Virtual power plant multi-objective robust optimization ai regulation method, system and device

Abstract

The application discloses a virtual power plant multi-target robust optimization AI regulation method, system and device, comprising: predicting multi-element uncertain information of a virtual power plant; based on the multi-element uncertain information, AI evaluating the multi-time scale adjustable capacity of the virtual power plant; constructing a virtual power plant capacity-power-carbon emission multi-dimensional coordination model, and utilizing the virtual power plant capacity-power-carbon emission multi-dimensional coordination model to coordinate the weight of the capacity, power and carbon emission budget; based on the evaluation result of the multi-time scale adjustable capacity of the virtual power plant and the weight obtained by coordination, constructing a learning enhanced virtual power plant multi-target robust optimization model, and utilizing the virtual power plant multi-target robust optimization model to optimize and regulate the virtual power plant. The application significantly improves the intelligent regulation and control capability of the virtual power plant under uncertainty and multi-target trade-off, realizes end-to-end closed-loop optimization from prediction, evaluation, coordination to execution, and promotes stable, efficient and green operation in the market.

Description

Technical Field

[0001] This invention relates to the field of power technology, specifically to a multi-objective robust optimization AI control method, system, and equipment for virtual power plants. Background Technology

[0002] With the rapid development of new power systems, virtual power plants (VPPs), as key carriers for integrating distributed energy and realizing coordinated operation of power generation, grid, load and storage, are gradually moving towards the core of engineering practice and market transactions.

[0003] Currently, virtual power plants face the dual challenges of various uncertainties and multi-objective optimization requirements in actual operation, particularly exhibiting significant technical bottlenecks in the following aspects: First, the operation of virtual power plants relies on heterogeneous data sources such as electricity prices, carbon emission factors, meteorological conditions, and external renewable energy output. This information is highly uncertain, nonlinear, and time-series oriented, making it difficult for traditional single-prediction models to capture their joint evolution patterns. Especially under extreme scenarios, they exhibit poor robustness, leading to strong dependence on control strategies and poor generalization. Second, virtual power plants possess typical "multi-device, multi-timescale" coupling characteristics. Adjustability depends not only on the performance of individual devices but also on multiple influences from inter-device coordination, network constraints, and scheduling mechanisms. Current mainstream methods struggle to achieve unified assessment of control capabilities across the day-ahead, intraday, and online timescales, especially lacking systematic modeling in areas such as multi-timescale capability continuity, response timeliness, and implementation. Furthermore, with the advancement of capacity markets and carbon trading mechanisms, the operational objectives of virtual power plants are gradually shifting from a single focus on economic optimization to a multi-dimensional coordination of objectives including capacity commitments, power generation plans, and carbon emission budgets. However, current regulatory mechanisms often suffer from a mismatch between capacity incentives and implementation effectiveness, lacking targeted capacity compensation mechanisms and failing to unleash the potential of capacity as a resource for system flexibility.

[0004] In summary, there is an urgent need for a method that can achieve full-process optimization control of virtual power plants under uncertain environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a multi-objective robust optimization AI control method, system, and equipment for virtual power plants, which can significantly improve the operational efficiency and market responsiveness of virtual power plants under multi-objective and multi-disturbance scenarios.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] This invention is a multi-objective robust optimization AI control method for virtual power plants, comprising:

[0008] Predicting diverse and uncertain information about virtual power plants;

[0009] Based on diverse and uncertain information, AI assesses the multi-timescale adjustability of virtual power plants;

[0010] Construct a multi-dimensional coordination model of virtual power plant capacity, electricity, and carbon emissions, and use the virtual power plant capacity, electricity, and carbon emissions multi-dimensional coordination model to coordinate the weights of capacity, electricity, and carbon emissions budgets;

[0011] Based on the evaluation results of the virtual power plant's multi-timescale adjustable capabilities and the weights obtained through coordination, a learning-enhanced multi-objective robust optimization model for the virtual power plant is constructed, and the virtual power plant is optimized and controlled using this model.

