A method, apparatus, terminal equipment, and storage medium for coordinated control of power generation, grid, load, and storage.
By calculating the distance of the source-grid-load-storage time-series tensor set and using a generative adversarial network model, a core extreme scenario set is generated, which solves the problem of overfitting of control strategies under extreme conditions in existing technologies and realizes reliable scheduling and real-time correction of the source-grid-load-storage system under extreme conditions.
Patent Information
- Application Number
- CN202611334841.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing source-grid-load-storage optimization control methods struggle to generate virtual scenarios covering different extreme event types, durations, and impact intensities under extreme conditions. This leads to overfitting of control strategies, difficulty in timely correction of real-time control during day-ahead scheduling, and improper reserve capacity settings, all of which affect the executability and reliability of control results.
By calculating the distance between the source-network-load-storage time-series tensor set and the normal dataset, a seed dataset is selected, a generative adversarial network model is trained, a core extreme scenario set is generated, and scheduling decisions are made in conjunction with the current multi-source prediction data. The model is optimized to consider risks in advance under extreme conditions and formulate reliable scheduling strategies.
It improves the reliability of source-grid-load-storage coordinated control, effectively copes with extreme combinations of changes not seen in historical samples, reduces computational load, and ensures the physical feasibility and real-time adaptability of control strategies.
Smart Images

Figure CN123121507A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of source-grid-load-storage coordinated control, and in particular to a source-grid-load-storage coordinated control method, apparatus, terminal equipment and storage medium. Background Technology
[0002] As the penetration rate of new energy sources such as wind power and photovoltaic power in the power system continues to increase, the uncertainty of power output on the power generation side and the volatility of load on the load side have significantly increased. Especially under extreme conditions such as extreme cold, heat waves, typhoons, strong convection, continuous rain, sudden cloud cover, and equipment failures, the output of new energy sources may drop sharply in a short period of time, while load demand may increase simultaneously, resulting in a large net load gap. If the various links of power generation, grid, load, and storage cannot be coordinated in a timely manner, problems such as power imbalance, insufficient state of charge of energy storage, equipment output exceeding limits, and reduced power supply reliability can easily occur. Existing power generation, grid, load, and storage optimization control methods are usually based on historical operating data, deterministic prediction results, or preset uncertainty boundaries, and operate by means of day-ahead scheduling or real-time control.
[0003] However, due to the low probability of extreme operating conditions, the number of extreme samples available for training and optimization in historical databases is limited, making it difficult to cover different extreme event types, durations, impact intensities, and combinations. Training a data-driven model directly with a small number of historical extreme samples easily leads to overfitting, causing the resulting control strategy to only adapt to a few historical operating conditions and failing to cope with extreme combinations not fully represented in historical samples. Generative adversarial networks (GANs) can expand the sample size to some extent, but traditional GANs primarily judge the authenticity of generated samples based on data distribution similarity, failing to fully consider physical limitations such as installed capacity, allowable power range, power ramping capability, and power balance between power sources, grids, loads, and storage. This may result in virtual scenarios exceeding equipment capacity, exceeding equipment ramping limits, or having unreasonable source-load power relationships. Directly using such virtual scenarios for control strategy training may cause the optimization model to solve control commands based on physically unrealizable data, affecting the executability of the control results. Furthermore, existing optimization control methods typically perform day-ahead scheduling or real-time control separately. Day-ahead scheduling struggles to promptly adjust control schemes using real-time operating data, while real-time control alone lacks advance preparation for future extreme risks. While some methods employ robust optimization to reserve backup capacity, they typically construct uncertain sets based on manually set single boundaries or limited historical scenarios, which can easily lead to problems such as setting the backup capacity too large or insufficient risk coverage. Summary of the Invention
[0004] This invention provides a source-grid-load-storage coordinated control method, apparatus, terminal equipment, and storage medium, which can solve the above-mentioned problems in the prior art and improve the reliability of source-grid-load-storage coordinated control.
[0005] This invention provides a source-grid-load-storage coordinated control method, comprising: Obtain the source-network-load-storage time-series tensor set and the source-network-load-storage normal dataset; Distance calculation is performed on the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and the source-grid-load-storage time series tensor set is filtered based on the anomaly score dataset to obtain a seed dataset; Based on the seed dataset, a seed condition vector set corresponding to the seed dataset is obtained, and a preset generative adversarial network model is trained based on the seed dataset and the seed condition vector set to obtain a target generative adversarial network model. Obtain the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set; Obtain recent multi-source forecast data; Based on the aforementioned multi-source forecast data and core extreme scenario set, the preset daytime and intraday control model is solved to obtain source-grid-load-storage scheduling data, which is then used for control.
[0006] In the above scheme, distance calculation is performed between the source-grid-load-storage time-series tensor set and the normal source-grid-load-storage dataset to reflect the deviation of the source-grid-load-storage time-series tensor set from the normal dataset and the correlation between various variables. This allows for the accurate extraction of representative seed datasets from a large number of source-grid-load-storage time-series tensors, overcoming the dilution problem of extreme features caused by conventional samples. Then, a pre-defined generative adversarial network (GAN) model is trained based on the selected seed dataset and its corresponding seed conditional vector set. This pre-defined GAN model can learn the temporal variation patterns of the extreme samples represented by the seed dataset and seed conditional vector set. Finally, the target conditional vector set is obtained and input into the trained target GAN model, enabling further processing. Without relying on a large amount of historical extreme measured data, a diverse set of core extreme scenarios is generated. Finally, by acquiring day-ahead multi-source prediction data, and based on the day-ahead multi-source prediction data and the core extreme scenario set, the preset day-ahead intraday control model is solved. This allows the optimization model to consider various representative high-risk disturbances that may occur in the future when making scheduling decisions. The final source-grid-load-storage scheduling data can provide a basis for day-ahead planning and intraday correction of the source-grid-load-storage system. This effectively solves the technical problems in the prior art where the lack of historical extreme samples leads to overfitting of control strategies and difficulty in dealing with extreme change combinations that have not fully appeared in historical samples. This significantly improves the reliability of source-grid-load-storage coordinated control.
[0007] Further, the step of calculating the distance between the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and then filtering the source-grid-load-storage time-series tensor set based on the anomaly score dataset to obtain a seed dataset, includes: Based on the source-grid-load-storage normal dataset, the mean value of the normal operating condition state vector and the covariance matrix of the normal operating condition state vector are obtained. Based on the preset Mahalanobis distance algorithm, the distance is calculated between the mean of the normal operating condition state vector, the covariance matrix of the normal operating condition state vector and the source-grid-load-storage time series tensor set to obtain the anomaly score dataset corresponding to the source-grid-load-storage time series tensor set. The seed dataset is obtained by filtering the source-network-load-storage time-series tensor set based on the anomaly score dataset and the preset anomaly threshold.
[0008] In the above scheme, the mean value of the normal operating condition state vector and the covariance matrix of the normal operating condition state vector are obtained based on the source-grid-load-storage normal dataset. The distance between the mean value of the normal operating condition state vector, the covariance matrix of the normal operating condition state vector, and the source-grid-load-storage time series tensor set is calculated based on the preset Mahalanobis distance algorithm. This can simultaneously consider the degree of deviation of each operating variable from the normal value and the correlation between each operating variable. The resulting anomaly score dataset can accurately reflect the combined abnormal state when the load increases rapidly and the output of new energy sources decreases rapidly at the same time. This allows for the precise selection of representative historical extreme sample fragments to form the seed dataset.
[0009] Further, the preset generative adversarial network model includes a preset generator and a preset discriminator. The step of obtaining a seed condition vector set corresponding to the seed dataset based on the seed dataset, and training the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model includes: The seed condition vector set and the preset noise distribution are input into the preset generator so that the preset generator outputs a virtual extreme sample set. The preset discriminator is trained based on the seed dataset and the virtual extreme sample set to obtain the target discriminator and the discriminator feedback result output by the target discriminator; Based on the virtual extreme sample set, the sample physical violation quantity set corresponding to the virtual extreme sample set is obtained, and the preset generator is updated based on the discriminator feedback result and the sample physical violation quantity to obtain the target generator; The target generation adversarial network model is obtained based on the target generator and the target discriminator.
[0010] In the above scheme, by inputting the seed condition vector set and the preset noise distribution into the preset generator to output a virtual extreme sample set, the generator learns the occurrence pattern of extreme events under the constraints of the condition vectors, while the introduction of random noise ensures the diversity of the generated scenarios. By training the preset discriminator based on the seed dataset and the virtual extreme sample set, the target discriminator and its output discriminator feedback results are obtained, enabling the discriminator to accurately distinguish between real extreme samples and generated samples, providing effective discrimination signals for the optimization of the generator. By obtaining the sample physical violation quantity set based on the virtual extreme sample set, and updating the preset generator based on the discriminator feedback results and the sample physical violation quantity, the generator is simultaneously constrained by the distribution similarity of the discriminator feedback and the physical constraints of equipment capacity and ramp-up ability represented by the physical violation quantity during the optimization process. This effectively reduces the possibility of the generated scenario violating the equipment operating boundary, so that the final target generator can generate virtual extreme scenarios that both conform to the distribution characteristics of extreme samples and satisfy the physical constraints of the power system.
[0011] Further, the step of obtaining the target condition vector set and inputting the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set includes: The target condition vector set is input into the target generative adversarial network model to obtain a set of virtual extreme scenarios; Based on preset tolerance values, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios; The subset of virtual extreme scenarios is further filtered based on the preset maximum available adjustable power to obtain the core extreme scenario set.
