Electric power spot transaction strategy risk dynamic assessment system and method based on real-time information

Abstract

The invention discloses an electric power spot transaction strategy risk dynamic assessment system and method based on real-time information, and the system comprises a data acquisition module which collects multi-source heterogeneous data including market power grid quotation and load prediction in real time through the deployment of an edge calculation node, and carries out the preprocessing of the data, and obtains a low-delay electric power data flow; the risk acquisition module is used for constructing market structure data through a specific incremental GAN architecture on the basis of the low-delay power data flow, and generating a dynamic income matrix and antagonistic risk data by using antagonistic machine learning on the basis of the market structure data; the intelligent decision-making module is used for constructing a risk stress field model based on the dynamic income matrix and the adversarial risk data, and introducing an online adversarial learning mechanism to optimize the risk stress field model to obtain a power transaction risk agent; and the risk prediction module is used for inputting the power data flow collected in real time into a power transaction risk agent, outputting a power risk quantitative index and realizing dynamic evaluation of power spot transaction strategy risks.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, and in particular to a dynamic risk assessment system and method for electricity spot trading strategies based on real-time information. Background Technology

[0002] The electricity spot market has gradually become the core platform for the allocation of electricity resources. Its trading model has shifted from traditional planned dispatch to dynamic game among multiple entities, such as power generation companies, electricity sales companies, and users. In order to cope with market uncertainty, risk assessment technology has evolved from static statistical analysis to dynamic model simulation, such as game theory and Monte Carlo simulation.

[0003] Current risk assessment methods for electricity trading strategies are limited to a single scenario, focusing primarily on routine market conditions. They lack sufficient simulation of extreme scenarios, such as cascading risks caused by extreme weather or grid failures, resulting in weak predictive capabilities for extreme risks. Furthermore, traditional methods are mostly based on offline batch data processing, making it difficult to handle real-time market information flows caused by sudden changes in renewable energy output or unexpected load fluctuations. Consequently, risk assessment results are lagging and cannot provide immediate decision support for high-frequency trading. Therefore, this paper proposes a dynamic risk assessment method for electricity spot trading strategies based on real-time information. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: Firstly, this application proposes a dynamic risk assessment system for electricity spot trading strategies based on real-time information, including: Data acquisition module: Collects multi-source heterogeneous data, including market grid quotations and load forecasts, in real time by deploying edge computing nodes, and preprocesses the data to obtain low-latency power data streams; Risk acquisition module: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrix and adversarial risk data. Intelligent decision-making module: Based on the dynamic profit matrix and adversarial risk data, a risk stress field model is constructed, and an online adversarial learning mechanism is introduced to optimize the risk stress field model to obtain an intelligent agent for power trading risk; Risk prediction module: Input the real-time collected power data stream into the power trading risk intelligent agent, output power risk quantitative indicators, and realize dynamic assessment of the risk of power spot trading strategy.

[0005] Optionally, the process of obtaining a low-latency power data stream is as follows: Edge computing devices are deployed to collect raw multi-source heterogeneous data. Outliers and missing values ​​are removed from the raw multi-source heterogeneous data. The min-max standardization method is used to standardize any value x in the dataset. Then, time series data from different sources are aligned by timestamp to form multi-dimensional data records indexed by time, resulting in a low-latency power data stream.

[0006] Optionally, the process for obtaining a specific incremental GAN ​​architecture is as follows: Based on the basic GAN architecture, a low-latency power data stream is input, and a sliding time window is constructed using the Flink streaming engine. The electric feature representation f is obtained through the Transformer temporal embedding layer, and f is input into the generator G and discriminator D of a specific incremental GAN ​​architecture. This paper introduces an online reinforcement learning Actor-Critic framework. The Actor takes the power feature representation f as input and the intermediate feature output of the generator G as input, outputting a scenario control vector z. This scenario control vector z is then input into the generator G. The Critic framework takes the output samples of the generator G and the real samples as input to calculate a risk discrimination index. The reward function is optimized and fed back to the Actor framework to drive the generator G. Finally, the loss function of the multi-task generator head is used to train the generator G and the discriminator D simultaneously to obtain a specific incremental GAN ​​architecture. The loss function of the multi-task generation head includes normal scene loss and extreme scene loss. The normal scene loss is represented by Wasserstein distance loss, while the extreme scene loss is represented by constraint loss combined with domain knowledge.

