Method for continuous switching of thermal power energy storage frequency regulation responsibility based on deep autoencoder

Abstract

The application belongs to the technical field of thermal power control, and specifically discloses a thermal power energy storage frequency regulation responsibility continuous switching method based on a deep autoencoder, which comprises the following steps: based on pre-acquired online resource amounts of multiple thermal power units, the member responsibility characteristics of each unit and energy storage are calculated; a deep autoencoding state space model is inputted, each member responsibility characteristic is mapped into a responsibility bearing representation, and is gathered into a responsibility hidden distribution; under constraints, topological stable segment evolution is performed on the responsibility hidden distribution to obtain a current responsibility hidden distribution; based on the online resource amounts, topological change determination is performed; when the topological change condition is met, a target support relationship is established according to the member responsibility characteristics of the residual units and energy storage, and a responsibility migration constraint is formed; under the constraint, non-equilibrium optimal transmission is performed on the current responsibility hidden distribution to generate a migration responsibility hidden distribution; the deep autoencoding state space model is inputted, and a responsibility share result is outputted; and based on the responsibility share result, an energy storage responsibility increment and a residual unit responsibility share are generated.

Description

Technical Field

[0001] This application belongs to the field of thermal power control technology, and more specifically, relates to a method for continuous switching of frequency regulation responsibility in thermal power energy storage based on a deep self-encoder. Background Technology

[0002] In existing technologies, the sharing of energy storage by multiple thermal power units for overall AGC or joint frequency regulation of the plant is typically achieved through a combination of station-level total control and member allocation. Common schemes first collect data on unit output, output limits, grid connection status, frequency deviation, and energy storage output. Then, they calculate the unit's adjustable margin, available energy storage power, and frequency regulation participation coefficient. Finally, they generate the unit and energy storage sharing results according to capacity ratio, margin ratio, or preset responsibility coefficient. Some schemes further input multi-moment operational data into a state-space model, recursive estimation model, or neural network model, directly outputting the unit power correction or energy storage power correction for the next moment. When faced with changes in member positions, existing schemes generally combine circuit breaker position signals and unit grid connection status to synchronously update the input structure, state variables, or responsibility ratios.

[0003] In the scenario of unit disconnection, most existing solutions focus on parameter recalculation, dimension reduction, or ratio updates, which fails to reflect the operational mechanism that the object to which frequency regulation responsibility is attached has changed. Due to the lack of responsibility status that is bound to the member's location and can be continuously transmitted before and after topology changes, the responsibility before disconnection is easily stuck at the location of the unit that has already been disconnected, or is roughly pushed into the remaining units and energy storage during the process of dimension deletion, zeroing, and re-normalization. This leads to distorted responsibility attribution, unclear acceptance boundaries, and instantaneous state jumps during disconnection, which further affects the continuous generation of shared energy storage capacity and the closed-loop execution of plant-side responsibility allocation, resulting in low accuracy of frequency regulation responsibility allocation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application aims to provide a method for continuous switching of frequency regulation responsibility based on deep autoencoders for thermal power energy storage, which addresses the problem of low accuracy in frequency regulation responsibility allocation caused by existing technologies focusing on parameter recalculation, dimension reduction, or proportional updates.

[0005] To achieve the above objectives, in a first aspect, this application provides a method for continuous switching of frequency regulation responsibility in thermal power storage based on a deep autoencoder, comprising: Based on the pre-acquired online resource quantities of multiple thermal power units, calculate the member responsibility characteristics of each unit and energy storage; The member responsibility features are input into a deep Koopman autoencoder state space model, and each member responsibility feature is mapped to a responsibility-bearing representation. The features are then aggregated into a responsibility latent distribution according to the current online resource set. Under the constraints of the current online resource set, the responsibility latent distribution is subjected to topologically stable segment evolution to obtain the current responsibility latent distribution. Based on the online resource quantity, topology change is determined. When the topology change condition is met, the remaining units and energy storage are screened according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, a target support relationship is established for the transfer of the current hidden distribution of responsibility to the set of online resource quantity after disconnection, thus forming a responsibility migration constraint. Under the responsibility migration constraint, an unbalanced optimal transfer is performed on the current responsibility hidden distribution, allowing partial responsibility migration and partial cancellation, thereby generating a migrated responsibility hidden distribution; The migration responsibility hidden distribution is input into the deep Koopman autoencoder state space model for demultiplexing, evolution and decoding, and the responsibility share result is output. Based on the responsibility share result, the energy storage responsibility increment and the remaining unit responsibility share are generated, and the shared energy storage frequency regulation undertaking amount is output.

[0006] This application combines a deep Koopman autoencoder state-space model with unbalanced optimal transmission, constructing a complete solution chain from online resource acquisition to shared energy storage frequency regulation load output around the unique intermediate object of the responsibility latent distribution. By defining the responsibility latent distribution as a unique latent state, the latent state simultaneously carries the responsibility magnitude, responsibility attribution, and load boundary, ensuring that the responsibility source object before and after unit disconnection maintains clear semantics at the member location level. This prevents responsibility from remaining in physically meaningless locations before disconnection, improving the continuity of shared energy storage frequency regulation responsibility identification and subsequent load calculation. After determining the topology change, a target support relationship is established and a responsibility migration constraint is formed, which is then applied during responsibility migration. The unbalanced optimal transmission under constraints allows for partial migration and cancellation of responsibilities, avoiding the misallocation of responsibilities and state jumps caused by mechanically pushing the responsibilities of decommissioned units into remaining resources in existing technologies. By implicitly distributing responsibilities throughout the entire process of responsibility generation, migration, post-decommissioning evolution, and the output of the received quantity, a closed-loop connection around the same responsible object is formed. This solves the problem in existing technologies where responsibility redistribution, state updates, and result output are scattered in multiple stages, making it difficult to maintain consistent transmission of responsibilities in scenarios with changes in resource sets. It can maintain the business closed loop of shared energy storage frequency regulation responsibility allocation and received quantity output in the scenario of unit decommissioning, improve the accuracy of frequency regulation responsibility allocation, and has clear engineering application value.

[0007] According to the method for continuous switching of frequency regulation responsibility based on deep autoencoder provided in this application, the responsibility hidden distribution is arranged according to fixed member positions, and each member position corresponds to a fixed dimension of responsibility bearing state. The responsibility bearing state is obtained by combining the responsibility bearing representation with the underlying numerical fields in the member responsibility characteristics, so that the responsibility hidden distribution carries the responsibility size, responsibility attribution and bearing boundary at the same time. Member positions that do not belong to the current online resource quantity set are written to a fixed zero state placeholder, and no responsibility quantity is introduced from other member positions.

[0008] This application employs a structure design that restricts the implicit distribution of responsibility to a fixed member position arrangement, with each member position corresponding to a fixed-dimensional responsibility-bearing state. It stipulates that member positions not belonging to the current online resource set are written into fixed zero-state placeholders without introducing responsibility from other member positions. This ensures that each dimension of the implicit distribution forms a fixed binding relationship with a specific member position, maintaining clear semantics at the member position level before and after decoupling, thus avoiding the loss of responsibility attribution information caused by variable dimensions or general compressed vectors. By combining the responsibility-bearing representation with the underlying numerical fields (margin, frequency difference response slope, etc.) in the member responsibility characteristics during convergence, the physical meaning of the responsibility-bearing state is traceable, providing a directly readable source of constraints for establishing target support relationships. The fixed zero-state placeholder method ensures that exited member positions do not introduce false responsibility, while maintaining the total dimension unchanged. This eliminates the need for dynamic adjustment of the encoding / decoding structure of the deep Koopman autoencoder state space model and the input / output dimensions of unbalanced optimal transmission throughout the entire runtime, reducing model switching complexity.

[0009] According to the method for continuous switching of frequency regulation responsibility based on deep autoencoder provided in this application, the step of establishing a target support relationship for the transfer of the current hidden distribution of responsibility to the set of online resources after the disconnection based on the member responsibility characteristics of the remaining units and energy storage includes: Based on the set of online resources after disconnection, extract the member responsibility characteristics of the remaining generating units and energy storage; For each migration path from the source member location to the target member location, the acceptable boundary is calculated. When the target member location is a generator unit, the up and down adjustment margin path is automatically selected. When the target member location is energy storage, the energy storage receiving margin path is automatically selected. The acceptable boundary of the disconnected member location is set to zero, thus forming the target support relationship.

[0010] This application automatically selects the up-and-down margin path when the target member location is a generating unit and the energy storage capacity path when it is energy storage. This allows for the differentiation and handling of the physical characteristics differences of different types of resources at the constraint level, avoiding constraint distortion caused by using a uniform scalar or preset ratio to describe the capacity of heterogeneous resources. By assigning zero to the receptive boundary of disconnected member locations, the application blocks the return path of responsibility to withdrawn resources from the source of constraints, eliminating the hidden danger that withdrawn units still retain false receiving capacity in the target support relationship. The application calculates the receptive boundary based on multi-dimensional physical quantities such as the saturation proximity of the remaining units and the short-time frequency difference response slope, so that the transfer upper limit of each migration path directly corresponds to the actual on-site operational constraints. This provides a physically interpretable constraint basis for unbalanced optimal transmission and improves the rationality of the responsibility migration results.

[0011] According to the method for continuous switching of frequency regulation responsibility based on deep autoencoder provided in this application, the step of performing unbalanced optimal transmission on the current hidden responsibility distribution under the responsibility migration constraint, allowing partial migration and partial cancellation of responsibility, and generating a migrated hidden responsibility distribution includes: Using the transfer direction and transfer upper limit in the responsibility migration constraint as constraints, perform unbalanced optimal transfer on the current hidden responsibility distribution; Using the cross-member location migration constraint coefficient and the responsibility cancellation constraint coefficient as regularization terms, the responsibility transfer amount and responsibility cancellation amount to each acceptable member location are solved item by item for each responsibility component of the source member location; The responsibility transfer amounts received by each receiving member position are accumulated according to the member position and reconstructed in a fixed order into a vector with the same dimension as the responsibility latent distribution, which is used as the migration responsibility latent distribution.

