Multi-wind farm bidding strategy cooperation optimization method and system based on differential privacy consensus
By employing a collaborative optimization method for bidding strategies in multiple wind farms using differential privacy consensus, and utilizing high-resolution integrated Kalman filtering and Gaussian process modeling, combined with Laplace noise and relaxation factors, the method addresses the issues of privacy protection and strategy theft prevention in collaborative optimization of multiple wind farms, thereby improving the economic operation and secure communication capabilities of wind farms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-10
AI Technical Summary
In existing technologies, the collaborative optimization mode of multiple wind farms has problems such as poor robustness to single-point failures, high sensitivity of centralized information processing, and insufficient privacy protection. It is difficult to balance the needs of privacy protection and collaborative optimization, which affects the anti-theft capability of wind farm bidding strategies and economic operation capability.
A collaborative optimization method for bidding strategies among wind farms based on differential privacy consensus is adopted. By modeling the uncertainty of wind power and spot price through high-resolution integrated Kalman filtering and Gaussian process, a smart agent for optimizing bidding strategies of wind farms is constructed. Laplace noise is applied in the distributed communication framework, and a relaxation factor is introduced for strategy aggregation. A full-process differential privacy budget monitoring mechanism is constructed to realize collaborative optimization among wind farms.
It has improved the anti-theft capabilities of wind farm bidding strategies, enhanced secure communication performance, improved the economic operation level and risk control capabilities of wind farms in the electricity spot market, and adapted to the high-frequency trading needs of the electricity spot market.
Smart Images
Figure CN122366748A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of renewable energy development and utilization technology, specifically relating to a collaborative optimization method and system for bidding strategies of multiple wind farms. Background Technology
[0002] As the proportion of new energy in the power system gradually increases, the role of the electricity spot market in ensuring power supply and promoting the consumption of new energy is becoming increasingly prominent. However, the randomness of spot market price fluctuations poses a severe challenge to the economic viability of wind farm bidding. Therefore, it is necessary to combine high-precision power prediction technology with bidding strategy optimization technology that considers market uncertainty to improve the risk control level and proactive profitability of wind farms participating in electricity spot market transactions. Traditional local optimization methods for wind farms are limited by the lack of local scenarios and data information, which restricts the generalization of bidding optimization models. Therefore, how to achieve collaborative optimization of bidding strategies among multiple wind farms while taking into account the privacy and confidentiality attributes of each wind farm's bidding strategy is the key to improving the profitability of each wind farm. There are two major challenges in this research area: 1) Existing centralized collaboration models have problems such as poor robustness to single-point failures and high sensitivity of centralized information processing, making it difficult to adapt to the needs of efficient privacy protection in multi-wind farm collaboration; 2) Existing distributed privacy-protected collaboration models often require attenuated noise modulus to achieve global consensus, resulting in the dissipation of privacy budget during communication collaboration, making it difficult to balance model convergence performance and privacy protection strength. Therefore, how to construct a consensus mechanism for distributed collaborative optimization of multiple wind farms, and improve the anti-theft capability of wind farm bidding strategies, the economic operation capability and secure communication level of wind farms participating in the spot market, without affecting the bidding and settlement revenue of wind farms, is the current research focus and challenge. Summary of the Invention
[0003] Purpose of the invention: To address the pain points mentioned above regarding balancing privacy protection and collaborative optimization among multiple wind farms, the purpose of this invention is to provide a collaborative optimization method and system for bidding strategies of multiple wind farms based on differential privacy consensus, thereby synergistically improving the anti-theft capabilities and secure communication performance of allied wind farms in bidding.
[0004] Technical Solution: To achieve the above-mentioned objectives, this invention proposes a collaborative optimization method for multi-wind farm bidding strategies based on differential privacy consensus, comprising the following steps:
[0005] (1) Input multi-source meteorological forecast information and power market source-load forecast information of the target wind farm, and use high-resolution integrated Kalman filter and Gaussian process to model the uncertainty of wind power and spot price as the basis for bidding strategy optimization;
[0006] (2) Using wind power uncertainty, spot price uncertainty and time information as state variables, construct a wind farm bidding strategy optimization agent, output the day-ahead bidding strategy of wind farm, and use incremental near-end strategy to optimize the agent;
[0007] (3) Considering the balance between improving the revenue of wind farms and protecting the privacy of bidding strategies, a fully distributed communication framework based on differential privacy consensus is constructed. By applying Laplace noise to the bidding strategies of wind farms, the anti-theft capability of the bidding strategy information of wind farms is improved. A relaxation factor is introduced to construct a differential aggregation mechanism for bidding strategies, so that each wind farm can simultaneously adopt the bidding strategies of other wind farms and balance the strategy gain obtained by training with local data.