[0007] Optionally, the process of obtaining the dynamic return matrix and adversarial risk data is as follows: Based on adversarial machine learning, this paper defines the game players in the electricity spot market, each with a strategy space. Using market structure data as the target, it simulates the strategy adversarial interactions between the players under extreme scenarios. For each type of market structure data, the payoff M of different players under their chosen strategies is calculated, resulting in a dynamic payoff matrix. , An index for market structure data, based on a dynamic return matrix. Calculate risk loss to obtain adversarial risk data.

[0008] Optionally, the calculation process for the returns M of different entities under the chosen strategies in the market structure data is as follows: Define the payoff function of each player in the electricity market as follows: The cost function is For the main body The j-th chosen strategy, the strategies chosen by other participants, and market structure data are used to calculate the return function. for To obtain the unit revenue under their combined effect, through the cost function. Acquisition subject The cost of implementing the j-th strategy is calculated by multiplying the unit revenue by the size of the entity when implementing the j-th strategy. Subtract the main body The cost of executing the j-th strategy and the profit M obtained.

[0009] Optionally, the process of constructing the risk stress field model is as follows: Define risk stress as , indicating the subject Market structure data Next selection strategy At that time, the rate of change in risk resulting from unit strategy adjustments will reflect the risk stress of each entity in market structure data. The following steps are performed to integrate and construct a risk stress field model. , ,in, This represents the risk stress weighting coefficient. This represents the output of the risk stress field model.

[0010] Optionally, the process of optimizing the risk stress field model is as follows: An online adversarial learning mechanism is introduced, inputting real-time collected low-latency power data streams into an incremental GAN ​​architecture, and applying this mechanism to each agent. Create a strategy-based evolution pool ; When incremental GAN ​​architecture generates new market structure data Then, the strategy adversarial evolution pool is invoked to generate a new dynamic payoff matrix and adversarial risk data. Based on the newly generated dynamic payoff matrix and adversarial risk data, a new risk stress is calculated. This leads to the new field strength function. Then, the gradient descent method is used to optimize the parameters of the risk stress field model, and the model parameters are optimized by calculating the loss function. gradient And update the parameters to obtain the power trading risk intelligent agent. ,in, This is the learning rate.

[0011] Optionally, a strategy-versus-strategy evolution pool Represented as: Strategy Battle Evolution Pool Including high-risk strategy pool and low-risk strategy pool ,in: High-risk strategy pool : Filter to meet strategy, The high-risk threshold is defined as the strategies that account for the top 30% of risk losses in historical confrontations. Low-risk strategy pool : Filter to meet strategy, The low-risk threshold is defined as the strategies that rank in the bottom 30% in terms of risk loss in historical confrontations.

[0012] Secondly, this application proposes a dynamic risk assessment method for electricity spot trading strategies based on real-time information, including: S1: By deploying edge computing nodes, multi-source heterogeneous data, including market grid quotations and load forecasts, are collected in real time and preprocessed to obtain low-latency power data streams; S2: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrices and adversarial risk data. The specific incremental GAN ​​architecture is based on existing GAN architectures, combined with the online reinforcement learning Actor-Critic framework, and the initial network structure and parameter implementation of the multi-task generator head are designed to determine the generator G and discriminator D. S3: Construct a risk stress field model based on a dynamic profit matrix and adversarial risk data, and introduce an online adversarial learning mechanism to optimize the risk stress field model, thereby obtaining an intelligent agent for power trading risk; S4: Input the real-time collected power data stream into the power trading risk intelligence agent, output power risk quantification indicators, and realize dynamic assessment of the risk of power spot trading strategies.

[0013] Thirdly, this application proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method proposed in the second aspect above.