[0012] This application directly incorporates the transfer direction and upper limit of the responsibility transfer constraint into the optimization solution process instead of post-processing correction, ensuring that each iteration of the unbalanced optimal transfer strictly adheres to physical constraints and avoids constraint violations or result distortion caused by external pruning after the solution. It introduces cross-member position transfer constraint coefficients and responsibility cancellation constraint coefficients as regularization terms, allowing partial transfer and cancellation while controlling the transfer magnitude and cancellation ratio to prevent extreme cases of excessive concentration or loss of responsibility during the transfer process. The transferred responsibility hidden distribution is reconstructed according to the same fixed member position order as the current responsibility hidden distribution, ensuring complete consistency of dimensional semantics before and after the transfer. This allows the evolution and decoding stages after unpacking to directly receive the transferred responsibility hidden distribution as input without additional dimensional adaptation or semantic remapping, maintaining the continuity of the entire solution chain.

[0013] According to the method for continuous switching of frequency regulation responsibility based on deep autoencoder provided in this application, the unbalanced optimal transmission is embedded in the hidden state switching initialization stage of the deep Koopman autoencoder state space model. The embedding position is located after the current hidden responsibility distribution completes the evolution of the topology stable segment and before the migration hidden responsibility distribution enters the post-disjoint evolution, so as to enhance the responsibility migration capability of the deep Koopman autoencoder state space model across resource sets when the topology changes. The unbalanced optimal transmission does not generate solution results independent of the deep Koopman autoencoder state-space model.

[0014] This application clarifies that the unbalanced optimal transfer addresses the shortcoming of the deep Koopman autoencoder state-space model in the lack of cross-resource set responsibility transfer capability during the hidden state switching initialization stage, rather than building an independent allocation module outside the model. This avoids the inconsistency of intermediate objects and solution path divergence problems that may arise from the dual-backbone parallel architecture. The output of the unbalanced optimal transfer directly regresses to the original evolution chain of the deep Koopman autoencoder state-space model, ensuring that the responsibility share result is generated entirely from a single model skeleton. This guarantees the consistency of the generation path of the output results before and after decoupling, facilitating engineering implementation and debugging verification. This embedding method strictly limits the handling of topology changes to the hidden state switching initialization stage. The evolution of topology-stable segments and the model structure after decoupling do not need to be changed, maximizing the preservation of the original state tracking capability of the deep Koopman autoencoder state-space model. At the same time, the unbalanced optimal transfer endows the model with the responsibility switching capability when the resource set changes.

[0015] According to the method for continuous switching of frequency regulation responsibility for thermal power energy storage based on a deep autoencoder provided in this application, the step of generating the energy storage responsibility increment and the remaining unit responsibility share based on the responsibility share result includes: Extract the energy storage responsibility share from the responsibility share results, calculate the difference between the energy storage responsibility share and the corresponding value of the energy storage member position in the current responsibility hidden distribution before the topology change, and obtain the energy storage responsibility increment. The energy storage responsibility increment retains the sign information and does not perform absolute value processing or double normalization. The remaining unit responsibility share is selected from the responsibility share results for units whose grid connection status is "connected".

[0016] This application directly compares the energy storage responsibility share obtained after decoupling with the energy storage responsibility value in the implicit distribution of responsibility before decoupling. This ensures that the incremental energy storage responsibility accurately reflects the net change in energy storage responsibility caused by topological changes, rather than the absolute responsibility of energy storage. This avoids misjudgment of the amount of responsibility taken on due to inconsistencies in the benchmarks before and after decoupling. By retaining symbolic information, the incremental energy storage responsibility can directly distinguish whether energy storage needs to increase or decrease its frequency regulation responsibility, providing a directional decision-making basis for the energy storage charging and discharging control strategy at the plant side. By not performing absolute value processing and secondary normalization, the incremental energy storage responsibility strictly follows the semantics of the same responsibility object established by the implicit distribution of responsibility, maintaining consistency with the responsibility share result and the current implicit distribution of responsibility in terms of numerical scale and physical meaning. This ensures the traceability and verifiability of the shared energy storage frequency regulation take-on amount throughout the entire business chain.

[0017] Secondly, this application provides a thermal power energy storage frequency regulation responsibility continuous switching device based on a deep self-encoder, comprising: The calculation module is used to calculate the member responsibility characteristics of each unit and energy storage based on the pre-acquired online resource quantities of multiple thermal power units; The evolution module is used to input the member responsibility features into the deep Koopman autoencoder state space model, map each member responsibility feature into a responsibility-bearing representation, and aggregate them into a responsibility latent distribution according to the current online resource quantity set. Under the constraints of the current online resource quantity set, the responsibility latent distribution is subjected to topologically stable segment evolution to obtain the current responsibility latent distribution. The construction module is used to determine topology changes based on the online resource quantity. When the topology change conditions are met, the remaining units and energy storage are screened according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, the target support relationship for the transfer of the current hidden distribution of responsibility to the set of online resource quantity after disconnection is established, forming a responsibility migration constraint. The generation module is used to perform unbalanced optimal transfer on the current hidden responsibility distribution under the responsibility migration constraint, allowing partial migration and partial cancellation of responsibility, and generating a migrated hidden responsibility distribution; The output module is used to input the migration responsibility hidden distribution into the deep Koopman autoencoder state space model for demultiplexing, evolution and decoding, output responsibility share results, generate energy storage responsibility increment and remaining unit responsibility share based on the responsibility share results, and output the shared energy storage frequency regulation undertaking amount.

[0018] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method for continuous switching of frequency regulation responsibility based on deep autoencoder for thermal power storage as described in the first aspect or any possible implementation of the first aspect.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder, as described in the first aspect or any possible implementation of the first aspect.

[0020] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to execute the method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder, as described in the first aspect or any possible implementation of the first aspect.

[0021] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0022] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application combines a deep Koopman autoencoder state-space model with unbalanced optimal transmission, constructing a complete solution chain from online resource acquisition to shared energy storage frequency regulation load output around the unique intermediate object of the responsibility latent distribution. By defining the responsibility latent distribution as a unique latent state, the latent state simultaneously carries the responsibility magnitude, responsibility attribution, and load boundary, ensuring that the responsibility source object before and after unit disconnection maintains clear semantics at the member location level. This prevents responsibility from remaining in physically meaningless locations before disconnection, improving the continuity of shared energy storage frequency regulation responsibility identification and subsequent load calculation. After determining the topology change, a target support relationship is established and a responsibility migration constraint is formed, which is then applied during responsibility migration. The unbalanced optimal transmission under constraints allows for partial migration and cancellation of responsibilities, avoiding the misallocation of responsibilities and state jumps caused by mechanically pushing the responsibilities of decommissioned units into remaining resources in existing technologies. By implicitly distributing responsibilities throughout the entire process of responsibility generation, migration, post-decommissioning evolution, and the output of the received quantity, a closed-loop connection around the same responsible object is formed. This solves the problem in existing technologies where responsibility redistribution, state updates, and result output are scattered in multiple stages, making it difficult to maintain consistent transmission of responsibilities in scenarios with changes in resource sets. It can maintain the business closed loop of shared energy storage frequency regulation responsibility allocation and received quantity output in the scenario of unit decommissioning, improve the accuracy of frequency regulation responsibility allocation, and has clear engineering application value. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the continuous switching method for frequency regulation responsibility of thermal power energy storage based on a deep autoencoder provided in an embodiment of this application. Figure 2 The unit provided in the embodiments of this application Schematic diagram of the timing evolution of energy storage frequency regulation responsibility; Figure 3 This is a block diagram of the plant AGC-shared energy storage coordinated frequency regulation structure provided in the embodiments of this application; Figure 4 This is a responsibility migration path and resource constraint matrix diagram under topology changes provided in the embodiments of this application; Figure 5 This is a comparison diagram of the stability covariance ellipse of the hidden space of responsibility provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the thermal power energy storage frequency regulation responsibility continuous switching device based on a deep self-encoder provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In this article, the term "and / or" describes 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, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0027] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0028] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0029] Next, combined Figures 1-5 This application introduces a method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder, as provided in the embodiments of this application.

[0030] Figure 1 This is a flowchart illustrating the continuous switching method for frequency regulation responsibility of thermal power storage based on a deep autoencoder provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: Step S1: Based on the pre-acquired online resource quantities of multiple thermal power units, calculate the member responsibility characteristics of each unit and energy storage; Optionally, the circuit breaker position signal, grid connection status, current output and upper and lower limits, frequency deviation and energy storage output of multiple thermal power units sharing energy storage services are obtained when the plant participates in AGC or joint frequency regulation as a whole, forming an online resource quantity with fixed member positions, unified time index and clear field boundaries.

[0031] In one embodiment of this application, the power plant includes four thermal power units and one energy storage system, and the set of locations of the unit members is denoted as... The location of the energy storage member is recorded as For any crew member's position Receive the original sequence of circuit breaker position signals respectively Original sequence of unit grid connection status The original sequence of current unit output and upper and lower limits Location of energy storage members Receive the original sequence of energy storage output ; the original sequence of frequency deviation received by the power plant Step S1 does not perform dynamic rearrangement of unit member positions, nor does it aggregate multiple thermal power units into a station-level total. The positions of the four unit members and one energy storage member remain in a fixed order throughout the data construction process. Subsequent unit disconnection is expressed through circuit breaker position signals and unit grid connection status, not through deleting member positions.

[0032] To ensure that circuit breaker position signals, unit grid connection status, unit current output and upper and lower limits, frequency deviation, and energy storage output can be directly combined at the same sampling time, a unified time axis is first constructed. ,in , The sampling period is calculated online for the plant. For any original sequence, discretization and alignment are performed according to a unified time axis: if the current sampling time... If multiple original records exist within the previous sampling period, the timestamp should be selected no later than [date missing]. The latest record; if the current sampling time If no new record is created in the previous sampling period, the record value that was already aligned in the previous sampling time is used. This rule yields the discrete sequence. , , , , , and This alignment method establishes a correspondence between frequency deviation and the positions of unit members and energy storage members under a unified sampling index. Based on this, cross-source reconfiguration is no longer required when calculating the short-time frequency difference response slope, and member position misalignment no longer occurs when performing topology change determination in step S3.

[0033] After discretization and alignment, the circuit breaker position signals and unit grid connection status are collected according to the corresponding member positions of each unit to construct... and will This is denoted as the topology state sequence. The topology state sequence is an 8-dimensional discrete vector sequence. The first four dimensions correspond to the circuit breaker position signals of the four thermal power units, and the last four dimensions correspond to the grid connection status of the four thermal power units. If a unit is disconnected from the grid, the positions of the unit's members remain unchanged; only the circuit breaker position signals and the unit's grid connection status at the corresponding member positions are updated to the disconnected state. This construction method directly provides the topology change determination input for step S3 based on member positions, and also provides the member position boundary for the current online resource set in step S2.

