A multi-domain economic dispatch method and system based on game theory and generative network

Abstract

This invention relates to the field of energy system optimization and intelligent scheduling technology, and particularly to a multi-domain economic scheduling method and system based on game theory and generative networks. The invention includes: constructing a master-slave game framework; using a conditional generative adversarial network to generate a sequence of predicted scenarios for future time periods and selecting typical scenarios; the slaves calculating their local optimal trading volume based on time-of-use pricing and typical scenarios; calculating a global consistency variable based on the local optimal trading volumes of all slaves, and dynamically adjusting the penalty factor of the alternating direction multiplier method according to scenario deviations; updating the dual variable based on the global consistency variable until the convergence condition is met, obtaining the converged trading result; evaluating the overall system benefit based on the converged trading result; if the preset target is not achieved, adjusting the master's strategy and re-triggering the game process until the preset target is achieved, outputting the final scheduling plan. This invention significantly improves the game model's ability to handle uncertainty and achieves high accuracy.

Description

Technical Field

[0001] This invention relates to the field of energy system optimization and intelligent scheduling technology, and in particular to a multi-domain economic scheduling method and system based on game theory and generative networks. Background Technology

[0002] In existing microgrid multi-domain collaborative energy storage dispatching technologies, a master-slave game (Stackelberg game) is often used to achieve hierarchical decision-making (e.g., patent application CN118611067A discloses a master-slave game optimization dispatching method for distribution network-microgrid groups based on multi-agent reinforcement learning algorithm); and distributed optimization is performed by combining the Alternating Direction Multiplier Method (ADMM) (e.g., patent application CN117791610A discloses a multi-microgrid energy dispatching optimization method based on ADMM algorithm). These methods achieve global collaboration by having an upper-level leader (such as a grid operator) set electricity price signals, and lower-level followers (each energy domain) respond and optimize their local strategies.

[0003] Meanwhile, generative network technologies (such as Generative Adversarial Networks (GAN) and Deep Reinforcement Learning (DRL)) have been applied in energy scenario prediction (for example, patent application CN114862312A discloses a reasonable inventory management method based on time series algorithm; patent application CN119787352A discloses an optimized scheduling method for integrated energy microgrid based on improved TD3 algorithm), but they are mostly limited to single-domain scenarios and have not been combined with game theory to achieve cross-domain dynamic collaboration.

[0004] The existing technology has the following significant drawbacks in practical applications:

[0005] 1. Poor adaptability to dynamic scenarios: Traditional game theory models rely on static scenario assumptions, which makes it difficult to handle uncertainties in multi-domain coupling (such as fluctuations in charging demand in the transportation domain and sudden changes in renewable energy output), and lacks the ability to generate and optimize dynamic scenarios driven by generative networks.

[0006] 2. Rigid profit distribution mechanism: The distribution of benefits in master-slave games is mostly based on fixed rules (such as static calculation of Shapley value), without taking into account the real-time changes in the contribution of each energy domain, which may lead to insufficient enthusiasm for collaboration.

[0007] 3. Low efficiency of distributed solution: ADMM's penalty factor is mostly a fixed value, which results in slow convergence speed in large-scale multi-domain scenarios, and it is not adjusted in conjunction with the prediction results of the generative network.

[0008] 4. Insufficient cross-domain privacy protection: In the existing ADMM interaction, some solutions need to share sensitive data within the domain (such as load curves), which poses a risk of privacy leakage. Summary of the Invention

[0009] The purpose of this invention is to solve the problems of poor adaptability to dynamic scenarios, rigid benefit allocation, low efficiency of distributed solution and insufficient privacy protection in the existing multi-domain collaborative microgrid scheduling, and to propose a multi-domain economic scheduling method and system based on game theory and generative networks.

[0010] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0011] A multi-domain economic scheduling method based on game theory and generative networks includes:

[0012] A master-slave game framework is constructed, defining the master as the microgrid operator and the slaves as multiple energy domains. The master publishes a master strategy that includes time-of-use pricing, cross-domain capacity constraints, and revenue distribution coefficients.

[0013] Historical data is acquired, and a conditional generative adversarial network is trained based on the preprocessed historical data. The conditional generative adversarial network is used to generate a sequence of predicted scenarios for future periods, and typical scenarios are selected. The historical data includes date type, season, load, actual output power of renewable energy, predicted output power of renewable energy, and meteorological data.

[0014] Based on time-of-use pricing and typical scenarios, Fang constructs a local optimization model with the goal of maximizing revenue within the region, and calculates the local optimal transaction volume.

[0015] Using the alternating direction multiplier method, a global consistency variable is calculated based on the local optimal transaction volume of all slaves. The scenario deviation between the predicted scenario generated by the conditional generative adversarial network and the actual running state is calculated in real time. The penalty factor of the alternating direction multiplier method is dynamically adjusted according to the scenario deviation, and the dual variable is updated based on the global consistency variable.

[0016] Repeatedly calculate the local optimal transaction volume and the globally consistent variable, and update the dual variable until the convergence condition is met, and obtain the converged transaction result.