[0014] The present invention has the following beneficial effects: 1. By introducing a specific incremental GAN ​​architecture and a multi-task generation head that combines online reinforcement learning, it is possible to generate extreme market scenarios that are both consistent with knowledge in the power sector and highly adversarial. At the same time, by constructing a dynamic payoff matrix and a strategy adversarial evolution pool, the subject's strategy selection is based on historical high and low risk strategies to choose better strategies, and the strategy library is updated in real time. This realistically reproduces the strategy iteration process of market subjects under risk feedback, making the dynamic payoff matrix and adversarial risk data closer to the actual game logic and improving the accuracy of risk quantification. Ultimately, the power trading risk intelligence agent can accurately simulate the strategy interaction and risk transmission of multiple subjects in extreme scenarios, making the risk assessment results closer to the real market environment full of games. 2. By constructing a risk stress field model, risk is quantified into calculable stress and field strength. This not only outputs the overall risk level of the market, but also accurately measures the risk exposure of a single entity's specific strategy and its contribution to the overall risk. This provides more targeted decision-making basis for different market participants, such as power generation companies and market regulatory agencies, and enhances the authenticity of risk assessment results. Attached Figure Description

[0015] Figure 1 This is a system block diagram of a dynamic risk assessment system for electricity spot trading strategies based on real-time information proposed in this invention.

[0016] Figure 2 This is a flowchart illustrating the steps of a method for dynamic risk assessment of electricity spot trading strategies based on real-time information, as proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: like Figure 1 As shown, the present invention proposes a dynamic risk assessment system for electricity spot trading strategies based on real-time information, comprising: Data acquisition module: Collects multi-source heterogeneous data, including market grid quotations and load forecasts, in real time by deploying edge computing nodes, and preprocesses the data to obtain low-latency power data streams; The deployment of edge computing devices covers multiple data sources, including market price release and power grid load monitoring. Market power grid price data includes real-time electricity price declarations from various power generation entities and capacity quotations for transmission lines. Load forecast data includes historical electricity consumption patterns of regional power users, real-time electricity consumption feedback, and meteorological factors such as the impact of temperature and humidity on air conditioning load. These multi-source heterogeneous data are collected in real time through edge computing nodes using IoT protocols such as MQTT and CoAP. The collected raw data exhibits multi-source heterogeneity and requires preprocessing to obtain a low-latency power data stream. The process is as follows: The collected raw data has multi-source heterogeneous characteristics. First, outliers in the price quotation data (such as electricity prices that are significantly outside the reasonable range) are removed, and missing values ​​in the load forecast data are filled in (using linear interpolation for missing values ​​in the time series). Then, data standardization is performed to convert price data with different units, such as yuan / MWh and load data (such as megawatts), into values ​​with a unified unit. The min-max standardization method is used to standardize any value x in the dataset. Then, time series data from different sources are aligned by timestamps to form multi-dimensional data records indexed by time, ensuring the time sequence consistency of subsequent processing and obtaining a low-latency power data stream. .

