Artificial intelligence-based industrial virtual power plant flexibility resource capacity aggregation method

Abstract

The application discloses an industrial virtual power plant flexibility resource capacity aggregation method based on artificial intelligence, belongs to the field of power system dispatching and energy management, and comprises the following steps: S1, generating dynamic model parameters containing constraint boundaries and output prediction curves; S2, scoring the generated dynamic model parameters containing constraint boundaries and output prediction curve results, and performing priority sorting according to the scoring results; S3, generating an optimal flexibility resource aggregation capacity distribution scheme based on the priority sorting results; S4, determining a feasible declaration scheme of the industrial virtual power plant; and S5, dynamically updating steps S2, S3 and S4 by using real-time operation data. The industrial virtual power plant flexibility resource capacity aggregation method based on artificial intelligence realizes accurate quantification and efficient aggregation of the industrial virtual power plant flexibility resource capacity by fusing an AI algorithm with multi-dimensional feature extraction and industrial scene constraint modeling.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and energy management technology, and in particular to an artificial intelligence-based method for aggregating the flexibility and capacity of industrial virtual power plants. Background Technology

[0002] As the new power system evolves towards "high proportion of renewable energy and high proportion of power electronic equipment", the fluctuation of the system's net load has intensified. Therefore, it is urgent to mobilize flexible resources on both the source and load sides to participate in regulation in order to maintain the balance between supply and demand.

[0003] As a core carrier for integrating dispersed and flexible resources in industrial scenarios, industrial virtual power plants aggregate resources such as photovoltaics, energy storage, and interruptible production lines to form a unified dispatch unit. They can provide ancillary services such as peak shaving and frequency regulation to the power grid, and help industrial enterprises reduce energy costs through electricity market transactions, thus becoming a key technological direction for resolving the contradiction between energy supply and demand. Unlike virtual power plants geared towards residential use, industrial virtual power plants have characteristics such as large scale, strong coupling, and complex constraints—their adjustable loads are deeply integrated with production processes, energy storage equipment needs to balance backup power supply and peak shaving needs, and electric vehicle charging behavior is highly correlated with plant transportation plans. This places stringent requirements on the accuracy, dynamic adaptability, and industrial constraint compatibility of capacity aggregation methods.

[0004] Currently, industrial virtual power plant flexibility resource capacity aggregation technology is mainly divided into three categories:

[0005] 1. Mathematical Programming-Based Aggregation Methods: These methods construct optimization models that incorporate constraints such as upper and lower limits of resource power, ramp rate, and state of charge (SOC) to solve for aggregated capacity with the objective of minimizing operating costs or maximizing revenue. These methods have the following inherent drawbacks: First, they rely on precise resource parameters and operational scenario assumptions. When uncertain events such as sudden changes in industrial production load, energy storage equipment failure, or temporary use of electric vehicles occur, the model deviates significantly from actual operation, leading to a substantial decrease in aggregation accuracy. Second, as the amount of resources increases, the complexity of model solving grows exponentially, and real-time response latency typically exceeds the second level, failing to meet the millisecond-level control requirements of industrial scenarios.

[0006] 2. Rule-based aggregation method: Capacity aggregation is achieved by setting preset resource response thresholds and scheduling rules. This type of method is simple to implement, but the rules are all set based on historical experience and cannot dynamically adapt to changes in industrial production patterns or fluctuations in electricity market prices. In addition, it has no self-repair capability for sudden disturbances. For example, when a production line is temporarily shut down and the load drops sharply, it is easy to cause the system-level response deviation to exceed the threshold and trigger grid penalties.

[0007] 3. Aggregation methods based on traditional artificial intelligence: These methods employ deep neural networks, support vector machines, and other models to train aggregation capacity prediction models based on historical data. However, they have the following significant limitations: First, most models are trained offline, lacking an online update mechanism. When new types of resources are added (e.g., new vanadium redox flow storage) or the operating scenario changes (e.g., production line expansion), the model's generalization ability is insufficient, requiring retraining and causing response interruptions. Second, they do not fully integrate industrial production constraints with real-time operating data. The aggregation schemes output by the model often conflict with equipment operating limits (e.g., SOH limits for energy storage) or production demands (e.g., minimum load guarantees), limiting their engineering practicality. Summary of the Invention

[0008] The purpose of this invention is to provide an artificial intelligence-based method for aggregating the flexibility and capacity of industrial virtual power plants, thereby solving the aforementioned technical problems.

[0009] To achieve the above objectives, this invention provides an artificial intelligence-based method for aggregating the flexibility and capacity of industrial virtual power plants, comprising the following steps:

[0010] S1. Based on the differences in physical characteristics and operational constraints of flexible resources in industrial virtual power plants, AI is used to predict and generate dynamic model parameters and output prediction curves that include constraint boundaries.

[0011] S2. Using multidimensional evaluation indicators, score the results generated in step S1 and prioritize them according to the scoring results.