[0012] In the above scheme, a set of virtual extreme scenarios is obtained by inputting the target condition vector set into the target generative adversarial network model. The generator then generates virtual scenarios corresponding to the extreme operating conditions to be analyzed under the guidance of the target condition vector. A subset of virtual extreme scenarios is obtained by initially screening the set of virtual extreme scenarios based on a preset tolerance value. Unusable scenarios with physical violations exceeding the tolerance value can be eliminated, ensuring that the retained scenarios meet hard physical constraints such as equipment capacity and ramping requirements. A core extreme scenario set is obtained by further screening the subset of virtual extreme scenarios based on a preset maximum available adjustment power. This allows for the further screening of representative high-risk scenarios while considering the actual adjustment capability of the system. This reduces the number of scenarios that the subsequent optimization model needs to process while retaining extreme risk characteristics, thus reducing the computational load of multi-scenario optimization.
[0013] Furthermore, the preliminary screening of the set of virtual extreme scenarios based on preset tolerance values yields a subset of virtual extreme scenarios, including: Based on the set of virtual extreme scenarios, obtain the set of virtual physical violations corresponding to the set of virtual extreme scenarios; Based on the preset tolerance value and the set of virtual physical violations, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios.
[0014] In the above scheme, by obtaining the virtual physical violation quantity set corresponding to the virtual extreme scenario set, and performing preliminary screening of the virtual extreme scenario set based on the preset tolerance value and the virtual physical violation quantity set, the quantitative physical violation quantity can be used as the screening basis to independently evaluate the degree of equipment capacity exceeding the limit and the degree of power ramping exceeding the limit for each virtual extreme scenario. Scenarios with physical violation quantities not exceeding the preset tolerance value are retained, while scenarios exceeding the tolerance value are eliminated or regenerated, so that each scenario in the retained virtual extreme scenario subset satisfies the equipment physical boundary conditions.
[0015] Furthermore, the secondary filtering of the virtual extreme scenario subset based on the preset maximum available adjustable power to obtain the core extreme scenario set includes: Based on the aforementioned subset of virtual extreme scenarios, we obtain the scenario load power set, the scenario photovoltaic power output set, and the scenario wind power output set; Based on the aforementioned scenario load power set, scenario photovoltaic power output set, and scenario wind power output set, the scenario net load set is obtained; A net load gap set is obtained based on the scenario net load set and the preset maximum available adjustable power; The set of net load gaps is processed by time accumulation to obtain a set of destructive force indicators; The core extreme scenario set is obtained by filtering the subset of virtual extreme scenarios based on the set of destructive power indicators.
[0016] In the above scheme, by obtaining the scenario load power set, scenario photovoltaic power output set, and scenario wind power output set based on the subset of virtual extreme scenarios, and obtaining the scenario net load set based on the above three sets, it is possible to quantify the net load demand that the system still needs to meet after the participation of renewable energy power generation in balancing at each moment in each scenario; by obtaining the net load gap set based on the scenario net load set and the preset maximum available adjustable power, it is possible to identify the moments and gap sizes in each scenario where the system's adjustment capacity is insufficient to meet the net load; by performing time accumulation processing on the net load gap set to obtain the destructive force index set, it is possible to comprehensively evaluate the size and duration of the net load gap in each scenario over the entire time period, avoiding judging the scenario risk based solely on the maximum gap at a single moment; by filtering the subset of virtual extreme scenarios based on the destructive force index set to obtain the core extreme scenario set, it is possible to retain core scenarios with greater destructive force and a higher degree of threat to the system.
[0017] Furthermore, the process of filtering the subset of virtual extreme scenarios based on the set of destructive power indicators to obtain a core set of extreme scenarios includes: Based on the numerical values of the set of destructive power indicators, the subset of virtual extreme scenarios is divided to obtain several initial virtual extreme cluster centers and additional virtual extreme scenarios. Based on the time series distance between several additional virtual extreme scenarios and several initial virtual extreme cluster centers, several target virtual extreme clusters are obtained by matching the several additional virtual extreme scenarios and several initial virtual extreme cluster centers. Each target virtual extreme cluster is filtered based on preset cluster center selection criteria to obtain the core extreme scenario; The core extreme scenario set is obtained based on all core extreme scenarios corresponding to several target virtual extreme clusters.
[0018] In the above scheme, the subset of virtual extreme scenarios is divided based on the numerical value of the set of destructive power indicators to obtain several initial virtual extreme cluster centers and additional virtual extreme scenarios. This allows for the selection of representative scenarios as initial cluster centers based on the destructive power indicators. Several target virtual extreme clusters are obtained by matching the time series distances between the additional virtual extreme scenarios and the initial virtual extreme cluster centers, allowing scenarios with similar change patterns to be grouped into the same cluster. Core extreme scenarios are obtained by filtering each target virtual extreme cluster based on preset cluster center selection criteria, ensuring that each cluster retains one actual, representative scenario as the core extreme scenario. This avoids weakening extreme peaks or creating curves that do not exist in the actual scenario set by calculating average curves for multiple scenarios, thereby effectively reducing the number of scenarios while preserving extreme risk characteristics and reducing the computational burden on subsequent optimization models.
[0019] Another embodiment of the present invention provides a source-grid-load-storage coordinated control device, comprising: The basic data acquisition module is used to acquire the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset; The seed dataset acquisition module is used to calculate the distance between the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and to filter the source-grid-load-storage time series tensor set based on the anomaly score dataset to obtain the seed dataset. The target generative adversarial network model generation module is used to obtain the seed condition vector set corresponding to the seed dataset based on the seed dataset, and to train the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model. The core extreme scenario set acquisition module is used to acquire the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set. The current day multi-source forecast data acquisition module is used to acquire current day multi-source forecast data; The source-grid-load-storage scheduling data acquisition module is used to solve the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and to perform control based on the source-grid-load-storage scheduling data.
[0020] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps of the source-grid-load-storage coordinated control method of the present invention.
[0021] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the source-grid-load-storage coordinated control method of the present invention. Attached Figure Description
[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0023] Figure 1 This is a schematic flowchart of a source-grid-load-storage coordinated control method provided in an embodiment of the present invention; Figure 2 This is a morphological comparison diagram between a generated scene and a real extreme scene provided in an embodiment of the present invention; Figure 3 This is a bar chart showing the physical consistency error statistics of the generated data provided in an embodiment of the present invention; Figure 4 This is a comparison chart of system frequency response under extreme operating conditions provided by an embodiment of the present invention; Figure 5 This is a comparison chart of load shedding and operating costs under extreme operating conditions provided by an embodiment of the present invention; Figure 6 This is a visualization of the diversity distribution of generated scenes provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a source-grid-load-storage coordinated control device provided in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.
[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.
[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).
[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.
[0031] See Figure 1 To address the aforementioned problems in the prior art and improve the reliability of source-grid-load-storage coordinated control, an embodiment of the present invention provides a source-grid-load-storage coordinated control method, comprising: Step S1: Obtain the source-network-load-storage time-series tensor set and the source-network-load-storage normal dataset; Step S2: Calculate the distance between the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and filter the source-grid-load-storage time series tensor set based on the anomaly score dataset to obtain a seed dataset; Step S3: Obtain the seed condition vector set corresponding to the seed dataset based on the seed dataset, and train the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model; Step S4: Obtain the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set; Step S5: Obtain day-ahead multi-source forecast data; Step S6: Solve the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and then perform control based on the source-grid-load-storage scheduling data.
[0032] In the above scheme, distance calculation is performed between the source-grid-load-storage time-series tensor set and the normal source-grid-load-storage dataset to reflect the deviation of the source-grid-load-storage time-series tensor set from the normal dataset and the correlation between various variables. This allows for the accurate extraction of representative seed datasets from a large number of source-grid-load-storage time-series tensors, overcoming the dilution problem of extreme features caused by conventional samples. Then, a pre-defined generative adversarial network (GAN) model is trained based on the selected seed dataset and its corresponding seed conditional vector set. This pre-defined GAN model can learn the temporal variation patterns of the extreme samples represented by the seed dataset and seed conditional vector set. Finally, the target conditional vector set is obtained and input into the trained target GAN model, enabling further processing. Without relying on a large amount of historical extreme measured data, a diverse set of core extreme scenarios is generated. Finally, by acquiring day-ahead multi-source prediction data, and based on the day-ahead multi-source prediction data and the core extreme scenario set, the preset day-ahead intraday control model is solved. This allows the optimization model to consider various representative high-risk disturbances that may occur in the future when making scheduling decisions. The final source-grid-load-storage scheduling data can provide a basis for day-ahead planning and intraday correction of the source-grid-load-storage system. This effectively solves the technical problems in the prior art where the lack of historical extreme samples leads to overfitting of control strategies and difficulty in dealing with extreme change combinations that have not fully appeared in historical samples. This significantly improves the reliability of source-grid-load-storage coordinated control.
[0033] It should be noted that a source-grid-load-storage coordinated control method is used to coordinate the control of power sources, grids, loads and energy storage units under extreme weather, sudden faults or abnormal fluctuations in renewable energy output.
[0034] In this embodiment, a source-grid-load-storage coordinated control method is applied to a source-grid-load-storage system, which may include a distributed power supply unit, an energy storage unit, a load unit, a grid connection unit, a data acquisition module, a scenario generation module, an optimization calculation module, and an execution control module.
[0035] The distributed power unit includes a photovoltaic (PV) array and a wind turbine generator. The PV array is connected to the AC bus via a PV inverter, and the wind turbine generator is connected to the AC bus via a wind power converter. Both the PV inverter and the wind power converter have power regulation interfaces that can receive power control commands output by the optimization calculation module.