[0019] Risk acquisition module: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrix and adversarial risk data. Traditional GAN ​​architectures include generators and discriminators. Specific incremental GAN ​​architectures are based on traditional GAN ​​architectures combined with the online reinforcement learning Actor-Critic framework and design multi-task generator heads to determine the initial network structure and parameters of generator G and discriminator D. The multi-task generator head includes normal scenario branches and extreme scenario branches. The extreme scenario branches introduce prior knowledge constraints of extreme events in the power market to determine the initial network structure and parameters of generator G and discriminator D. The process of obtaining a specific incremental GAN ​​architecture is as follows: Based on low-latency power data streams Input Flink streaming engine to build sliding time window (The window size is adaptively adjusted in real time based on the volatility of the electricity market, and the window is reduced when the volatility is high.) After passing through the Transformer temporal embedding layer, the embedding layer introduces a vocabulary specific to the electricity field to enhance the correlation between temporal features and market structure data features, and obtains the electricity feature representation f. f is then input into the generator G and discriminator D of a specific incremental GAN ​​architecture. Then, the online reinforcement learning Actor-Critic framework is introduced. The Actor takes the power feature representation f as input and the intermediate feature output of the generator G as input, and outputs the scenario control vector z. Then, the scenario control vector z is input into the generator G. Specifically, the scenario control vector z is the decision signal for the scenario type, such as the control code for extreme electricity price scenarios and normal load scenarios, which enables the generator G to generate market scenario samples with different risk levels. Furthermore, the output process of the scenario control vector z is represented as follows: ,in, This represents the intermediate feature output of generator G; The Critic framework uses the output samples of the generator G and real samples as inputs to calculate a risk discrimination index. To optimize the reward function, the reward is fed back to the Actor framework to drive the generator G to autonomously learn the highest extreme scenario of the reward function; Furthermore, the calculation process for the risk differentiation index is expressed as follows: ,in, This represents the output sample of generator G. This represents the actual samples of the generator G; Specifically, the risk discrimination index is the difference in risk distribution between generated samples and real samples, which is used to transform it into a reward function. The higher the risk discrimination of the generated samples, the greater the reward, thereby driving the Actor to optimize the scenario control strategy. The reward function is expressed as: ,in, and Indicates the reward weight. Other reward items include the degree to which the generated scenario aligns with the real-world dynamics of the electricity market; Then, the loss function of the multi-task generator head is used to train the generator G and the discriminator D simultaneously to obtain a specific incremental GAN ​​architecture. Among them, the loss function of the multi-task generation head Including losses in normal scenarios and loss in extreme scenarios ; The normal scene branch uses Wasserstein distance loss to ensure that the generated normal scene data closely approximates the distribution of real normal scene data. Gradient penalty is used to guarantee the stability of the training process. The normal scene loss... Represented as: ; in, It is the discriminator's discrimination output on the output samples of the generator G. It is the discriminator's output that distinguishes between real samples from the generator G. It is the gradient penalty coefficient. Represents the mathematical expectation; Extreme scenario losses are represented using constraint losses that incorporate domain knowledge. This ensures that the generated extreme scenario data meets the knowledge constraints of the power industry regarding extreme risks, such as electricity price fluctuations needing to reach expert-set extreme thresholds. This ensures that the generated extreme scenarios are both challenging and consistent with industry realities. Represented as: ; in, These are thresholds for extreme electricity price fluctuations set by experts. It represents the electricity price fluctuation range generated from the actual samples of generator G. These are constraint weights. It is a backup loss item; After obtaining a specific incremental GAN ​​architecture, input real-time low-latency power data streams. And the scenario control vector z used to generate extreme scenarios, outputting market structure data containing extreme scenarios. l is the index for market structure data; The process of obtaining dynamic return matrix and adversarial risk data is as follows: Based on adversarial machine learning, the game players in the electricity spot market are defined as power generation companies, electricity sales companies, and grid operators, denoted as set. where i is the number of subjects, and each subject contains a policy space. ; The strategy space corresponding to each subject This includes predefined decisions for power generation companies' pricing strategies and power sales companies' load dispatching strategies; Market structure data As the target entity, simulate the strategic confrontation between various entities under extreme scenarios, for each type of market structure data. Calculate the market structure data of different entities' chosen strategies. The following calculation process is used to determine the profit M: Define the payoff function of each player in the electricity market as follows: The cost function is ,main body Choosing the j-th strategy is denoted as The strategies chosen by other subjects are Market structure data is Then the formula for each entity's profit M is: ; in, Representing the subject Choose the j-th strategy Strategies for other subjects to choose Market structure data The unit revenue below Representing the subject Execution strategy The cost, for The corresponding scale, such as power generation and electricity sales; Then the acquired revenue data Represented as a matrix with dimensions of number of subjects × number of strategies × number of market structure data, i.e. , where j is the number of strategies, to obtain a dynamic payoff matrix; based on Calculate risk loss and obtain adversarial risk data, assuming the subject Execution strategy Market structure data The following adversarial risk data is set as ,but: ,in, as the main body The initial expected revenue threshold; Specifically, this formula is used to calculate adversarial risk data. It is the subject's expected return threshold, obtained by comparing expected returns with actual returns from the dynamic return matrix. In contrast, when the actual return is lower than expected, the difference between the two is the risk loss; if the actual return is not lower than expected, the risk loss is 0. This quantifies the risk exposure of the subject under different strategies and market structure data, and better reflects the risk characteristics in adversarial scenarios.