[0034] Subsequently, following a fixed order of the positions of four unit crew members and one energy storage crew member, the current output and upper and lower limits of the units and the energy storage output are aggregated and constructed as follows: and will This is denoted as the output boundary sequence. The output boundary sequence is a 13-dimensional discrete vector sequence, where each thermal power unit occupies 3 dimensions, representing the unit's current output, upper output limit, and lower output limit, respectively. The location of the energy storage member occupies 1 dimension, representing the energy storage output. Frequency deviation. For common operational quantities at the plant level, they are not split according to member location, but are bound to the topology state sequence and output boundary sequence at the same sampling time when constructing time resource segments. Based on the same sampling time... ,structure and will This is denoted as a time-resource fragment. Under the current implementation configuration, the time-resource fragment is a 22-dimensional vector. The circuit breaker position signal, unit grid connection status, unit current output and upper and lower limits, frequency deviation, and energy storage output are all completely preserved without any station-level compression.

[0035] Finally, the time-series resource fragments from consecutive sampling moments are concatenated in chronological order to form the online resource volume. In the implementation configuration consistent with the input configuration of the deep Koopman autoencoder state-space model, the starting point of the window is taken. ,Will This is denoted as an online resource quantity window, with a window length of 6 sampling times. Each resource segment within the online resource quantity window retains a fixed member position order. Therefore, the calculation basis for grid connection availability, distance from current output to upper and lower limits, upper and lower adjustment margins, short-time frequency difference response slope, and energy storage capacity can be directly extracted from the online resource quantity window according to the unit member position and energy storage member position. Step S2 further inputs the member responsibility characteristics corresponding to the 5 member positions into the deep Koopman autoencoder state space model, forming a responsibility-bearing representation and responsibility implicit distribution bound to the member position. This completes the construction from multi-source raw records to online resource quantities and retains a fixed member position basis for the continuous switching of the responsibility implicit distribution before and after decoupling.

[0036] In the above embodiments, the field information of each online resource quantity is shown in Table 1 below: Table 1 Field Information Table

[0037] Figure 2 The unit provided in the embodiments of this application A schematic diagram illustrating the timing evolution of energy storage frequency regulation responsibilities, as shown below. Figure 2 As shown in one embodiment of this application, two typical moments are compared horizontally to illustrate the changes in responsibility share before and after all units are connected to the grid and unit 3 is disconnected. Each unit's output is represented by an upper rectangle, and its lower gray bar represents its available frequency regulation responsibility. In moment one on the left, all units and energy storage are in a normal sharing state; in moment two on the right, unit 3 is disconnected, its responsibility bar disappears, and arrows indicate that responsibility smoothly migrates from the disconnected unit to the remaining units and energy storage units over time. This demonstrates the continuous responsibility sharing capability of the method provided in this application over the time dimension.

[0038] Then, based on the online resource quantity, the grid connection availability, the distance from the current output to the upper and lower limits, the upper and lower adjustment margins, the short-time frequency difference response slope, and the energy storage capacity are calculated separately for the generating units and energy storage, forming member responsibility characteristics. The purpose is to organize the online resource quantity into member responsibility characteristics that can be directly entered into the encoding end of the deep Koopman autoencoder state space model, so that each member position retains a clear responsibility attachment boundary.

[0039] In one embodiment of this application, an online resource quantity window has been formed. ,in This represents six consecutive sampling times. Let the member position index be... ,in Corresponding to the positions of the 4 crew members, Corresponding to the location of the energy storage member. Construct the member state variable for each sampling time and each member location. Regarding crew member locations and crew status variables... It consists of circuit breaker position signal, unit grid connection status, unit current output, unit output upper limit, unit output lower limit, frequency deviation, and energy storage output; for energy storage member position, member status quantity. It consists of energy storage output, frequency deviation, circuit breaker position signals corresponding to the positions of the four unit members, and the grid connection status of the units. This construction method ensures that the positions of the energy storage members do not exist independently of the multiple thermal power units sharing the energy storage service, but are always bound to the set of currently online resources.

[0040] After obtaining the member state quantity Then, the adjustment boundary quantity is first formed. For unit member positions, if the circuit breaker position signal is ON and the unit is in grid-connected status, then grid-connected availability is recorded as valid; if the circuit breaker position signal is OFF or the unit is in disconnected grid-connected status, then grid-connected availability is recorded as invalid, and the distance from the current output to the upper limit, the distance from the current output to the lower limit, the upward adjustment margin, and the downward adjustment margin are all set to 0. For unit member positions with valid grid-connected availability, the distance from the current output to the upper limit is determined based on the difference between the current output and the upper limit of the unit's output, and the distance from the current output to the lower limit is determined based on the difference between the current output and the lower limit of the unit's output. The distance from the current output to the upper limit is written into the upward adjustment margin, and the distance from the current output to the lower limit is written into the downward adjustment margin. For the location of energy storage members, grid connection availability is determined jointly based on the grid connection availability of the four unit member locations. As long as there is a unit member location with effective grid connection availability, the energy storage member location is marked as available. The distance from the current output to the upper and lower limits and the upper and lower adjustment margins corresponding to the energy storage member location do not use independent off-site quantities, but use the summation result of the corresponding values ​​of all effective unit member locations at the current sampling time. This directly binds the responsibility boundary of the energy storage member location to the current adjustable space of multiple thermal power units sharing energy storage services.

[0041] In forming the adjustment boundary quantity Then, the responsibility calculation is formed. The short-time frequency difference response slope is calculated using a short-window differential method, taking two adjacent sampling intervals as the calculation window in the current implementation configuration. For the unit member location, the current output and frequency deviation of the unit at the current sampling time and the previous sampling time are read. When the frequency deviation change meets the frequency deviation threshold parameter, the change in the current output of the unit meets the output change threshold parameter, and the grid availability is in an effective state, the direction of the change in the current output of the unit is correlated with the direction of the change in the frequency deviation to obtain the short-time frequency difference response slope; when the threshold conditions are not met, the short-time frequency difference response slope is recorded as 0. For the energy storage member location, the short-time frequency difference response slope is calculated using the same window change relationship between the energy storage output and the frequency deviation. The energy storage capacity is calculated based on the energy storage output and the adjustment boundary quantity. Specifically, the upward and downward adjustment margins of all effective unit member locations are read first, and then the responsibility space that the energy storage can continue to undertake at the current sampling time is determined in combination with the current energy storage output direction to form a scalar. Since the energy storage capacity reflects the responsibility space that shared energy storage provides for all member locations to share, it will... The locations of four unit members and one energy storage member are simultaneously written as boundary constraint inputs shared by all member locations at the same sampling time.

[0042] Finally, grid connection availability, distance from current output to upper limit, distance from current output to lower limit, upward adjustment margin, downward adjustment margin, short-time frequency difference response slope, and energy storage capacity are combined according to member location to form member responsibility characteristics. In the current implementation configuration, For a 7-dimensional vector, write as ,in Indicates grid availability. and This indicates the distance from the current output to the upper or lower limit. and Indicates the margin for adjustment. Indicates the slope of the short-time frequency difference response. This indicates the energy storage capacity. Five sets of member responsibility characteristic sequences can be obtained by following the fixed member position order. The organization of the member responsibility features corresponds to the input structure of the encoder of the deep Koopman autoencoder state-space model. The encoder performs a two-layer fully connected mapping on each 7-dimensional member responsibility feature to generate a responsibility-bearing representation, which is then aggregated into a responsibility latent distribution according to the current online resource quantity set. Thus, the output member responsibility features retain the fixed semantics of the unit member positions and energy storage member positions, while also incorporating shared energy storage constraints into all member positions. This provides direct input for step S2 to generate the responsibility latent distribution and for step S3 to establish the target support relationship using up and down adjustment margins, short-time frequency difference response slopes, and energy storage capacity.

[0043] In the above embodiments, the relevant information regarding the responsibility characteristics of each member is shown in Table 2 below: Table 2 Feature-related information table

[0044] Figure 3 This is a block diagram of the plant AGC-shared energy storage coordinated frequency regulation structure provided in the embodiments of this application, such as... Figure 3 As shown in one embodiment of this application, Figure 3 The upper and lower parts correspond to two operating states: all units are connected to the grid and Unit 3 is disconnected from the grid. Figure 3 The plant AGC or joint frequency regulation control unit serves as the upper control source, and frequency regulation commands are issued to each unit and energy storage through the bus. After unit 3 is disconnected, its connection with the bus is broken, while the connection between energy storage and other units is thickened, indicating that under the constraints of target support relationship and responsibility transfer, energy storage and the remaining units jointly undertake the frequency regulation task originally undertaken by unit 3.

[0045] Step S2: Input the member responsibility features into the deep Koopman autoencoder state space model, map each member responsibility feature to a responsibility-bearing representation, and aggregate them into a responsibility latent distribution according to the current online resource quantity set. Under the constraint of the current online resource quantity set, perform topological stable segment evolution on the responsibility latent distribution to obtain the current responsibility latent distribution. Although the member responsibility characteristics have decomposed the online resource volume to the unit and energy storage levels, they still describe the current state and cannot be directly used as the responsibility object continuously transferred before and after decoupling in the responsibility share generation process. This is because when multiple thermal power units share energy storage services and the entire plant participates in AGC or joint frequency regulation, the real change is not a single power value, but rather which online resources the responsibility is attached to and the boundaries of responsibility borne by each online resource.

[0046] Existing technologies typically use station-level total input, fixed number input, or static responsibility coefficient input in this stage, and then the time series model directly outputs the power correction amount. In the scenario of unit disconnection, the disconnected unit dimension in the fixed number will lose its physical meaning, and the resource set corresponding to the static responsibility coefficient will also change. Continuing to use the original state will mistakenly leave the responsibility before disconnection at the location of the disconnected unit, or after deleting the dimension, the responsibility will be roughly distributed to the remaining units and energy storage, resulting in the disconnection of the responsibility object from the current online resource set.

[0047] Therefore, this application performs two-layer scenario-based processing on member responsibility features within the deep Koopman autoencoder state space model: first, the member responsibility features are mapped to responsibility-bearing representations, so that each unit and energy storage no longer participates in the calculation with the original power value, but enters the model in the state of undertaking frequency regulation responsibility; then, the responsibility-bearing representations are aggregated according to the current online resource quantity set to form a responsibility implicit distribution.