[0012] S3. Based on the priority ranking results obtained in step S2, and combined with market data, a multi-agent proximal strategy optimization algorithm with shared parameters is used to generate the optimal flexible resource aggregation capacity allocation scheme.

[0013] S4. Based on the optimal flexible resource aggregation capacity allocation scheme and market uncertainty parameters, and by integrating robust optimization and stochastic game theory, determine the feasibility application scheme for industrial virtual power plants.

[0014] S5. Dynamically update steps S2, S3, and S4 using real-time running data.

[0015] Therefore, the above-mentioned artificial intelligence-based industrial virtual power plant flexibility resource capacity aggregation method has the following beneficial effects:

[0016] 1. Significantly improved aggregation accuracy: Through in-depth mining of multi-source data such as energy storage SOC / SOH, EV travel patterns, and industrial load process constraints, the prediction error of aggregation capacity is controlled within 5%, which is more than 40% lower than the traditional linear aggregation method, accurately depicting the actual adjustment potential of resources.

[0017] 2. Enhanced compatibility with industrial production: By introducing constraints, dynamic adaptation of "adjustment instructions - production demand" is achieved in aggregate scheduling, reducing the risk of industrial equipment downtime by 90% during resource response and completely avoiding interference with production continuity.

[0018] 3. Win-win situation for market revenue and grid value: By relying on AI models to predict market prices and grid congestion, the revenue from participating in the ancillary services market can be increased by 25%-30%. At the same time, through precise peak shaving, the peak-valley difference of the regional power grid can be reduced by 15%, and the renewable energy consumption rate can be increased by 8%-12%.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the AI-based industrial virtual power plant flexibility resource capacity aggregation method of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0022] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, the AI-based method for aggregating the flexibility and capacity of industrial virtual power plants includes the following steps:

[0025] S1. Based on the differences in physical characteristics and operational constraints of flexible resources in industrial virtual power plants, AI is used to predict and generate dynamic model parameters and output prediction curves that include constraint boundaries.

[0026] Step S1 specifically includes the following steps:

[0027] S11. Data Acquisition: Real-time acquisition of energy storage data, mobile data, and industrial load to obtain raw datasets. After data cleaning, interpolation, and standardization, a standardized feature matrix is ​​obtained. Among them, energy storage data includes state of charge, health status, real-time charging and discharging power, and charging and discharging efficiency; mobility data includes real-time charging power, state of charge, travel time, mileage, and charging interface connection status of electric vehicles; and industrial load includes operating parameters of industrial equipment.

[0028] S12. Based on the standardized feature matrix, and using demand-side resource heterogeneous modeling logic-industrial load constraint modeling, construct resource dynamic model, energy storage resource dynamic model and adjustable industrial load dynamic model respectively.

[0029] The dynamic resource model expression is as follows:

[0030] ;

[0031] In the formula, and They represent the first The minimum and maximum charging and discharging power of an electric vehicle; express Time of the first The response power of an electric vehicle; Indicates the first The state-of-charge travel threshold for electric vehicles, and , Indicates the first The average daily travel distance of an electric vehicle Indicates the standard deviation of travel distance. Indicates the first The driving range of a fully charged electric vehicle; express Time of the first The state of charge of an electric vehicle, and its dynamic update equation is: , express Time of the first The state of charge of an electric vehicle. and They represent the first The charging and discharging efficiency of an electric vehicle. and They represent Time of the first The charging and discharging power of an electric vehicle Indicates the time step. Indicates the first The rated capacity of the battery of an electric vehicle; This indicates the upper limit of the state of charge of an electric vehicle;

[0032] The dynamic model expression for energy storage resources is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] In the formula, and They represent Time and The first moment The state of charge of each energy storage unit; and They represent the first The charging and discharging efficiency of each energy storage unit; and They represent The first moment The charging and discharging power of each energy storage unit; Indicates the first The rated capacity of each energy storage unit; and They represent The time and the first time of the initial time The health status of each energy storage unit; Indicates the attenuation coefficient; and They represent the first During the first charge and discharge cycle The charging and discharging power of each energy storage unit; Indicates the first Rated power of each energy storage unit; express The first moment The actual response power of each energy storage unit;

[0037] The dynamic model expression for adjustable industrial load is as follows:

[0038] ;

[0039] and ;

[0040] ;

[0041] In the formula, and They represent Time of the first Actual power and reference power of each industrial load; express Time of the first Adjustable power for each industrial load; Indicates the first Maximum adjustable power for each industrial load; Indicates the first Minimum operating power for each industrial load; express Time of the first Adjustable power for each industrial load; Indicates the first Maximum power change rate of each industrial load;

[0042] S13. Using the resource dynamic model, energy storage resource dynamic model and adjustable industrial load dynamic model constructed in step S12 as constraint boundaries, construct a GNMTL-LSTM model, integrate resource operating status, environmental factors and production plan, and predict the output curve and error range of each resource.