[0020] Intelligent decision-making module: Based on the dynamic profit matrix and adversarial risk data, a risk stress field model is constructed, and an online adversarial learning mechanism is introduced to optimize the risk stress field model to obtain an intelligent agent for power trading risk; Based on dynamic return matrix and opposing risk data The core idea of ​​constructing a risk stress field model is to regard the risks of each entity in the electricity spot market as stresses, and the interaction of risks formed by the confrontation of different entities' strategies constitutes a field. The overall risk is assessed by quantifying this stress field. Define risk stress as , indicating the subject Market structure data Next selection strategy The formula for calculating the rate of change in risk resulting from a unit strategy adjustment is as follows: That is, adversarial risk data strategy The partial derivatives of the equation reflect the sensitivity of the strategy to risk. Then, the risk stress of each entity is applied to market structure data. The following steps are performed to integrate and construct a risk stress field model. : , This represents the risk stress weighting coefficient; Specifically, risk stress field model This indicates the market structure data of the entire electricity spot market. The distribution and interaction of risks under the following conditions, , Dynamic return matrix Individual revenue data within; An online adversarial learning mechanism is introduced, in which real-time low-latency power data streams are input into an incremental GAN ​​architecture, which then generates new market structure data. It also simulates the strategic competition among various entities under this market structure data, generating a new dynamic payoff matrix. and adversarial risk data The implementation process is as follows: Construct a strategy adversarial evolution pool, classify the high-risk and low-risk strategies of each subject in history, and make the strategy selection of the subject in the new scenario clear about the value of the strategy in the strategy adversarial evolution pool. Specifically, based on the generated new dynamic return matrix and adversarial risk data For each subject Create a strategy category pool: High-risk strategy pool : Filter to meet strategy , The high-risk threshold is set at 0.7, which represents the top 30% of strategies in terms of risk loss in historical confrontations. Low-risk strategy pool : Filter to meet strategy , The low-risk threshold is set at 1.3, which represents the strategies that rank in the bottom 30% in terms of risk loss in historical confrontations. Then the strategy-versus-strategy evolution pool is represented as follows: ; When incremental GAN ​​architecture generates new market structure data Then, the strategy-based evolution pool is invoked, and the main body... The policy candidate set in the new scenario must satisfy: from 60% of the strategies are selected, prioritizing those with manageable risk. 30% of the strategies are randomly selected from non-extreme risk strategies to ensure diversity. Only 10% of the strategies are allowed to come from [the following sources]. By strictly limiting the use of high-risk strategies, a new dynamic return matrix is ​​ultimately generated. and adversarial risk data ; Based on the newly generated dynamic return matrix and adversarial risk data Calculate the new risk stress This leads to the new field strength function. ; Specifically, this process enables the risk stress field model to adapt to new market changes in real time, maintaining the accuracy of risk characterization; Then, the gradient descent method is used to optimize the parameters of the risk stress field model, and a loss function is defined. loss function It measures the difference between the risk field strength predicted by the model and the actual adversarial risk data; Finally, the model parameters are evaluated by calculating the loss function. gradient and in accordance with Update parameters, Using the learning rate, after multiple iterations (preset to 50), a power trading risk agent capable of accurately characterizing the risks of power spot trading is finally obtained. .

[0021] Risk prediction module: Inputs real-time collected power data streams into the power trading risk intelligent agent, outputs quantitative indicators of power risk, and realizes dynamic assessment of the risks of power spot trading strategies; The low-latency power data stream, collected in real time and preprocessed, is continuously input into the power trading risk intelligence agent. In China, intelligent agents for power trading risk After receiving the data stream, the internal incremental GAN ​​architecture is invoked to generate market structure data for current and potential extreme scenarios. Next, based on this market structure data and the generation logic of the dynamic payoff matrix, the strategy confrontation of each game player in a real-time scenario is simulated. The intelligent agent runs the risk stress and the field strength function of the risk stress field according to this data, and finally outputs a series of quantitative indicators of power risk. ,include: Single Entity Risk Indicators For example, the risk value under a specific pricing strategy of a power generation company; Overall market risk indicators Such as the total field strength of the market risk stress field; Extreme Scenario Risk Indicators Such as the maximum possible risk loss under the most extreme market structure; Specifically, single-entity risk indicators To quantify the risk level of a single market participant under its current strategy portfolio from the perspective of an individual market participant, and to help the participant identify the risk level of its own strategy; Overall market risk indicators From a macro perspective of the entire electricity spot market, the risk values ​​of all participants are integrated to reflect the overall risk distribution and interaction intensity of the market, providing market regulators with a quantitative overview of overall risk. Extreme Scenario Risk Indicators Focusing on extreme scenarios in the electricity market, such as the chain of risks caused by extreme weather and grid failures, we quantify the maximum potential risk loss in such scenarios. These quantitative indicators enable real-time acquisition of the risk exposure of electricity spot trading strategies under current and potential extreme scenarios, achieving dynamic assessment of the risks of electricity spot trading strategies and providing quantitative basis for trading entities to adjust their strategies and for market regulators to monitor risks.