[0008] (4) Construct a full-process differential privacy budget monitoring mechanism. By monitoring the total differential privacy budget of the entire communication process, determine the status of key parameter sequence rights, thereby realizing end-to-end multi-wind farm bidding strategy collaborative optimization.
[0009] Furthermore, in step (1), the uncertainty of wind power fluctuations is modeled using a high-resolution integrated Kalman filter, specifically including:
[0010] Input low-resolution multi-source weather forecast fields and meteorological reanalysis field ,in and These represent the width and height of the weather forecast field, respectively. The number of multi-source weather forecasts;
[0011] Multiple random sampling points are generated using the random weighting method. , To integrate the number of filters, a high-resolution random sample set is obtained using spatial quadratic interpolation.
[0012]
[0013] in and These are low-resolution and high-resolution meteorological fields, respectively. For high-resolution time intervals, and These are the pixel velocity matrix and acceleration matrix during the simulation process, respectively. Represent a weighted matrix;
[0014] The derivation of high-resolution random sampling sets is as follows:
[0015]
[0016] in, Let be the covariance matrix at time t. and These are respectively the high-resolution weather forecast field and the high-resolution ensemble mean field;
[0017] Using the Kalman filter method, the optimal high-resolution meteorological estimation field The derivation is as follows:
[0018]
[0019] in, To integrate the weighted analytical fields, For the integration weights that need to be determined, and These are the observation matrix and the perturbation matrix, respectively;
[0020] By analyzing high-resolution reanalysis fields With the best high-resolution meteorological estimation field The spatiotemporal deviation pattern was analyzed, and the integrated weighting factor was determined using the least squares method. :
[0021]
[0022] in, The window length for the integrated Kalman filter backtracking.
[0023] Furthermore, in step (1), modeling the uncertainty of spot prices using a Gaussian process specifically includes:
[0024] The uncertainty of spot prices is modeled as a Gaussian function of the supply and demand vector disclosed by the market:
[0025]
[0026] in, and These are the vectors for spot price and market supply and demand, respectively. , , and These represent the regional load, regional wind power, regional centralized photovoltaic, and regional distributed photovoltaic forecast information disclosed in the spot market, respectively. Represents a mapping function. Let represent the applied observation noise; then the joint Gaussian distribution model of the spot price uncertainty and the mapping function is as follows:
[0027]
[0028] in, and These represent historical supply and demand information in the spot market, as well as supply and demand forecasts disclosed recently. This represents the covariance matrix, where each element records any two supply and demand relationship vectors. The covariance is expressed as The covariance matrix is estimated using radial basis function kernels, thus yielding a probability prediction of spot prices characterized by a normal distribution. :
[0029] .
[0030] Furthermore, in step (2), a wind farm bidding strategy optimization agent is constructed, and the incremental near-end strategy is used to optimize the solution agent, specifically including:
[0031] Configure the action as a strategy to optimize the number of wind farm applications. ,in and These are the original wind power forecast and the adjusted reported volume, respectively. Let be the action function at time t. The maximum allowable declaration volume adjustment range, Set the installed capacity of the wind farm; set state variables. As follows:
[0032]
[0033] in, This represents the original wind power prediction from time t+1 to time t+f. This represents the standard deviation of the original wind power prediction; The time step for the proposed bid; and These are the intraday time sine values and the year-round date sine values, respectively.
[0034] The reward function takes into account both the spot settlement revenue of wind farms and the penalty for energy imbalance caused by reporting deviations, and is set as follows:
[0035]
[0036] in, For the reward function, Let be the spot price at time t. This indicates the discrepancy between the reported number of wind farms and the actual number of wind farms connected to the grid. and These represent the auxiliary regulation coefficients for market electricity purchases and the consumption coefficients for surplus electricity, respectively. This indicates the penalty value for malicious bidding. This indicates the threshold for penalties for malicious bidding. This is a function for calculating spot settlement revenue. Represents the ReLU function;
[0037] Each wind farm constructs an agent-based solution method based on near-end policy optimization using local data. The objective function for optimization is defined as:
[0038]
[0039] in, Indicates by parameters Indicated and These represent the parameters to be updated. and original parameters The policy function, policy The update coefficient, This means clipping the original gradient to... The clipping function, This is the gradient clipping factor. Represents the time difference function. For the dominant function, To reward the weakening factor, Indicates by strategy The resulting value function Represents the longest step size for interaction with the environment;
[0040] Taking into account the time-varying characteristics of the spot market trading environment, the agent parameters are updated using layer-by-layer near-end replay and nearest-neighbor samples:
[0041]
[0042] in, and These are the bidding strategies before and after the update. and These are the gradient update step size and the near-end policy regularization step size, respectively. Indicates the first The activation value of the layer, Indicates the nearest sample. Indicates the number of network layers. Used to measure the deviation of the bidding model activation value before and after the update.