[0012] A further improvement of the present invention lies in: predicting multivariate uncertain information of a virtual power plant, including:

[0013] The historical measurement data sequence of the virtual power plant is aligned and cleaned with the external feature sequence. Point predictions of uncertain information are obtained through a multimodal time series model. A diffusion model is used to generate a full-time-domain scenario including extreme disturbances. Combining the Wasserstein-DRO and CVaR methods, a robust tail risk boundary is formed, expressed as:

[0014] ;

[0015] ;

[0016] ;

[0017] In the formula, For the future Multivariate point prediction Index for the current discrete time. For multimodal time series models, This is a historical measurement data sequence for a virtual power plant. For external feature sequences, Forward stride length, For the first The generated full-time-domain multivariate scenario trajectory, For the number of scenarios, To generate samples according to a given generative model. For parameters diffusion model, include and , The predicted quantity is the historical measurement data sequence of the virtual power plant. For robust sets, For candidate true distributions, For the empirical distribution prediction, The first-order Wasserstein distance. The radius is the Wasserstein radius. For confidence level of The conditional value at risk operator, where Z is the random loss. To calibrate the quantile threshold, To test the candidate true distribution within the robust set Take the expected value. The positive part operator is T, where T is the total number of time periods for horizon prediction. This is a full horizon point prediction sequence. For a set of scenarios, , It is a unified data package that includes point prediction, scenario set, robust set, and conditional value of risk operator.

[0018] A further improvement of this invention lies in: AI-based assessment of the multi-timescale adjustable capability of a virtual power plant, based on multivariate uncertain information. Specifically, this includes: constructing an adjustable capability assessment model for three timescales—day-ahead, intraday, and online—based on predicted multivariate uncertain information; evaluating the boundaries of the adjustable capability at each timescale; and obtaining the executable capability envelope, expressed as:

[0019] ;

[0020] ;

[0021] ;

[0022] In the formula, For parameters Adjustable capability evaluator , , These are respectively: today, today, and online. These represent capabilities for the day before, intraday, and online. For scale The lower bound of the confidence level of the capability. For scale The confidence level of the capability For quantile operations, To assess capabilities based on scenario sets and robustness sets, It provides indexes for three time scales: day-to-day, intraday, and online. To adjust the capacity upwards or downwards, The longest continuously adjustable duration. This represents the maximum available total battery capacity. For the maximum climbing speed, Let be the probability of the candidate true distribution within the robust set. As a condition, The confidence threshold. These represent the minimum capabilities for the day before, intraday, and online, respectively. This is the set of lower limits for capabilities within the current day, intraday, and online. For robust sets, For candidate true distributions, For a set of scenarios, This is a predicted sequence of points across the entire horizon.

[0023] A further improvement of this invention is that the virtual power plant capacity-electricity-carbon emission multidimensional coordination model includes an upper layer and a lower layer. The upper layer learns capacity, electricity and carbon emission price signals, while the lower layer solves for the optimal response based on a given price and jointly determines the capacity commitment, electricity and carbon emission budget. Through KKT conditions and implicit differentiation, the lower layer optimization is embedded into the upper layer to form an end-to-end "price-supply" sensitive mapping. Furthermore, by combining inverse reinforcement learning to back-infer weights from historical measurement data sequences, compatible capacity, electricity and carbon emission budgets and price weights are obtained.

[0024] A further improvement of this invention is that the expression for the multi-objective robust optimization model of the virtual power plant is:

[0025] (7);

[0026] (8);

[0027] (9);

[0028] (10);