[0017] The master party evaluates the overall system benefit based on the converged transaction results. If the preset target is not achieved, the master party adjusts the time-of-use electricity price and revenue distribution coefficient in its strategy and re-triggers the game process until the preset target is achieved. Finally, it outputs the final scheduling plan to guide the energy domain in executing charging, discharging and trading actions.

[0018] Furthermore, the step of acquiring historical data, training a conditional generative adversarial network (GAN) based on the preprocessed historical data, and using the GAN to generate a predicted scene sequence for future time periods includes:

[0019] The conditional generative adversarial network includes a generator and a discriminator;

[0020] Eradicate time dimension differences in multi-source data by aligning historical data with timestamps and standardizing the format; remove data noise by handling missing values, outliers, and duplicate values ​​to obtain conditional vectors.

[0021] The generator is fed random noise and conditional vectors, and temporal features are extracted using a long short-term memory network structure to output a predicted scene sequence.

[0022] The predicted scene sequence and the real historical scene are input into the discriminator, and the parameters of the adversarial network are generated by optimizing the conditions using the adversarial loss function.

[0023] Output the predicted scene sequence generated by the trained generator.

[0024] Furthermore, the adversarial loss function is expressed by the formula:

[0025] ;

[0026] in, Represents the adversarial loss function. [・] indicates mathematical expectation. This indicates that the discriminator is effective in real historical scenarios. The probability output of the truth. The generator is based on random noise. and condition vector The generated predicted scene sequence, This indicates that the discriminator predicts the scene sequence. The probability output of the truth. Indicates random noise. This represents the condition vector.

[0027] Furthermore, the typical scenarios selected include:

[0028] Calculate the KL divergence between each scene in the predicted scene sequence and the historical distribution;

[0029] Scenarios with KL divergence less than a preset threshold are selected as typical scenarios.

[0030] Furthermore, the formula for calculating the globally consistent variable is as follows:

[0031] ;

[0032] in, Represents globally consistent variables. Indicates the first Local optimal trading volume for each energy domain This indicates the number of energy domains.

[0033] Furthermore, the deviation between the predicted scene generated by the real-time computational conditional generative adversarial network and the actual running scene is dynamically adjusted according to the scene deviation, including:

[0034] Calculate multi-domain load deviation, renewable energy output deviation, and meteorological correlation deviation, and then weight and fuse them to obtain the scenario deviation;

[0035] Calculate the original and dual residuals of the alternating direction multiplier method, and obtain the scaling factor for adjusting the penalty factor based on the original and dual residuals;

[0036] The uncertainty scaling factor is obtained based on the scene deviation;

[0037] The unsmoothed target value is calculated based on the scaling factor used to adjust the penalty factor, the uncertainty scaling factor, and the current penalty factor;

[0038] The unsmoothed target value is smoothed by an exponential moving average to obtain the penalty factor for the next scheduling period.

[0039] Furthermore, the update formula for the dual variable is as follows:

[0040] ;

[0041] in, Indicates the first Dual variables of each energy domain, Indicates the first The dual variables of the updated energy domain This represents the dynamically adjusted penalty factor. Indicates the first Local optimal trading volume for each energy domain This represents a globally consistent variable.

[0042] Furthermore, the convergence condition is: ,in, Indicates the first Local optimal trading volume for each energy domain This represents a globally consistent variable.

[0043] A multi-domain economic scheduling system based on game theory and generative networks includes a master layer, a generative network layer, an intermediate coordination layer, and a slave layer.

[0044] The master layer is used to construct the master-slave game framework. The master is defined as the microgrid operator and the slaves are multiple energy domains. The master publishes a master strategy that includes time-of-use pricing, cross-domain capacity constraints and revenue distribution coefficients. The total system benefit is evaluated based on the converged transaction results. If the preset target is not achieved, the master strategy is adjusted and the game process is retried until the preset target is achieved, and the final scheduling plan is output.

[0045] Generative network layers are used to acquire historical data, train conditional generative adversarial networks based on preprocessed historical data, generate predicted scene sequences for future periods using conditional generative adversarial networks, and select typical scenes.

[0046] The intermediate coordination layer is used to calculate the global consistency variable based on the local optimal transaction volume of all slaves using the alternating direction multiplier method, and to calculate the scenario deviation between the predicted scenario generated by the generative network and the actual running state in real time. The penalty factor of the alternating direction multiplier method is dynamically adjusted according to the scenario deviation, and the dual variable is updated based on the global consistency variable.

[0047] From the perspective of the layer, it is used to build a local optimization model based on time-of-use electricity pricing and typical scenarios, with the goal of maximizing revenue within the domain, and to calculate the local optimal transaction volume.

[0048] Compared with existing technologies, the present invention has the following significant advantages: (1) Improved dynamic adaptability: Generative networks are used to generate multi-domain dynamic scenarios with high accuracy, significantly improving the game model's ability to handle uncertainty. (2) Significantly improved solution efficiency: By dynamically adjusting the penalty factor of ADMM based on the prediction bias of generative networks, the convergence speed is significantly accelerated (as mentioned in the embodiment, it is accelerated by 40%), meeting the real-time scheduling requirements. (3) Optimization of economy and motivation: The dynamic benefit distribution mechanism combined with game optimization reduces the total operating cost of the system while improving the average revenue of each energy domain. (4) Enhanced privacy protection: Only non-sensitive coordination variables are interacted, effectively protecting the load and energy storage privacy within each energy domain. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall architecture of a multi-domain economic scheduling system based on game theory and generative networks according to the present invention.