[0043] Furthermore, in step (3), constructing a multi-wind farm distributed communication framework based on differential privacy consensus specifically includes:
[0044] Based on the privacy protection requirements during the collaborative bidding data exchange process of wind farms, Laplace noise is applied to the bidding strategies to be used for collaborative bidding to enhance the anti-theft capability of the bidding strategies:
[0045]
[0046] in, and They are respectively The encrypted iteration rounds and the original bidding strategy; express Laplace noise applied in each iteration round, The scale representing Laplace noise;
[0047] Each wind farm freely exchanges encrypted bidding strategies with other wind farms. A relaxation difference algorithm is introduced to aggregate the bidding strategies of multiple wind farms. The aggregated bidding strategies are then adaptively aggregated with the bidding strategies optimized from local data. The adaptively aggregated strategy is used as the initial value for the next iteration. The bidding strategy aggregation method is expressed as follows:
[0048]
[0049] in, Indicates the relaxation factor. This represents the local training strategy balance factor. This represents the weighting factor for the differences between the various strategies. For the first A set of bidding strategies collected by each wind farm. Indicates the first The first wind farm station The gradient of descent in each iteration round; in the communication of each iteration round, the relaxation factor and the local policy balance factor satisfy... , as well as The Laplace noise scale satisfies .
[0050] Furthermore, in step (4), the privacy budget quantification monitoring mechanism for differential privacy consensus specifically includes:
[0051] Privacy budgeting for single-step communication rounds ,in For the first The sensitivity of round-based communication; thus enabling the measurement of the first round. Upper bound of the deviation between the encryption of each communication round and the original bidding strategy:
[0052]
[0053]
[0054] in, Indicates weighting factor The aggregated results then yield the following average differential privacy budget estimate for the entire process:
[0055]
[0056] in, To the total number of iterations, express The supremum;
[0057] Privacy Budget Based on Average Difference Determine the relaxation factor in each communication round. Gradient update factor and Laplace noise scale .
[0058] A collaborative optimization system for bidding strategies across multiple wind farms based on differential privacy consensus includes:
[0059] The data preparation module is used to input multi-source meteorological forecast information and source-load forecast information released by the electricity market for the target wind farm. It uses high-resolution integrated Kalman filtering and Gaussian process to model the uncertainty of wind power and spot price, which serves as the decision basis for optimizing the bidding strategy.
[0060] The bidding strategy construction module is used to construct a wind farm bidding strategy optimization agent by taking wind power uncertainty, spot price uncertainty and time information as state variables. It is used to output the day-ahead bidding strategy of wind farms and use incremental near-end strategy optimization to solve the agent.
[0061] A distributed communication framework construction module is used to consider the balance between improving wind farm revenue and protecting the privacy of bidding strategies. It constructs a fully distributed communication framework based on differential privacy consensus. By applying Laplace noise to the wind farm bidding strategies, it improves the anti-theft capability of wind farm bidding strategy information. It introduces a relaxation factor to construct a differential aggregation mechanism for bidding strategies, so that each wind farm can simultaneously adopt the bidding strategies of other wind farms and balance the strategy gains obtained by training with local data.
[0062] The collaboration parameter determination module is used to build a privacy budget quantification and monitoring mechanism for differential privacy consensus. By monitoring the total differential privacy budget of the entire communication process, it determines the weighting status of key parameter sequences, thereby realizing end-to-end collaborative optimization of bidding strategies for multiple wind farms.
[0063] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described above.
[0064] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described above.
[0065] The present invention also provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described above.
[0066] Beneficial effects: (1) The high-resolution integrated Kalman filter method of this invention can accurately estimate the trend and uncertainty of wind power fluctuations, effectively transform low-resolution original multi-source weather forecasts into high-resolution wind power prediction results, and successfully adapt to the needs of high-frequency trading in the electricity spot market; (2) This invention proposes a collaborative optimization framework for multi-wind farm bidding strategies based on differential privacy consensus, ensuring that wind farms participating in this framework can improve their bidding risk management capabilities and spot settlement benefits through information sharing of bidding strategies among multiple wind farms without disclosing their own bidding strategies; (3) This invention constructs a full-process monitoring mechanism for differential privacy budgets, effectively improving the anti-theft capability of wind farm bidding strategies in each communication round to prevent the dissipation problem of differential privacy budgets; (4) This invention fits the actual needs of electricity spot market trading, significantly improving the secure communication level between multi-wind farm collaborative optimizations without requiring additional hardware modifications, providing decision support for mitigating the trading risks caused by the uncertainty of spot price fluctuations, and improving the economic operation level of wind farms in the electricity spot market, which has a good promoting effect. Attached Figure Description
[0067] Figure 1 This invention presents a high-resolution integrated Kalman filtering method.
[0068] Figure 2 This invention presents an incremental proximal strategy optimization method.