[0036] The energy storage unit includes an energy storage battery pack, a battery management system, and a bidirectional energy storage converter. The battery management system is used to collect the energy storage unit's state of charge, allowable charging power, allowable discharging power, and operating status, while the bidirectional energy storage converter is used to execute charging or discharging according to control commands.
[0037] Load units include conventional loads and controllable loads. Controllable loads can be industrial flexible loads or electric vehicle charging facilities, which have remote control interfaces and can perform load reduction, load transfer, or load restoration according to control commands.
[0038] The grid connection unit includes transformers, grid-connected switches, smart circuit breakers, and relay protection devices, which are used to realize the exchange of electrical energy between the source-grid-load-storage system and the upper-level grid, and to ensure the safety of system connection and operation.
[0039] The data acquisition module collects system operation data through a synchronous phasor measurement device, smart meter, photovoltaic inverter, wind power converter, battery management system, and micro-weather station. The collected data includes load power, photovoltaic output, wind power output, energy storage charging and discharging power, energy storage state of charge, bus voltage, system frequency, line power flow, switch status, irradiance, ambient temperature, and wind speed.
[0040] The data acquisition module connects to the edge gateway via industrial Ethernet, fiber optic communication network, or wireless private network. The edge gateway performs time synchronization, missing value processing, outlier removal, and normalization on the acquired data, and then sends the processed data to the optimization calculation module.
[0041] The scene generation module and optimization calculation module can be located in the central control server. The server can be configured with a graphics processing unit (GPU) and a central processing unit (CPU). The GPU is used for training and inference of the generative adversarial network (GAN) model, while the CPU is used to solve the target day-to-day control model, which includes a day-to-day pre-scheduling model and a real-time predictive control model.
[0042] Furthermore, regarding step S1: obtaining the source-network-load-storage time-series tensor set and the source-network-load-storage normal dataset, specifically: It should be noted that the data on load power, photovoltaic output, wind power output, and their rate of change of the power generation, grid, load, and storage system are collected over multiple historical operating cycles and time-aligned according to a unified sampling period. For cases where different data sources use different sampling frequencies, all types of data are converted to the same time interval to ensure that data at the same sampling moment correspond to each other.
[0043] Missing values are filled and outlier measurements are removed from the time-aligned data. Outlier measurements refer to unreasonable data caused by sensor failure, communication interruption, or data recording errors, excluding extreme operating data such as sudden drops in photovoltaic power or load surges that actually occur in the system.
[0044] In the At each sampling time point, a state vector is constructed that includes load, photovoltaic output, wind power output, and corresponding rates of change. : in, Indicates the first Load power at each sampling time; Indicates photovoltaic power output; Indicates wind power output; , and These represent the rates of change of load, photovoltaic power output, and wind power output, respectively.
[0045] Power values reflect the current operating state of the system, while the rate of change reflects the rate of change between adjacent sampling times. For example, a moderate level of photovoltaic output does not necessarily indicate an abnormal state, but a rapid decrease in photovoltaic output within a short period may indicate sudden cloud cover or extreme weather events.
[0046] The state vectors from multiple consecutive sampling moments are arranged in chronological order to form a multidimensional time-series feature tensor, named the source-grid-load-storage time-series tensor set. Thus, a sample includes not only the power state at a specific moment, but also the continuous changes before, during, and after an extreme event.
[0047] Then, data from historical operational data showing the system in a normal state are selected as the source-grid-load-storage normal dataset.
[0048] In another embodiment, the step of calculating the distance between the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and then filtering the source-grid-load-storage time-series tensor set based on the anomaly score dataset to obtain a seed dataset, includes: Based on the source-grid-load-storage normal dataset, the mean value of the normal operating condition state vector and the covariance matrix of the normal operating condition state vector are obtained. Based on the preset Mahalanobis distance algorithm, the distance is calculated between the mean of the normal operating condition state vector, the covariance matrix of the normal operating condition state vector and the source-grid-load-storage time series tensor set to obtain the anomaly score dataset corresponding to the source-grid-load-storage time series tensor set. The seed dataset is obtained by filtering the source-network-load-storage time-series tensor set based on the anomaly score dataset and the preset anomaly threshold.
[0049] It should be noted that data from historical operational data showing the system in a normal state is selected as the normal source-grid-load-storage dataset, and the mean of this normal dataset is calculated. Covariance Matrix And calculate the state vector at each sampling time in the source-network-load-storage time-series tensor set. The Mahalanobis distance is calculated using the following pre-defined algorithm: in, Indicates the first Anomaly score at each sampling time; This represents the state vector at the corresponding moment; This represents the mean of the normal operating condition state vector; The covariance matrix represents the normal operating condition state vector.
[0050] At each sampling time The results are then aggregated to obtain the anomaly score dataset.
[0051] The Mahalanobis distance considers both the deviation of each operating variable from its normal value and the correlation between these variables. Therefore, when a rapid increase in load and a rapid decrease in renewable energy output occur simultaneously, even if a single variable has not exceeded the equipment's limits, the combined state may still achieve a high anomaly score.
[0052] After calculating the anomaly score for each sampling moment, a preset anomaly threshold is set. When the anomaly score corresponding to the state vector at a certain sampling moment exceeds the preset anomaly threshold, that sampling moment is marked as an anomaly moment, indicating that the system's operating state at that moment has significantly deviated from normal operating conditions. Since a single anomaly moment often only reflects a momentary state in the development process of an extreme event and is insufficient to characterize the complete dynamic evolution law of the extreme event, this embodiment uses continuously distributed anomaly moments as the basis for forward and backward expansion along the time axis. Specifically, taking the time interval formed by the continuous anomaly moments as the center, a preset number of sampling points are expanded forward and backward respectively. The state vectors corresponding to all the sampling moments covered after the expansion are arranged in chronological order to form a continuous time series segment containing the initial state before the event, the power change trajectory during the event's development, and the recovery process after the event. This segment is taken as an independent historical extreme sample segment. In the above way, each historical extreme sample segment is composed of state vectors from multiple consecutive sampling moments arranged in chronological order, preserving the complete time dependency relationship of the extreme event from before the event to after recovery, thereby providing a training basis containing dynamic evolution features for the subsequent generative adversarial network model.
[0053] After marking and expanding the abnormal moments, when multiple consecutive sampling moments exceed a preset anomaly threshold, the corresponding historical extreme sample segments are merged into a single historical extreme event. This avoids repeatedly dividing the same event into multiple highly overlapping samples, ensuring that each extreme event appears only once in the seed dataset. This prevents overfitting or training bias during subsequent model training due to highly similar samples. All the selected historical extreme sample segments form the seed dataset. Each sample in the seed dataset is a complete time series segment of an extreme event, composed of state vectors from multiple consecutive sampling moments arranged chronologically. This preserves the complete evolution of the extreme event from before its occurrence to after recovery and provides a real data foundation containing time dependencies and dynamic changes for training the generative adversarial network model in subsequent steps. This completes the construction of the seed dataset.
[0054] In another embodiment, the preset generative adversarial network model includes a preset generator and a preset discriminator. The step of obtaining a seed condition vector set corresponding to the seed dataset based on the seed dataset, and training the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain a target generative adversarial network model includes: The seed condition vector set and the preset noise distribution are input into the preset generator so that the preset generator outputs a virtual extreme sample set. The preset discriminator is trained based on the seed dataset and the virtual extreme sample set to obtain the target discriminator and the discriminator feedback result output by the target discriminator; Based on the virtual extreme sample set, the sample physical violation quantity set corresponding to the virtual extreme sample set is obtained, and the preset generator is updated based on the discriminator feedback result and the sample physical violation quantity to obtain the target generator; The target generation adversarial network model is obtained based on the target generator and the target discriminator.
[0055] It should be noted that corresponding conditional labels are set for each seed sample in the seed dataset. These labels can include extreme event type, impact intensity, duration, time of occurrence, meteorological conditions, system topology, and load trends prior to the event. The meteorological feature matrix... System topology status and historical load trend series Input the conditional encoder to obtain the seed condition vector: in, Represents a condition vector; This represents a meteorological feature matrix consisting of temperature, irradiance, wind speed, and their rates of change. This represents the system topology state, which consists of the line on / off status, unit operating status, and energy storage availability status. Represents a historical load trend sequence; This represents the conditional feature extraction function.
[0056] It should be noted that the conditional feature extraction function can be implemented by the encoder of a variational autoencoder, used to convert high-dimensional meteorological, topological, and load trend data into conditional vectors of fixed dimension. Thus, each seed sample in the seed dataset has a corresponding seed conditional vector, which describes the meteorological, load, and topological conditions under which each extreme scenario occurs, providing conditional constraint information for subsequent generative adversarial network models.
[0057] Next, a physically constrained Wasserstein generative adversarial network (GAN) model is constructed, namely, a pre-defined GAN model. This model includes a pre-defined generator and a pre-defined discriminator. The input to the pre-defined generator includes a random noise vector z and a seed condition vector c. The output is a virtual extreme scenario with the same time length and feature dimension as the seed sample, expressed as: in, Indicates a generator; This represents a random noise vector extracted from a preset noise distribution; This represents the seed condition vector corresponding to a specific extreme working condition; This represents the generated virtual extreme scenario.
[0058] It should be noted that the seed condition vector is used to limit the basic type of the generated scene. For example, when the condition vector represents a cold wave, high load, and low photovoltaic output, the generator should generate a power change curve corresponding to the above conditions. The random noise vector is used to enable the generator to produce different scene change forms under the same conditions, so that different virtual scenes differ in terms of occurrence time, change amplitude, duration, and recovery speed.