[0048] The implicit responsibility distribution corresponds to the responsibility allocation state on the current set of online resources, simultaneously representing responsibility attribution and the boundary of acceptance. Therefore, it can serve as the state object for the evolution of topologically stable segments. The deep Koopman autoencoder state-space model does not search for general low-dimensional states in the conventional way, but instead identifies the implicit responsibility distribution as the unique hidden state, ensuring that the evolution of topologically stable segments always revolves around it. The current implicit responsibility distribution formed through this process retains the physical constraint source in the member responsibility characteristics and explicitly binds responsibility to the current set of online resources, becoming a necessary source object before the formation of responsibility migration constraints. Without the implicit responsibility distribution, unbalanced optimal transmission cannot obtain a source distribution with clear responsibility significance, and responsibility migration constraints cannot be applied to specific responsibility objects. There would be no continuously referential technical interface between the deep Koopman autoencoder state-space model and unbalanced optimal transmission. Therefore, state object modification is made around the failure mechanism of the change in the responsibility attachment object during unit disconnection.

[0049] Step S3: Determine topology changes based on online resource quantity. When the topology change conditions are met, screen the remaining units and energy storage according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, establish the target support relationship for the transfer of the current hidden distribution of responsibility to the set of online resource quantity after disconnection, and form responsibility migration constraints. The current responsibility hidden distribution can characterize the responsibility attachment relationship before the disconnection. However, when the topology change condition is met, the current responsibility hidden distribution is still defined on the current online resource quantity set before the disconnection and cannot be directly sent to the evolution and decoding stage after the disconnection.

[0050] Existing technologies typically employ three approaches to address this issue: first, directly zeroing out or deleting the corresponding dimension of the disconnected unit; second, re-normalizing it according to the remaining unit capacity ratio or a preset responsibility ratio; and third, resetting the model state and re-estimating from a new topology. All three approaches treat the problem as a parameter update problem, failing to address the fundamental change in the supporting object of the implicit responsibility distribution. This leads to issues such as responsibility disappearance, misallocation of responsibility, or instantaneous state jumps during disconnection.

[0051] This application sets up step S3, which first determines the topology change based on the circuit breaker position signal and the grid connection status of the units in the online resource quantity. When the topology change conditions are met, the current responsibility implicit distribution is not directly rewritten. Instead, based on the member responsibility characteristics of the remaining units and energy storage, namely the adjustment margin, short-time frequency difference response slope, saturation proximity, and energy storage capacity, a target support relationship is established to further form responsibility migration constraints. The target support relationship defines which resources in the online resource quantity set after decoupling can receive the responsibility migrated from the current responsibility implicit distribution and the receiving boundary. Therefore, the target support relationship is not an independent result, but rather the constraint basis for the unbalanced optimal transmission before entering the state switching position of the deep Koopman autoencoder state space model.

[0052] Step S4: Under the responsibility migration constraint, perform unbalanced optimal transfer on the current responsibility hidden distribution, allowing partial migration and partial cancellation of responsibilities, and generate a migrated responsibility hidden distribution; Under the constraint of responsibility migration, an unbalanced optimal transfer is performed on the current responsibility implicit distribution to generate a migrated responsibility implicit distribution. The reason for using unbalanced optimal transfer is that after the decommissioned units are removed, the original responsibility does not require mechanical conservation in the remaining units and energy storage. Part of the responsibility will be cancelled due to resource withdrawal, and only the other part will migrate to the set of online resources after decommissioning. If a conservation-based migration is used, the responsibility of the decommissioned units will be forcibly pushed into the remaining resources, which will still cause migration distortion. After the migrated responsibility implicit distribution is generated, the deep Koopman autoencoder state-space model has the legal initial state required for evolution and decoding after decommissioning.

[0053] Step S5 involves demultiplexing and decoding the hidden distribution of migration responsibility into the deep Koopman autoencoder state space model, outputting the responsibility share result, and generating the energy storage responsibility increment and the remaining unit responsibility share based on the responsibility share result, and outputting the shared energy storage frequency regulation undertaking amount.

[0054] Based on the completion of the responsibility object support switch in step S4, the migrated responsibility hidden distribution is sent back to the original evolution chain of the deep Koopman autoencoder state space model to generate responsibility share results.

[0055] In one embodiment of this application, the migration responsibility hidden distribution obtained in step S4 is received. and the set of online resources after unblocking determined in step S3 ,in This represents the current inference moment. Transfer responsibility hidden distribution. The vector is 20-dimensional, arranged in a fixed order of 4 unit member positions and 1 energy storage member position, with every 4 consecutive dimensions corresponding to 1 member position. Step S5 does not reconstruct the station-level input, does not re-estimate the model state under the new topology, and does not rewrite the migration responsibility hidden distribution into other intermediate objects; instead, it directly uses the migration responsibility hidden distribution. It serves as the sole hidden state input for the evolutionary stage after unblocking in the deep Koopman autoencoder state-space model.

[0056] Before entering the evolutionary layer, first determine the set of online resources after unpacking. Hidden distribution of migration responsibility Constraining the positions of disconnected members in the equation forms the evolutionary input after disassembly. The specific approach is as follows: Distribute the migration responsibility implicitly according to the order of member positions. The data is divided into 5 groups of 4-dimensional member position and responsibility states; the online resource quantity set after decoupling is also considered. For member positions, retain the corresponding 4-dimensional member position responsibility status values ​​unchanged; for those not belonging to the online resource quantity set after unpacking... The disconnected member position is written into a fixed zero-occupying state, corresponding to the 4-dimensional member position's responsibility state. This constraint action only occurs before the hidden states of the deep Koopman autoencoder state-space model enter the evolutionary layer; it does not change the total number of member positions, delete disconnected member positions, or reassign the responsibilities corresponding to disconnected member positions to other member positions. The resulting post-disjoint evolutionary input... It still maintains the 20-dimensional structure and the semantic consistency of one member position for every four dimensions.

[0057] The decoupled input will be evolved. The input is fed into the evolutionary layer of a deep Koopman autoencoder state-space model. The evolutionary layer employs a linear mapping from 20 linear neurons to 20 linear neurons, and the input is evolved after demultiplexing the 20-dimensional model. Performing a single decoupling state transfer yields a 20-dimensional decoupling-responsibility hidden state. Due to the evolutionary input after demultiplexing. The online resource quantity has been set according to the ungrouped data. For the constraint on the disconnected member positions, the evolutionary layer in the current implementation configuration is only responsible for continuing the responsibility distribution along the existing hidden state chain, and does not undertake the member position support switching function. The member position support switching function has been completed in steps S3 and S4, so the evolutionary layer in step S5 continues to use the original 20-dimensional hidden state linear evolution structure of the deep Koopman autoencoder state space model, without adding a parallel model or an independent reconstruction path.

[0058] After being released from the state of responsibility, the hidden state remains unchanged. Afterwards, the hidden responsibility status will be delisted. Input the decoder of the deep Koopman autoencoder state-space model. The decoder and encoder each have two fully connected layers. The first layer deconcatenates the 20-dimensional hidden responsibility state. The mapping is done as a 16-dimensional intermediate vector. The first layer has 16 neurons, and the nonlinear units are rectified linear units. The second layer maps the 16-dimensional intermediate vector into a 5-dimensional output vector, and the second layer has 5 linear neurons. The 5-dimensional output vector is arranged in a fixed order according to the member positions, corresponding to the positions of 4 unit members and 1 energy storage member, forming the responsibility share result. .in, to The results of the responsibility share for the four crew member positions are as follows. The results of the responsibility share corresponding to the position of the energy storage member.

[0059] Finally, the results of the liability share will be determined. The corresponding values ​​of the four unit member positions and the one energy storage member position are directly determined as the responsibility share result. No member position rearrangement is performed after decoding, no capacity ratio redistribution is performed in the responsibility share result, and no topology correction is added to the responsibility share result. Therefore, step S5 continuously feeds the migration responsibility hidden distribution obtained in step S4 back to the decoupling evolution and decoding chain of the deep Koopman autoencoder state space model, ensuring that the responsibility share result always maintains a fixed correspondence with the four unit member positions and the one energy storage member position, and providing direct input for generating the energy storage responsibility increment and the remaining unit responsibility share.

[0060] Then, without changing the deep Koopman autoencoder state-space model and the unbalanced optimal transmission master solution chain, the output responsibility share results are converted into the energy storage responsibility increment (i.e. the shared energy storage frequency regulation undertaking amount) and the remaining unit responsibility share.

[0061] This application provides a method for continuous switching of frequency regulation responsibility for thermal power and energy storage based on a deep autoencoder. It combines a deep Koopman autoencoder state-space model with unbalanced optimal transmission, constructing a complete solution chain from online resource acquisition to shared energy storage frequency regulation acceptance output, centered around the unique intermediate object of the responsibility hidden distribution. By defining the responsibility hidden distribution as a unique hidden state, the hidden state simultaneously carries the responsibility magnitude, responsibility attribution, and acceptance boundary, ensuring that the responsibility source object before and after unit disconnection maintains clear semantics at the member location level. This suppresses responsibility from remaining in physically meaningless locations before disconnection, improving the continuity of shared energy storage frequency regulation responsibility identification and subsequent acceptance calculation. After determining topology changes, a target support relationship is established and... By establishing responsibility migration constraints and performing unbalanced optimal transmission under these constraints, partial responsibility migration and cancellation are permitted. This avoids the misallocation of responsibilities and state jumps caused by mechanically pushing the responsibilities of decommissioned units into remaining resources, as is common in existing technologies. Through implicit responsibility distribution throughout the entire process of responsibility generation, migration, post-decommissioning evolution, and the output of the assigned quantity, a closed-loop connection around the same responsibility object is formed. This solves the problem in existing technologies where responsibility redistribution, state updates, and result output are scattered across multiple stages, making it difficult to maintain consistent transmission of responsibility objects in scenarios with changing resource sets. It can maintain the business closed loop of shared energy storage frequency regulation responsibility allocation and the output of the assigned quantity in the scenario of unit decommissioning, improve the accuracy of frequency regulation responsibility allocation, and has clear engineering application value.

[0062] In some embodiments, the implicit distribution of responsibility in step S2 is arranged according to fixed member positions. Each member position corresponds to a fixed dimension of responsibility bearing state. The responsibility bearing state is obtained by combining the responsibility bearing representation with the underlying numerical fields in the member responsibility characteristics. This allows the implicit distribution of responsibility to carry responsibility size, responsibility attribution, and bearing boundary at the same time. Member positions that do not belong to the current online resource quantity set are written to a fixed zero state placeholder, and no responsibility quantity is introduced from other member positions.