[0069] Figure 3 This invention presents a collaborative optimization framework for wind farm bidding strategies based on differential privacy consensus.
[0070] Figure 4 This is the high-resolution integrated Kalman filter result in the embodiment of the present invention;
[0071] Figure 5 This is the result of the optimized bidding strategy for wind farms in this embodiment of the invention;
[0072] Figure 6 This is a comparative analysis of bidding performance under different privacy protection levels in the embodiments of the present invention. Detailed Implementation
[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0074] This invention proposes a collaborative optimization method for multi-wind farm bidding strategies based on differential privacy consensus, comprising the following steps:
[0075] Step S1: Input multi-source meteorological forecast information and power market source-load forecast information for the target wind farm. Use high-resolution integrated Kalman filtering and Gaussian process to model the uncertainty of wind power and spot price as the basis for bidding strategy optimization.
[0076] Step S2: Using wind power uncertainty, spot price uncertainty, and time information as state variables, construct a wind farm bidding strategy optimization agent, output the day-ahead bidding strategy for wind farms, and use incremental near-end strategy optimization to solve the agent.
[0077] Step S3: Considering the balance between improving wind farm revenue and protecting the privacy of bidding strategies, a fully distributed communication framework based on differential privacy consensus is constructed. By applying Laplace noise to the wind farm bidding strategies, the anti-theft capability of wind farm bidding strategy information is improved. A relaxation factor is introduced to construct a differential aggregation mechanism for bidding strategies, enabling each wind farm to simultaneously adopt the bidding strategies of other wind farms and balance the strategy gains obtained through local data training.
[0078] Step S4: Construct a full-process differential privacy budget monitoring mechanism. By monitoring the total differential privacy budget of the entire communication process, determine the status of key parameter sequence rights, thereby achieving end-to-end multi-wind farm bidding strategy collaborative optimization.
[0079] like Figure 1 As shown, in step S1, regarding the modeling of wind power uncertainty, a low-resolution multi-source weather forecast field is input. and meteorological reanalysis field ,in and These represent the width and height of the weather forecast field, respectively. This represents the number of multi-source weather forecasts. First, multiple random sampling points are generated using a random weighting method. , To integrate the number of filters, a high-resolution random sample set is obtained using spatial quadratic interpolation.
[0080]
[0081] in, and These are low-resolution and high-resolution meteorological fields, respectively. For high-resolution time intervals, and These are the pixel velocity matrix and acceleration matrix during the simulation process, respectively. Let represent a weighting matrix. Secondly, the high-resolution random sample set can be derived as:
[0082]
[0083] in, Let be the covariance matrix at time t. and These are the high-resolution weather forecast field and the high-resolution ensemble mean field, respectively.
[0084] Next, using the Kalman filtering method, the optimal high-resolution meteorological estimation field is obtained. This can be deduced as:
[0085]
[0086] in, To integrate the weighted analytical fields, For the integration weights that need to be determined, and These are the observation matrix and the perturbation matrix, respectively.
[0087] Finally, by analyzing the high-resolution reanalysis field With the best high-resolution meteorological estimation field The spatiotemporal deviation pattern was analyzed, and the integrated weighting factor was determined using the least squares method. :
[0088]
[0089] in, This is the window length for the integrated Kalman filter backtracking. When the optimal estimated field is obtained... The local wind power fluctuation trend can then be obtained through the wind speed-power conversion model:
[0090]
[0091] in, This represents the wind speed-wind power conversion model. and These represent the power prediction result and uncertainty estimation result for the i-th wind farm, respectively. Its coordinates, This refers to the time step for the proposed bid.
[0092] Regarding the understanding of spot price trends, considering that spot prices are mainly driven by the supply and demand relationship in the spot market, the uncertainty of spot prices can be modeled as a Gaussian function of the supply and demand vector disclosed by the market:
[0093]
[0094] in, and These are the vectors for spot price and market supply and demand, respectively. , , and These represent the regional load, regional wind power, regional centralized photovoltaic, and regional distributed photovoltaic forecast information disclosed in the spot market, respectively. Represents a mapping function. This represents the applied observation noise. Therefore, the joint Gaussian distribution of the spot price uncertainty and the mapping function can be modeled as:
[0095]
[0096] in, and These represent historical supply and demand information in the spot market, as well as supply and demand forecasts disclosed recently. This represents the covariance matrix, where each element records any two supply and demand relationship vectors. The covariance is expressed as Furthermore, the covariance matrix can be estimated using the radial basis function kernel, thereby obtaining the probability prediction result of spot prices characterized by a normal distribution. :
[0097]
[0098] like Figure 2 As shown, in step S2, the agent is optimized using an incremental proximal strategy, specifically including:
[0099] Configure the action as a strategy to optimize the number of wind farm applications. ,in Indicates the action at time t. and These are the original wind power forecast and the adjusted reported volume, respectively. The maximum allowable declaration volume adjustment range, Set the installed capacity of the wind farm. Set state variables. As follows:
[0100]
[0101] in, and These are the sine values for intraday times and year-round dates, respectively. The calculation methods for the intraday times and year-round dates are as follows:
[0102]
[0103] in, and These are the hour and date positions of the bidding start time, respectively.