[0059] The preset generator employs a bidirectional long short-term memory (LSTM) network structure. This network processes input information in both the forward and backward directions of time, then combines the hidden states from both directions to learn the temporal dependencies between the events preceding, during, and after extreme events. This bidirectional processing is used to generate a complete virtual scene offline and does not imply the use of future experimental data during real-time control.
[0060] The pre-defined discriminator employs a one-dimensional convolutional neural network combined with a self-attention mechanism. It receives both real historical extreme samples and virtual extreme samples output by the generator, along with the corresponding condition vectors. The one-dimensional convolutional neural network extracts local change features between adjacent sampling times, such as identifying whether photovoltaic output drops rapidly in a short period, whether the load increases continuously, and whether there are unreasonable high-frequency fluctuations in the power curve. The self-attention mechanism establishes connections between sampling times separated by a considerable distance, enabling the discriminator to determine whether there is a reasonable continuity between the state before the event, the state at the peak of the event's impact, and the subsequent recovery state. The discriminator distinguishes between real and generated samples based on the curve shape, temporal correlation, and correspondence between the input scene and the condition vector.
[0061] The above completes the construction of the network structure for the generator and discriminator, that is, defining the structure of the generator and discriminator and the corresponding input and output data. The model parameters will be trained using the seed dataset in subsequent steps.
[0062] Then, using the seed dataset and seed condition vector set, adversarial training is performed on the pre-constructed generator and pre-constructed discriminator. During each training iteration, a batch of real extreme samples and their corresponding seed condition vectors are extracted from the seed dataset. Simultaneously, random noise vectors are extracted from a pre-constructed noise distribution. These random noise vectors and a batch of seed condition vectors are input into the pre-constructed generator to obtain generated samples corresponding to the real samples (i.e., the real extreme samples), i.e., the virtual extreme sample set. First, the generator parameters are fixed, and the discriminator parameters of the pre-constructed discriminator are updated using the real samples and generated samples, gradually improving the discriminator's ability to identify the differences in the distributions of the two types of samples. Then, the discriminator parameters are fixed again, and the generator parameters are updated based on the results returned by the discriminator (i.e., the discriminator feedback results) and the physical violation amount of the samples in the virtual extreme sample set.
[0063] The total loss of the generative adversarial network model is a weighted sum of the Wasserstein distance loss term, the gradient penalty term, and the physical constraint penalty term. The Wasserstein distance loss term measures the difference between the distribution of generated samples and the distribution of real samples. When the difference between generated samples and real samples is large, this loss can provide a clear direction for parameter adjustment in the generator. As the generated samples gradually approach the real samples, the distribution difference gradually decreases.
[0064] The gradient penalty term is used to constrain the discriminator to satisfy the Lipschitz continuity condition, preventing the discriminator output from changing too drastically with the input data, thereby improving the training stability of the generative adversarial network. The gradient penalty mainly acts on the discriminator and does not directly replace the device ramping constraint.
[0065] The physical constraint penalty term is used to calculate the device capacity exceeding the limit and the power ramp exceeding the limit in the generated scenario. For the i-th power variable in the generated scenario, its physical constraint penalty is expressed as: in, This represents the total number of power variables included in the physical constraint penalty calculation in the generated scene. Indicates the physical constraint penalty term; Indicates the first in the generated scene The power variable at the th ... Power values at each sampling time; and These represent the upper and lower limits of the allowable power variable, respectively. This indicates the maximum allowable ramp rate for the corresponding equipment; Indicates the sampling time interval; Indicates the climbing constraint weight; Denotes the positive part function, that is: = max(0, a); when a≤0, [a]+ = 0; when a>0, [a]+ = a. This function is used to penalize only the portion exceeding the corresponding physical constraint.
[0066] When the generated power value is within the allowable range of the equipment, no capacity over-limit penalty will be generated. When the generated power value exceeds the installed capacity of the equipment or is lower than the allowable lower limit of the equipment, the excess part will be included in the loss. No ramp-up penalty occurs when the power change between adjacent sampling times does not exceed the maximum allowable change of the device. When the power change exceeds the ramp-up limit of the device, the excess will be included in the loss.
[0067] When the generator outputs complete power generation, energy storage power, external grid exchange power, and load power, power balance residuals can be added to the physical constraint penalty term to further limit the supply and demand relationship between power variables at the same sampling time. If the generator only generates external disturbance variables such as load, photovoltaic output, and wind power output, then equipment capacity constraints and ramp-up constraints are the main contents of the physical constraint penalty.
[0068] Wasserstein distance loss is used to make the generated scene approximate the distribution of real extreme samples, gradient penalty is used to ensure the stability of the training process, and physical constraint penalty is used to reduce the possibility of the generated scene violating the device's operating boundaries. The three solve different problems and work together to participate in model training.
[0069] The preset discriminator and generator are alternately updated according to a preset training epoch until the generator loss and discriminator loss tend to stabilize over multiple consecutive training epochs, or until a preset maximum number of training epochs is reached, resulting in a trained target generator and target discriminator. The trained target generator is used to generate virtual extreme scenarios, and the trained target discriminator is used to monitor the real-time data distribution during subsequent closed-loop feedback updates. Thus, a target generative adversarial network model is constructed based on the target generator and the target discriminator.
[0070] In another embodiment, obtaining the target condition vector set and inputting the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set includes: The target condition vector set is input into the target generative adversarial network model to obtain a set of virtual extreme scenarios; Based on preset tolerance values, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios; The subset of virtual extreme scenarios is further filtered based on the preset maximum available adjustable power to obtain the core extreme scenario set.
[0071] It should be noted that after obtaining the trained generator, a set of virtual extreme scenarios is generated and then the scenarios are reduced. Specifically, target condition labels are set according to the extreme operating conditions to be analyzed. The target condition labels can represent cold waves, high temperatures, typhoons, continuous rain, sudden cloud cover, rapid decline in renewable energy, or line failures. They can also further include event intensity, duration, time of occurrence, load level, and system topology status.
[0072] Using the same condition processing method as described above for generating seed condition vectors, the target condition label for a specific extreme operating condition is input into the condition encoder and converted into a target condition vector. After obtaining the target condition vector, this target condition vector, along with different random noise vectors repeatedly extracted from a preset noise distribution, is input into the target generator in the trained target generative adversarial network model. It should be noted that the target condition vector remains unchanged throughout this process; only the values of the random noise vectors are changed. Each time the target generator receives a set of random noise vectors and the target condition vector, it outputs a multi-dimensional time series containing variables such as load power, photovoltaic output, and wind power output. By repeatedly inputting different random noise vectors, multiple virtual extreme scenarios with different amplitudes, durations, and recovery processes can be obtained under the same extreme operating condition type, thus obtaining a set of virtual extreme scenarios. This effectively expands the coverage of extreme scenarios while preserving the basic characteristics of extreme events.
[0073] In addition, within the reasonable range of conditions covered by the seed sample, continuous condition variables can be combined or interpolated. For example, a high load condition in one historical sample can be combined with a rapid decline in new energy in another historical sample to generate a composite extreme scenario that has not appeared in the history, thereby further enriching the diversity of virtual extreme scenarios.
[0074] It should be noted that the condition combinations and interpolations should be within the range of the model's learned conditions or their reasonable vicinity. Simultaneously, the generated results remain subject to the aforementioned device capacity and ramp-up constraints to avoid generating scenarios that arbitrarily exceed the system's physical boundaries. By inputting the target condition vector set into the trained target generative adversarial network model using the above method, a set of physically feasible virtual extreme scenarios covering various extreme operating conditions can be obtained.
[0075] In another embodiment, the preliminary screening of the set of virtual extreme scenarios based on a preset tolerance value to obtain a subset of virtual extreme scenarios includes: Based on the set of virtual extreme scenarios, obtain the set of virtual physical violations corresponding to the set of virtual extreme scenarios; Based on the preset tolerance value and the set of virtual physical violations, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios.
[0076] It should be noted that after obtaining the set of virtual extreme scenarios, a further physical feasibility verification is performed on each scenario. Specifically, for each virtual extreme scenario in the set, the device capacity violation and power ramp violation are recalculated to obtain the corresponding virtual physical violation amount for each scenario. It should be noted that the calculation method for capacity and ramp violations used here is consistent with the calculation logic of the physical constraint penalty term mentioned above, but the purpose is different: the sample physical violation amount calculated above is used in the training phase, participating in the generator parameter update as part of the loss function; while the virtual physical violation amount calculated in this step is used in the post-verification phase as a screening indicator to measure the physical compliance of the generated scenarios. Specifically, for the i-th power variable in the scenario, it is determined whether the generated power at each time exceeds the device's allowed upper limit or falls below the device's allowed lower limit, and whether the power change at adjacent time moments exceeds the device's allowed maximum ramp rate. All violations are weighted and summed according to preset weights to obtain the comprehensive virtual physical violation amount for that scenario. After calculating the virtual physical violation amount for each scenario, the virtual physical violation amount for each scenario is compared with the preset allowable value. Scenarios with virtual physical violations not exceeding a preset tolerance value are deemed physically feasible and retained; scenarios with virtual physical violations exceeding the preset tolerance value are deemed physically infeasible and are either eliminated or regenerated. Through this preliminary screening based on the preset tolerance value, unusable scenarios that do not meet the physical boundary conditions of the equipment are eliminated. This ensures that the retained scenarios, while satisfying the similarity requirement of the generative adversarial network distribution, further meet the hard constraints of power system equipment capacity and ramp-up capability, thereby improving the executability of scenarios in subsequent optimization calculations. All scenarios that meet the requirements constitute a subset of virtual extreme scenarios. This completes the preliminary screening process for virtual physical violations based on the preset tolerance value.