[0043] The GNMTL-LSTM model described in step S13 includes a first shared feature backbone network and a dedicated detection head. The first shared feature backbone network includes two LSTM layers for extracting global temporal features. , express The hidden layer output of the LSTM layer at any given time. express The state of the hidden layer is always hidden. Represents the normalized feature matrix The feature values ​​in the data; the dedicated detection head includes three independent fully connected layers corresponding to resource operation data, production constraint data, and environmental auxiliary data, respectively. The expressions for the three independent fully connected layers are as follows:

[0044] ;

[0045] ;

[0046] ;

[0047] In the formula, , and They represent Real-time energy storage data, mobile data, and power output forecasts for industrial loads; , and All represent network parameters;

[0048] The expression for the multi-objective loss function is as follows:

[0049] ;

[0050] in,

[0051] ;

[0052] ;

[0053] In the formula, Indicates the multi-objective loss value; and Both represent loss weights; express The predicted output value at any given time, and ; express Real-time output value; Indicates the total number of time steps; and These represent the mean square error and the root mean square error, respectively.

[0054] The expression for the error interval is as follows:

[0055] ;

[0056] In the formula, express Time prediction error; Indicates the standard deviation of the prediction error; The Z-value corresponds to the 95% confidence interval.

[0057] S2. Using multidimensional evaluation indicators, score the results generated in step S1 and prioritize them according to the scoring results.

[0058] Step S2 specifically includes the following steps:

[0059] S21. Construct a multi-dimensional evaluation system consisting of response capability indicators, response reliability indicators, and response flexibility indicators;

[0060] The expression for the response capability index is as follows:

[0061] ;

[0062] ;

[0063] In the formula, express Time Resources Maximum effective response power; This represents a 0-1 variable, where 0 represents a resource. exist If a resource is not responding at all, 1 indicates that the resource is not responding. exist Responding at all times; This represents a 0-1 variable, where 1 represents a resource. exist Always conforming to industrial production constraints, 0 represents resources exist It is constantly inconsistent with the constraints of industrial production; Representing resources In the The power of the next historical response; Representing resources The shortest duration to maintain an effective response; and Representing resources No. The start and end times of each valid response;

[0064] The expression for the response reliability index is as follows:

[0065] ;

[0066] ;

[0067] In the formula, express Time Resources The standard deviation of the response power; express Time Resources The historical average response power, and ; This represents the number of historical response samples; Representing resources Historical effective response completion rate; Representing resources The number of historical valid response completions; Representing resources Total number of valid historical responses;

[0068] The expression for the response flexibility index is as follows:

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] In the formula, Representing resources Response speed; Representing resources The number of valid response tests; Indicates the first The adjustable power range of an industrial load; Indicates the first The adjustable power range of an electric vehicle; Indicates the state of charge limit of an electric vehicle; Indicates the first The adjustable power range of each energy storage unit; and Describe the upper and lower limits of the state of charge of energy storage, respectively; Indicates the first The state of charge of each energy storage unit;

[0074] S22. Weighted scoring combining subjective and objective factors:

[0075] ;

[0076] in,

[0077] ;

[0078] In the formula, Representing resources Flexibility score; Representing resources No. The quantitative value of each evaluation indicator, and , respectively representing the response capability index, response reliability index, and response flexibility index; This indicates the weighting of the subjective-objective integration. Indicates subjective weight, and , Experts indicated their opinion on the first Scores for each evaluation indicator; Indicates objective weight, and , Indicates the first Information entropy of each evaluation indicator , , Representing resources No. The normalized value of each evaluation indicator. For constant terms, Indicates the total number of resources; and They represent the first The minimum and maximum values ​​of each evaluation indicator across all resources; Indicates the subjective bias coefficient;

[0079] S23. Score the flexibility obtained in step S22 from high to low. Sort the data and use the flexibility score as the priority ranking result.

[0080] S3. Based on the priority ranking results obtained in step S2, and combined with market data, a multi-agent proximal strategy optimization algorithm with shared parameters is used to generate the optimal flexible resource aggregation capacity allocation scheme.