[0050] Figure 2 This is a structural block diagram of the generative network scene generation module of the present invention;

[0051] Figure 3 This is a flowchart of the master-slave game and ADMM interaction process of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Example 1.

[0054] This invention provides a multi-domain economic scheduling method based on game theory and generative networks, which can be applied to, for example... Figure 1 The system shown is hierarchically structured as follows: Master Layer (microgrid operator), Generative Network Layer, Intermediate Coordination Layer, and Slave Layer (multi-energy domain side). Each layer forms a closed-loop coordination mechanism through data interaction. The core functional modules of each layer include:

[0055] Main layer: Configures the global optimization module and strategy formulation unit, and is responsible for coordinating the output of global control strategies.

[0056] This layer comprises software-defined functional modules, relying on the existing data server hardware deployment of the power grid operator, and implementing functional logic through algorithm code. The global optimization module, based on power grid operation objectives (optimal efficiency, cross-domain stability, etc.), integrates the characteristics of multiple energy domains to construct a mathematical optimization model. The strategy formulation unit generates time-of-use pricing based on the optimal solutions of decision variables within the domain. Cross-domain capacity constraints and profit distribution coefficient Through the intermediate coordination layer data interaction unit, policy instructions are sent to the slave layer to provide a global control benchmark for multi-energy domain scheduling.

[0057] Generative network layers: such as Figure 2 As shown, the system integrates a data preprocessing unit, a cGAN scene generation module (including generator G and discriminator D), and a scene filtering unit to achieve multi-domain dynamic scene construction and filtering.

[0058] The data preprocessing unit is a core component connecting raw data acquisition and cGAN scene generation module training. Its role is to transform multi-source, heterogeneous raw data into high-quality, noise-free temporal feature data that meets model input requirements, directly determining the accuracy of subsequent scene prediction by the generator G and the optimization effect of the discriminator D. First, timestamp alignment and format standardization eliminate temporal dimension differences between multi-source data, laying the foundation for subsequent temporal feature concatenation. Then, missing value processing, outlier processing, and duplicate value processing remove data noise, ensuring that the data input to the model conforms to the actual operating rules of the microgrid. The preprocessed data yields a conditional vector.

[0059] Both the generator G and discriminator D are algorithmic models built on deep learning frameworks (such as PyTorch) and deployed on an AI server with GPU computing power. The scene selection unit is a post-processing logic module that runs on a CPU server. The generator G uses random noise. Conditional vector The system takes as input features (including date type, season, load, renewable energy, and meteorological data) and combines them with a 3-layer LSTM network structure (Layer 1: LSTM unit with 128 neurons, using the Tanh activation function, mainly responsible for extracting macroscopic time features from the original conditional sequence; Layer 2: LSTM unit with 64 neurons, introducing Dropout (ratio 0.2) to suppress overfitting, focusing on capturing the subtle coupling relationship between load and renewable energy fluctuations; Layer 3: LSTM unit with 32 neurons) to learn the patterns of historical load, meteorological, and renewable energy data (extracted from the microgrid historical monitoring, control, and data acquisition (SCADA) system, covering at least 2 years of operating history) and outputs a predicted scene sequence. The discriminator D uses a CNN+fully connected network, receiving generated scenes and real historical scenes, and outputs the probability of authenticity (0-1 value range) to form an adversarial training with the generator G (loss function: ,in, Represents the adversarial loss function. [・] indicates mathematical expectation. This indicates that the discriminator is effective in real historical scenarios. The probability output of the truth. The generator is based on random noise. and condition vector The generated predicted scene sequence, This indicates that the discriminator predicts the scene sequence. The scene selection unit optimizes the quality of scene generation by calculating the KL divergence and similarity to historical distributions from the 50 candidate scenes output by the generator. Ten typical multi-domain dynamic scenes (KL divergence < 0.05) are then selected and synchronized to the slave layer via the intermediate coordination layer, providing scene input for multi-energy domain scheduling. Historical distribution refers to the distribution characteristics obtained by learning the temporal correlation between "load-weather-new energy output" in historical data, based on historical data, to generate a predicted scene sequence that conforms to actual operating logic, providing realistic scene input for subsequent scheduling.

[0060] Intermediate Coordination Layer: Equipped with ADMM coordination module and data interaction unit, it undertakes cross-layer information transfer and consistency calculation tasks.

[0061] This layer is a software functional module deployed on a server with multi-threaded computing capabilities, relying on the ADMM algorithm library and data communication protocol code to implement its functions. The data interaction unit bidirectionally transfers information, receiving strategy instructions from the master layer and scenario data from the generative network layer and synchronizing them to the slave layer; it also collects and transmits locally optimal trading volume data from the slave layer to support a global control closed loop. The ADMM coordination module, based on the collected locally optimal trading volume from the slave layer, executes global consistency variables. Calculate and update the dual variable. And it feeds back to the sub-layer, driving iterative optimization of the local optimal trading volume of the sub-layer until the convergence condition is met. At the same time, the converged transaction results are sent back to the main layer to assist in strategy iteration and optimization.