[0077] In another embodiment, the secondary filtering of the virtual extreme scenario subset based on a preset maximum available adjustable power to obtain the core extreme scenario set includes: Based on the aforementioned subset of virtual extreme scenarios, we obtain the scenario load power set, the scenario photovoltaic power output set, and the scenario wind power output set; Based on the aforementioned scenario load power set, scenario photovoltaic power output set, and scenario wind power output set, the scenario net load set is obtained; A net load gap set is obtained based on the scenario net load set and the preset maximum available adjustable power; The set of net load gaps is processed by time accumulation to obtain a set of destructive force indicators; The core extreme scenario set is obtained by filtering the subset of virtual extreme scenarios based on the set of destructive power indicators.
[0078] In another embodiment, the step of filtering the subset of virtual extreme scenarios based on the set of destructive power indicators to obtain a core set of extreme scenarios includes: Based on the numerical values of the set of destructive power indicators, the subset of virtual extreme scenarios is divided to obtain several initial virtual extreme cluster centers and additional virtual extreme scenarios. Based on the time series distance between several additional virtual extreme scenarios and several initial virtual extreme cluster centers, several target virtual extreme clusters are obtained by matching the several additional virtual extreme scenarios and several initial virtual extreme cluster centers. Each target virtual extreme cluster is filtered based on preset cluster center selection criteria to obtain the core extreme scenario; The core extreme scenario set is obtained based on all core extreme scenarios corresponding to several target virtual extreme clusters.
[0079] It should be noted that due to the large number of virtual extreme scenarios, inputting them all into the day-to-day control model would significantly increase the computational load and affect the solution speed of the rolling control. Therefore, it is necessary to reduce the subset of virtual extreme scenarios. Specifically, for scenario s in the subset of virtual extreme scenarios, the load power at each sampling time in that scenario is first determined as follows: in, Representing a scene At any moment The net load, i.e., the scenario net load; This indicates the load power, i.e., the scenario load power; This indicates the output of photovoltaic power, specifically the photovoltaic output of a given scenario. This indicates the wind power output, specifically the wind power output in the given scenario.
[0080] It should be noted that net load reflects the load demand that the system still needs to meet from dispatchable power sources, energy storage, or the upstream grid after renewable energy generation participates in the balancing process. Simultaneously, based on the current available output of dispatchable power sources, the available discharge power of energy storage, and the input power of the upstream grid, the maximum available regulating power that the system can provide at a given time is determined. .
[0081] When the net load exceeds the maximum available regulating power, the difference between the two constitutes the net load gap, expressed as: in, Representing a scene At any moment The net load gap. This means that if the net load does not exceed the maximum available regulating power, then the net load gap is zero.
[0082] Then, the net load gap throughout the entire time period of the scenario is accumulated over time according to the sampling period Δt to obtain the destructive power index of the scenario: in, Representing a scene The destructive power index; Indicates the number of sampling points in the scene; Indicates the sampling period.
[0083] This destructive force index takes into account both the size and duration of the net load gap, which avoids judging the risk of a scenario based solely on the maximum gap at a single moment, thus providing a more comprehensive assessment of the potential impact of various extreme scenarios on the system.
[0084] After calculating the destructive power index of all virtual extreme scenarios, they are sorted from largest to smallest, and the virtual extreme scenario subset is reduced based on the K-medoids clustering algorithm.
[0085] It should be noted that the K-medoids clustering algorithm is a clustering method that uses actual scenes in the scene set as cluster centers. Unlike the K-means algorithm, K-medoids does not use the method of calculating the average curve to determine the cluster centers, thereby avoiding the averaging process from weakening extreme peaks or forming curves that do not exist in the actual scene set.
[0086] Specifically, firstly, several real-world scenarios with relatively high destructive power indices are selected as initial cluster centers, i.e., initial virtual extreme cluster centers. The remaining scenarios in the subset of virtual extreme scenarios, excluding the initial virtual extreme cluster centers, are then used as additional virtual extreme scenarios. It is understandable that the specific number of real-world scenarios with relatively high destructive power indices can be determined by setting a predefined destructive power index selection threshold, i.e., several real-world scenarios with destructive power indices greater than the predefined destructive power index selection threshold are selected as initial cluster centers.
[0087] Then, the time-series distance between each additional virtual extreme scenario and each initial virtual extreme cluster center is calculated, and several additional virtual extreme scenarios and several initial virtual extreme cluster centers are matched. Specifically, based on the time-series distance, each additional virtual extreme scenario is assigned to the nearest initial virtual extreme cluster center, forming several target virtual extreme clusters. For each target virtual extreme cluster, according to the preset cluster center screening criteria, that is, selecting the scenario with the largest or most representative destructive index within the cluster, the core extreme scenario of the cluster is determined. At the same time, the weight of the corresponding core extreme scenario is determined according to the number of scenarios or the sum of scenario probabilities contained in each cluster. All core extreme scenarios and their weights together constitute the core extreme scenario set. This scenario set is used to represent a variety of representative high-risk disturbances that may occur in the future and serves as the input for subsequent day-ahead pre-scheduling and real-time model predictive control. Thus, the process of secondary screening of the subset of virtual extreme scenarios based on the preset maximum available adjustment power and obtaining the core extreme scenario set is completed.
[0088] Further, regarding step S5: obtaining day-ahead multi-source forecast data; and step S6: solving the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and then performing control based on the source-grid-load-storage scheduling data, specifically: It should be noted that after obtaining the core extreme scenario set, a multi-timescale rolling optimization control model is established, namely, a pre-set day-ahead and intra-day control model, and control commands are executed. This step includes two parts: day-ahead pre-scheduling and real-time model predictive control. The pre-set day-ahead and intra-day control model includes a day-ahead pre-scheduling model and a real-time model predictive control model. Day-ahead pre-scheduling is used to determine the overall operation plan and reserve level in advance for the next scheduling cycle, while real-time model predictive control is used to continuously revise specific control commands based on the actual operating status.
[0089] During the day-ahead pre-schedule phase, day-ahead multi-source forecast data is first acquired. This day-ahead multi-source forecast data includes conventional load forecast data and renewable energy output forecast data. The renewable energy output forecast data further includes photovoltaic output forecast data and wind power output forecast data.
[0090] The day-ahead multi-source forecast data and the core extreme scenario set are used as inputs to the day-ahead pre-scheduling model. The inputs also include equipment capacity parameters, equipment ramp-up parameters, energy storage operation parameters, upstream grid switching power limits, and various operating cost parameters.
[0091] The day-ahead pre-dispatch model uses the next 24 hours or other preset dispatch period as the optimization time domain and divides it into multiple dispatch periods. The decision variables of the day-ahead pre-dispatch model include the planned output of dispatchable power sources, the charging and discharging power of energy storage, the target state of charge of energy storage, the planned exchange power with the upstream grid, the flexible load adjustment amount, and the reserve capacity.
[0092] The current pre-scheduling model employs a distributed bar optimization framework, whose objective function is to minimize the sum of the expected basic operating cost and the conditional risk value, expressed as: in, This represents the pre-scheduled target value for the day; This represents the basic operating cost, including the operating cost of dispatchable power sources, electricity purchase cost, energy storage loss cost, and flexible load adjustment cost; Indicates confidence level Conditional Value at Risk (VaR); This represents the risk aversion coefficient.
[0093] It should be noted that the Conditional Value at Risk (VQR) is calculated based on the core extreme scenario set and is used to represent the average loss in the tail scenario where the loss is greater. Risk loss can include load shedding loss, voltage over-limit penalty, frequency deviation penalty, and equipment over-limit penalty.
[0094] When the scheduling scheme, scenario set, scenario probability, and confidence level are all fixed. It is a definite scalar, but during the optimization process, the scenario loss changes with the scheduling decision, therefore It is a function of scheduling decisions, not a fixed constant.
[0095] in, Let be the conditional value of risk (VaR) of day-ahead scheduling scheme u at confidence level α. u is the day-ahead scheduling decision vector, also known as the day-ahead scheduling scheme. η is an auxiliary optimization variable characterizing the VaR threshold; α is the confidence level, e.g., 0.95; s is the core extreme scenario number; N is the total number of core extreme scenarios. For the probability or normalized weight of scenario s; The risk loss incurred by executing the day-ahead scheduling scheme u in scenario s; (·)+ is the positive part function, which takes the larger value between the value in parentheses and zero.
[0096] By adopting conditional value of risk, even if the probability of a certain extreme scenario is low, as long as it may cause significant losses, it will still affect the day-ahead pre-scheduling results, thereby reserving a certain amount of reserve capacity and energy storage regulation capability in the optimization results.
[0097] The larger the risk aversion coefficient, the more the optimization model values the risks of extreme operating conditions, and usually reserves more backup resources; the smaller the risk aversion coefficient, the more the optimization model values the normal operating costs. By setting the risk aversion coefficient, the relationship between operating economy and the ability to defend against extreme operating conditions can be adjusted.
[0098] The pre-dispatch should also meet the following constraints: power balance constraints, dispatchable power supply capacity constraints, dispatchable power supply ramp-up constraints, energy storage charging and discharging power constraints, energy storage state of charge constraints, tie-line switching power constraints, and flexible load adjustment constraints.