[0081] Step S3 specifically includes the following steps:

[0082] S31. Construct an objective function that balances maximizing market returns and minimizing deviation risk:

[0083] ;

[0084] in,

[0085] ;

[0086] ;

[0087] In the formula, Indicates the target optimization value; and All represent preference weights; This refers to the potential market benefits of an aggregate (specifically, an Industrial Virtual Power Plant (IVPP) that forms a unified response and trading entity by aggregating multiple types of flexible resources (such as energy storage, EVs, and adjustable industrial loads). It is a 0-1 variable, where 1 represents The aggregated capacity meets market entry requirements at all times; express Real-time market electricity prices; This represents the set of resources participating in the aggregation; express Time allocated to resources The polymerization capacity; This indicates the total deviation penalty risk of the aggregate; and Risk weights, representing the range and degree of volatility, respectively; express Penalty price for grid deviation at any time;

[0088] S32. Construct a parameter-sharing PS-PPO network, wherein the PS-PPO network includes a second shared backbone network and a dedicated policy head. The second shared backbone network includes two independent fully connected layers activated by the ReLU function for global feature extraction. , This represents the global feature vector output by the second shared backbone network; express Time Resources The state of charge; the dedicated policy header includes 3 independent fully connected layers, and the output aggregate capacity accounts for the proportion of the response capacity. , Representing resources The proportion of polymerization capacity to response capacity;

[0089] The loss function expression is as follows:

[0090] ;

[0091] In the formula, Indicates the loss value due to policy updates; Expressing expected value based on experience; Indicates current policy In state Choose below The probability of; Representing historical policies In state Choose below general; Denotes the dominance function, and , Representing state Execute The value of the action, and regard it as The estimated value, Representing state The state value, and , Let the value network be represented by the loss function. , Indicates value network loss. Indicates the value of the target state; Indicates the cutting factor;

[0092] S33, the resource dynamic model, the energy storage resource dynamic model, and the adjustable industrial load dynamic model, set the following boundary conditions for aggregated capacity:

[0093] Power balance constraints: ;

[0094] Resource output constraints: ;

[0095] Energy storage SOC constraints:

[0096] ;

[0097] Industrial load production constraints: ;

[0098] In the formula, express Real-time virtual power plant aggregated total capacity;

[0099] S34. Input the standardized feature matrix obtained in step S11 into the PS-PPO network, train the PS-PPO network with the constraints added in step S33 based on the Ray RLlib distributed framework until convergence, and input the output prediction curve results into the converged PS-PPO network according to the priority sorting in step S2 to obtain the optimal flexible resource aggregation capacity allocation scheme.

[0100] S4. Based on the optimal flexible resource aggregation capacity allocation scheme and market uncertainty parameters, and by integrating robust optimization and stochastic game theory, determine the feasibility application scheme for industrial virtual power plants.

[0101] Step S4 specifically includes the following steps:

[0102] S41. Market Uncertainty Parameter Modeling:

[0103] Market price uncertainty: ;

[0104] Response bias uncertainty: ;

[0105] In the formula, express The actual market settlement price at any given time; express Real-time market price forecasts, which can be obtained from electricity market platforms; express The price fluctuation range is half-width at any given moment; Indicates the price volatility coefficient; Indicates the robustness coefficient; express Time Resources The half-width of the response deviation interval; Indicates the response deviation coefficient;

[0106] S42. Construct a robust transaction decision model and use a column constraint generation algorithm to decompose the robust decision problem into determining the optimal reporting strategy and searching for the worst scenario. Iterate the solution until convergence to obtain the robust optimal reporting strategy.

[0107] S43. Based on stochastic game theory, the industrial virtual power plant, power grid dispatching agency and resource users are set as the players. By constructing the three-party payoff function, the Nash equilibrium is solved to obtain the equilibrium strategy of the robust-stochastic game.

[0108] S44. Transform the equilibrium strategy of the robust-random game into a standardized declaration scheme, and verify the feasibility of the standardized declaration scheme to obtain a feasible declaration scheme.

[0109] Step S42 specifically includes the following steps:

[0110] S421. Define the robust objective function of the robust trading decision model:

[0111] ;

[0112] in,

[0113] ;

[0114] In the formula, They represent The virtual power plant at Time Factory declares its capacity and price to the market; and Let these represent the set of price uncertainty and the set of response bias uncertainty, respectively. In This means converting maximizing net profit into minimizing negative net profit; This represents the net revenue of the virtual power plant in the factory over 24 hours. and These represent the upper limit of the effective response power ratio and the lower limit of the deviation penalty ratio, respectively. express Time deviation penalty price; This indicates that the factory's virtual power plant is supplying resources. The acquisition price;

[0115] S422. Using a column constraint generation algorithm, the main problem and subproblems are solved iteratively. The main problem is to search for the optimal reporting strategy in a fixed worst-case scenario, and the subproblems are to search for the worst-case scenario in a fixed reporting strategy, until the worst-case scenario no longer changes. At this point, it is determined that convergence to the robust optimal solution has been achieved. ;

[0116] The main problem optimization model expression is as follows:

[0117] st AND ;

[0118] In the formula, This represents the net profit under the worst-case scenario, where... , and These represent the optimal bid price and bid capacity under the worst-case scenario, respectively; Representing resources Maximum response capacity;

[0119] The subproblem optimization model expression is as follows:

[0120] ;

[0121] In the formula, and They represent The lower and upper limits of the electricity market clearing price at any given time; Representing resources Minimum response capacity;

[0122] Convergence condition: ;

[0123] In the formula, and They represent the first The second iteration and the first The worst-case net gain of the next iteration.

