consensus algorithm is used to achieve privacy-preserving aggregation of resource-side data sharing levels. Federated learning and Monte Carlo simulation are used to achieve a safe assessment of operational benefit gradients. A reinforcement learning (Actor-Critic) mechanism is then used to iteratively update the leader's strategic parameters. Finally, a convergence judgment verifies the Stackelberg equilibrium state, ensuring data market clearing, risk-reward balance, and high strategic-operational synergy. The module outputs the globally optimal solution for VPP data procurement, security protection, privacy protection, and energy scheduling, providing precise guidance for actual VPP operations.

[0393] In summary, the VPP privacy-preserving energy-data collaborative optimization system constructs a master-slave game system with follower and leader models in the outer game module to output strategic parameters. The inner operation optimization module integrates three types of models: day-ahead split-Blu-ray optimization, intraday federated learning-model predictive control, and real-time differential privacy control. Based on the strategic parameters, it generates operational data. The coupling feedback module establishes a two-way coupling feedback and Stackelberg equilibrium mechanism to achieve deep linkage between strategy and operation. The collaborative solution module adopts a distributed multi-agent collaborative method to solve for the global optimal solution. Ultimately, it achieves deep integration and collaboration between data elements and energy resources, significantly improving the overall net profit and resource utilization efficiency of VPP, enhancing the scientific nature of decision-making and risk resistance. Through quantitative modeling, it makes privacy protection and data security controllable and balancing, achieving a dynamic balance between development and security. At the same time, it provides a general optimization paradigm for the deep integration of digital elements and physical systems in the energy internet.

[0394] In this embodiment, the specific processing of a VPP privacy-preserving energy-data collaborative optimization system and its resulting technical effects can be referred to separately. Figure 1 The relevant descriptions of steps S1, S2, S3 and S4 in the corresponding embodiments will not be repeated here.

[0395] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0396] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0397] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0398] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0399] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A VPP privacy-preserving energy-data collaborative optimization method, characterized in that, include: S1. Establish an outer-layer master-slave game model and obtain the strategic parameters output by the outer-layer master-slave game model; S2. Based on the strategic parameters output by the outer master-slave game model and combined with the inner multi-timescale operation optimization model, obtain the operation data output by the inner multi-timescale operation optimization model. S3. Using the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model, a two-way coupling feedback mechanism is constructed, and a Stackelberg equilibrium mechanism is established. S4. Based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism, a distributed collaborative solution method is adopted to obtain the VPP privacy-preserving energy-data collaborative optimization results; The outer master-slave game model includes a follower model and a leader model, and the inner multi-timescale operation optimization model includes a day-ahead split-bar optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model.

2. The VPP privacy-preserving energy-data collaborative optimization method according to claim 1, characterized in that, S1. Establish an outer-layer master-slave game model and obtain the strategic parameters output by the outer-layer master-slave game model, including: Acquire virtual power plant operators and distributed resources in the outer master-slave game; A follower model is constructed, the level of willingness to share data is defined as a decision variable for distributed resources, and the optimal response function of distributed resources is obtained based on privacy costs and data purchase prices. Based on the optimal response function of the distributed resources, the aggregated data level of the virtual power plant operator is obtained by using the relative value of the distributed resource data as weights. A leader model is constructed, defining data procurement price, data security protection investment level, and privacy budget as decision variables for virtual power plant operators, and constructing the objective function of virtual power plant operators in combination with the aggregated data level of the virtual power plant operators; Based on the decision variables and objective function of the virtual power plant operator, set the leader constraint conditions; By combining the optimal reaction function of the distributed resources with the aggregated data level of the virtual power plant operator and the leader constraint, the strategic parameters output by the outer master-slave game model are obtained.