[0099] Solve the day-ahead pre-schedule model to obtain the day-ahead pre-schedule results, namely the scheduled output of dispatchable power sources, the energy storage charging and discharging plan, the energy storage target state of charge, the external grid exchange plan, and the reserve capacity plan for the next scheduling cycle. These results serve as the baseline plan for real-time model predictive control.
[0100] During the real-time model predictive control phase, as the system operates in real time, the latest load power, photovoltaic output, wind power output, energy storage state of charge, dispatchable power output, system frequency, and bus voltage are collected according to a preset control cycle, and the current system state is updated based on the latest measurements. The discrete state-space equations of the power source-grid-load-storage system are established: in, Indicates the first The system state vector at each control moment; These represent control variables, including energy storage charging and discharging commands, dispatchable power output commands, and flexible load adjustment commands. This indicates dynamic disturbances such as load changes and changes in new energy output. Represents the state transition matrix; Represents the control input matrix; This represents the disturbance input matrix. It's understandable that load changes and renewable energy output changes within the core extreme scenario cluster serve as dynamic disturbances. The real-time model predictive control model is input and applied to the system state through the disturbance input matrix E.
[0101] Within each rolling window, one or more scenarios matching the current meteorological conditions, load level, and system state are selected from the core extreme scenarios, and the load changes and new energy output changes in the corresponding scenarios are used as dynamic disturbance inputs to the state space equation.
[0102] Using the pre-scheduled results as the baseline plan and the current measured state as the initial state of the rolling window, the optimal control sequence is solved within a preset prediction time domain. The control sequence should ensure that the system still meets power balance, equipment capacity, equipment ramp-up, energy storage state of charge, voltage, and frequency safety constraints under dynamic disturbances corresponding to core extreme scenarios.
[0103] The objective function of real-time model predictive control includes the adjustment cost of control commands deviating from the day-ahead plan, energy storage usage cost, load shedding penalty, voltage deviation penalty, and frequency deviation penalty.
[0104] After obtaining a control sequence covering the entire prediction window each time, only the first control period instruction in the control sequence is executed. After entering the next control time, the actual operating data is collected again, the system state and disturbance prediction are updated, and a new control sequence is solved again.
[0105] It should be noted that by adopting a rolling optimization method that solves multiple time periods, executes the first time period, and solves again at the next time period, subsequent control schemes can be continuously corrected based on the latest operating data, avoiding the accumulation of prediction errors over multiple consecutive control time periods.
[0106] The optimization calculation module sends control commands to each execution device. The energy storage converter performs charging or discharging according to the control commands. The dispatchable power controller adjusts the target output. The photovoltaic inverter and wind power converter perform power limiting or power tracking. The flexible load controller performs load reduction, load transfer or load restoration.
[0107] After each device completes its adjustment, it feeds back the actual execution power and device status to the data acquisition module, which serves as the basis for updating the status in the next scrolling window.
[0108] Day-ahead pre-scheduling primarily addresses how to allocate resources in advance for the next scheduling cycle, while real-time model predictive control (MMDC) mainly addresses how to promptly correct deviations from the day-ahead plan when actual operating conditions deviate. Day-ahead pre-scheduling provides the overall plan and reserve level, while MMDC generates specific execution instructions based on real-time data. Together, they constitute multi-timescale rolling optimization control. Thus, based on day-ahead multi-source prediction data and a core extreme scenario set, source-grid-load-storage scheduling data is obtained by solving the multi-timescale rolling optimization control model, and control is executed based on this scheduling data.
[0109] Furthermore, in the process of solving the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and in the process of executing the control steps based on the source-grid-load-storage scheduling data, a trained discriminator is used to continuously monitor the real-time data stream.
[0110] Following the same sampling period, feature dimension, and normalization method as the sampling source network load storage time series tensor set, continuous real-time running data is constructed into real-time data segments, and these data segments and their condition vectors are input into the discriminator.
[0111] Based on the features extracted from the discriminator output or internally by the discriminator, the difference between the real-time data distribution and the historical seed data distribution is calculated. This distribution difference can be represented using the Mahalanobis distance or Wasserstein distance in the feature space.
[0112] When the difference between the distribution characteristics of real-time data and the historical distribution does not exceed the preset value, continue to use the existing generator, discriminator and core extreme scenario set for rolling optimization.
[0113] When the distribution difference exceeds a preset value, it indicates that the current actual operating data may belong to a new operating condition that was not fully covered by the original seed dataset. After ruling out sensor malfunctions and communication anomalies, the corresponding real-time data segments and their meteorological information, system topology status, and condition vectors are added to the seed dataset as new samples.
[0114] Using the trained parameters of the original generator and discriminator as initial parameters, the generative adversarial network model is fine-tuned online through transfer learning using new samples and some original seed samples. During the fine-tuning process, the Wasserstein distance loss, gradient penalty, and physical constraint penalty mentioned above are continued to be used to ensure that the training logic before and after the model update is consistent.
[0115] Retaining some of the original seed samples to participate in fine-tuning can reduce the possibility that the model will forget the original extreme scene features when learning new working conditions.
[0116] After the model is fine-tuned, an updated set of virtual extreme scenarios is generated, the virtual physical violation amount and destructive power index of each scenario are recalculated, and K-medoids scenario reduction is performed again to obtain an updated set of core extreme scenarios.
[0117] The updated core extreme scenario set is used for subsequent day-ahead pre-scheduling and real-time model predictive control. This forms a closed-loop update process: real-time data monitoring, identification of distribution changes, addition of new samples, fine-tuning of the generated model, updating of core scenarios, and re-optimization of control.
[0118] In summary, this embodiment first filters real extreme samples from a large amount of historical operating data and establishes the correspondence between extreme samples and meteorological conditions, system topology and load trends.
[0119] Subsequently, a conditional Wasserstein generative adversarial network was established, which uses a bidirectional long short-term memory network as the generator and a network combining a one-dimensional convolutional neural network and a self-attention mechanism as the discriminator.
[0120] The generator and discriminator are then trained using real extreme samples. Wasserstein distance loss gradually brings the generated scene closer to the distribution of real extreme samples, gradient penalty is used to improve training stability, and physical constraint penalty is used to reduce device capacity overruns and hill climb overruns.
[0121] Input specific extreme operating conditions and different random noises into the trained generator to generate multiple virtual extreme scenarios with different change patterns. Then, calculate the destructive force index based on the integral value of the net load gap of each scenario throughout the time period, and obtain a small number of representative core extreme scenario sets through K-medoids clustering.
[0122] The core extreme scenario set is used for day-ahead pre-scheduling and real-time model predictive control. Day-ahead pre-scheduling determines the basic operating plan and reserve level in advance, while real-time model predictive control continuously adjusts the current control commands based on the latest operating status.
[0123] After the system executes control commands, the discriminator continues to monitor the actual operating data. When the actual operating data deviates significantly from the original historical distribution, it is used as a new sample to update and generate the adversarial network and the core extreme scenario set.
[0124] Therefore, this embodiment forms a complete technical chain, including historical extreme sample screening, generative model construction and training, virtual extreme scenario generation, core risk scenario reduction, day-ahead and real-time collaborative control, and actual operation feedback updates. The output of each step serves as the input of the next step, and there is a clear data correspondence between them.
[0125] Furthermore, to verify the feasibility of the method of the present invention, a comparative experiment was conducted in a source-grid-load-storage microgrid simulation system. The simulation system includes photovoltaic power generation units, wind power generation units, dispatchable power sources, energy storage units, flexible loads, and upstream grid connection units.
[0126] Five consecutive years of historical operating data from a local power grid were selected, and 50 historical extreme samples containing characteristics of cold waves, severe convective weather, and sudden changes in photovoltaic output were chosen as the seed dataset. The mean and covariance matrix of the normal state were calculated using the remaining regular operating data.
[0127] In this embodiment, a bidirectional long short-term memory network is used as the generator, and a one-dimensional convolutional neural network and a self-attention mechanism are used as the discriminator. Fifty historical extreme samples and their conditional vectors are input into the generative adversarial network for 5000 rounds of alternating training. After training, different random noise vectors are extracted for preset extreme conditions to generate 1000 virtual extreme scene curves.
[0128] Set the following comparison method: Method A is the method of this invention, which uses a conditional generative adversarial network that includes Wasserstein distance loss, gradient penalty and physical constraint penalty, and adopts multi-timescale rolling optimization based on the core extreme scenario set.
[0129] Method B employs a traditional generative adversarial network without physical constraint penalties and uses conventional stochastic programming.
[0130] Method C employs traditional robust optimization based on historical operational boundaries, without utilizing generative models to expand extreme scenarios.
[0131] Figure 2 This is a morphological comparison diagram between the generated scene and the real extreme scene provided in the embodiments of the present invention. Figure 2 The horizontal axis represents time, in minutes; the vertical axis represents power, in megawatts (MW). Figure 2 The diagram shows three curves: the real historical data curve (black line), the data generated by the method of this invention (red line, corresponding to method A), and the traditional unconstrained data generation curve (blue dashed line, corresponding to method B).