[0104] In addition, the reward function takes into account both the spot settlement revenue of wind farms and the energy imbalance penalty caused by reporting deviations, and is set as follows:
[0105]
[0106] in, For the reward function, Let be the spot price at time t. This indicates the discrepancy between the number of wind farms reported and the actual number of wind farms connected to the grid. and These represent the auxiliary regulation factors for electricity purchases from the market and the consumption factors for surplus electricity, respectively. This indicates the penalty value for malicious bidding. This indicates the threshold for penalties for malicious bidding. This is a function for calculating spot settlement revenue. This represents the ReLU function.
[0107] Next, each wind farm constructs an agent-based solution method based on local data and near-end policy optimization. The objective function for optimization is defined as:
[0108]
[0109] in, Indicates by parameters Representation strategy The update coefficient, and These represent the parameters to be updated. and original parameters The policy function is represented. This means clipping the original gradient to... The clipping function, This is the gradient clipping factor. Represents the time difference function. For the dominant function, To reward the weakening factor in order to balance immediate and long-term returns, Indicates by strategy The resulting value function Represents the longest step size for interaction with the environment.
[0110] Ultimately, considering the time-varying nature of the spot market trading environment, historically accumulated market environment and bidding strategy samples are difficult to adapt to real-time bidding and trading needs. Therefore, it is necessary to combine the nearest sample... The original bidding model is incrementally updated to ensure it can efficiently adapt to changes in the spot market trading environment. A layer-by-layer near-end replay method is used to incrementally update the agent parameters.
[0111]
[0112] in, and These are the bidding strategies before and after the update. and These are the gradient update step size and the proximal policy regularization step size, respectively. Indicates the first The activation value of the layer, Indicates the number of network layers. Used to measure the deviation of the bidding model activation value before and after the update.
[0113] like Figure 3 As shown, in step S3, a fully distributed communication framework based on differential privacy consensus is constructed to achieve collaborative optimization of bidding strategies for multiple wind farms, specifically including:
[0114] First, based on the privacy protection requirements during the collaborative bidding data exchange process of wind farms, Laplace noise is applied to the bidding strategies to be used for collaborative bidding to enhance the anti-theft capability of the bidding strategies:
[0115]
[0116] in, and They are respectively After encryption of the iteration rounds and the original bidding strategy, express Laplace noise applied in each iteration round, The scale representing Laplace noise.
[0117] Secondly, each wind farm can freely exchange encrypted bidding strategies with other wind farms. A relaxation difference algorithm is introduced to aggregate the bidding strategies of multiple wind farms. The aggregated bidding strategies are then adaptively aggregated with bidding strategies optimized from local data, thereby balancing performance in terms of local training, privacy protection, and global consensus convergence. This strategy is then used as the initial value for the next iteration. The bidding strategy aggregation method can be expressed as:
[0118]
[0119] in, Indicates the relaxation factor. This represents the local training strategy balance factor. This represents the weighting factor for the differences between the various strategies. For the first A set of bidding strategies collected by each wind farm. Indicates the first The first wind farm station The gradient of descent in each iteration. In each iteration's communication, the relaxation factor and the local policy balance factor need to satisfy... , as well as The Laplace noise scale also needs to meet the requirements. .
[0120] In step S4, a full-process differential privacy budget monitoring mechanism is constructed, specifically including:
[0121] First, construct a privacy budget for each single-step communication round. ,in For the first The sensitivity of each round of communication. This allows us to measure the sensitivity of the first round. Upper bound of the deviation between the encryption of each communication round and the original bidding strategy:
[0122]
[0123]
[0124] in, Indicates weighting factor The aggregation results. Therefore, the overall average differential privacy budget can be estimated as:
[0125]
[0126] in, To the total number of iterations, express The upper bound. Through rigorous monitoring of the average differential privacy budget for each round. The relaxation factor can be determined in each communication round. Gradient update factor and Laplace noise scale .
[0127] This provides quantitative calculation results for privacy-preserving communications across the entire framework by accurately estimating round-by-round and cumulative differential privacy budgets.
[0128] To verify the performance of the method proposed in this invention, the following experiments were conducted in this embodiment. The data used for testing in this embodiment comes from the Australian Energy Management Market, which coordinates and schedules spot electricity trading in southeastern Australia. Each wind farm is required to report its bidding results for the next day (288 points) every 5 minutes, and the Energy Management Market clears bids in 30-minute increments. Since June 2024, the Australian Energy Management Market has allowed multiple wind farms to become energy integration providers through approval, thereby legally optimizing bidding strategies through collaboration among wind farms.