[0029] In the formula, for Distributed conventional power supply output at all times for Real-time renewable energy output for Net exchange power with the main power grid at all times. For energy storage charging and discharging power, The total load of the virtual power plant. This refers to the amount of interruptible load reduction. This represents the time shift of peak load. For robust sets, For candidate true distributions, In energy storage state, For charging and discharging efficiency, These are the lower and upper limits of energy storage status. To ensure that charging and discharging are mutually exclusive, For a set of time windows, In order to be in The amount of calls is increased at any time. In order to increase capacity, For time step, For electricity quota planning, In order to be in Carbon emissions at any given time For carbon emissions budget, Comp represents the total capacity compensation revenue. Price based on capacity, In order to be in The constant increase in capacity, and the risk cost of short-term capacity shortages, The penalty coefficient for short supply, In order to be in The capacity demand proposed by external entities at any given moment. For the capacity to be delivered, For confidence level of Conditional Value at Risk (VaR) operator, For positive part operators, For economic goals, For electricity purchase and sales prices, For negative part operators, For the cost of energy storage degradation, For the operating costs of each distributed conventional power source, For green goals, A trade-off factor for green goals. for The amount of wind and solar energy abandoned at any given moment. For the proportion of green electricity, For the goal of low carbon, For the first Emission factors of distributed conventional power sources in Taiwan for Time of the first Taiwan distributed conventional power supply output, The marginal carbon factor of purchased electricity For scheduling decision vector, The optimal scheduling solution under the comprehensive objective is... Indexed to the three objectives of economy, green and low carbon. For the first One objective function, For the first The weight of each objective, As a reference to the optimal value, The smoothing coefficient for the second term. , Scale the target for Tchebycheff. The internal shadow price of electricity and carbon emissions. As weight, for A collection of settlement results and trajectory data.

[0030] The present invention provides a virtual power plant multi-objective robust optimization AI control system, comprising:

[0031] The information prediction module is used to predict the diverse and uncertain information of the virtual power plant;

[0032] The capability assessment module uses AI to evaluate the adjustable capabilities of a virtual power plant across multiple time scales, based on diverse and uncertain information.

[0033] The weight coordination module is used to construct a multi-dimensional coordination model of virtual power plant capacity, electricity, and carbon emissions, and to coordinate the weights of capacity, electricity, and carbon emission budgets using the virtual power plant capacity, electricity, and carbon emission multi-dimensional coordination model.

[0034] The model building module is used to construct a learning-enhanced multi-objective robust optimization model for virtual power plants based on the evaluation results of the multi-timescale adjustable capabilities of virtual power plants and the weights obtained through coordination. The model is then used to optimize and regulate the virtual power plant.

[0035] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described virtual power plant multi-objective robust optimization AI control method.

[0036] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described virtual power plant multi-objective robust optimization AI control method.

[0037] The computer program product of the present invention includes computer instructions, which, when executed by a processor, perform the steps of the above-described virtual power plant multi-objective robust optimization AI control method.

[0038] The beneficial effects of this invention are as follows: This invention integrates multimodal prediction and extreme disturbance modeling to generate a unified and callable data package, which can accurately assess the adjustable boundaries of virtual power plants across multiple time scales. Furthermore, this invention enables multi-objective scheduling of virtual power plants, including economic, green energy, and low-carbon goals, ensuring safe operation under disturbances and timely and full capacity delivery, thus solving the problems of disconnect between planning and execution, difficulty in fulfilling targets in extreme scenarios, and difficulty in simultaneously achieving multiple objectives. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0041] like Figure 1 As shown, the virtual power plant multi-objective robust optimization AI control method of this embodiment includes:

[0042] Step 1: Predict diverse and uncertain information about virtual power plants, including electricity prices, carbon emissions per kilowatt-hour, weather, and external renewable energy sources;

[0043] Step 2: Based on diverse and uncertain information, AI assesses the multi-timescale adjustability of the virtual power plant;

[0044] Step 3: Construct a multi-dimensional coordination model of virtual power plant capacity, electricity, and carbon emissions, and coordinate the weights of capacity, electricity, and carbon emissions budgets.

[0045] Step 4: Based on the evaluation results of the virtual power plant's multi-timescale adjustable capabilities and the weights obtained through coordination, construct a learning-enhanced multi-objective robust optimization model for the virtual power plant, and use the virtual power plant multi-objective robust optimization model to optimize and regulate the virtual power plant.