[0132] according to Figure 2 As can be seen, the actual historical data curves present the trajectory of power changes in the power generation, grid, load, and storage system during a certain extreme event. The overall trend is a slight initial increase followed by a period of sharp fluctuations and a rapid drop in power, then a gradual recovery and stabilization. The entire process exhibits significant non-stationarity and abrupt changes. The data curves generated by the method of this invention maintain a high degree of consistency with the actual historical data in terms of overall trend, accurately reproducing the key stages of power increase, decrease, fluctuation, and recovery in the actual data. Furthermore, the power changes between adjacent sampling times are smooth and continuous, without any unreasonable phenomena such as power abrupt changes or high-frequency oscillations. That is, while maintaining the downward trend of extreme output, the changes between adjacent sampling points are constrained by ramp-up constraints. In contrast, although the traditional unconstrained data curves have some similarity to the actual historical data in general trend, significant power jumps occur between multiple adjacent sampling times. The rate of power change in some sections significantly exceeds the ramp-up capability allowed by the equipment, and the curve shape exhibits irregular sawtooth fluctuations. This lacks the physical feasibility of a real power system; that is, the photovoltaic output curve shows significant jumps between adjacent sampling points, accompanied by high-frequency fluctuations.
[0133] The results demonstrate that physical constraint penalties can reduce the probability of unreasonable abrupt changes in the generation curve. This invention introduces physical constraint penalties during the training process of the generative adversarial network (GAN), subjecting the generator to device capacity and ramp-up constraints while it learns the distribution characteristics of real-world extreme samples. This effectively avoids physically unreasonable phenomena such as power and ramp-up exceeding limits in the generation curve, thereby generating virtual extreme scenarios that are statistically close to real-world extreme scenarios and physically executable.
[0134] Figure 3 A bar chart showing the physical consistency error of the generated data provided in this embodiment of the invention. Figure 3 The vertical axis represents the violation rate, expressed as a percentage (%); the horizontal axis represents different violation types, including energy balance violation rate and hill climb violation rate. Figure 3 The figure shows two sets of bar data: the violation rate statistics of the traditional unconstrained generative adversarial network method (i.e., the blue part, representing method B) and the violation rate statistics of the method of the present invention (i.e., the red part, representing method A). Each statistical group includes two bars: energy balance violation rate and ramp-over-limit rate.
[0135] according to Figure 3 It is known that traditional unconstrained generative adversarial networks (GANs) have an energy balance violation rate of approximately 14.5% and a ramp-up violation rate of approximately 14.5%, indicating that the virtual extreme scenarios generated by traditional GANs exhibit a significant proportion of power imbalances and ramp-up violations. If these physically infeasible scenarios are directly used for training or decision-making in subsequent optimization control models, they will severely impact the executability of control commands. In contrast, the method of this invention achieves both an energy balance violation rate and a ramp-up violation rate of approximately 0.01%, significantly lower than traditional unconstrained GAN methods.
[0136] The above comparison results show that by incorporating energy balance constraints, device capacity constraints, and ramp constraints into the generator loss function as penalty terms during the training process of the generative adversarial network, the present invention subjectes the generator to the mandatory constraints of physical laws during the optimization process. This significantly reduces the probability of physical violations such as power imbalance, power exceeding limits, and ramp exceeding limits in the generated scene, thereby generating virtual extreme scenes with high physical consistency and providing reliable scene input for subsequent day-ahead pre-scheduling and real-time model predictive control.
[0137] In the closed-loop simulation, an extreme disturbance is applied at the 10th second, causing the photovoltaic output to drop from 80% to 10% of the rated output within 5 seconds, while simultaneously increasing the load demand by 20%. The source-grid-load-storage system is controlled using various methods, and the system frequency, energy storage power, load shedding amount, and operating cost are recorded.
[0138] Figure 4A comparison chart of system frequency response under extreme operating conditions provided for embodiments of the present invention. Figure 4 The horizontal axis represents time, in seconds (s); the vertical axis represents the system frequency, in hertz (Hz). Figure 4 The figure shows three frequency response curves: the frequency response curve of the control strategy of the present invention (i.e., method A), the frequency response curve of the traditional robust optimization (i.e., method C), and the frequency response curve of the conventional model predictive control.
[0139] according to Figure 4 It can be seen that before the disturbance occurred, the system frequency under all three control strategies was stable at around the rated frequency of 50Hz. After the extreme disturbance was applied at t=10 seconds, a sudden photovoltaic power deficit occurred, and the frequency response curve of the conventional model predictive control dropped sharply, with its lowest frequency point below 49.5Hz, reaching the unsafe operating region. Moreover, the frequency recovery speed was slow, indicating that the conventional model predictive control, which did not utilize virtual extreme scenarios for pre-scheduling, lacked sufficient reserve capacity and energy storage regulation capability when facing extreme disturbances, making it difficult to maintain the frequency within the safe operating range. Although the frequency response curve of the traditional robust optimization did not drop below 49.5Hz, its lowest frequency point was still significantly lower than that of the control strategy of this invention, and there were certain fluctuations during the frequency recovery process. This indicates that although the traditional robust optimization reserved a certain reserve by setting a fixed uncertainty boundary, the construction of its uncertainty set depends on a manually set single boundary or a limited historical scenario, resulting in a deviation between the control resource allocation and the actual disturbance process, making it difficult to achieve precise regulation resource scheduling. The frequency response curve of the control strategy of this invention has the smallest frequency drop after a disturbance occurs, and the lowest frequency value is always kept within the preset safe range (above 49.5Hz). Moreover, the frequency can recover to the rated frequency of around 50Hz in a short time and remain stable.
[0140] The above comparison results show that the present invention can pre-schedule the system day-ahead by using the core extreme scenario set and its corresponding conditional risk value. This allows for the early determination of reasonable energy storage state of charge targets and reserve capacity schemes. At the same time, it can reallocate energy storage and dispatchable power according to the actual operating status within a real-time rolling window. This enables the system to have better frequency response capability and faster recovery speed under extreme disturbances, effectively improving the power supply reliability and frequency security level of the power source-grid-load-storage system under extreme operating conditions.
[0141] like Figure 5 As shown, Method C reduces the risk of load shedding by setting a larger fixed uncertainty boundary, but the corresponding reserve and operating costs are higher; Method B is affected by physical violations in the generated scenario, and some control strategies cannot fully adapt to actual extreme disturbances, resulting in a relatively large load shedding amount; Method A determines the risk reserve through the core extreme scenario and conditional risk value, and under this simulation condition, it simultaneously achieves a lower load shedding amount and a lower operating cost, specifically: Figure 5 A comparison chart of load shedding and operating costs under extreme operating conditions provided for embodiments of the present invention. Figure 5 The horizontal axis represents the economic cost of operation, in ten thousand yuan; the vertical axis represents the load shedding under extreme scenarios, in megawatt-hours (MWh). Figure 5 The comparison results of three control strategies are shown: deterministic optimization strategy, traditional robust optimization strategy (i.e., method C), and the control strategy of the present invention (i.e., method A). Figure 5 Each data point represents the combined result of operating economic cost and load shedding amount obtained by running a complete simulation cycle under specific parameter configuration.
[0142] See Figure 5 It can be seen that the data points of the deterministic optimization strategy ( Figure 5 The load shedding (represented by triangles) is distributed in the upper-middle region of the graph, with a load shedding range of 50 to 60 megawatt-hours and an operating economic cost range of 1 million to 2 million yuan. It should be noted that the deterministic optimization strategy corresponds to a combination of traditional unconstrained generative adversarial networks and conventional stochastic programming (i.e., Method B). Because the virtual extreme scenarios generated by this method contain physical violations, the control strategy trained based on unexecutable scenarios cannot adequately adapt to actual extreme disturbances, resulting in large fluctuations in load shedding under different parameter configurations and an overall high load shedding rate.
[0143] Traditional robust optimization strategies (data points) Figure 5 The load shedding (represented by circular markers) is distributed in the lower right region of the figure, with a load shedding range of 5 to 20 MWh and an operating economic cost range of 5 to 6 million yuan. As the operating economic cost increases (i.e., the reserved reserve capacity increases), the load shedding shows a downward trend. However, even under conditions of high operating economic cost, the load shedding remains around 5 to 20 MWh, failing to decrease further. This indicates that while this method uses a large fixed uncertainty boundary to cover extreme operating conditions and reduce load shedding risk, the overly conservative construction of the uncertainty set leads to excessively large reserve capacity settings, significantly increasing operating costs, and exhibiting technical limitations in completely eliminating load shedding risk.
[0144] The multiple data points of the control strategy of this invention (represented by star marks in the figure) are distributed in the lower left area of the figure, with a load shedding range of 0 to 10 megawatt-hours and an operating economic cost range of 1.5 million to 3 million yuan.
[0145] Further comparison shows that, under the same load shedding conditions (e.g., both 10 MWh), the operating economic cost of the method of this invention is approximately 2 million yuan, while the operating economic cost of the traditional robust optimization strategy is approximately 5.5 million yuan. The method of this invention reduces costs at the same load shedding level. Under the same operating economic cost conditions, the load shedding amount of the method of this invention is approximately 5 MWh, while the load shedding amount of the deterministic optimization strategy is approximately 50 to 60 MWh. The method of this invention reduces the load shedding amount at the same cost level.
[0146] The above comparison results show that the present invention determines risk reserves by using the core extreme scenario set and its corresponding conditional risk value. It can reserve reasonable adjustment capabilities for high-risk extreme scenarios at a lower cost during the day-ahead pre-scheduling stage. At the same time, it can accurately allocate adjustment resources according to the actual operating status within the real-time rolling window. Thus, under the simulation conditions, it can simultaneously obtain a lower load shedding amount and a lower operating economic cost, achieving a better balance between operating economy and extreme condition defense capabilities.
