A large model driven multi-type energy storage multi-scale collaborative control method and system

Abstract

This invention discloses a large-scale collaborative control method and system for multiple types of energy storage driven by a large model, belonging to the field of intelligent control technology for energy storage systems. By collecting power grid operating status and state-of-charge data of each energy storage unit, a spatiotemporal coding strategy is used to map continuously changing physical signals into semantically encoded token sequences. These token sequences are input into a finely tuned large-scale power model to identify the current power grid operating conditions and generate multi-timescale power decomposition commands based on the remaining cycle life of the energy storage units. Based on these power decomposition commands, the output benchmark and compensation components of each type of energy storage unit are determined, and the commands are validated for feasibility through a physical constraint projection layer to filter out commands that violate physical extrema and fall outside the feasible domain. The droop coefficient of the underlying controller is adjusted in real time according to the validated commands, controlling each energy storage unit to inject virtual inertia and damping into the power grid, thus completing dynamic response adjustment.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for energy storage systems, and in particular to a large-model-driven multi-scale collaborative control method and system for multiple types of energy storage. Background Technology

[0002] With the integration of a high proportion of renewable energy, the power system exhibits significant low inertia and weak damping characteristics, posing severe challenges to frequency and voltage stability when the grid is subjected to load fluctuations or fault impacts. Traditional synchronous machine frequency regulation methods are no longer sufficient to meet the rapid dynamic response requirements at the microsecond to minute level. Although hybrid energy storage systems composed of power-type energy storage (such as supercapacitors and flywheels) and energy-type energy storage (such as lithium batteries and flow batteries) have complementary advantages, in practical applications, the output allocation of multiple types of energy storage often faces complex nonlinear coupling challenges. Furthermore, traditional control algorithms such as model predictive control (MPC) rely heavily on accurate physical models and are difficult to adapt to parameter drift caused by aging of energy storage cells and random changes in grid operating conditions. In addition, existing multi-timescale scheduling methods often suffer from high computational overhead, response lag, or overly rigid control logic when dealing with the coordinated balancing of millisecond-level transient stability and hourly-level energy balance, failing to balance operational economy and response sensitivity. Therefore, by introducing a large model with strong logical reasoning and generalization learning capabilities, and combining reinforcement learning and adaptive dynamic programming techniques, it is possible to achieve intelligent sensing and hierarchical collaborative control of multiple types of energy storage resources without the need for precise modeling, thereby effectively improving the safe and stable operation of the power grid under multi-scale dynamic disturbances. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] To address this, this invention proposes a large-scale collaborative control method for multi-type energy storage driven by a large model. This method collects data on the grid's operating status and the state of charge (SOC) of each energy storage unit, and uses a spatiotemporal coding strategy to map continuous physical signals into semantic token sequences. These token sequences are then input into a finely tuned large-scale power model to identify grid operating conditions and generate multi-timescale power decomposition commands based on the remaining cycle life of the energy storage units. Based on these power decomposition commands, the output benchmark and compensation components for each type of energy storage unit are determined, and the feasibility of the commands is verified through a physical constraint projection layer, filtering out commands that violate physical extrema and fall outside the feasible domain. The droop coefficient of the underlying controller is adjusted in real time based on the verified commands, controlling each energy storage unit to inject virtual inertia and damping into the grid, thus achieving dynamic response adjustment.

[0005] Another objective of this invention is to propose a large-scale collaborative control system for multi-type energy storage driven by a large model.

[0006] To achieve the above objectives, this invention proposes a large-model-driven multi-type energy storage multi-scale collaborative control method, comprising:

[0007] Collect power grid operation status data and state of charge data of each energy storage unit, and use a spatiotemporal coding strategy to map continuously changing physical signals into semantic token sequences; The token sequence is input into a finely tuned power model to identify the current grid operating conditions and generate power decomposition instructions at multiple time scales based on the remaining cycle life of the energy storage unit. Based on the power decomposition instructions, the output benchmark and compensation components of each type of energy storage unit are determined, and the instructions are verified for feasibility through the physical constraint projection layer to filter out non-feasible domain instructions that violate physical extreme values. Based on the verified instructions, the droop coefficient of the underlying controller is corrected in real time, and each energy storage unit is controlled to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

[0008] In one embodiment of the present invention, the step of collecting power grid operation status data and state-of-charge data of each energy storage unit, and mapping continuously changing physical signals into semantically encoded token sequences using a spatiotemporal coding strategy, includes: Real-time acquisition of grid frequency deviation, voltage fluctuation and state of charge of each energy storage unit as input data; A multidimensional linear quantization algorithm is used to perform a mapping operation from continuous physical quantities to discrete semantic space on the input data, generating a semantic token sequence containing electromagnetic transient features and power grid status information; The semantic token sequence is transformed into a context vector that the large model can understand, and used as input for subsequent condition recognition and instruction generation.