[0124] Step S43 specifically includes the following steps:

[0125] S431. Define the following three-party revenue function:

[0126] Revenue function for industrial virtual power plants: ;

[0127] Revenue function of power grid dispatching agency: ;

[0128] Resource user revenue function: ;

[0129] In the formula, , and These represent the revenues of the industrial virtual power plant, the power grid dispatching agency, and the resource user, respectively. Indicates the coefficient of synergy; express The capacity required for the flexibility of the power grid dispatching agency at all times; This indicates the cost of expanding power grid capacity to address congestion. express Real-time grid congestion forecast capacity can be provided by the grid dispatching agency; This indicates the compensation price paid by the power grid dispatching agency to industrial virtual power plants. Indicates the retail electricity price for users;

[0130] S432, Solving for Nash equilibrium;

[0131] S4321. Initialize the robust optimal solution ;

[0132] S4322, In the robust optimal solution Under the premise of adjustment until To maximize the optimal action of the power grid dispatching agency;

[0133] S4323. Under the premise of optimal action of the power grid dispatching agency, adjust... until The maximum value is obtained, which yields the optimal action for the resource user.

[0134] S4324. Under the premise of optimal action of resource users, adjust , making maximum;

[0135] S4325. Repeat steps S4322-S4324 until the strategies of the three parties remain unchanged, thus obtaining the Nash equilibrium strategy. , These represent the capacity and price that the virtual power plant, after being processed by the Nash equilibrium strategy, declares to the market.

[0136] The feasibility verification described in step S44 is to verify the Nash equilibrium strategy. Net profit under a fixed worst-case scenario Does it meet the following verification conditions:

[0137] Market capacity rule verification: If violated, then... ;

[0138] Aggregate capacity constraint verification: If violated, then... ;

[0139] Verification of non-negative returns: If violated, the value will be increased by the set interval. Recalculate , until the non-negative return verification condition is met;

[0140] In the formula, This indicates the maximum number of applications a single market entity can submit; This represents the capacity that the industrial virtual power plant, after undergoing aggregated capacity constraint verification, declares to the electricity market.

[0141] S5. Dynamically update steps S2, S3, and S4 using real-time running data.

[0142] Simulation Experiment

[0143] Scenario setting: Based on a virtual power plant in a chemical industrial park, it includes three types of core flexible resources: a 10MWh energy storage system (including 50 battery clusters), 80 in-plant logistics EVs, and 120 adjustable industrial loads (fans, water pumps, and waste heat boilers), simulating a day-ahead and real-time two-stage scheduling scenario.

[0144] Data sources: Energy storage data: 1-second data such as SOC (5%~95%) and charging / discharging power (±500kW) collected by the BMS system, totaling 7200 records / day; EV (electric vehicle) data: 5-minute data such as charging power (0~120kW), travel time (6:00-22:00), and historical mileage (0~200km) recorded by smart charging piles, totaling 288 records / vehicle / day; Industrial load data: 10-second data such as real-time power (10~200kW), process temperature (80~150℃), and production shifts (3 shifts) uploaded by the PLC system, totaling 8640 records / unit / day; External data: Congestion capacity prediction (0~5MW) and electricity market price range (0.3~1.2 yuan / kWh) provided by the power grid dispatch center.

[0145] Comparison of approaches: Approach 1: Traditional virtual battery aggregation method (ignoring resource heterogeneity); Approach 2: Inner approximation aggregation method based on Zonotope (no industrial constraint adaptation).

[0146] Key metrics: aggregated capacity prediction error, computation time, resource response success rate, and ancillary service revenue;

[0147] Experimental environment: Intel i9-13900K processor, 32GB memory, MATLAB R2023b simulation platform, using CVX optimization toolkit to solve the constraint problem.

[0148] Table 1 Simulation Experiment Results

[0149]

[0150] As shown in Table 1, the error of the present invention is reduced by 66.4% compared with comparative scheme 1 and by 46.2% compared with comparative scheme 2. This is because the attention mechanism accurately captures key features such as industrial load process constraints and EV travel randomness, avoiding the "oversimplification" or "conservative estimation" of traditional methods.

[0151] The computation time of this invention is only 14.7% of that of comparative scheme 1 and 25.8% of that of comparative scheme 2. This is due to the feature point set optimization algorithm reducing the complexity of Minkowski and computation, while the parallel computing capability of the AI ​​model is adapted to large-scale resource scenarios.

[0152] The invention achieved a response success rate of 98.7% with no production downtime, proving that the introduction of constraint parameters can effectively avoid conflicts between scheduling instructions and industrial production, and solve the pain point of traditional aggregation methods that "emphasize adjustment and neglect production".