[0063] In one embodiment of this application, the current inference window endpoint is taken as... The member responsibility characteristics at six consecutive sampling times are recorded as follows: ,in The system consists of member responsibility characteristics corresponding to 5 member positions, with 4 unit member positions and 1 energy storage member position arranged in a fixed order. Step S2 inputs the member responsibility characteristics independently for each member position, without compressing the 4 unit member positions and 1 energy storage member position into a station-level total, nor deleting any disconnected member positions before encoding. This input organization method is adopted because step S3, establishing the target support relationship, requires directly reading the member positions to which the responsible object is attached, and step S4, performing unbalanced optimal transmission, requires directly reading the member positions from which the responsible object is relocated. If the member responsibility characteristics were rewritten as station-level quantities or variable-dimensional quantities in step S2, the responsible attached objects would not be able to maintain a correspondence with the current online resource quantity set.

[0064] The member responsibility feature sequence is input into the encoder of the deep Koopman autoencoder state-space model. ,parameter This represents the weights and biases at the encoding end. For any member position and any sampling time, the corresponding 7-dimensional member responsibility feature is first input into the first fully connected layer of shared parameters. The first fully connected layer has 16 neurons, performing weight and bias calculations, with rectified linear units used for the nonlinear units, resulting in a 16-dimensional intermediate representation. This 16-dimensional intermediate representation is then input into the second fully connected layer of shared parameters. The second fully connected layer has 12 neurons, resulting in a 12-dimensional responsibility-bearing representation. Since the four generating unit locations and one energy storage location share the same set of coding parameters, the 12-dimensional responsibility-bearing representation has a unified representation scale across different member locations. Because the input retains the member location index, the 12-dimensional responsibility-bearing representation does not lose responsibility attribution information. The coding end's role in this step is not to extract general time-series features, but rather to rewrite grid availability, the distance from the current output to the upper and lower limits, the upper and lower adjustment margins, the short-time frequency difference response slope, and the energy storage capacity as responsibility-bearing states, providing a unified representation basis for the subsequent implicit distribution of responsibility.

[0065] Before the model is put into online computation, the parameters of the encoder, evolution layer, and decoder of the deep Koopman autoencoder state-space model are determined using historical operating data. Historical operating data includes online resource windows for multiple thermal power units during normal grid connection, unit disconnection, and post-disconnection stabilization processes; the corresponding unit grid connection status, energy storage output, plant AGC regulation, and responsibility share calibration results obtained from frequency regulation assessment records or manual verification rules. Following the same method as step S1, the historical operating data is constructed into a member responsibility feature sequence, and the responsibility share calibration results after the same time window are used as the supervision target. The error between the responsibility share output from the decoder and the responsibility share calibration results, the change in the hidden distribution of responsibility at adjacent sampling times, and the residual responsibility at the disconnected member location constitute the training target. The weights and biases of the encoder, evolution layer, and decoder are iteratively updated. When the training target meets the preset convergence condition, the aforementioned weights and biases are fixed as the model parameters called during online execution of steps S2 to S5. This parameter determination process is only used to obtain the executable parameters of the deep Koopman autoencoder state-space model, and does not change the main solution chain that generates, migrates, uncoalesces, evolves, and outputs around the responsibility hidden distribution during online execution of this application.

[0066] After obtaining the 12-dimensional responsibility-bearing characterization Then, the current set of online resources is determined based on the grid availability corresponding to the 12-dimensional responsibility-bearing representation. The specific procedure is as follows: Read For the grid connection availability field in the corresponding source member responsibility characteristics, when the grid connection availability is in an effective state, the corresponding member's location will be included in the current online resource quantity set. When the grid availability is invalid, the corresponding member location will be excluded from the current online resource set. In addition, based on the current collection of online resources. From the 12-dimensional responsibility-bearing representations corresponding to all member positions, member position representations belonging to the current online resource quantity set are selected to form an online responsibility representation set. This selection process does not change the main chain of the deep Koopman autoencoder state-space model, and the selection results are directly entered into the member position aggregation calculation within the encoder. The reason for using the current online resource quantity set to constrain the online responsibility representation set is that, in the unit disconnection scenario, the actual failure location occurs during the change of responsibility attachment object. If offline member positions still participate in responsibility aggregation together with online member positions, the implicit distribution of responsibility will retain physically meaningless responsibility traces at the member positions that have withdrawn.

[0067] When aggregating the online responsibility representation set by member position, the 12-dimensional responsibility bearing representation corresponding to each member position is first divided into 4 groups in sequence, with each group containing 3 consecutive dimensions. Then, the underlying numerical fields in the member responsibility features are combined to generate a 4-dimensional responsibility bearing state. In this embodiment, the set of currently online resources is... The member positions are aggregated using the following formula:

[0068] In the formula, Indicates the first The member position is in the... The 4-dimensional responsibility-bearing state corresponding to each sampling moment; Indicates the member position index; Indicates the sampling time index; Indicates the first The set of current online resources corresponding to each sampling time; Indicates the first The member position is in the... The 12-dimensional responsibility-bearing representation corresponding to the sampling time is the first... Dimensional components; This represents the component index within the 12-dimensional responsibility-bearing representation; Indicates the first Upward margin corresponding to each member position; Indicates the first Downsizing margin corresponding to each member position; Indicates the first The distance from the current output to the upper limit corresponding to each member's position; Indicates the first The distance from the current output to the lower limit corresponding to each member's position; Indicates the first The short-time frequency difference response slope corresponding to each member position; Indicates the first The energy storage capacity corresponding to each sampling moment; This represents the index of the sampling time within the current inference window; Indicates the starting point of the current inference window; Indicates the current inference window endpoint; This represents the threshold parameter used to limit the lower bound of the denominator for the distance and margin terms; This represents the threshold parameter used to limit the lower bound of the denominator of the short-time-frequency-difference response slope; This represents the threshold parameter used to limit the lower bound of the denominator for energy storage capacity. The aggregation rule operates at the single point of "aggregating the online responsibility representation set by member position," directly converting the 12-dimensional responsibility capacity representation into a 4-dimensional responsibility capacity state. It explicitly writes the current online resource quantity set, upper and lower adjustment margins, distance from the current output to the upper and lower limits, short-time frequency difference response slope, and energy storage capacity into the formation process of the implicit responsibility distribution. For member positions that do not belong to the current online resource quantity set, step S2 writes a fixed zero-state placeholder at the corresponding member position, without introducing responsibility from other member positions.

[0069] Subsequently, following a fixed order of four unit member positions and one energy storage member position, the four-dimensional responsibility-bearing states corresponding to the five member positions are concatenated into a 20-dimensional responsibility latent distribution. Each four dimensions in the 20-dimensional responsibility latent distribution correspond to only one member position; therefore, the 20-dimensional responsibility latent distribution simultaneously carries the magnitude of responsibility, responsibility attribution, and acceptance boundary. Step S2 determines the 20-dimensional responsibility latent distribution as the unique latent state of the deep Koopman autoencoder state-space model, without setting up additional abstract low-dimensional states. The reason for using a unique latent state is that when forming the target support relationship and responsibility migration constraints in step S3, the responsibility source object must be directly read at the member position level; if the latent state is rewritten as a general compressed vector without member position semantics, the unbalanced optimal transmission will lose a continuously referable source distribution.

[0070] After forming the 20-dimensional hidden distribution of responsibility at each sampling time, take the current sampling time as the basis. A continuous sub-window that ends and whose current online resource set remains unchanged is considered a topologically stable segment. The responsibility hidden distribution sequence within the sub-window is input into the evolutionary layer. ,parameter This represents the weights and biases of the evolutionary layer. The evolutionary layer uses a linear mapping from 20 linear neurons to 20 linear neurons, performing state transfer on the 20-dimensional hidden responsibility distribution at each sampling time to obtain the evolved hidden responsibility distribution. Step S3 only performs this linear evolution within the topologically stable segment, preventing the evolutionary layer from crossing the boundary of the current online resource quantity set change. The reason for this constraint is that the deep Koopman autoencoder state-space model can continuously track the hidden responsibility distribution under a stable topology, but when the supporting object of the member position changes, directly using the same hidden state evolution chain will mistakenly leave the responsibility before the separation at the exiting member position. The responsibility switching after the topology change is handled by steps S3 and S4, and step S2 is responsible for generating a valid source state within the topologically stable segment. Finally, the current sampling time is... The corresponding evolutionary responsibility hidden distribution is determined as the current responsibility hidden distribution, and the current responsibility hidden distribution is used as the source object for establishing the target support relationship and responsibility transfer constraint in step S3.

[0071] In some embodiments, the step S3, which establishes the target support relationship for the transfer of the current implicit distribution of responsibilities to the set of online resources after delisting based on the member responsibility characteristics of the remaining units and energy storage, specifically includes: Step S31: Based on the set of online resources after disconnection, extract the member responsibility characteristics of the remaining generating units and energy storage. Step S32: Calculate the acceptable boundary for each migration path from the source member location to the target member location. When the target member location is a generator unit, automatically select the up and down adjustment margin path. When the target member location is energy storage, automatically select the energy storage bearing capacity path. Set the acceptable boundary of the disconnected member location to zero to form the target support relationship.

[0072] The current hidden responsibility distribution generated in step S2 is still defined on the current online resource quantity set before unblocking. Step S3 needs to determine whether the member position support object has changed without rewriting the numerical structure of the current hidden responsibility distribution, and then construct responsibility migration constraints around the current hidden responsibility distribution.

[0073] In one embodiment of this application, the current inference time is taken as Read the discrete sequence of circuit breaker position signals corresponding to the positions of the four unit members. Discrete sequence of grid connection status of generating units ,in For each crew member's location, compare the current sampling time. Compared with the previous sampling time The circuit breaker position signal is used to obtain the circuit breaker position signal change sequence; the current sampling time is compared. Compared with the previous sampling time The unit's grid connection status is determined, and a sequence of changes in the unit's grid connection status is obtained. Step S3 compares the circuit breaker position signal change sequence with the unit's grid connection status change sequence one by one according to the member's position. Only changes in the same direction from grid connection to grid disconnection are included in the topology change judgment quantity; individual measurement point jumps are not included. The topology change judgment quantity is then compared with a preset threshold. Comparison: When the topology change determination value is not less than a preset threshold At that time, it is determined that the topological change condition is satisfied.