[0062] The dual variable, through the continuous accumulation of deviation errors, generates increasing pressure / incentives in subsequent iterations, driving the local trading volume of each slave. Eventually converges to a globally consistent level (i.e., satisfies) The larger the penalty factor, the more efficient the deviation error is in converting into dual pressure.

[0063] From the perspective of the energy domain: the power domain (MG1), transportation domain (MG2) and thermal domain (MG3) run in parallel. Each energy domain has a built-in local optimization module, energy storage control unit and transaction decision unit to perform local scheduling and transaction decisions.

[0064] The local optimization module and transaction decision unit are algorithm logic modules deployed on the local server / smart terminal in the energy domain. The energy storage control unit is adaptable to hardware such as lithium battery energy storage and flywheel energy storage, and collaborates with other units through control circuits and communication modules. The local optimization module receives the master strategy forwarded by the intermediate coordination layer (…). Based on the prediction scenarios of generative networks and other technologies, a local optimization model is constructed by integrating local load characteristics and energy storage status. The energy storage control unit, based on the local optimization results (including charging / discharging mode commands, target charging / discharging power, and SOC target trajectory), regulates the charging and discharging actions of energy storage (such as power domain energy storage charging / discharging power and energy storage charging / discharging periods) to maintain local energy balance. It is then uploaded to the intermediate coordination layer to participate in global collaborative scheduling.

[0065] like Figure 3 As shown, the present invention provides a multi-domain economic scheduling method based on game theory and generative networks, the specific steps of which are as follows:

[0066] Step S1: Construct a master-slave game framework.

[0067] S101: Define the master (microgrid operator) and slave (microgrid power domain, microgrid traffic domain, microgrid thermal domain, etc.).

[0068] The primary objective is to minimize the total system operating cost (electricity purchase cost + network loss + penalty cost). The secondary objective is to maximize the domain revenue (electricity sales revenue - energy storage cost - electricity purchase cost). In other words, it maximizes the difference between the total system benefit and the preset target benefit.

[0069] S102: Design master-slave strategy variables.

[0070] The principal strategy variables include time-of-use electricity price λ(t) and cross-domain capacity constraints. and profit distribution coefficient Among them, cross-domain capacity constraints The fixed strategy issued by the main party is embedded in the local optimization model and is not dynamically adjusted.

[0071] Step S2: Generative network dynamic scene generation.

[0072] S201: Training a Conditional Generative Adversarial Network (cGAN).

[0073] Historical data is collected, encompassing date type, season, load, renewable energy, and meteorological data. This historical data, along with random noise, is input into a generator. The generator outputs predicted scenarios for future periods across multiple energy domains (e.g., charging demand curves in the transportation domain, photovoltaic output fluctuations), forming a sequence of predicted scenarios. A discriminator optimizes the generator's output quality through adversarial training, ensuring the predicted scenarios better reflect actual operational patterns while maintaining a KL divergence of less than 0.05 between the predicted scenarios and historical distributions. The specific implementation process involves: first, training the generator model; then, generating a sufficient number of candidate scenarios using the trained model; quantifying the fit between each candidate scenario and the historical data distribution using KL divergence; and finally, selecting scenarios with KL divergence meeting the threshold requirements. Simultaneously, considering the diversity of scenario coverage, a set of typical scenarios reflecting different operational states across multiple domains is determined.

[0074] Predicting scene sequences For length is Multivariate time series vector sequences: .in vector of time It includes key variables in several domains (electricity load L, photovoltaic output PV, wind power W, transportation charging demand EV, meteorological data θ, etc.). Indicates the first One scenario, This indicates the number of time slots within the predicted period (e.g., 15 minutes / time slot, or T=96 for 24 hours).

[0075] The adversarial loss function is expressed by the formula:

[0076] ;

[0077] in, Represents the adversarial loss function. [・] indicates mathematical expectation. This indicates that the discriminator is effective in real historical scenarios. The probability output of the truth. The generator is based on random noise. and condition vector The generated predicted scene sequence, This indicates that the discriminator predicts the scene sequence. The probability output of the authenticity.

[0078] S202: Filter typical scenario sets.

[0079] This step uses a reproducible automated filtering process, eliminating the need for manual selection of each item. The specific process is as follows:

[0080] 1. The trained cGAN generates N candidate scene sequences based on the conditional vector and random noise.

[0081] 2. Calculate the similarity between each candidate and the historical distribution (e.g., calculate the KL divergence under unified binning or KDE, or directly calculate the historical distribution likelihood), and use authenticity as a hard threshold to remove unqualified samples (the initial threshold can be set to KL<0.05, and if the number of samples after screening is too small, it can be gradually relaxed to 0.07-0.1).

[0082] 3. Perform physical legitimacy checks and corrections on candidate scenarios that pass the authenticity screening (for nighttime photovoltaic output values ​​that are not 0, force them to be set to 0 and no longer participate in subsequent calculations); the discriminator confidence D(S) can be used as an auxiliary criterion.