[0022] Example 2: like Figure 2 As shown in the figure, this embodiment presents a method for dynamic risk assessment of electricity spot trading strategies based on real-time information, including: S1: By deploying edge computing nodes, multi-source heterogeneous data, including market grid quotations and load forecasts, are collected in real time and preprocessed to obtain low-latency power data streams; S2: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrices and adversarial risk data. The specific incremental GAN ​​architecture is based on existing GAN architectures, combined with the online reinforcement learning Actor-Critic framework, and the initial network structure and parameter implementation of the multi-task generator head are designed to determine the generator G and discriminator D. S3: Construct a risk stress field model based on a dynamic profit matrix and adversarial risk data, and introduce an online adversarial learning mechanism to optimize the risk stress field model, thereby obtaining an intelligent agent for power trading risk; S4: Input the real-time collected power data stream into the power trading risk intelligence agent, output power risk quantification indicators, and realize dynamic assessment of the risk of power spot trading strategies.

[0023] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0024] Example 3: This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method proposed in Embodiment 2 above.

[0025] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic risk assessment system for electricity spot trading strategies based on real-time information, characterized in that, include: Data acquisition module: Collects multi-source heterogeneous data, including market grid quotations and load forecasts, in real time by deploying edge computing nodes, and preprocesses the data to obtain low-latency power data streams; Risk acquisition module: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrix and adversarial risk data. Intelligent decision-making module: Based on the dynamic profit matrix and adversarial risk data, a risk stress field model is constructed, and an online adversarial learning mechanism is introduced to optimize the risk stress field model to obtain an intelligent agent for power trading risk; Risk prediction module: Input the real-time collected power data stream into the power trading risk intelligent agent, output power risk quantitative indicators, and realize dynamic assessment of the risk of power spot trading strategy.

2. The dynamic risk assessment system for electricity spot trading strategies based on real-time information as described in claim 1, characterized in that, The process of obtaining low-latency power data streams is as follows: Edge computing devices are deployed to collect raw multi-source heterogeneous data. Outliers and missing values ​​are removed from the raw multi-source heterogeneous data. The min-max standardization method is used to standardize any value x in the dataset. Then, time series data from different sources are aligned by timestamp to form multi-dimensional data records indexed by time, resulting in a low-latency power data stream.

3. The dynamic risk assessment system for electricity spot trading strategies based on real-time information according to claim 2, characterized in that, The process of obtaining a specific incremental GAN ​​architecture is as follows: Based on the basic GAN architecture, a low-latency power data stream is input, and a sliding time window is constructed using the Flink streaming engine. The electric feature representation f is obtained through the Transformer temporal embedding layer, and f is input into the generator G and discriminator D of a specific incremental GAN ​​architecture. This paper introduces an online reinforcement learning Actor-Critic framework. The Actor takes the power feature representation f as input and the intermediate feature output of the generator G as input, outputting a scenario control vector z. This scenario control vector z is then input into the generator G. The Critic framework takes the output samples of the generator G and the real samples as input to calculate a risk discrimination index. The reward function is optimized and fed back to the Actor framework to drive the generator G. Finally, the loss function of the multi-task generator head is used to train the generator G and the discriminator D simultaneously to obtain a specific incremental GAN ​​architecture. The loss function of the multi-task generation head includes normal scene loss and extreme scene loss. The normal scene loss is represented by Wasserstein distance loss, while the extreme scene loss is represented by constraint loss combined with domain knowledge.