[0147] Figure 6 A visualization of the diversity distribution of generated scenarios provided in this embodiment of the invention. Figure 6 The horizontal axis represents the first feature dimension after dimensionality reduction, and the vertical axis represents the second feature dimension after dimensionality reduction. Both dimensions are dimensionless units and are obtained by dimensionality reduction of the feature vectors of the intermediate layer of the generative adversarial network. They are used to visualize the relative positional relationship of samples in the high-dimensional feature space on a two-dimensional plane. Figure 6 The diagram shows the distribution of two types of sample points, where the cross markers indicate the positions of historical real extreme samples in the dimension-reduced feature space, and the circle markers indicate the positions of the virtual extreme scenes generated by the method of this invention (i.e., method A) in the dimension-reduced feature space.
[0148] See Figure 6 It is known that the distribution of historical extreme samples in the dimensionality-reduced feature space is relatively sparse, with sample points existing only in a few local regions. The spacing between sample clusters is large, and the boundary range covered by the sample clusters is limited. This indicates that the number of usable extreme samples in real historical data is limited, and the feature differences between samples are significant, making it difficult to cover different types of extreme working conditions and their combinations. The virtual extreme scenes generated by the method of this invention not only cover the regions between multiple historical sample clusters in the dimensionality-reduced feature space, filling the blank areas in the historical sample distribution, but also extend to some edge regions of the historical sample distribution, forming a more continuous distribution pattern. This shows that the virtual scenes generated by the method of this invention, while retaining the basic characteristics of historical extreme samples, produce diverse scenes with differences in variation amplitude, duration, and recovery process by inputting different random noise vectors under the same condition vector.
[0149] The above comparison results show that by introducing Wasserstein distance loss and gradient penalty terms during the training process of generative adversarial networks, the present invention can effectively reduce the risk of mode collapse, enabling the generator to produce different extreme scenarios under different random noise inputs, avoiding the generator from repeatedly generating only a few similar scenarios. This expands the coverage of extreme scenarios while maintaining the diversity of generated samples, providing richer risk scenario inputs for subsequent scenario reduction and optimization control.
[0150] Under the dataset, equipment parameters, and extreme perturbation conditions used in this embodiment, Method A reduces the risk of load loss by more than 40% compared to traditional robust optimization based solely on historical boundaries. The above experimental results illustrate that a complete data transfer relationship can be formed between scene generation, physical constraints, scene reduction, and multi-timescale rolling control described in this invention. The specific degree of improvement in practical applications depends on the scale of the source-grid-load-storage system, equipment adjustment capabilities, the number of seed samples, and model parameter settings.
[0151] like Figure 7 As shown, based on the above method embodiments, corresponding apparatus embodiments are provided; An embodiment of the present invention provides a source-grid-load-storage coordinated control device, comprising: The basic data acquisition module is used to acquire the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset; The seed dataset acquisition module is used to calculate the distance between the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and to filter the source-grid-load-storage time series tensor set based on the anomaly score dataset to obtain the seed dataset. The target generative adversarial network model generation module is used to obtain the seed condition vector set corresponding to the seed dataset based on the seed dataset, and to train the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model. The core extreme scenario set acquisition module is used to acquire the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set. The current day multi-source forecast data acquisition module is used to acquire current day multi-source forecast data; The source-grid-load-storage scheduling data acquisition module is used to solve the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and to perform control based on the source-grid-load-storage scheduling data.
[0152] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the source-grid-load-storage coordinated control method provided by any of the above-described method embodiments of the present invention.
[0153] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.
[0154] Based on the above-described embodiment of the source-grid-load-storage coordinated control method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a source-grid-load-storage coordinated control method according to any embodiment of the present invention.
[0155] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.
[0156] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0157] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.
[0158] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the source-grid-load-storage coordinated control method described in any of the above-described method embodiments of the present invention.
[0159] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
[0160] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A source-grid-load-storage coordinated control method, characterized in that, include: Obtain the source-network-load-storage time-series tensor set and the source-network-load-storage normal dataset; Distance calculation is performed on the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and the source-grid-load-storage time series tensor set is filtered based on the anomaly score dataset to obtain a seed dataset; Based on the seed dataset, a seed condition vector set corresponding to the seed dataset is obtained, and a preset generative adversarial network model is trained based on the seed dataset and the seed condition vector set to obtain a target generative adversarial network model. Obtain the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set; Obtain recent multi-source forecast data; Based on the aforementioned multi-source forecast data and core extreme scenario set, the preset daytime and intraday control model is solved to obtain source-grid-load-storage scheduling data, which is then used for control.
2. The source-grid-load-storage coordinated control method according to claim 1, characterized in that, The step of calculating the distance between the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and then filtering the source-grid-load-storage time-series tensor set based on the anomaly score dataset to obtain a seed dataset, includes: Based on the source-grid-load-storage normal dataset, the mean value of the normal operating condition state vector and the covariance matrix of the normal operating condition state vector are obtained. Based on the preset Mahalanobis distance algorithm, the distance is calculated between the mean of the normal operating condition state vector, the covariance matrix of the normal operating condition state vector and the source-grid-load-storage time series tensor set to obtain the anomaly score dataset corresponding to the source-grid-load-storage time series tensor set. The seed dataset is obtained by filtering the source-network-load-storage time-series tensor set based on the anomaly score dataset and the preset anomaly threshold.
3. The source-grid-load-storage coordinated control method according to claim 1, characterized in that, The preset generative adversarial network model includes a preset generator and a preset discriminator. The step of obtaining a seed condition vector set corresponding to the seed dataset based on the seed dataset, and training the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model includes: The seed condition vector set and the preset noise distribution are input into the preset generator so that the preset generator outputs a virtual extreme sample set. The preset discriminator is trained based on the seed dataset and the virtual extreme sample set to obtain the target discriminator and the discriminator feedback result output by the target discriminator; Based on the virtual extreme sample set, the sample physical violation quantity set corresponding to the virtual extreme sample set is obtained, and the preset generator is updated based on the discriminator feedback result and the sample physical violation quantity to obtain the target generator; The target generation adversarial network model is obtained based on the target generator and the target discriminator.
4. The source-grid-load-storage coordinated control method according to claim 1, characterized in that, The process of obtaining the target condition vector set and inputting it into the target generative adversarial network model to obtain the core extreme scenario set includes: The target condition vector set is input into the target generative adversarial network model to obtain a set of virtual extreme scenarios; Based on preset tolerance values, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios; The subset of virtual extreme scenarios is further filtered based on the preset maximum available adjustable power to obtain the core extreme scenario set.
5. The source-grid-load-storage coordinated control method according to claim 4, characterized in that, The preliminary screening of the set of virtual extreme scenarios based on preset tolerance values yields a subset of virtual extreme scenarios, including: Based on the set of virtual extreme scenarios, obtain the set of virtual physical violations corresponding to the set of virtual extreme scenarios; Based on the preset tolerance value and the set of virtual physical violations, the set of virtual extreme scenarios is initially screened to obtain a subset of virtual extreme scenarios.
6. The source-grid-load-storage coordinated control method according to claim 4, characterized in that, The process of performing a secondary screening on the subset of virtual extreme scenarios based on a preset maximum available adjustable power to obtain the core extreme scenario set includes: Based on the aforementioned subset of virtual extreme scenarios, we obtain the scenario load power set, the scenario photovoltaic power output set, and the scenario wind power output set; Based on the aforementioned scenario load power set, scenario photovoltaic power output set, and scenario wind power output set, the scenario net load set is obtained; A net load gap set is obtained based on the scenario net load set and the preset maximum available adjustable power; The set of net load gaps is processed by time accumulation to obtain a set of destructive force indicators; The core extreme scenario set is obtained by filtering the subset of virtual extreme scenarios based on the set of destructive power indicators.
7. The source-grid-load-storage coordinated control method according to claim 6, characterized in that, The process of filtering the subset of virtual extreme scenarios based on the set of destructive power indicators yields a core set of extreme scenarios, including: Based on the numerical values of the set of destructive power indicators, the subset of virtual extreme scenarios is divided to obtain several initial virtual extreme cluster centers and additional virtual extreme scenarios. Based on the time series distance between several additional virtual extreme scenarios and several initial virtual extreme cluster centers, several target virtual extreme clusters are obtained by matching the several additional virtual extreme scenarios and several initial virtual extreme cluster centers. Each target virtual extreme cluster is filtered based on preset cluster center selection criteria to obtain the core extreme scenario; The core extreme scenario set is obtained based on all core extreme scenarios corresponding to several target virtual extreme clusters.
8. A source-grid-load-storage coordinated control device, characterized in that, include: The basic data acquisition module is used to acquire the source-grid-load-storage time-series tensor set and the source-grid-load-storage normal dataset; The seed dataset acquisition module is used to calculate the distance between the source-grid-load-storage time series tensor set and the source-grid-load-storage normal dataset to obtain an anomaly score dataset, and to filter the source-grid-load-storage time series tensor set based on the anomaly score dataset to obtain the seed dataset. The target generative adversarial network model generation module is used to obtain the seed condition vector set corresponding to the seed dataset based on the seed dataset, and to train the preset generative adversarial network model based on the seed dataset and the seed condition vector set to obtain the target generative adversarial network model. The core extreme scenario set acquisition module is used to acquire the target condition vector set and input the target condition vector set into the target generative adversarial network model to obtain the core extreme scenario set. The current day multi-source forecast data acquisition module is used to acquire current day multi-source forecast data; The source-grid-load-storage scheduling data acquisition module is used to solve the preset day-ahead intraday control model based on the day-ahead multi-source forecast data and the core extreme scenario set to obtain source-grid-load-storage scheduling data, and to perform control based on the source-grid-load-storage scheduling data.
9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a source-grid-load-storage coordinated control method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a source-grid-load-storage coordinated control method as described in any one of claims 1-7.