[0046] In step 1, this embodiment proposes a multimodal virtual power plant uncertainty information prediction method. In the virtual power plant scenario, this embodiment simultaneously focuses on heterogeneous data including electricity price, carbon emissions per kilowatt-hour, weather, and external renewable energy sources, addressing the problems of information dispersion and lack of characterization of extreme operating conditions. The external renewable energy sources refer to the available grid-connected power of wind / solar power, i.e., available wind and solar power. The specific prediction steps are as follows:

[0047] Step 1.1: Align and clean the historical measurement data sequence and external feature sequence of the virtual power plant, and obtain point predictions of uncertain information through a multimodal time series model. In this embodiment, the historical measurement data sequence includes load, energy storage status, available wind and solar power, electricity price and carbon emission factor per kilowatt-hour, etc., and the external feature sequence includes weather, holidays, maintenance and market events, etc. The multimodal time series model is a selective state-space model (Mamba), a temporal fusion transformer (TFT), or a patch time series transformer (PatchTST).

[0048] Step 1.2: Based on the external feature sequence, use the diffusion model (DDPM) to generate a full-time-domain scenario containing extreme perturbations.

[0049] Step 1.3: Quantitative / conformal predictions produce confidence intervals. Combining the Wasserstein-DRO and CVaR methods, robust tail risk boundaries are formed, resulting in a unified data package containing point predictions, scenario sets, robust sets, and conditional value of risk operators.

[0050] The prediction of multivariate uncertain information in step 1 is shown in expressions (1) to (3). Expression (1) provides point predictions of electricity price, carbon emission factor per kilowatt-hour, available wind and solar power, load, and weather at various future times, as well as multiple full-time-domain scenarios covering normal and extreme fluctuations, for use by virtual power plant (VPP) scheduling, capacity application, and settlement. Expression (2) uses a first-order Wasserstein sphere to enclose the range that the actual distribution may deviate from and focuses on the most unfavorable tails such as wind and solar power gaps, load surges, and price anomalies. Expression (3) is a unified data package containing point predictions, scenario sets, robust sets, and conditional value at risk operators.

[0051] ;

[0052] ;

[0053] ;

[0054] In the formula, For the future Multivariate point prediction Index the current discrete time point, configure as needed, such as in 15-minute intervals. For multimodal time series models, This is a historical measurement data sequence for a virtual power plant. For external feature sequences, This is the forward sight step length, i.e., the lead time for prediction. For the first The generated full-time-domain multivariate scenario trajectory, For the number of scenarios, To generate samples according to a given generative model. For parameters diffusion model, include and , The predicted quantity is the historical measurement data sequence of the virtual power plant. For robust sets, For candidate true distributions, For the empirical distribution prediction, The first-order Wasserstein distance. The radius is the Wasserstein radius. For confidence level of The conditional value at risk operator is used to measure tail risk, where Z represents stochastic losses, including losses or deviation costs due to power outages, capacity shortages caused by load surges, etc. To calibrate the quantile threshold, To test the candidate true distribution within the robust set Take the expected value. The positive part operator is T, where T is the total number of time periods for horizon prediction. This is a full horizon point prediction sequence. For a set of scenarios, , A unified robust data package is a unified data package that includes point prediction, scenario set, robust set, and conditional value at risk operator (risk operator).

[0055] In step 2, this embodiment proposes an AI evaluation method for the multi-timescale adjustable capabilities of virtual power plants that considers multiple uncertain information. Based on the predicted multiple uncertain information, an adjustable capability evaluation model is constructed for the three timescales of day-ahead, day-intraday, and online. The model is based on a Physical Information Long Short-Term Memory (PSM) network and a Graph Attention (GAT) network. The adjustable capacity assessment model is based on physical information LSTM fusion of device and network constraints. The network constraints include power limits, SOC upper and lower limits, ramping, and response delay. A graph attention network (GAT) is used to aggregate the coupling potential of distributed power sources, energy storage, and adjustable loads to assess adjustable capacity indicators, namely, the up-up capacity, down-down capacity, sustainability, total available power, and maximum ramping rate at each time scale. The resulting executable capacity envelope is shown in expressions (4) to (6). Expression (4) indicates that the capacity evaluator transforms the prediction and scenario in step 1 into a three-scale capacity set. Each set contains the maximum up-up power and down-down power, sustainability, available power, ramping rate, and other boundaries, while taking confidence quantiles. Gain the ability to safeguard, ensuring that promises can be kept;

[0056] Expression (5) is defined with probabilistic constraints: in At least with confidence level The executable tuple represents the capacity increase, capacity decrease, maximum continuously adjustable duration, and maximum available total power.