[0153] The ancillary service revenue of this invention is 30.7% higher than that of the comparative solution 1. This is mainly because the AI ​​model accurately predicts market price fluctuations and grid congestion periods, enabling the aggregation of resources to achieve revenue superposition in multiple markets such as frequency regulation and reserve, thus verifying the commercial potential of the solution.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for aggregating the flexibility resource capacity of an industrial virtual power plant based on artificial intelligence, characterized by: Includes the following steps: S1. Based on the differences in physical characteristics and operational constraints of flexible resources in industrial virtual power plants, AI is used to predict and generate dynamic model parameters and output prediction curves that include constraint boundaries. S2. Using multidimensional evaluation indicators, score the results generated in step S1 and prioritize them according to the scoring results. S3. Based on the priority ranking results obtained in step S2, and combined with market data, a multi-agent proximal strategy optimization algorithm with shared parameters is used to generate the optimal flexible resource aggregation capacity allocation scheme. S4. Based on the optimal flexible resource aggregation capacity allocation scheme and market uncertainty parameters, and by integrating robust optimization and stochastic game theory, determine the feasibility application scheme for industrial virtual power plants. Step S1 specifically includes the following steps: S11. Data Acquisition: Real-time acquisition of energy storage data, mobile data, and industrial load to obtain raw datasets. After data cleaning, interpolation, and standardization, a standardized feature matrix is ​​obtained. Among them, energy storage data includes state of charge, health status, real-time charging and discharging power, and charging and discharging efficiency; mobility data includes real-time charging power, state of charge, travel time, mileage, and charging interface connection status of electric vehicles; and industrial load includes operating parameters of industrial equipment. S12. Based on the standardized feature matrix, and using demand-side resource heterogeneous modeling logic-industrial load constraint modeling, construct resource dynamic model, energy storage resource dynamic model and adjustable industrial load dynamic model respectively. The dynamic resource model expression is as follows: ； In the formula, and They represent the first The minimum and maximum charging and discharging power of an electric vehicle; express Time of the first The response power of an electric vehicle; Indicates the first The state-of-charge travel threshold for electric vehicles, and , Indicates the first The average daily travel distance of an electric vehicle Indicates the standard deviation of travel distance. Indicates the first The driving range of a fully charged electric vehicle; express Time of the first The state of charge of an electric vehicle, and its dynamic update equation is: , express Time of the first The state of charge of an electric vehicle. and They represent the first The charging and discharging efficiency of an electric vehicle. and They represent Time of the first The charging and discharging power of an electric vehicle Indicates the time step. Indicates the first The rated capacity of the battery of an electric vehicle; This indicates the upper limit of the state of charge of an electric vehicle; The dynamic model expression for energy storage resources is as follows: ； ； ； In the formula, and They represent Time and The first moment The state of charge of each energy storage unit; and They represent the first The charging and discharging efficiency of each energy storage unit; and They represent The first moment The charging and discharging power of each energy storage unit; Indicates the first The rated capacity of each energy storage unit; and They represent The time and the first time of the initial time The health status of each energy storage unit; Indicates the attenuation coefficient; and They represent the first During the first charge and discharge cycle The charging and discharging power of each energy storage unit; Indicates the first The rated power of each energy storage unit; express The first moment The actual response power of each energy storage unit; The dynamic model expression for adjustable industrial load is as follows: ； and ; ； In the formula, and They represent Time of the first Actual power and reference power of each industrial load; express Time of the first Adjustable power for each industrial load; Indicates the first Maximum adjustable power for each industrial load; Indicates the first Minimum operating power for each industrial load; express Time of the first Adjustable power for each industrial load; Indicates the first Maximum power change rate of each industrial load; S13. Using the resource dynamic model, energy storage resource dynamic model and adjustable industrial load dynamic model constructed in step S12 as constraint boundaries, construct a GNMTL-LSTM model, integrate resource operating status, environmental factors and production plan, and predict the output curve and error range of each resource.

2. The method for aggregating flexible resource capacity in an industrial virtual power plant based on artificial intelligence according to claim 1, characterized in that: The GNMTL-LSTM model described in step S13 includes a first shared feature backbone network and a dedicated detection head. The first shared feature backbone network includes two LSTM layers for extracting global temporal features. , express The hidden layer output of the LSTM layer at any given time. express The state of the hidden layer is always hidden. Represents the normalized feature matrix The feature values ​​in the data; the dedicated detection head includes three independent fully connected layers corresponding to resource operation data, production constraint data, and environmental auxiliary data, respectively. The expressions for the three independent fully connected layers are as follows: ； ； ； In the formula, , and They represent Real-time energy storage data, mobile data, and power output forecasts for industrial loads; , and All represent network parameters; The expression for the multi-objective loss function is as follows: ； in, ； ； In the formula, Indicates the multi-objective loss value; and Both represent loss weights; express The predicted output value at any given time, and ; express Real-time output value; Indicates the total number of time steps; and These represent the mean square error and the root mean square error, respectively. The expression for the error interval is as follows: ； In the formula, express Time prediction error; Indicates the standard deviation of the prediction error; The Z-value corresponds to the 95% confidence interval.