[0129] This invention employs six wind farms from five administrative states in Australia for collaborative optimization of bidding strategies. Multi-source weather forecast data utilizes the global high-resolution ensemble forecast (HRES) published by the European Centre for Medium- and Long-Term Weather Forecasts (ECMWF) and the Global Weather Forecast (GFS) product published by the National Oceanic and Atmospheric Administration (NOAA). The original temporal resolution for both is 1 hour, and the spatial resolution is 0.25°.
[0130] To highlight the superiority of the fully distributed collaborative framework proposed in this invention, comparative experiments were conducted using both independent training and centralized training methods in this embodiment. The sample set from January 2023 to December 2024 was selected as the training set, and the sample set for the entire year of 2025 was selected as the test set. The parameters of the collaborative bidding strategy optimization framework are shown in Table 1. Based on the parameter settings in Table 1, the verification results of the collaborative bidding strategy for six wind farms are shown in Table 2. Figure 4 The optimal weather forecast estimation results obtained based on the high-resolution integrated Kalman filter of this invention are presented, and it is found that this invention can effectively integrate the advantages of multi-source weather forecasts and obtain more reliable weather forecast data. Figure 5 The results of the bidding strategy at wind farm 1 are presented, demonstrating that the present invention can effectively identify high price risks and adopt a robust strategy to protect returns, thereby improving the profitability of wind farms participating in the spot market. Figure 6 The study demonstrates the bidding settlement revenue levels of wind farms under different privacy protection levels. When the privacy budget is small (indicating a high level of privacy protection), the performance of the bidding strategy model degrades rapidly. However, when the privacy budget exceeds the given model setting, the bidding revenue is essentially the same as the revenue obtained from free data communication. This indicates that the proposed method can guarantee users' key privacy information without jeopardizing participants' revenue. Furthermore, it proves that the information theft prevention capability is ensured throughout the entire process and does not decrease with each communication round, thereby improving the secure communication capability of distributed collaboration among multiple wind farms.
[0131] Table 1. Parameter Settings for the Wind Farm Collaborative Bidding Framework Based on Differential Privacy Consensus
[0132]
[0133] Table 2 Results of collaborative bidding for wind farms
[0134] method Current level Independent training Intensive training This invention Wind farm station 1 27.85 30.74 31.58 34.83 Wind farm station 2 54.77 60.19 62.01 68.90 Wind farm station 3 40.88 44.72 43.39 45.37 Wind farm station 4 39.86 40.70 41.69 45.67 Wind farm station 5 56.40 57.94 57.65 59.11 Wind farm station 6 23.41 25.99 28.32 32.81
[0135] In summary, the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus designed in this invention firstly obtains accurate wind power uncertainty estimation results by processing multi-source weather forecast fields using high-resolution integrated Kalman filtering; secondly, it models the uncertainty of electricity spot prices using the Gaussian process method; furthermore, it constructs a deep reinforcement learning agent based on incremental near-end policy optimization to optimize bidding strategies using local data from wind farms; next, it builds a multi-wind farm bidding strategy collaborative optimization framework based on differential privacy consensus, using Laplace noise to protect the privacy of participants' bidding strategies, and introducing a relaxation factor during strategy aggregation to ensure the accuracy of strategy solution, thereby improving the privacy protection capabilities of participants without affecting their returns; finally, it constructs a differential privacy budget monitoring mechanism to determine the overall privacy protection strength by monitoring the divergence of differential privacy. This invention can guide wind farms to prevent price risks, enhance trading initiative, and provide decision support for the economic operation of wind farms in the spot market.
[0136] Based on the same technical concept as the method embodiments, the present invention also provides a multi-wind farm bidding strategy collaborative optimization system based on differential privacy consensus, including:
[0137] The data preparation module is used to input multi-source meteorological forecast information and source-load forecast information released by the electricity market for the target wind farm. It uses high-resolution integrated Kalman filtering and Gaussian process to model the uncertainty of wind power and spot price, which serves as the decision basis for optimizing the bidding strategy.
[0138] The bidding strategy construction module is used to construct a wind farm bidding strategy optimization agent by taking wind power uncertainty, spot price uncertainty and time information as state variables. It is used to output the day-ahead bidding strategy of wind farms and use incremental near-end strategy optimization to solve the agent.
[0139] A distributed communication framework construction module is used to consider the balance between improving wind farm revenue and protecting the privacy of bidding strategies. It constructs a fully distributed communication framework based on differential privacy consensus. By applying Laplace noise to the wind farm bidding strategies, it improves the anti-theft capability of wind farm bidding strategy information. It introduces a relaxation factor to construct a differential aggregation mechanism for bidding strategies, so that each wind farm can simultaneously adopt the bidding strategies of other wind farms and balance the strategy gains obtained by training with local data.