[0083] 4. To ensure both realism and physical feasibility in the candidate set, and to balance diversity and representativeness, each scenario is first standardized in time sequence and then embedded in low dimension (PCA or autoencoder, retaining d=3~5 dimensions or retaining 90% variance), and then the distance matrix is ​​calculated in the embedding space.

[0084] 5. Use clustering-centering or greedy forward selection to automatically select K typical scenarios.

[0085] 6. Assign probability weights to each selected scenario (obtained by historical likelihood or candidate sample density normalization), and reserve several extreme / risk scenarios for stress testing as needed.

[0086] The aforementioned automated process ensures that the selected scenarios are located in a high-probability region, consistent with historical distributions (realism constraint), and cover different operating modes (diversity constraint). The selected scenario set and its probability weights will serve as input constraints and parameters in subsequent master-slave game theory (slave-side local optimization and ADMM coordination), driving the parallel or weighted solution of local subproblems under different uncertainties. The thresholds and algorithms in this process (such as N, K, KL thresholds, dimensionality reduction, distance metrics, and clustering methods) are adjustable parameters, and the optimal settings can be determined through historical backtesting to improve robustness and reproducibility.

[0087] Step S3: ADMM distributed solution.

[0088] S301: Local optimization from the local side.

[0089] Based on the principal strategy λ(t) and typical scenario sets, the objective of solving the local subproblem in each energy domain is to maximize the domain revenue, with constraints including the remaining energy storage capacity constraint. , This represents the remaining battery power at time t. Indicates the minimum remaining battery power. This involves defining the maximum remaining power capacity and power balance constraints (load consumption = local generation supply + external transaction input - energy storage charging consumption + energy storage discharging supply). After solving the local subproblem in each energy domain (which can be achieved using linear programming (LP)), the optimal local transaction volume is output. .

[0090] S302: Intermediate layer coordination.

[0091] Collect the local optimal trading volume of all slaves Calculate the globally consistent variables:

[0092] ;

[0093] in, Represents globally consistent variables. Indicates the first Local optimal trading volume for each energy domain This indicates the number of energy domains.

[0094] The ADMM's penalty factor is dynamically adjusted to address the deviation between the predicted scenario and the actual operating scenario based on generative network layer prediction. (The greater the deviation, The larger the (i.e., the larger). This embodiment uses a scheduling cycle of 15 minutes (i.e., ... Time slicing is performed, and the scene deviation is calculated once after each scheduling cycle, which is used as the penalty factor for the next scheduling cycle. Adjustments were made to ensure that deviations were synchronized with the scheduling rhythm.

[0095] Generative network layers calculate multi-domain load deviations, renewable energy output deviations, and meteorological correlation deviations, and then weight and fuse them to obtain directly driveable data. Adjusted scene deviation (Value range 0~1, The larger the value, the higher the system uncertainty. The specific steps are as follows:

[0096] 1. Calculate multi-domain load deviation (Reflecting the difference between load forecast and actual load) Load is the core of microgrid energy demand. Load fluctuations in different energy domains (electricity, transportation, heat) directly affect inter-domain transaction volume. The load deviation of all slaves must be covered, as shown in the following formula:

[0097]

[0098] in, Indicates the number of energy domains. Indicates the first Load weights for each energy domain Indicates the first The energy domain in the first Actual load power during the scheduling cycle (unit: MW, real-time data collected from smart meters and charging pile controllers). Indicates the first The energy domain in the first The predicted load power (unit: MW) for the scheduling cycle is taken from the load value of the corresponding scenario in the "typical scenario set" predicted in the previous cycle on the generative network. If there are 10 typical scenarios, the "high probability scenario" is taken. Indicates the first Historical average load power of each energy domain (unit: MW, calculated based on data from the same period of the past 3 months to avoid distortion due to differences in load magnitude).

[0099] The calculation logic for multi-domain load deviation is to eliminate the differences in load magnitude between different domains through weighted average and normalization, highlighting the load deviation impact of high-proportion domains.

[0100] 2. Calculate the output deviation of renewable energy sources. (Reflecting the discrepancy between energy supply forecasts and actual levels) Renewable energy (solar, wind power) is the main clean energy source for microgrids, and its output randomness is the core source of system uncertainty. The deviation between predicted output and actual output needs to be quantified, and the formula is as follows:

[0101]

[0102] in, Indicates the number of renewable energy types. Indicates the first The weight of renewable energy categories (set according to their installed capacity as a percentage of total renewable energy installed capacity). Indicates the first Renewable energy in the first Actual power output during the scheduling cycle (unit: MW, real-time data collected from photovoltaic inverters and wind power converters). Indicates the first Renewable energy in the first The predicted output power of the scheduling cycle (unit: MW, taken from the output value of the corresponding scenario in the typical scenario set predicted in the previous cycle on the generative network, taking the average of the high probability scenarios). Indicates the first The installed capacity of renewable energy (unit: MW) is used for normalization to avoid deviations and distortions caused by differences in installed capacity.

[0103] The calculation logic for renewable energy output deviation focuses on supply-side uncertainties and accurately reflects the degree of fluctuation in renewable energy output through installed capacity normalization (the higher the deviation ratio, the greater the pressure on system power balance).