3. The method for aggregating flexible resource capacity in an industrial virtual power plant based on artificial intelligence according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Construct a multi-dimensional evaluation system consisting of response capability indicators, response reliability indicators, and response flexibility indicators; The expression for the response capability index is as follows: ； ； In the formula, express Time Resources Maximum effective response power; This represents a 0-1 variable, where 0 represents a resource. exist If a resource is not responding at all, 1 indicates that the resource is not responding. exist Responding at all times; This represents a 0-1 variable, where 1 represents a resource. exist Always conforming to industrial production constraints, 0 represents resources exist It is constantly inconsistent with the constraints of industrial production; Representing resources In the The power of the next historical response; Representing resources The shortest duration to maintain an effective response; and Representing resources No. The start and end times of each valid response; The expression for the response reliability index is as follows: ； ； In the formula, express Time Resources The standard deviation of the response power; express Time Resources The historical average response power, and ; This represents the number of historical response samples; Representing resources Historical effective response completion rate; Representing resources The number of historical valid response completions; Representing resources Total number of valid historical responses; The expression for the response flexibility index is as follows: ； ； ； ； In the formula, Representing resources Response speed; Representing resources The number of valid response tests; Indicates the first The adjustable power range of an industrial load; Indicates the first The adjustable power range of an electric vehicle; Indicates the state of charge limit of an electric vehicle; Indicates the first The adjustable power range of each energy storage unit; and Describe the upper and lower limits of the state of charge of energy storage, respectively; Indicates the first The state of charge of each energy storage unit; S22. Weighted scoring combining subjective and objective factors: ； in, ； In the formula, Representing resources Flexibility score; Representing resources No. The quantitative value of each evaluation indicator, and , respectively representing the response capability index, response reliability index, and response flexibility index; This indicates the weighting of the subjective-objective integration. Indicates subjective weight, and , Experts indicated their opinion on the first Scores for each evaluation indicator; Indicates objective weight, and , Indicates the first Information entropy of each evaluation indicator , , Representing resources No. The normalized value of each evaluation indicator. For constant terms, Indicates the total number of resources; and They represent the first The minimum and maximum values ​​of each evaluation indicator across all resources; Indicates the subjective bias coefficient; S23. Score the flexibility obtained in step S22 from high to low. Sort the data and use the flexibility score as the priority ranking result.

4. The method for aggregating flexible resource capacity in an industrial virtual power plant based on artificial intelligence according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31. Construct an objective function that balances maximizing market returns and minimizing deviation risk: ； in, ； ； In the formula, Indicates the target optimization value; and All represent preference weights; This indicates the potential market benefits of the polymer; It is a 0-1 variable, where 1 represents The aggregated capacity meets market entry requirements at all times; express Real-time market electricity price; This represents the set of resources participating in the aggregation; express Time allocated to resources The polymerization capacity; This indicates the total deviation penalty risk of the aggregate; and Risk weights, representing the range and degree of volatility, respectively; express Penalty price for grid deviation at any time; S32. Construct a parameter-sharing PS-PPO network, wherein the PS-PPO network includes a second shared backbone network and a dedicated policy head. The second shared backbone network includes two independent fully connected layers activated by the ReLU function for global feature extraction. , This represents the global feature vector output by the second shared backbone network; express Time Resources The state of charge; the dedicated policy header includes 3 independent fully connected layers, and the output aggregate capacity accounts for the proportion of the response capacity. , Representing resources The proportion of polymerization capacity to response capacity; The loss function expression is as follows: ； In the formula, Indicates the loss value due to policy updates; Expressing expected value based on experience; Indicates current policy In state Choose below The probability of; Representing historical policies In state Choose below general; Denotes the dominance function, and , Representing state Execute The value of the action, and regard it as The estimated value, Representing state The state value, and , Let the value network be represented by the loss function. , Indicates value network loss. Indicates the value of the target state; Indicates the cutting factor; S33, the resource dynamic model, the energy storage resource dynamic model, and the adjustable industrial load dynamic model, set the following boundary conditions for aggregated capacity: Power balance constraints: ; Resource output constraints: ; Energy storage SOC constraints: ； Industrial load production constraints: ; In the formula, express Real-time virtual power plant aggregated total capacity; S34. Input the standardized feature matrix obtained in step S11 into the PS-PPO network, train the PS-PPO network with the constraints added in step S33 based on the Ray RLlib distributed framework until convergence, and input the output prediction curve results into the converged PS-PPO network according to the priority sorting in step S2 to obtain the optimal flexible resource aggregation capacity allocation scheme.