[0074] After the topology change conditions are met, based on the current sampling time in the unit's grid connection status... The remaining generating units and energy storage are screened based on their connection status to form a set of online resources after disconnection. Subsequently, the focus shifted to the set of online resources after the decoupling. Extract resource constraints. For each set of online resource quantities after demultiplexing... For the generator crew positions, the system reads the upward adjustment margin, downward adjustment margin, short-time frequency difference response slope, and the distance from the current output to the upper and lower limits obtained in step S1. For the energy storage crew positions, it reads the energy storage capacity and the corresponding short-time frequency difference response slope. Step S3 maps the distance from the current output to the upper limit to the upward adjustment margin, and the distance from the current output to the lower limit to the downward adjustment margin, and converts the side with greater restriction to the lower limit according to the preset boundary standard. to The degree of saturation proximity within the interval forms a resource constraint quantity expanded according to the member position. This resource constraint quantity maintains the same member position order as the current responsibility implicit distribution, allowing for direct invocation for responsibility component matching.

[0075] The current responsibility implicit distribution is a 20-dimensional member position splicing state. Step S3, according to a fixed order of 4 unit member positions and 1 energy storage member position, splits the current responsibility implicit distribution into 5 groups of 4-dimensional responsibility components, denoted as... Then, based on the grid connection status of the generating units, The responsibility components are divided into those corresponding to units disconnected from the grid and those corresponding to the remaining units and energy storage. Step S3 does not directly set the responsibility components corresponding to units disconnected from the grid to zero, nor does it delete the corresponding member positions from the current implicit responsibility distribution. Instead, it associates each responsibility component with the set of online resources after decoupling. All candidate members are matched to form a responsibility support matching quantity. Responsibility support matching quantity Each matching record includes the source member location, target member location, responsibility component identifier, adjustment direction identifier, upward and downward adjustment margins corresponding to the target member location, short-time frequency difference response slope, saturation proximity, and energy storage capacity. For responsibility components whose grid connection status is disconnected from the generating unit, the responsibility support matching quantity is... Only the set of online resources after unpacking is allowed. The member positions in the set are used as candidate targets; for the set of online resources still in the ungrouped state... The responsibility component and the matching quantity of responsibility support. Preserve the original inheritance path from the source member location to the same member location.

[0076] After obtaining the matching amount of responsibility support Then, step S3 calculates the acceptable boundary for each matching path from the source member location to the target member location. And based on this, a target support relationship is formed. For any source member position... and the location of any target member Receive boundary Determine using the following formula:

[0077] In the formula, Indicates the current sampling time Lower source member position To the target member's location The acceptable boundary when transferring responsibility components; Indicates the source member position index; Indicates the position index of the target member; Indicates the current inference time; Indicates the position of the target member Does the current sampling time belong to the set of online resources after demultiplexing? The member position indicator is set to 1 when the member belongs and 0 when the member does not belong. This indicates the type of target member location. It is 1 when the target member location is an energy storage member location and 0 when the target member location is a generator unit member location. This indicates the remaining energy storage capacity at the current sampling moment; Indicates the amount of responsibility supporting the matching quantity Zhongyuan member position To the target member's location The adjustment direction indicator is set to 1 for upward adjustment and 0 for downward adjustment. Indicates the position of the target member The upsizing margin corresponding to the current sampling time; Indicates the position of the target member The downsizing margin corresponding to the current sampling time; Indicates the position of the target member The short-frequency difference response slope corresponding to the current sampling moment; Indicates the position of the target member The degree of saturation proximity at the current sampling time; This represents the index of the sampling time within the current inference window; Indicates the starting point of the current inference window; The threshold parameter represents the lower bound of the normalized denominator of the short-time frequency difference response slope. This represents the threshold parameter corresponding to the lower bound of the normalized denominator of the energy storage capacity capacity.

[0078] When using this acceptable boundary calculation method, if the target member's location is the same as the crew member's location, then... Automatically select the upside and downside margin path; when the target member location is the energy storage member location, it is determined by... Automatically select energy storage to take over the remaining capacity path; the target member's location does not belong to the set of online resources after the disconnection. At that time, by The acceptable boundary is directly compressed to zero. Step S3 will then... Organized according to a 5x5 member position table, forming a target support relationship. Target Support Relationship Zero entries in the China-Africa context indicate the location of target members allowed to receive responsibility components and their corresponding upper limits; zero entries indicate prohibited migration paths. Finally, based on the target support relationship... Constraining the transfer direction and upper limit of each responsibility component in the current implicit distribution of responsibility forms a responsibility migration constraint. and constrain the transfer of responsibility. The input constraints for performing unbalanced optimal transmission under the current responsibility hidden distribution are fed into step S4.

[0079] Figure 4 This is a responsibility migration path and resource constraint matrix diagram under topology changes provided in the embodiments of this application, such as... Figure 4 As shown in one embodiment of this application, Figure 4 The left side shows a 5x5 matrix from the source member location to the target member location. Rows represent responsible output parties, and columns represent responsible receiver parties. Solid arrows indicate allowed migration paths, while blank cells implicitly prohibit migration. Rows containing exiting units have no self-supporting arrows, pointing only to other units and energy storage, highlighting the principle of "no further self-supporting upon exit." The right side of the matrix provides resource constraints such as up / down adjustment margins, short-time frequency difference response slopes, and energy storage capacity. These constraints act on the entire matrix through connections, limiting the maximum acceptable responsibility for each path. This diagram depicts the target support relationships and responsibility migration constraint mechanisms at the member location level.

[0080] In some embodiments, step S4 specifically includes: Step S41: Using the transfer direction and transfer upper limit in the responsibility transfer constraint as constraints, perform unbalanced optimal transfer on the current hidden responsibility distribution; Step S42: Using the cross-member location migration constraint coefficient and responsibility cancellation constraint coefficient as regularization terms, solve for the responsibility transfer amount and responsibility cancellation amount to each acceptable member location for each dimension of the responsibility component of the source member location. Step S43: The responsibility transfer amount received by each receiving member position is accumulated according to the member position and reconstructed in a fixed order into a vector with the same dimension as the responsibility latent distribution, which is used as the migration responsibility latent distribution.

[0081] The purpose of step S4 is to directly inject the responsibility transfer constraint formed in step S3 into the hidden state switching initialization position of the current responsibility hidden distribution, so that the unbalanced optimal transfer works around the current responsibility hidden distribution as the only intermediate object, rather than generating independent results outside the deep Koopman autoencoder state space model.

[0082] In one embodiment of this application, the current inference time is taken as When the topology change condition is met, the current responsibility hidden distribution is received. and responsibility transfer constraints Step S4 does not implicitly distribute the current responsibility. Instead of performing dimension deletion, zeroing, or re-normalization, the 20-dimensional member position representation determined in step S2 is kept unchanged. The current responsibility hidden distribution is then performed according to a fixed order of 4 unit member positions and 1 energy storage member position. It is divided into 5 groups of 4-dimensional responsibility components to form the source responsibility component set. .in, Indicates the first The 4-dimensional responsibility component corresponding to the position of each member at the current sampling time. Indicates the first The position corresponding to the first member position Dimensional responsibility component value, This splitting action only changes the data organization structure and does not alter the current implicit distribution of responsibilities. The semantics of the member positions are not changed, nor is the arrangement order of the 4-dimensional responsibility bearing states within each member position altered. Therefore, step S5 can still receive the migration responsibility hidden distribution in such a way that each 4-dimensional position corresponds to one member position.

[0083] After obtaining the source responsibility component set Then, based on the responsibility transfer constraint The target support relationship Read the allowed acceptable member positions and corresponding transfer limits of each source responsibility component to form a support matching set. Support matching set Each matching record must contain at least the source member location index, the target member location index, the transfer direction identifier, and the transfer limit. Step S4 re-verifies whether the target member location is within the set of online resources after disconnection based on the unit's grid connection status. For member locations corresponding to units with a disconnected grid connection status, they are directly deemed unacceptable, with no return path or in-situ retention path retained. For unit member locations and energy storage member locations still within the set of online resources after disconnection, step S4 sets the target support relationship... The zero-entry point in the China-Africa region was identified as an acceptable path, and the target support relationship was established. Zero entries are identified as prohibited paths. Subsequently, based on the source responsibility component set... and support matching set Forming a migration candidate set Step S4 involves each source responsibility component. The four internal dimensions are processed synchronously to read the supporting matching set. Position of middle and source members For each available member location, a responsibility migration candidate is generated, using the corresponding path's transfer upper limit as the boundary. If the responsibility component corresponding to a disconnected member location is not fully included in the available path, it is not included in the support matching set. A portion of these was determined as a candidate for liability cancellation.

[0084] For the migration candidate set When performing unbalanced optimal transfer calculations, step S4 calculates the responsibility transfer amount and actual deregistration amount item by item, based on the source member location, target member location, and internal dimension. In the current implementation configuration, unbalanced optimal transfer is solved using the following formula: ; satisfy ; In the formula, Indicates the number of samples taken at the current sampling time. The first member position Dimensional responsibility to the first The amount of responsibility transferred by the transfer of each member's location; Indicates the number of samples taken at the current sampling time. The first member position The actual number of cancellations corresponding to the responsibility component; Indicates the source member position index; Indicates the position index of the target member; This represents the internal dimension index of the 4-dimensional responsibility component; Indicates the current inference time; Indicates the supporting matching set Zhongyuan member position Point to the target member's location The acceptable path indicator is set to 1 when the path is acceptable and 0 when it is not. Indicates the target support relationship Zhongyuan member position Point to the target member's location The upper limit of transfer; An indicator that shows whether the source member position and the target member position are the same; it takes the value 1 if they are the same and 0 if they are different. Represents the source responsibility component set The Middle The position corresponding to the first member position Dimensional responsibility component values; Indicates the cross-member location migration constraint coefficient; Indicates the liability cancellation constraint coefficient; This represents the threshold parameter corresponding to the lower bound of the denominator of the upper bound of the transition; This represents the threshold parameter corresponding to the lower bound of the denominator of the responsibility component. This solution form directly incorporates the transfer direction and upper limit of the responsibility transfer constraint into the unbalanced optimal transfer process, and treats the responsibility cancellation candidate quantity as a constraint variable solved in parallel with the responsibility transfer quantity, instead of performing external corrections after the unbalanced optimal transfer is completed.