[0057] Expression (6) is the output obtained, namely the capability envelope.

[0058] ;

[0059] ;

[0060] ;

[0061] In the formula, For parameters Adjustable capability evaluator for Three items, , , These are respectively: today, today, and online. These represent capabilities for the day before, intraday, and online. For scale The lower bound of the confidence level of the capability. For scale The confidence level of the capability For quantile operations, To assess capabilities based on scenario sets and robustness sets, It provides indexes for three time scales: day-to-day, intraday, and online. To adjust the capacity upwards or downwards, The longest continuously adjustable duration. This represents the maximum available total battery capacity. For the maximum climbing speed, Let be the probability of the candidate true distribution within the robust set. As a condition, The confidence threshold. These represent the minimum capabilities for the day before, intraday, and online, respectively. This is the set of lower limits for capabilities within the current day, intraday, and online. For robust sets, For candidate true distributions, For a set of scenarios, This is a predicted sequence of points across the entire horizon.

[0062] In step 3, this embodiment proposes a multi-dimensional coordination model for virtual power plant capacity, electricity, and carbon emissions. This model comprises an upper layer and a lower layer. The upper layer, i.e., the trading or operating coordinator, learns capacity, electricity, and carbon emission price signals. The lower layer, i.e., the virtual power plant, solves for the optimal response based on a given price, jointly determining capacity commitments, electricity plans, and carbon emission budgets. Through KKT conditions and implicit differentiation (OptNet / JAXopt), the lower-layer optimization is sub-embedded into the upper layer, forming an end-to-end price-supply sensitive mapping. The KKT conditions for the lower layer are:

[0063] ;

[0064] Lower-level optimal solution With the upper level , and Correlated, the Jacobian matrix is ​​obtained by taking the differential:

[0065] ;

[0066] Then, the gradient of the upper-level target W is calculated: .

[0067] By combining inverse reinforcement learning to deduce weights from historical measurement data sequences, compatible capacity, electricity, and carbon emission budget and price weights are obtained. As shown in expressions (7) and (8), expression (7) hierarchically coordinates the virtual power plant capacity-electricity-carbon emission multidimensional model. The upper layer selects the internal shadow prices of capacity price, electricity, and carbon emission to maximize operating welfare, while the lower layer selects the internal shadow prices of capacity price, electricity, and carbon emission within the feasible region. Within this framework, the committable capacity, electricity consumption plan, and carbon emission budget are jointly determined to achieve consistent calibration of capacity, electricity consumption, and carbon emissions. Expression (8) is the result obtained and is used as input for step 4.

[0068] ;

[0069] ;

[0070] In the formula, Price based on capacity, The internal shadow price of electricity and carbon emissions. For the operation of welfare functions, For external capacity requirements, An externally provided carbon emissions budget cap. For electricity quota planning, For carbon emissions budget, of which Returning from the lower level to the upper level, These are the committed capacity, electricity consumption, and carbon emissions, forming the lower-level decision-making tripartite. The feasible region is defined by the prior day, intraday, and online capability lower limits obtained in step 2, i.e., the constraint set synthesized from the three scales. This represents the maximum available total battery capacity. To increase the maximum capacity, The optimal solution set that maximizes the objective.