[0140] The collaboration parameter determination module is used to build a privacy budget quantification and monitoring mechanism for differential privacy consensus. By monitoring the total differential privacy budget of the entire communication process, it determines the weighting of key parameter sequences, thereby achieving a balance between collaborative optimization of multi-wind farm bidding strategies and privacy protection performance.
[0141] It should be understood that the multi-wind farm bidding strategy collaborative optimization system based on differential privacy consensus in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.
[0142] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described above.
[0143] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described above.
Claims
1. A collaborative optimization method for bidding strategies of multiple wind farms based on differential privacy consensus, characterized in that, Includes the following steps: (1) Input multi-source meteorological forecast information and power market source-load forecast information of the target wind farm, and use high-resolution integrated Kalman filter and Gaussian process to model the uncertainty of wind power and spot price as the basis for bidding strategy optimization; (2) Using wind power uncertainty, spot price uncertainty and time information as state variables, construct a wind farm bidding strategy optimization agent to output the day-ahead bidding strategy of wind farms, and use incremental near-end strategy to optimize the agent; (3) Considering the balance between improving the revenue of wind farms and protecting the privacy of bidding strategies, a fully distributed communication framework based on differential privacy consensus is constructed. By applying Laplace noise to the bidding strategies of wind farms, the anti-theft capability of the bidding strategy information of wind farms is improved. A relaxation factor is introduced to construct a differential aggregation mechanism for bidding strategies, so that each wind farm can simultaneously adopt the bidding strategies of other wind farms and balance the strategy gain obtained by training with local data. (4) Construct a privacy budget quantitative monitoring mechanism for differential privacy consensus. By monitoring the total differential privacy budget of the entire communication process, determine the status of key parameter sequence confirmation, thereby realizing end-to-end multi-wind farm bidding strategy collaborative optimization.
2. The method according to claim 1, characterized in that, In step (1), the uncertainty of wind power fluctuations is modeled using a high-resolution integrated Kalman filter, specifically including: Input low-resolution multi-source weather forecast fields and meteorological reanalysis field ,in and These represent the width and height of the weather forecast field, respectively. The number of multi-source weather forecasts; Multiple random sampling points are generated using the random weighting method. , To integrate the number of filters, a high-resolution random sample set is obtained using spatial quadratic interpolation. in and These are low-resolution and high-resolution meteorological fields, respectively. For high-resolution time intervals, and These are the pixel velocity matrix and acceleration matrix during the simulation process, respectively. Represent a weighted matrix; The derivation of high-resolution random sampling sets is as follows: in, Let be the covariance matrix at time t. and These are respectively the high-resolution weather forecast field and the high-resolution ensemble mean field; Using the Kalman filter method, the optimal high-resolution meteorological estimation field The derivation is as follows: in, To integrate the weighted analytical fields, For the integration weights that need to be determined, and These are the observation matrix and the perturbation matrix, respectively; By analyzing high-resolution reanalysis fields With the best high-resolution meteorological estimation field The spatiotemporal deviation pattern was analyzed, and the integrated weighting factor was determined using the least squares method. : in, The window length for the integrated Kalman filter backtracking.
3. The method according to claim 2, characterized in that, In step (1), the uncertainty of spot prices is modeled using a Gaussian process, specifically including: The uncertainty of spot prices is modeled as a Gaussian function of the supply and demand vector disclosed by the market: in, and These are the vectors for spot price and market supply and demand, respectively. , , and These represent the regional load, regional wind power, regional centralized photovoltaic, and regional distributed photovoltaic forecast information disclosed in the spot market, respectively. Represents a mapping function. Let represent the applied observation noise; then the joint Gaussian distribution model of the spot price uncertainty and the mapping function is as follows: in, and These represent historical supply and demand information in the spot market, as well as supply and demand forecasts disclosed recently. This represents the covariance matrix, where each element records any two supply and demand relationship vectors. The covariance is expressed as The covariance matrix is estimated using radial basis function kernels, thus yielding a probability prediction of spot prices characterized by a normal distribution. : 。 4. The method according to claim 3, characterized in that, In step (2), a wind farm bidding strategy optimization agent is constructed, and the incremental near-end strategy is used to optimize and solve the agent, specifically including: Configure the action as a strategy to optimize the number of wind farm applications. ,in and These are the original wind power forecast and the adjusted reported volume, respectively. Let be the action function at time t. The maximum allowable declaration volume adjustment range, Set the installed capacity of the wind farm; set state variables. As follows: in, This represents the original wind power prediction from time t+1 to time t+f. This represents the standard deviation of the original wind power prediction; The time step for the proposed bid; and These are the intraday time sine values and the year-round date sine values, respectively. The reward function takes into account both the spot settlement revenue of wind farms and the penalty for energy imbalance caused by reporting deviations, and is set as follows: in, For the reward function, Let be the spot price at time t. This indicates the discrepancy between the reported number of wind farms and the actual number of wind farms connected to the grid. and These represent the auxiliary regulation coefficients for market electricity purchases and the consumption coefficients for surplus electricity, respectively. This indicates the penalty value for malicious bidding. This indicates the threshold for penalties for malicious bidding. This is a function for calculating spot settlement revenue. Represents the ReLU function; Each wind farm constructs an agent-based solution method based on near-end policy optimization using local data. The objective function for optimization is defined as: in, Indicates by parameters Indicated and These represent the parameters to be updated. and original parameters The policy function, policy The update coefficient, This means clipping the original gradient to... The clipping function, The gradient clipping factor is... Represents the time difference function. For the dominant function, To reward the weakening factor, Indicates by strategy The resulting value function Represents the longest step size for interaction with the environment; Taking into account the time-varying characteristics of the spot market trading environment, the agent parameters are updated using layer-by-layer near-end replay and nearest-neighbor samples: in, and These are the bidding strategies before and after the update. and These are the gradient update step size and the near-end policy regularization step size, respectively. Indicates the first The activation value of the layer, Indicates the nearest sample. Indicates the number of network layers. Used to measure the deviation of the bidding model activation value before and after the update.