[0085] Seeking and Then, step S4 will assign the source responsibility component. The original symbols are attached to the corresponding responsibility transfer quantity and the actual cancellation quantity, and then the responsibility component is updated according to the member position to form the responsibility transfer result. For each target member position in the set of online resource quantities after decoupling, the part of the source responsibility component retained in the original member position is written first, and then the part migrated from the remaining units and the part migrated from the energy storage are accumulated; for the member position corresponding to the disconnection status with the unit grid connection, no receiving quantity is written. Step S4 obtains 5 sets of 4-dimensional target responsibility components to form the target responsibility distribution. Then, it is reconstructed into a 20-dimensional responsibility representation in a fixed order of 4 unit member positions and 1 energy storage member position. and will express 20-dimensional responsibility. The migration responsibility hidden distribution is determined. When the topology change condition is not met, step S4 skips the unbalanced optimal transport calculation and directly assigns the current responsibility hidden distribution. It was determined to be a hidden distribution of migration responsibility.

[0086] Figure 5 This is a comparison diagram of the responsibility hidden space stability covariance ellipse provided in the embodiments of this application, such as... Figure 5 As shown in one embodiment of this application, the hidden states of responsibility under different operating conditions are projected onto a two-dimensional space using the first and second principal components of PCA as coordinate axes. Solid dots and solid-line ellipses correspond to the hidden space distribution under normal topology, while hollow cubes and dashed-line ellipses represent the results of using traditional methods when the topology changes, with dispersed point clouds and large elliptical ranges. Solid triangles and dashed-line ellipses represent the distribution when using the method of this invention, with concentrated point clouds, compact ellipses, and smaller offsets compared to traditional methods, indicating that the hidden states of responsibility remain stable and smoothly transition under topology changes. This figure verifies the effectiveness of the deep Koopman state space modeling proposed in this application from a statistical feature perspective.

[0087] In some embodiments, unbalanced optimal transmission is embedded in the hidden state switching initialization stage of the deep Koopman autoencoder state space model. The embedding position is after the current responsibility hidden distribution completes the evolution of the topology stable segment and before the migration responsibility hidden distribution enters the post-disjoint evolution. This is used to enhance the responsibility migration capability of the deep Koopman autoencoder state space model across resource sets when the topology changes. Unbalanced optimal transmission does not generate solution results independent of the deep Koopman autoencoder state-space model.

[0088] The unbalanced optimal transmission here does not independently solve for the shared energy storage frequency regulation capacity, but rather embeds the existing link of switching from the current hidden responsibility distribution to the migrating hidden responsibility distribution, thus strengthening the shortcoming of the deep Koopman autoencoder state space model in lacking the ability to migrate responsibility across resource sets when the topology change conditions are met. If the target support relationship and responsibility migration constraints are removed, the unbalanced optimal transmission can neither construct responsibility objects nor form responsibility share results.

[0089] The deep Koopman autoencoder state-space model is responsible for the generation and evolution of the responsibility latent distribution, while the unbalanced optimal transport is responsible for the support and reconstruction of the responsibility latent distribution when the resource set changes. The two are continuously connected in the same main solution chain around the same responsibility object.

[0090] In one embodiment of this application, a power plant includes four thermal power units and one energy storage system. The model receives online resource data for six consecutive sampling times. The online resource data at each sampling time includes circuit breaker position signals, unit grid connection status, current unit output and upper / lower limits, frequency deviation, and energy storage output. Five member responsibility feature vectors are formed based on the online resource data, corresponding to the positions of the four unit members and the position of the one energy storage member. Each member responsibility feature vector is seven-dimensional, corresponding to the following dimensions: grid availability (1 dimension), distance from current output to upper limit (1 dimension), distance from current output to lower limit (1 dimension), upward adjustment margin (1 dimension), downward adjustment margin (1 dimension), short-time frequency difference response slope (1 dimension), and energy storage capacity (1 dimension). The circuit breaker position signals and unit grid connection status simultaneously enter the topology change condition determination stage, with the circuit breaker position signals occupying four dimensions and the unit grid connection status occupying four dimensions.

[0091] The deep Koopman autoencoder state-space model employs a single-skeleton structure consisting of an encoder, an evolutionary layer, and a decoder. The encoder performs a two-layer fully connected mapping with shared parameters on each 7-dimensional member responsibility feature vector. The first layer has 16 neurons, each containing 7 weights and 1 bias, with rectified linear units as the nonlinear units. The second layer has 12 neurons, each containing 16 weights and 1 bias, resulting in a 12-dimensional responsibility-bearing representation. The 12-dimensional responsibility-bearing representations of the five member positions are then converged under the constraint of the current online resource set, compressing each member position into a 4-dimensional responsibility-bearing representation. The five member positions are then concatenated in a fixed order to form a 20-dimensional responsibility latent distribution. The evolutionary layer uses a linear mapping from 20 linear neurons to 20 linear neurons, each containing 20 weights and 1 bias, to perform topologically stable fragment evolution on the 20-dimensional responsibility latent distribution.

[0092] The scenario-based transformation of the deep Koopman autoencoder state-space model occurs at the input organization and hidden state definition stages of the encoder. Conventional deep Koopman autoencoder state-space models typically compress the total station-level power or a uniform dimensional sequence into a general hidden state. While this general hidden state can describe continuous changes, it cannot indicate which unit or energy storage member location the frequency regulation responsibility is attached to. During unit disconnection, it's easy to retain the pre-disconnection responsibility at a member location that has lost its physical meaning. To suppress this failure, the encoder does not receive the total station-level power but instead receives member responsibility characteristics; the hidden state is not defined as an abstract compressed vector but as a 20-dimensional responsibility hidden distribution, where every 4 dimensions correspond to the responsibility-bearing state of one member location. The current online resource set directly participates in the convergence of responsibility-bearing representations, enabling the responsibility hidden distribution to simultaneously carry responsibility magnitude, responsibility attribution, and acceptance boundaries. Both target support relationships and responsibility migration constraints use the 20-dimensional responsibility hidden distribution as input; therefore, the responsibility hidden distribution constitutes a necessary intermediate object for embedding non-equilibrium optimal transmission into the deep Koopman autoencoder state-space model.

[0093] The unbalanced optimal transfer is embedded in the hidden state switching initialization stage of the deep Koopman autoencoder state-space model. The embedding position is after the current responsibility hidden distribution completes its topologically stable segment evolution and before the migrating responsibility hidden distribution enters the post-disconnection evolution. When the topological change conditions are met, the model does not perform dimension deletion, zeroing, or re-normalization on the 20-dimensional current responsibility hidden distribution. Instead, it establishes a 5x5 target support relationship based on the remaining unit up / down adjustment margins, short-time frequency difference response slope, saturation proximity, and energy storage capacity. This target support relationship then forms the responsibility migration constraint. Under the responsibility migration constraint, the unbalanced optimal transfer performs migration calculations on the positions of the five source members and the target member positions in the post-disconnection online resource set, outputting a 20-dimensional migrating responsibility hidden distribution. The key role of the unbalanced optimal transfer reinforcement is in the state initialization during hidden state updates. The unbalanced optimal transfer does not generate independent solution results or participate in the responsibility share result decoding. The unbalanced optimal transfer changes the way the current responsibility hidden distribution enters the post-disconnection evolution layer.

[0094] The decoding and encoding ends each have two fully connected layers. The first layer maps the 20-dimensional migration responsibility latent distribution to a 16-dimensional intermediate vector, with 16 neurons, each containing 20 weights and 1 bias. The nonlinear unit uses a rectified linear unit. The second layer has 5 linear neurons, each containing 16 weights and 1 bias, resulting in a 5-dimensional responsibility share. The 5-dimensional responsibility share corresponds to the responsibility share of 4 thermal power units and 1 energy storage unit according to the member's position. The member's position value corresponding to the grid-connected state constitutes the remaining unit's responsibility share. The difference between the energy storage member's position value and the energy storage responsibility value in the current responsibility latent distribution is calculated to form a 1-dimensional energy storage responsibility increment. The energy storage responsibility increment directly corresponds to the shared energy storage frequency regulation capacity.

[0095] In this structure, the deep Koopman autoencoder state-space model is responsible for generating the responsibility latent distribution from the member responsibility features and completing the evolution of the topology-stable segment. The unbalanced optimal transfer is responsible for transforming the current responsibility latent distribution into a migrating responsibility latent distribution when the topology change conditions are met. The responsibility latent distribution is always the only intermediate object, and the target support relationship and responsibility migration constraints always act on the member position expression of the responsibility latent distribution. Therefore, there are no parallel dual backbones or independent post-processing paths within the model. This structure limits the failure mechanism of the change of responsibility attachment object during unit decoupling to the latent state switching initialization stage for correction, and the correction result still returns to the original evolution chain of the deep Koopman autoencoder state-space model.

[0096] In some embodiments, the step S5 of generating the energy storage responsibility increment and the remaining unit responsibility share based on the responsibility share results specifically includes: Step S51: Extract the energy storage responsibility share from the responsibility share results, calculate the difference between the energy storage responsibility share and the corresponding value of the energy storage member position in the current responsibility hidden distribution before the topology change, and obtain the energy storage responsibility increment. The energy storage responsibility increment retains the sign information and does not perform absolute value processing or secondary normalization. Step S52: Select the value corresponding to the unit with grid connection status as connected from the responsibility share results as the remaining unit responsibility share.

[0097] In one embodiment of this application, the current inference time is taken as Receive the responsibility share result output in step S5 ,in to These correspond to the positions of the four crew members. This corresponds to the location of one energy storage member. It also receives the grid connection status of the generating unit at the current sampling time. The current hidden responsibility distribution formed by step S2 Liability share results With the current hidden distribution of responsibility The fixed order of 4 unit members and 1 energy storage member is maintained, so member position rearrangement, station-level aggregation, and introduction of new responsible parties are not performed.

[0098] First, regarding the results of the liability share. Split by member position. to Extracted as a subset of unit responsibility share ,Will Extracted as a share of energy storage responsibility This splitting operation only reads the index according to the fixed member position order and does not change the responsibility share result. The value. Then, based on the unit's grid connection status. Subset of unit responsibility share The process involves filtering. For crew member positions that meet the "on" status, the corresponding crew responsibility share is retained; for crew member positions that do not meet the "on" status, the corresponding crew responsibility share is not written into the final output. This results in the remaining crew responsibility shares. Remaining unit liability share Maintaining the member location index relationship facilitates the direct mapping of responsibility shares to the remaining units at the plant / station side.