[0104] 3. Calculate meteorological correlation deviation (Reflecting the discrepancy between weather forecasts and actual conditions, indirectly affecting load and energy output) Meteorological data (temperature, sunshine, wind speed) is the core input for generative network forecasting of load and renewable energy output. Weather forecast deviations are the root cause of "load deviations and output deviations," therefore... The accuracy of weather forecasts needs to be quantified, and the formula is as follows:

[0105]

[0106] in, Indicates the number of key meteorological parameters. Indicates the first The weights of each meteorological parameter (set according to its impact on the microgrid). Indicates the first The meteorological parameter in the first The actual value of the scheduling period. Indicates the first The meteorological parameter in the first Predicted values ​​of scheduling cycles (taken from the generative network) The meteorological values ​​for typical scenarios in the scheduling cycle prediction set are taken as the average of the high-probability scenarios. Indicates the first The historical variation range of each meteorological parameter is used for normalization, converting the deviation of meteorological parameters in different units into dimensionless values ​​of 0 to 1.

[0107] The calculation logic for meteorological correlation bias is to trace the root cause of uncertainty. The smaller the meteorological bias, the more reliable the predictive basis of the generative network and the lower the system uncertainty; conversely, adjustments are needed. Enhance ADMM's immunity to disturbances.

[0108] 4. , , The scene deviation is obtained by weighting and fusing the results according to their importance in affecting system coordination. (0≤) ≤1), the formula is as follows:

[0109] ;

[0110] The weighting is based on the load deviation. Directly impacting cross-regional transaction volume demand, it is a core object of ADMM collaboration, with a weight set at 0.4. Renewable energy output deviation. Directly impacting energy supply, it, along with load deviation, determines power balance pressure; its weight is set at 0.4. Meteorological correlation deviation. This is an indirect source of deviation, with a weaker impact than direct load and output deviations; its weight is set at 0.2. The above weights can be adjusted according to the actual characteristics of the microgrid (e.g., for microgrids with a high proportion of renewable energy, the weight can be increased). The weight is set to 0.5, but the sum of the weights must be 1.

[0111] In each scheduling cycle At the end, calculate and update the penalty factor for the next cycle according to the following engineering rules. :

[0112] 1. Preparatory Calculation: Calculate scene deviation. ∈[0,1]; Calculate ADMM residuals: original residuals Dual residual .in, Indicates the first The original residual for each scheduling cycle, Indicates the first The dual residual of each scheduling cycle Indicates the first A mapping matrix from an energy domain to a globally consistent quantity space. Indicates the first Globally consistent variables for each scheduling cycle Indicates the first Globally consistent variables for each scheduling cycle It is the first The penalty factor for each scheduling cycle.

[0113] 2. Uncertainty mapping: by Obtain the uncertainty scaling factor It can be either a linear or exponential mapping, such as linear. ,in, This is the magnification factor; or the exponent. ,in, This represents the lower bound of the uncertainty measure. It represents the upper bound of the uncertainty measure.

[0114] 3. Residual balance correction: Compare according to empirical rules. and (Set ratio threshold) and multiples ):like ,but (increase) );like but (reduced) );otherwise .in, Indicates the use of adjusting the penalty factor The scaling factor.

[0115] 4. Target and Cutoff: Calculate the unsmoothed target value ,in, This indicates that the range is limited to a preset interval. , This represents the minimum permissible penalty factor. This represents the maximum allowed penalty factor.

[0116] 5. Smooth Updates: To avoid Severe oscillations, smoothed using exponential moving average: .in, Indicates the first The penalty factor for each scheduling cycle. Indicates the smoothing factor, and suggests ).

[0117] S303: Dual variable update.

[0118] Each energy domain is based on globally consistent variables Update the dual variable, and update the formula as follows:

[0119] ;

[0120] in, Indicates the first Dual variables of each energy domain, Indicates the first The dual variables of the updated energy domain This represents the dynamically adjusted penalty factor. Indicates the first The local optimal trading volume for each energy domain.

[0121] The updated dual variables, globally consistent variables, and dynamically adjusted penalty factors are fed back to the slave. The slave iteratively performs local optimization calculations based on the received feedback parameters, outputting the updated local optimal trading volume, until the convergence condition is met (the convergence condition is...). The converged trading results are obtained, including the local optimal trading volume. Dual variables Globally consistent variables .

[0122] Step S4: Dynamic feedback and strategy iteration.

[0123] S401: The principal evaluates the overall system benefit based on the transaction results after ADMM convergence. .

[0124]

[0125] in, This represents the baseline operating cost, which is the estimated total expenditure of the system under conditions where multi-domain game-theoretic collaborative scheduling is not performed (e.g., each domain operates independently, and there is no energy storage optimization). The actual total cost of this round is represented by the following formula:

[0126] in, express Real-time electricity purchase price per unit of the large power grid during the specified time period. This indicates the feedback sent to the master after each energy domain converges through ADMM iteration. Total power purchase of the entire system during the time period express The system network loss cost for a given period is converted into monetary value. express Punitive expenditures during certain periods (such as fines for curtailment of wind and solar power, and economic penalties for insufficient reserve capacity).