3. The VPP privacy-preserving energy-data collaborative optimization method according to claim 2, characterized in that, The leader constraints include privacy budget allocation constraints and strategic variable boundary constraints; The privacy budget allocation constraint is as follows: e day +e intra +e real ≤e total The boundary constraints of the strategic variables are: 0≤π data ≤π max S min ≤S≤S max e day ≥0.e intra ≥0.e real ≥0 Where, ε day For the privacy budget allocated to the day-ahead phase, ε intra For the privacy budget allocated to the intraday phase, ε real For the privacy budget allocated to the real-time phase, ε total For the total privacy budget, π data For data procurement price, π max S is the upper limit for data procurement price. min S represents the minimum level of investment in data security protection. max To maximize the level of investment in data security protection.

4. The VPP privacy-preserving energy-data collaborative optimization method according to claim 1, characterized in that, S2. Based on the strategic parameters output by the outer master-slave game model and combined with the inner multi-timescale operation optimization model, obtain the operational data output by the inner multi-timescale operation optimization model, including: Based on the strategic parameters output by the outer master-slave game model, the aggregation data level, data security protection investment level, and privacy budget are obtained. The privacy budget includes the day-ahead privacy budget, the intraday privacy budget, and the real-time privacy budget. By leveraging the aggregated data level and the level of investment in data security protection, a set of uncertainties in data-energy coupling is constructed; Based on the uncertainty set of the data-energy coupling and the day-ahead privacy budget combined with physical system and market constraints, a day-ahead sub-Bruker optimization model is constructed; The basic scheduling and bidding plan are obtained using the aforementioned day-ahead Brussels bar optimization model; Federated learning is used to update the global prediction model, and the optimized prediction model is obtained by combining the intraday privacy budget. Based on the aforementioned optimized prediction model, a model prediction and control optimization problem is constructed, and combined with intraday stage-specific constraints, an intraday federated learning-model prediction and control model is constructed. Based on the intraday federated learning-model prediction control model, intraday adjustment instructions are obtained; Based on the data security protection investment level, a communication encryption scheme is determined, and a linearized system state space model is established in conjunction with the real-time privacy budget. Based on the linearized system state-space model, a privacy-preserving output feedback mechanism is established. Based on the privacy protection output feedback mechanism, a control objective and privacy cost function are established, and a real-time differential privacy control model is constructed by combining real-time frequency modulation resource dynamics and allocation constraints. Real-time control commands are obtained using the aforementioned real-time differential privacy control model; Based on the basic scheduling and bidding plan, the intraday adjustment instructions, and the real-time control instructions, the operational data output by the inner multi-timescale operation optimization model is obtained.

5. The VPP privacy-preserving energy-data collaborative optimization method according to claim 4, characterized in that, The physical system and market constraints include power balance constraints, heat / cold power balance constraints, CHP unit operation constraints, energy storage dynamics and operation constraints, heat pump and electric boiler coupling constraints, demand response constraints, ancillary service capacity feasibility constraints, frequency regulation capacity constraints, and resource operation domain coupling constraints. The intraday phase-specific constraints include rolling status update constraints, frequency modulation preparation status maintenance constraints, and indoor temperature dynamic constraints. The real-time frequency regulation resource dynamics and allocation constraints include frequency regulation energy storage dynamic constraints, frequency regulation power allocation constraints, frequency regulation ramp rate constraints, and frequency regulation control end-to-end time delay constraints.

6. The VPP privacy-preserving energy-data collaborative optimization method according to claim 1, characterized in that, S3. Using the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operational optimization model, a two-way coupled feedback mechanism is constructed, and a Stackelberg equilibrium mechanism is established, including: Based on the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model, the bidirectional coupling variables of the virtual power plant master-slave participants are obtained. A bidirectional coupling feedback channel is established based on the bidirectional coupling variables of the virtual power plant master-slave participants. A bidirectional coupling feedback mechanism is constructed using the aforementioned bidirectional coupling feedback channel; The Stackelberg master-slave game subject boundaries are defined based on the bidirectional coupling feedback mechanism, and a Stackelberg equilibrium mechanism is established.