[0071] In step 4, this embodiment proposes a multi-objective AI robust optimization and control method for virtual power plants that considers capacity compensation. Using the capacity envelope obtained in step 2 and the coordination results obtained in step 3 as inputs, a learning-reinforced multi-objective robust optimization model for virtual power plants is constructed. Under power balance, energy storage status, equipment and network constraints, and capacity contracts, considering capacity compensation and short-supply penalties, an adaptive Tchebycheff unified economic dispatch (minimum overall cost), green dispatch (maximum renewable energy consumption / minimum wind and solar curtailment), and low-carbon dispatch (minimum emissions) are used. The MILP / MISOCP algorithm is employed to solve the multi-objective robust optimization model for virtual power plants. A neural hot-start algorithm is used to accelerate the convergence speed of the MILP / MISOCP algorithm, obtaining intraday and online rolling plans and time-of-use capacity results. Using safety reinforcement learning and behavior cloning, the plan is transformed into fine-grained control commands, maintaining the energy storage status (SOC), ramp-up, and over-limit safety during disturbances, and delivering capacity on time and in the required quantity. This solves the problems of disconnect between planning and execution, difficulty in fulfilling targets in extreme scenarios, and difficulty in balancing multiple objectives. As shown in expressions (9) to (12), the first two lines of expression (9) ensure that power balance and energy storage state changes are robust to the set. The third line is robust and feasible, ensuring that the scheduling aligns with the capacity-energy-carbon emission targets of step 3 and is constrained by the lower limit of the capabilities of step 2, thus preventing "over-commitment".

[0072] Expression (10) promises that increased capacity will be compensated for the shortfall. In extreme scenarios, insufficient delivery will be compensated for the tail supply based on CVaR and penalized accordingly. Expression (11) scalarizes the three objectives (green, economic, and low-carbon). The three objectives are combined into a solvable objective through adaptive Tchebycheff, with the weights induced by the shadow price in step 3. Or, given the operational preferences, the optimal joint scheduling is obtained. Expression (12) indicates that the output feedback of step 4 forms a closed loop with step 1. Through optimal scheduling, actual commitment and fulfillment data, it is used to update the data packet of step 1 and the capacity envelope of step 2, realizing the closed-loop self-learning of "prediction → capacity assessment → three-layer coordination of capacity, power and carbon emissions → robust optimization".

[0073] (9);

[0074] (10);

[0075] (11);

[0076] (12);

[0077] In the formula, for Distributed conventional power supply output at all times for Real-time renewable energy output for Net exchange power with the main power grid at all times. For energy storage charging and discharging power, The total load of the virtual power plant. This refers to the amount of interruptible load reduction. This represents the time shift of peak load. For robust sets, For candidate true distributions, In energy storage state, For charging and discharging efficiency, These are the lower and upper limits of energy storage status. To ensure that charging and discharging are mutually exclusive, For a set of time windows, In order to be in The amount of calls is increased at any time. In order to increase capacity, For time step, For electricity quota planning, In order to be in Carbon emissions at any given time For carbon emissions budget, Comp represents the total capacity compensation revenue. Price based on capacity, In order to be in The constant increase in capacity, and the risk cost of short-term capacity shortages, The penalty coefficient for short supply, In order to be in The capacity demand proposed by external entities at any given moment. For the capacity to be delivered, For economic goals, For electricity purchase and sales prices, For negative part operators, For the cost of energy storage degradation, For the operating costs of each distributed conventional power source, For green goals, A trade-off factor for green goals. for The amount of wind and solar energy abandoned at any given moment. For the proportion of green electricity, For the goal of low carbon, For the first Emission factors of distributed conventional power sources in Taiwan for Time of the first Taiwan distributed conventional power supply output, The marginal carbon factor of purchased electricity The scheduling decision vector includes , , , , , and the promised capacity r, The optimal scheduling solution under the comprehensive objective is... Indexed to the three objectives of economy, green and low carbon. For the first One objective function, For the first The weight of each objective, As a reference to the optimal value, The smoothing coefficient for the second term. Scale the target for Tchebycheff. The weights are determined based on shadow price induction or operational preferences. for The data consists of a set of settlement results and trajectory data. The settlement results are the actual delivery of data during the execution of the virtual power plant, including revenue, penalty mechanism processing, capacity fulfillment, carbon emission quotas, and power delivery deviations. The trajectory data are the time-series data trajectories during the regulation and operation process.