[0127] S402: If the overall system benefit does not reach the preset target. (Right now The principal agent combines the scene prediction of the generative network and readjusts its strategy (e.g., λ(t)). The specific steps are as follows:

[0128] 1. Generate new network scenarios: The master layer triggers the generative network layer, and based on the latest historical data and the real-time running data of the previous round of games, the cGAN is retrained to generate 50 candidate scenarios, and 10 new typical scenarios with KL divergence < 0.05 are selected.

[0129] 2. Targeted adjustment strategy:

[0130] according to The adjustment rules are used to calculate the new electricity price for each time period; the specific steps are as follows:

[0131] (1) Definition of input quantity:

[0132] The overall system benefits (scalar) of this round of evaluation; : Preset target benefit (scalar);

[0133] Time period Net load imbalance (positive for power shortage, negative for excess). Normalized reference power;

[0134] The scene bias given by generative networks ∈[0,1];

[0135] The time-of-use electricity price issued in this cycle; The proposed electricity price for the next cycle.

[0136] (2) Calculation steps and formulas:

[0137] Calculating driving factors: Based on the results of this round of evaluation and local / global indicators, two types of driving factors are calculated, including: economic drivers (reflecting...) - Direction and intensity) and balance drive (net load imbalance B in each time period) (As a balance driver).

[0138] Constructing candidate adjustment quantities: Combining economic and balance drivers yields candidate electricity price change rates for each time period or candidate allocation proportions for each sector (energy domain). For ease of implementation, simple multiplication or proportional forms can be used; for example, electricity prices can be expressed as... ,in, , The normalized economic drivers and balance drivers.

[0139] S403: Repeat steps S1-S4 until the master-slave game reaches Stackelberg equilibrium (both sides have no incentive to deviate), output the final scheduling plan, guide the multi-energy domains to perform charging, discharging, and trading actions, achieve the goal of multi-energy domain coordinated scheduling, and realize the optimal efficiency and stable operation of the power grid and multi-energy domains. The final scheduling plan includes the updated local optimal trading volume, energy storage charging and discharging periods, energy storage charging and discharging power, and the control range of remaining energy storage capacity (minimum remaining capacity). and maximum remaining power ).

[0140] This invention does not simply combine master-slave game theory, ADMM, and generative network techniques. The generative network not only outputs typical scenarios but also calculates the deviation between the scenario and the actual running scenario in real time, using this deviation as the core basis for adjusting ADMM parameters. This breaks through the limitation of traditional ADMM's fixed penalty factor, utilizing the scenario deviation predicted by the generative network to optimize the penalty factor in real time. This allows the solution process of ADMM to resonate with the dynamic characteristics of the system. The time-of-use electricity price and revenue distribution coefficient of the principal are not only based on the initial prediction scenario, but also combined with the transaction results after the convergence of ADMM. The scenario correction strategy is regenerated through a generative network, forming a closed-loop optimization mechanism of strategy, solution, feedback and correction.

[0141] Example 2.

[0142] This embodiment provides a multi-domain economic scheduling system based on game theory and generative networks, including a master layer, a generative network layer, an intermediate coordination layer, and a slave layer.

[0143] The master layer is used to construct the master-slave game framework. The master is defined as the microgrid operator and the slaves are multiple energy domains. The master publishes a master strategy that includes time-of-use pricing, cross-domain capacity constraints and revenue distribution coefficients. The total system benefit is evaluated based on the converged transaction results. If the preset target is not achieved, the master strategy is adjusted and the game process is retried until the preset target is achieved, and the final scheduling plan is output.

[0144] Generative network layers are used to acquire historical data, train conditional generative adversarial networks based on the historical data, generate predicted scene sequences for future periods using the conditional generative adversarial networks, and select typical scenes.

[0145] The intermediate coordination layer is used to calculate the global consistency variable based on the local optimal transaction volume of all slaves using the alternating direction multiplier method, and dynamically adjust the penalty factor of the alternating direction multiplier method according to the scenario deviation, and update the dual variable based on the global consistency variable.

[0146] From the perspective of the primary strategy and typical scenarios, a local optimization model is built with the goal of maximizing the revenue within the domain, and the optimal local transaction volume is calculated.

[0147] For specific limitations on multi-domain economic scheduling systems based on game theory and generative networks, please refer to the limitations on multi-domain economic scheduling methods based on game theory and generative networks mentioned above, which will not be repeated here.