[0099] In forming the remaining unit responsibility share Subsequently, from the current distribution of hidden responsibilities Read the location corresponding value of the energy storage member. Current responsibility hidden distribution. The 20-dimensional member position concatenation state is defined, with one member position corresponding to every four consecutive dimensions. Following the same member position partitioning method as in step S2, the current responsibility hidden distribution is then configured. The system is divided into 5 groups of 4-dimensional member location responsibility states, with the last group corresponding to the energy storage member locations. In the current implementation configuration, the deep Koopman autoencoder state-space model maintains a fixed internal arrangement for each group of 4-dimensional member location responsibility states, allowing the decoding end to read the result output in alignment with member locations. Based on this, the pre-fixed energy storage member location corresponding value for responsibility share alignment is read from the energy storage member location responsibility state, denoted as... Subsequently, the share of energy storage responsibility will be... Value corresponding to the location of energy storage members Perform difference calculations to generate incremental energy storage responsibility. Incremental energy storage responsibility In the current step, symbol information is preserved, absolute value processing is not performed, scaling is not performed, and quadratic normalization is not performed, thereby maintaining the incremental energy storage responsibility. Results of liability share and the current distribution of hidden responsibilities The semantics of the same responsible object.

[0100] Finally, the incremental responsibility for energy storage will be increased. The shared energy storage frequency regulation capacity will be directly determined and combined with the remaining unit responsibility share. As the final output. In this output chain, the responsibility share result. This represents the member position responsibility result after the migration responsibility hidden distribution is demultiplexed, evolved, and decoded; the current responsibility hidden distribution. This indicates the state of the responsibility source that was generated before the topology change, and the incremental energy storage responsibility. Based on the share of liability Energy storage responsibility share and current hidden responsibility distribution The value corresponding to the location of the energy storage member is obtained by direct comparison. Therefore, the shared energy storage frequency regulation capacity always follows the same responsibility object established around the hidden distribution of responsibility in steps S2 to S5.

[0101] The following describes the thermal power energy storage frequency regulation responsibility continuous switching device based on deep autoencoder provided in this application. The thermal power energy storage frequency regulation responsibility continuous switching device based on deep autoencoder described below can be referred to in correspondence with the thermal power energy storage frequency regulation responsibility continuous switching method based on deep autoencoder described above.

[0102] Figure 6 This is a schematic diagram of a thermal power energy storage frequency regulation responsibility continuous switching device based on a deep self-encoder, as provided in an embodiment of this application. Figure 6 As shown, the device 600 includes: Calculation module 610 is used to calculate the member responsibility characteristics of each unit and energy storage based on the pre-acquired online resource quantities of multiple thermal power units; Evolution module 620 is used to input member responsibility features into a deep Koopman autoencoder state space model, map each member responsibility feature to a responsibility-bearing representation, and aggregate them into a responsibility latent distribution according to the current online resource set. Under the constraints of the current online resource set, the responsibility latent distribution is subjected to topologically stable segment evolution to obtain the current responsibility latent distribution. Module 630 is used to determine topology changes based on online resource quantities. When the topology change conditions are met, the remaining units and energy storage are screened according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, the target support relationship for the transfer of the current hidden distribution of responsibility to the set of online resource quantities after disconnection is established, forming responsibility migration constraints. The generation module 640 is used to perform unbalanced optimal transfer on the current hidden responsibility distribution under the responsibility migration constraint, allowing partial migration and partial cancellation of responsibility, and generating a migrated hidden responsibility distribution; The output module 650 is used to deconcatenate and decode the deep Koopman autoencoder state space model into which the hidden distribution of migration responsibility is input, output the responsibility share result, generate the energy storage responsibility increment and the remaining unit responsibility share based on the responsibility share result, and output the shared energy storage frequency regulation undertaking amount.

[0103] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0104] Based on the methods in the above embodiments, Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown in the illustration, this application provides an electronic device that may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740. The processor 710, communication interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions from the memory 730 to execute the continuous switching method for frequency regulation responsibility of thermal power storage based on a deep autoencoder described in the above embodiment.

[0105] Furthermore, the logic instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the continuous switching method for frequency regulation responsibility based on a deep autoencoder for thermal power energy storage described in the various embodiments of this application.

[0106] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program. When the computer program runs on a processor, it causes the processor to execute the thermal power energy storage frequency regulation responsibility continuous switching method based on a deep autoencoder in the above embodiments.

[0107] Based on the methods in the above embodiments, this application provides a computer program product that, when running on a processor, causes the processor to execute the thermal power energy storage frequency regulation responsibility continuous switching method based on a deep autoencoder in the above embodiments.

[0108] It is understood that the processor in the embodiments of this application 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, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0109] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0110] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0111] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0112] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for continuous switching of frequency regulation responsibility in thermal power storage based on a deep autoencoder, characterized in that, include: Based on the pre-acquired online resource quantities of multiple thermal power units, calculate the member responsibility characteristics of each unit and energy storage; The member responsibility features are input into a deep Koopman autoencoder state space model, and each member responsibility feature is mapped to a responsibility-bearing representation. The features are then aggregated into a responsibility latent distribution according to the current online resource set. Under the constraints of the current online resource set, the responsibility latent distribution is subjected to topologically stable segment evolution to obtain the current responsibility latent distribution. Based on the online resource quantity, topology change is determined. When the topology change condition is met, the remaining units and energy storage are screened according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, a target support relationship is established for the transfer of the current hidden distribution of responsibility to the set of online resource quantity after disconnection, thus forming a responsibility migration constraint. Under the responsibility migration constraint, an unbalanced optimal transfer is performed on the current responsibility hidden distribution, allowing partial responsibility migration and partial cancellation, thereby generating a migrated responsibility hidden distribution; The migration responsibility hidden distribution is input into the deep Koopman autoencoder state space model for demultiplexing, evolution and decoding, and the responsibility share result is output. Based on the responsibility share result, the energy storage responsibility increment and the remaining unit responsibility share are generated, and the shared energy storage frequency regulation undertaking amount is output.

2. The method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder according to claim 1, characterized in that, The implicit responsibility distribution is arranged according to fixed member positions. Each member position corresponds to a fixed-dimensional responsibility-bearing state. The responsibility-bearing state is obtained by combining the responsibility-bearing representation with the underlying numerical fields in the member responsibility characteristics. This allows the implicit responsibility distribution to simultaneously carry the responsibility size, responsibility attribution, and acceptance boundary. Member positions that do not belong to the current online resource quantity set are written to a fixed zero-state placeholder, and no responsibility quantity is introduced from other member positions.

3. The method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder according to claim 1, characterized in that, The establishment of a target support relationship based on the member responsibility characteristics of remaining generating units and energy storage to transfer the current implicit distribution of responsibilities to the set of online resources after decoupling includes: Based on the set of online resources after disconnection, extract the member responsibility characteristics of the remaining generating units and energy storage; For each migration path from the source member location to the target member location, the acceptable boundary is calculated. When the target member location is a generator unit, the up and down adjustment margin path is automatically selected. When the target member location is energy storage, the energy storage receiving margin path is automatically selected. The acceptable boundary of the disconnected member location is set to zero, thus forming the target support relationship.

4. The method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder according to claim 1, characterized in that, Under the responsibility migration constraint, the unbalanced optimal transfer is performed on the current responsibility implicit distribution, allowing partial responsibility migration and partial deregistration, to generate a migrated responsibility implicit distribution, including: Using the transfer direction and transfer upper limit in the responsibility migration constraint as constraints, perform unbalanced optimal transfer on the current hidden responsibility distribution; Using the cross-member location migration constraint coefficient and the responsibility cancellation constraint coefficient as regularization terms, the responsibility transfer amount and responsibility cancellation amount to each acceptable member location are solved item by item for each responsibility component of the source member location; The responsibility transfer amounts received by each receiving member position are accumulated according to the member position and reconstructed in a fixed order into a vector with the same dimension as the responsibility latent distribution, which is used as the migration responsibility latent distribution.

5. The method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder according to claim 1, characterized in that, The unbalanced optimal transmission is embedded in the hidden state switching initialization stage of the deep Koopman autoencoder state space model. The embedding position is after the current responsibility hidden distribution completes the evolution of the topology stable segment and before the migration responsibility hidden distribution enters the post-disjoint evolution. It is used to enhance the responsibility migration capability of the deep Koopman autoencoder state space model across resource sets when the topology changes. The unbalanced optimal transmission does not generate solution results independent of the deep Koopman autoencoder state-space model.

6. The method for continuous switching of frequency regulation responsibility for thermal power storage based on a deep autoencoder according to claim 1, characterized in that, The generation of energy storage responsibility increment and remaining unit responsibility share based on the responsibility share results includes: Extract the energy storage responsibility share from the responsibility share results, calculate the difference between the energy storage responsibility share and the corresponding value of the energy storage member position in the current responsibility hidden distribution before the topology change, and obtain the energy storage responsibility increment. The energy storage responsibility increment retains the sign information and does not perform absolute value processing or double normalization. The remaining unit responsibility share is selected from the responsibility share results for units whose grid connection status is "connected".

7. A continuous switching device for frequency regulation responsibility of thermal power energy storage based on a deep self-encoder, characterized in that, include: The calculation module is used to calculate the member responsibility characteristics of each unit and energy storage based on the pre-acquired online resource quantities of multiple thermal power units; The evolution module is used to input the member responsibility features into the deep Koopman autoencoder state space model, map each member responsibility feature into a responsibility-bearing representation, and aggregate them into a responsibility latent distribution according to the current online resource quantity set. Under the constraints of the current online resource quantity set, the responsibility latent distribution is subjected to topologically stable segment evolution to obtain the current responsibility latent distribution. The construction module is used to determine topology changes based on the online resource quantity. When the topology change conditions are met, the remaining units and energy storage are screened according to the connection status at the current sampling time in the grid connection status of the units. Based on the member responsibility characteristics of the remaining units and energy storage, the target support relationship for the transfer of the current hidden distribution of responsibility to the set of online resource quantity after disconnection is established, forming a responsibility migration constraint. The generation module is used to perform unbalanced optimal transfer on the current hidden responsibility distribution under the responsibility migration constraint, allowing partial migration and partial cancellation of responsibility, and generating a migrated hidden responsibility distribution; The output module is used to input the migration responsibility hidden distribution into the deep Koopman autoencoder state space model for demultiplexing, evolution and decoding, output responsibility share results, generate energy storage responsibility increment and remaining unit responsibility share based on the responsibility share results, and output the shared energy storage frequency regulation undertaking amount.

8. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the continuous switching method for frequency regulation responsibility of thermal power storage based on a deep autoencoder as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, the processor performs the continuous switching method for frequency regulation responsibility of thermal power storage based on a deep autoencoder as described in any one of claims 1-6.

10. A computer program product, characterized in that, When the computer program product is run on the processor, the processor executes the continuous switching method for frequency regulation responsibility of thermal power storage based on a deep autoencoder as described in any one of claims 1-6.

Images