5. The method for aggregating the flexibility resource capacity of an industrial virtual power plant based on artificial intelligence according to claim 4, characterized in that: Step S4 specifically includes the following steps: S41. Market Uncertainty Parameter Modeling: Market price uncertainty: ; Response bias uncertainty: ; In the formula, express The actual market settlement price at any given time; express Forecasted market price at any given time; express The price fluctuation range is half-width at any given moment; Indicates the price volatility coefficient; Indicates the robustness coefficient; express Time Resources The half-width of the response deviation interval; Indicates the response deviation coefficient; S42. Construct a robust transaction decision model and use a column constraint generation algorithm to decompose the robust decision problem into determining the optimal reporting strategy and searching for the worst scenario. Iterate the solution until convergence to obtain the robust optimal reporting strategy. S43. Based on stochastic game theory, the industrial virtual power plant, power grid dispatching agency and resource users are set as the players. By constructing the three-party payoff function, the Nash equilibrium is solved to obtain the equilibrium strategy of the robust-stochastic game. S44. Transform the equilibrium strategy of the robust-random game into a standardized declaration scheme, and verify the feasibility of the standardized declaration scheme to obtain a feasible declaration scheme.

6. The method for aggregating the flexibility resource capacity of an industrial virtual power plant based on artificial intelligence according to claim 5, characterized in that: Step S42 specifically includes the following steps: S421. Define the robust objective function of the robust trading decision model: ； in, ； In the formula, They represent The virtual power plant at Time Factory declares its capacity and price to the market; and Let these represent the set of price uncertainty and the set of response bias uncertainty, respectively. In This means converting maximizing net profit into minimizing negative net profit; This represents the net revenue of the virtual power plant in the factory over 24 hours. and These represent the upper limit of the effective response power ratio and the lower limit of the deviation penalty ratio, respectively. express Time deviation penalty price; This indicates that the factory's virtual power plant is supplying resources. The acquisition price; S422. Using a column constraint generation algorithm, the main problem and subproblems are solved iteratively. The main problem is to search for the optimal reporting strategy in a fixed worst-case scenario, and the subproblems are to search for the worst-case scenario in a fixed reporting strategy, until the worst-case scenario no longer changes. At this point, it is determined that convergence to the robust optimal solution has been achieved. ; The main problem optimization model expression is as follows: st AND ； In the formula, This represents the net profit under the worst-case scenario, where... , and These represent the optimal bid price and bid capacity under the worst-case scenario, respectively; Representing resources Maximum response capacity; The subproblem optimization model expression is as follows: ； In the formula, and They represent The lower and upper limits of the electricity market clearing price at any given time; Representing resources Minimum response capacity; Convergence condition: ; In the formula, and They represent the first The second iteration and the first The worst-case net gain of the next iteration.

7. The method for aggregating flexible resource capacity in an industrial virtual power plant based on artificial intelligence according to claim 6, characterized in that: Step S43 specifically includes the following steps: S431. Define the following three-party revenue function: Revenue function for industrial virtual power plants: ; Revenue function of power grid dispatching agency: ; Resource user revenue function: ; In the formula, , and These represent the revenues of the industrial virtual power plant, the power grid dispatching agency, and the resource user, respectively. Indicates the coefficient of synergy; express The capacity required for the flexibility of the power grid dispatching agency at all times; This indicates the cost of expanding power grid capacity to address congestion. express Real-time power grid congestion capacity prediction; This indicates the compensation price paid by the power grid dispatching agency to industrial virtual power plants. Indicates the retail electricity price for users; S432, Solving for Nash equilibrium; S4321. Initialize the robust optimal solution ; S4322, In the robust optimal solution Under the premise of adjustment until To maximize the optimal action of the power grid dispatching agency; S4323. Under the premise of optimal action of the power grid dispatching agency, adjust... until The maximum value is obtained, which yields the optimal action for the resource user. S4324. Under the premise of optimal action of resource users, adjust , making maximum; S4325. Repeat steps S4322-S4324 until the strategies of the three parties remain unchanged, thus obtaining the Nash equilibrium strategy. , These represent the capacity and price that the virtual power plant, after being processed by the Nash equilibrium strategy, declares to the market.

8. The method for aggregating flexible resource capacity in an industrial virtual power plant based on artificial intelligence according to claim 7, characterized in that: The feasibility verification described in step S44 is to verify the Nash equilibrium strategy. Net profit under a fixed worst-case scenario Does it meet the following verification conditions: Market capacity rule verification: If violated, then... ; Aggregate capacity constraint verification: If violated, then... ; Verification of non-negative returns: If violated, the interval value will be increased. Recalculate , until the non-negative return verification condition is met; In the formula, This indicates the maximum number of applications a single market entity can submit; This represents the capacity that the industrial virtual power plant, after undergoing aggregated capacity constraint verification, declares to the electricity market.

9. The method for aggregating the flexibility resource capacity of an industrial virtual power plant based on artificial intelligence according to claim 8, characterized in that: Step S4 is followed by step S5, which uses real-time running data to dynamically update steps S2, S3 and S4.

Images