4. The dynamic risk assessment system for electricity spot trading strategies based on real-time information as described in claim 1, characterized in that, The process of obtaining the dynamic return matrix and adversarial risk data is as follows: Based on adversarial machine learning, this paper defines the game players in the electricity spot market, each with a strategy space. Using market structure data as the target, it simulates the strategy adversarial interactions between the players under extreme scenarios. For each type of market structure data, the payoff M of different players under their chosen strategies is calculated, resulting in a dynamic payoff matrix. , An index for market structure data, based on a dynamic return matrix. Calculate risk loss to obtain adversarial risk data.

5. The dynamic risk assessment system for electricity spot trading strategies based on real-time information according to claim 4, characterized in that, The calculation process for the returns M of different agents under market structure data for their chosen strategies is as follows: Define the payoff function of each player in the electricity market as follows: The cost function is For the main body The selected j-th strategy, the strategies chosen by other participants, and market structure data are analyzed using a return function. for To obtain the unit revenue under their combined effect, through the cost function. Acquisition subject The cost of implementing the j-th strategy is calculated by multiplying the unit revenue by the size of the entity when implementing the j-th strategy. Subtract the main body The cost of executing the j-th strategy and the profit M obtained.

6. The dynamic risk assessment system for electricity spot trading strategies based on real-time information according to claim 5, characterized in that, The process of constructing the risk stress field model is as follows: Define risk stress as , indicating the subject Market structure data Next selection strategy At that time, the rate of change in risk resulting from unit strategy adjustments will reflect the risk stress of each entity in market structure data. The following steps are performed to integrate and construct a risk stress field model. , ,in, This represents the risk stress weighting coefficient. This represents the output of the risk stress field model.

7. The dynamic risk assessment system for electricity spot trading strategies based on real-time information according to claim 6, characterized in that, The process of optimizing the risk stress field model is as follows: An online adversarial learning mechanism is introduced, inputting real-time collected low-latency power data streams into an incremental GAN ​​architecture, and applying this mechanism to each agent. Create a strategy-based evolution pool ; When incremental GAN ​​architecture generates new market structure data Then, the strategy adversarial evolution pool is invoked to generate a new dynamic payoff matrix and adversarial risk data. Based on the newly generated dynamic payoff matrix and adversarial risk data, a new risk stress is calculated. This leads to the new field strength function. Then, the gradient descent method is used to optimize the parameters of the risk stress field model, and the model parameters are optimized by calculating the loss function. gradient And update the parameters to obtain the power trading risk intelligent agent. ,in, This is the learning rate.

8. The dynamic risk assessment system for electricity spot trading strategies based on real-time information according to claim 7, characterized in that, Strategy Battle Evolution Pool Represented as: Strategy Battle Evolution Pool Including high-risk strategy pool and low-risk strategy pool ,in: High-risk strategy pool : Filter to meet strategy, The high-risk threshold is defined as the strategies that account for the top 30% of risk losses in historical confrontations. Low-risk strategy pool : Filter to meet strategy, The low-risk threshold is defined as the strategies that rank in the bottom 30% in terms of risk loss in historical confrontations.

9. A method for dynamic risk assessment of electricity spot trading strategies based on real-time information, characterized in that, include: S1: By deploying edge computing nodes, multi-source heterogeneous data, including market grid quotations and load forecasts, are collected in real time and preprocessed to obtain low-latency power data streams; S2: Market structure data is constructed based on low-latency power data streams using a specific incremental GAN ​​architecture. Based on the market structure data, adversarial machine learning is used to generate dynamic payoff matrices and adversarial risk data. The specific incremental GAN ​​architecture is based on existing GAN architectures, combined with the online reinforcement learning Actor-Critic framework, and the initial network structure and parameter implementation of the multi-task generator head are designed to determine the generator G and discriminator D. S3: Construct a risk stress field model based on a dynamic profit matrix and adversarial risk data, and introduce an online adversarial learning mechanism to optimize the risk stress field model, thereby obtaining an intelligent agent for power trading risk; S4: Input the real-time collected power data stream into the power trading risk intelligence agent, output power risk quantification indicators, and realize dynamic assessment of the risk of power spot trading strategies.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images