[0148] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A multi-domain economic scheduling method based on game theory and generative networks, characterized in that, The multi-domain economic scheduling method based on game theory and generative networks includes: A master-slave game framework is constructed, defining the master as the microgrid operator and the slaves as multiple energy domains. The master publishes a master strategy that includes time-of-use pricing, cross-domain capacity constraints, and revenue distribution coefficients. Historical data is acquired, and a conditional generative adversarial network is trained based on the preprocessed historical data. The conditional generative adversarial network is used to generate a sequence of predicted scenarios for future periods, and typical scenarios are selected. The historical data includes date type, season, load, actual output power of renewable energy, predicted output power of renewable energy, and meteorological data. Based on time-of-use pricing and typical scenarios, Fang constructs a local optimization model with the goal of maximizing revenue within the region, and calculates the local optimal transaction volume. Using the alternating direction multiplier method, a global consistency variable is calculated based on the local optimal transaction volume of all slaves. The scenario deviation between the predicted scenario generated by the conditional generative adversarial network and the actual running state is calculated in real time. The penalty factor of the alternating direction multiplier method is dynamically adjusted according to the scenario deviation, and the dual variable is updated based on the global consistency variable. Repeatedly calculate the local optimal transaction volume and the globally consistent variable, and update the dual variable until the convergence condition is met, and obtain the converged transaction result; The master party evaluates the overall system benefit based on the converged transaction results. If the preset target is not achieved, the master party adjusts the time-of-use electricity price and revenue distribution coefficient in its strategy and re-triggers the game process until the preset target is achieved. Finally, it outputs the final scheduling plan to guide the energy domain in executing charging, discharging and trading actions.

2. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The process of acquiring historical data, training a conditional generative adversarial network (GAN) based on the preprocessed historical data, and using the GAN to generate a predicted scene sequence for future time periods includes: The conditional generative adversarial network includes a generator and a discriminator; Eradicate time dimension differences in multi-source data by aligning historical data with timestamps and standardizing the format; remove data noise by handling missing values, outliers, and duplicate values ​​to obtain conditional vectors. The generator is fed random noise and conditional vectors, and temporal features are extracted using a long short-term memory network structure to output a predicted scene sequence. The predicted scene sequence and the real historical scene are input into the discriminator, and the parameters of the adversarial network are generated by optimizing the conditions using the adversarial loss function. Output the predicted scene sequence generated by the trained generator.

3. The multi-domain economic scheduling method based on game theory and generative networks according to claim 2, characterized in that, The adversarial loss function is expressed by the following formula: ； in, Represents the adversarial loss function. [・] indicates mathematical expectation. This indicates that the discriminator is effective in real historical scenarios. The probability output of the truth. The generator is based on random noise. and condition vector The generated predicted scene sequence, This indicates that the discriminator predicts the scene sequence. The probability output of the truth. Indicates random noise. This represents the condition vector.

4. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The selected typical scenarios include: Calculate the KL divergence between each scenario in the predicted scenario sequence and the historical distribution; Scenarios with KL divergence less than a preset threshold are selected as typical scenarios.

5. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The formula for calculating the globally consistent variable is as follows: ； in, Represents globally consistent variables. Indicates the first Local optimal trading volume for each energy domain This indicates the number of energy domains.

6. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The real-time computational conditional generative adversarial network generates a predicted scene that deviates from the actual running scene. Based on this deviation, the penalty factor of the alternating direction multiplier method is dynamically adjusted, including: Calculate multi-domain load deviation, renewable energy output deviation, and meteorological correlation deviation, and then weight and fuse them to obtain the scenario deviation; Calculate the original and dual residuals of the alternating direction multiplier method, and obtain the scaling factor for adjusting the penalty factor based on the original and dual residuals; The uncertainty scaling factor is obtained based on the scene deviation; The unsmoothed target value is calculated based on the scaling factor used to adjust the penalty factor, the uncertainty scaling factor, and the current penalty factor; The unsmoothed target value is smoothed by an exponential moving average to obtain the penalty factor for the next scheduling period.

7. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The update formula for the dual variable is as follows: ； in, Indicates the first Dual variables of each energy domain, Indicates the first The dual variables of the updated energy domain This represents the dynamically adjusted penalty factor. Indicates the first Local optimal trading volume for each energy domain This represents a globally consistent variable.

8. The multi-domain economic scheduling method based on game theory and generative networks according to claim 1, characterized in that, The convergence condition is: ,in, Indicates the first Local optimal trading volume for each energy domain This represents a globally consistent variable.

9. A multi-domain economic scheduling system based on game theory and generative networks, characterized in that: It includes a master layer, a generative network layer, an intermediate coordination layer, and slave layers. The master layer is used to construct the master-slave game framework. The master is defined as the microgrid operator and the slaves are multiple energy domains. The master publishes a master strategy that includes time-of-use pricing, cross-domain capacity constraints and revenue distribution coefficients. The total system benefit is evaluated based on the converged transaction results. If the preset target is not achieved, the master strategy is adjusted and the game process is retried until the preset target is achieved, and the final scheduling plan is output. Generative network layers are used to acquire historical data, train a conditional generative adversarial network based on the preprocessed historical data, use the conditional generative adversarial network to generate a sequence of predicted scenes for future periods, and select typical scenes. The intermediate coordination layer is used to calculate the global consistency variable based on the local optimal transaction volume of all slaves using the alternating direction multiplier method, and to calculate the scenario deviation between the predicted scenario generated by the generative network and the actual running state in real time. The penalty factor of the alternating direction multiplier method is dynamically adjusted according to the scenario deviation, and the dual variable is updated based on the global consistency variable. From the perspective of the layer, it is used to build a local optimization model based on time-of-use electricity pricing and typical scenarios, with the goal of maximizing revenue within the domain, and to calculate the local optimal transaction volume.

Images