5. The method according to claim 1, characterized in that, In step (3), a distributed communication framework for multiple wind farms based on differential privacy consensus is constructed, specifically including: Based on the privacy protection requirements during the collaborative bidding data exchange process of wind farms, Laplace noise is applied to the bidding strategies to be used for collaborative bidding to enhance the anti-theft capability of the bidding strategies: in, and They are respectively The encrypted iteration rounds and the original bidding strategy; express Laplace noise applied in each iteration round, The scale representing Laplace noise; Each wind farm freely exchanges encrypted bidding strategies with other wind farms. A relaxation difference algorithm is introduced to aggregate the bidding strategies of multiple wind farms. The aggregated bidding strategies are then adaptively aggregated with the bidding strategies optimized from local data. The adaptively aggregated strategy is used as the initial value for the next iteration. The bidding strategy aggregation method is expressed as follows: in, Indicates the relaxation factor. This represents the local training strategy balance factor. This represents the weighting factor for the differences between the various strategies. For the first A set of bidding strategies collected by each wind farm. Indicates the first The first wind farm station The gradient of descent in each iteration round; in the communication of each iteration round, the relaxation factor and the local policy balance factor satisfy... , as well as The Laplace noise scale satisfies .
6. The method according to claim 5, characterized in that, In step (4), the privacy budget quantification monitoring mechanism for differential privacy consensus specifically includes: Privacy budgeting for single-step communication rounds ,in For the first The sensitivity of round-based communication; thus enabling the measurement of the first round. Upper bound of the deviation between the encryption of each communication round and the original bidding strategy: in, Indicates weighting factor The aggregated results then yield the following average differential privacy budget estimate for the entire process: in, To the total number of iterations, express The supremum; Privacy Budget Based on Average Difference Determine the relaxation factor in each communication round. Gradient update factor and Laplace noise scale .
7. A collaborative optimization system for bidding strategies of multiple wind farms based on differential privacy consensus, characterized in that, include: The data preparation module is used to input multi-source meteorological forecast information and source-load forecast information released by the electricity market for the target wind farm. It uses high-resolution integrated Kalman filtering and Gaussian process to model the uncertainty of wind power and spot price, which serves as the decision basis for optimizing the bidding strategy. The bidding strategy construction module is used to construct a wind farm bidding strategy optimization agent by taking wind power uncertainty, spot price uncertainty and time information as state variables. It is used to output the day-ahead bidding strategy of wind farms and use incremental near-end strategy optimization to solve the agent. A distributed communication framework construction module is used to consider the balance between improving wind farm revenue and protecting the privacy of bidding strategies. It constructs a fully distributed communication framework based on differential privacy consensus. By applying Laplace noise to the wind farm bidding strategies, it improves the anti-theft capability of wind farm bidding strategy information. It introduces a relaxation factor to construct a differential aggregation mechanism for bidding strategies, so that each wind farm can simultaneously adopt the bidding strategies of other wind farms and balance the strategy gains obtained by training with local data. The collaboration parameter determination module is used to build a privacy budget quantification and monitoring mechanism for differential privacy consensus. By monitoring the total differential privacy budget of the entire communication process, it determines the weighting status of key parameter sequences, thereby realizing end-to-end collaborative optimization of bidding strategies for multiple wind farms.
8. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described in any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-wind farm bidding strategy collaborative optimization method based on differential privacy consensus as described in any one of claims 1-6.