[0078] The above method is implemented through the virtual power plant multi-objective robust optimization AI control system in this embodiment, the system comprising:

[0079] The information prediction module is used to predict the diverse and uncertain information of the virtual power plant;

[0080] The capability assessment module uses AI to evaluate the adjustable capabilities of a virtual power plant across multiple time scales, based on diverse and uncertain information.

[0081] The weight coordination module is used to construct a multi-dimensional coordination model of virtual power plant capacity, electricity, and carbon emissions, and to coordinate the weights of capacity, electricity, and carbon emission budgets using the virtual power plant capacity, electricity, and carbon emission multi-dimensional coordination model.

[0082] The model building module is used to construct a learning-enhanced multi-objective robust optimization model for virtual power plants based on the evaluation results of the multi-timescale adjustable capabilities of virtual power plants and the weights obtained through coordination. The model is then used to optimize and regulate the virtual power plant.

[0083] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0084] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-objective robust optimization AI control method for virtual power plants, characterized by: The method comprises the following steps: forecasting multi-element uncertain information of a virtual power plant; AI evaluating multi-time scale adjustable capacity of the virtual power plant based on the multi-element uncertain information; constructing a virtual power plant capacity-power-carbon emission multi-dimensional coordination model, and coordinating weights of capacity, power and carbon emission budget by using the virtual power plant capacity-power-carbon emission multi-dimensional coordination model; constructing a learning-enhanced virtual power plant multi-objective robust optimization model based on evaluation results of the virtual power plant multi-time scale adjustable capacity and the coordinated weights, and optimizing regulation and control of the virtual power plant by using the virtual power plant multi-objective robust optimization model; the virtual power plant capacity-power-carbon emission multi-dimensional coordination model comprises an upper layer and a lower layer, the upper layer learns capacity, power and carbon emission price signals, the lower layer solves an optimal response according to a given price to jointly determine capacity commitment, power plan and carbon emission budget, an optimization micro-embedding of the lower layer is formed into the upper layer through KKT conditions and implicit differentiation to form an end-to-end "price-supply" sensitive mapping, and inverse reinforcement learning is combined to back-propagate weights from a historical measurement data sequence to obtain compatible capacity, power and carbon emission quotas and price weights; an expression of the virtual power plant multi-objective robust optimization model is as follows: （7） （8） （9） （10） wherein, is the distributed conventional power output at time t, is the renewable actual output at time t, is the net exchange power with the large power grid at time t, is the charging and discharging power of the energy storage, is the total load of the virtual power plant, is the interruptible load curtailment, is the time-shifting amount of the movable load, is the robust set, is the candidate real distribution, is the energy storage state, is the charging and discharging efficiency, is the lower and upper limits of the energy storage state, is the charging and discharging exclusivity, is the set of time windows, is the upward call amount at time t, is the upward capacity, is the time step, is the electricity planning quota, is the carbon emission amount at time t, is the carbon emission budget, and Comp is the total capacity compensation revenue, is the capacity price, is the upward capacity at time t, and Short is the risk cost of capacity shortage, is the shortage penalty coefficient, is the externally proposed capacity demand at time t, is the delivered capacity, is the conditional value at risk operator with a confidence level of , is the positive part operator, is the economic objective, is the electricity purchase and sale price, is the negative part operator, is the energy storage degradation cost, is the operation cost of each distributed conventional power source, is the green objective, is the weighting coefficient of the green objective, is the wind and light curtailment amount at time t, is the green electricity proportion, is the low-carbon objective, is the emission factor of the ith distributed conventional power source, is the output of the ith distributed conventional power source at time t, is the marginal carbon factor of the purchased electricity, is the dispatching decision vector, is the optimal dispatching solution under the comprehensive objective, for economic, green and low carbon three target indexes, for the th objective function, for the weight of the th objective, for the reference optimal value, for the subterm smoothing coefficient, for the Tchebycheff scalarization objective, for the internal shadow price of electricity and carbon emission, for the weight, for , settlement result and trajectory data set.