7. The VPP privacy-preserving energy-data collaborative optimization method according to claim 1, characterized in that, S4. Based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism, a distributed collaborative solution method is used to obtain the VPP privacy-preserving energy-data collaborative optimization results, including: Based on the bidirectional coupling feedback mechanism, a multimodal intelligent agent is obtained, which includes a data market intelligent agent, a distributed resource intelligent agent, a day-ahead scheduling intelligent agent, an intraday rolling optimization intelligent agent, and a real-time control intelligent agent. Using the multimodal agent, initialize the multimodal agent strategy, which includes a data market agent strategy, a distributed resource agent strategy, and operational layer agent decision variables; Based on the aforementioned multimodal intelligent agent strategy, a distributed collaborative solution method is adopted to obtain a VPP privacy-preserving energy-data collaborative optimization scheme; The convergence judgment is performed based on the VPP privacy-preserving energy-data collaborative optimization scheme and the Stackelberg equilibrium mechanism to obtain the VPP privacy-preserving energy-data collaborative optimization result.

8. The VPP privacy-preserving energy-data collaborative optimization method according to claim 7, characterized in that, Based on the aforementioned multimodal agent strategy, a distributed collaborative solution method is employed to obtain a VPP privacy-preserving energy-data collaborative optimization scheme, including: The distributed resource agent is updated by performing an outer game agent update based on the multimodal agent strategy to obtain a distributed consensus result. Based on the distributed consensus result and the multimodal agent strategy, the inner layer operation agent is updated for the day-ahead scheduling agent, the intraday rolling optimization agent and the real-time control agent to obtain the inner layer operation agent optimization result; Monte Carlo simulation was performed using the optimization results of the inner layer operating agent to obtain global gradient estimation. Based on the global gradient estimation and the multimodal agent policy, the Actor-Critic method is used to perform leader policy updates on the data market agent to obtain a VPP privacy-preserving energy-data collaborative optimization scheme.

9. A VPP privacy-preserving energy-data collaborative optimization method according to claim 8, characterized in that, The convergence determination is performed based on the VPP privacy-preserving energy-data collaborative optimization scheme and the Stackelberg equilibrium mechanism to obtain the VPP privacy-preserving energy-data collaborative optimization result, including: The convergence residual is obtained using the aforementioned VPP privacy-preserving energy-data collaborative optimization scheme; Determine whether the convergence residual satisfies the convergence judgment condition. If it does, obtain the VPP privacy-preserving energy-data collaborative optimization result based on the VPP privacy-preserving energy-data collaborative optimization scheme combined with the Stackelberg equilibrium mechanism. Otherwise, return to execute the first operation. The first operation is as follows: updating the outer game agent of the distributed resource agent according to the multimodal agent strategy, and obtaining the distributed consensus result.

10. A VPP privacy-preserving energy-data collaborative optimization system, implementing the method as described in any one of claims 1-9, characterized in that, include: The module consists of an outer game theory module, an inner operation optimization module, a coupling feedback module, and a collaborative solution module. The outer game module is used to establish an outer master-slave game model and obtain the strategic parameters output by the outer master-slave game model. The outer master-slave game model includes a follower model and a leader model. The inner layer operation optimization module is used to obtain the operation data output by the inner layer multi-time scale operation optimization model based on the strategic parameters output by the outer layer master-slave game model and the inner layer multi-time scale operation optimization model. The inner layer multi-time scale operation optimization model includes a day-ahead sub-Blu-ray optimization model, an intraday federated learning-model prediction control model, and a real-time differential privacy control model. The coupling feedback module is used to construct a two-way coupling feedback mechanism and establish a Stackelberg equilibrium mechanism by utilizing the strategic parameters output by the outer master-slave game model and the operational data output by the inner multi-timescale operation optimization model. The collaborative solution module is used to obtain the VPP privacy-preserving energy-data collaborative optimization results by adopting a distributed collaborative solution method based on the bidirectional coupling feedback mechanism and the Stackelberg equilibrium mechanism.

Images