2. The virtual power plant multi-objective robust optimization AI regulation method according to claim 1, characterized in that: forecasting multi-element uncertain information of a virtual power plant, which comprises the following steps: aligning and cleaning historical measurement data sequences and external feature sequences of the virtual power plant, obtaining point prediction of uncertain information by using a multi-modal time series model, generating a full-time domain scenario containing extreme disturbance by using a diffusion model, combining a Wasserstein-DRO and a CVaR method to form a robust tail risk boundary, and an expression is as follows: wherein, is a future multivariate point prediction, is a current discrete time index, is a multi-modal time-series model, is a historical measurement data sequence of a virtual power plant, is an external feature sequence, is a look-ahead horizon, is a full time horizon multivariate scenario trajectory generated by the th scenario, is a number of scenarios, is generated by sampling according to a given generative model, is a diffusion model with parameters , includes and , is a predicted measurement of the historical measurement data sequence of the virtual power plant, is a robust set, is a candidate true distribution, is an empirical distribution prediction, is a first-order Wasserstein distance, is a Wasserstein radius, is a conditional value-at-risk operator with a confidence level of , and Z is a random loss, is a calibrated quantile threshold, is an expectation of the candidate true distribution within the robust set, is a positive part operator, and T is a total time horizon number of prediction horizons, is a full-horizon point prediction sequence, is a scenario set, , is a unified data package including a point prediction, a scenario set, a robust set, and a conditional value-at-risk operator. 3.The virtual power plant multi-objective robust optimization AI regulation method of claim 1, wherein: AI evaluating multi-time scale adjustable capacity of the virtual power plant based on the multi-element uncertain information, which specifically comprises the following steps: based on the predicted multi-element uncertain information, constructing an adjustable capacity evaluation model facing day-ahead, intra-day and online three time scales, evaluating boundaries of adjustable capacity of each time scale, and obtaining an executable capacity envelope, and an expression is as follows: In the formula, is an adjustable capability evaluator with parameters , , , are day-ahead, intra-day, online, respectively, are day-ahead, intra-day, online capability, respectively, is a lower bound of capability confidence with scale , is a capability confidence with scale , is a quantile operation, is to evaluate capability conditioned on scenario set and robust set, are day-ahead, intra-day, online time scale indices, are up, down capacity, is the longest duration of continuous adjustability, is the upper limit of total available power, is the maximum ramp rate, is the probability about candidate true distribution within robust set, is a condition, is a confidence threshold, are day-ahead, intra-day, online capability lower bound, respectively, is a set of day-ahead, intra-day, online capability lower bound, is a robust set, is a candidate true distribution, is a scenario set, is a full horizon point forecast sequence.

4. The virtual power plant multi-objective robust optimization AI regulation system based on the method of any one of claims 1 to 3, characterized in that: The system comprises: an information prediction module configured to forecast multi-element uncertain information of a virtual power plant; a capacity evaluation module configured to AI evaluate multi-time scale adjustable capacity of the virtual power plant based on the multi-element uncertain information; a weight coordination module configured to construct a virtual power plant capacity-power-carbon emission multi-dimensional coordination model, and coordinate weights of capacity, power and carbon emission budget by using the virtual power plant capacity-power-carbon emission multi-dimensional coordination model; a model construction module configured to construct a learning-enhanced virtual power plant multi-objective robust optimization model based on evaluation results of the virtual power plant multi-time scale adjustable capacity and the coordinated weights, and optimize regulation and control of the virtual power plant by using the virtual power plant multi-objective robust optimization model.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to realize the steps of the virtual power plant multi-objective robust optimization AI regulation and control method in any one of claims 1 to 3.

6. A computer readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the virtual power plant multi-objective robust optimization AI regulation and control method in any one of claims 1 to 3.

7. A computer program product comprising computer instructions, characterized in that: The computer instructions are executed by the processor to perform the steps of the virtual power plant multi-objective robust optimization AI regulation and control method in any one of claims 1 to 3.

Images