[0009] In one embodiment of the present invention, a multidimensional linear quantization algorithm is used to perform a mapping operation from continuous physical quantities to discrete semantic space on the input data, generating a semantic token sequence containing electromagnetic transient features and power grid status information, including: Based on timestamps, spatiotemporal alignment processing is performed on grid frequency deviation, voltage fluctuation and the state of charge of each energy storage unit to construct a multidimensional input vector; The initial semantic token is obtained by applying a quantization function to perform a linear transformation and discretization on the multidimensional input vector. The initial semantic tokens are assembled into a complete semantic token sequence in chronological order. ,in .

[0010] In one embodiment of the present invention, the step of inputting the token sequence into a fine-tuned power large-scale model, identifying the current grid operating conditions, and generating multi-timescale power decomposition instructions based on the remaining cycle life of the energy storage unit includes: Semantic token sequences are input into a fine-tuned large power model, and the attention mechanism of the large model is used to identify the current power grid conditions, including load surges, line trips, and damped oscillations. Based on the identified current grid operating conditions and the remaining cycle life of each energy storage unit, the decomposition coefficients at multiple time scales are dynamically calculated. Based on the decomposition coefficients, a long-scale energy-type energy storage output benchmark and a short-scale power-type energy storage compensation component are generated, forming a multi-time-scale power decomposition command.

[0011] In one embodiment of the present invention, a long-scale energy-type energy storage output benchmark and a short-scale power-type energy storage compensation component are generated based on the decomposition coefficients to form a multi-time-scale power decomposition command, including: Determine the output benchmark of energy storage system ESS-E over a long timescale. To ensure grid energy balance and determine the compensation component of power-type energy storage ESS-P within a short timescale. To quickly smooth out power grid fluctuations; According to the formula Verify the total power distribution relationship, where the decomposition coefficient α varies with the remaining cycle life of the energy storage. Adjustments are made in real time to reflect changes; The output includes the power reference. and compensation components Multi-timescale power decomposition instructions are sent to the underlying execution stage.

[0012] In one embodiment of the present invention, the step of determining the output benchmark and compensation component of each type of energy storage unit based on the power decomposition command, and performing feasibility verification of the command through a physical constraint projection layer to filter out infeasible domain commands that violate physical extreme values, includes: The power decomposition command is parsed to obtain the target output reference and compensation component of each type of energy storage unit, and the target output reference and compensation component are input to the physical constraint projection layer; In the physical constraint projection layer, the target command is compared with the power overload threshold and energy limit exceedance threshold of the energy storage unit to identify infeasible domain commands that may violate physical extremes. The identified infeasible domain instructions are subjected to projection correction operations to generate feasible domain instructions that satisfy physical constraints, and non-compliant instructions that cannot be corrected are filtered out.

[0013] In one embodiment of the present invention, the step of real-time correction of the droop coefficient of the underlying controller according to the verified instructions, and controlling each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment, includes: Extract the control parameters contained in the verified instructions, and use a large model to calculate the corrected droop coefficient based on real-time operating conditions. and virtual inertia parameters ; The corrected droop coefficient and virtual inertia parameters Substituting the improved droop control equation, the real-time output power of each energy storage unit is calculated. ; Based on the real-time output power The system controls the injection of millisecond-level virtual inertia and damping into the grid from each energy storage unit, where... .

[0014] In one embodiment of the present invention, the method further includes constructing a closed-loop feedback mechanism: The generalization learning capability of large models is used to predict future power grid trends, and a closed-loop feedback mechanism is constructed by combining real-time feedback data on energy storage status. Based on future power grid trends and real-time feedback data, the power decoupling frequency point is dynamically corrected to optimize the power weight allocation logic of various types of energy storage in the dynamic response process, achieving dynamic optimal allocation of resources without the need for precise physical modeling.

[0015] In one embodiment of the present invention, based on the identified current grid operating conditions and the remaining cycle life of each energy storage unit, the decomposition coefficients at multiple time scales are dynamically calculated, further including: Collect the remaining cycle life and corresponding degradation cost function of each energy storage unit, and convert the degradation cost function into the constraint conditions in the large model prompt words; The constraints and the semantic token sequence of the current power grid operating conditions are input together into the attention mechanism of the power big model to dynamically calculate the multi-timescale decomposition coefficients that enable the operating economy and power grid support performance to achieve Pareto optimality. The multi-timescale decomposition coefficients are output to the power decomposition instruction generation stage to adaptively adjust the power allocation weights between the long-scale energy-type energy storage output benchmark and the short-scale power-type energy storage compensation component.

[0016] This invention also proposes a large-model-driven multi-type energy storage multi-scale collaborative control system, comprising: The data acquisition and spatiotemporal coding module is used to acquire power grid operation status data and state of charge data of each energy storage unit, and to map continuously changing physical signals into semantic token sequences using a spatiotemporal coding strategy. The large model decision module is used to input the token sequence into the fine-tuned power large model, identify the current power grid operating conditions, and generate power decomposition instructions at multiple time scales based on the remaining cycle life of the energy storage unit. The physical constraint projection module is used to determine the output benchmark and compensation component of each type of energy storage unit based on the power decomposition command, and to perform feasibility verification of the command through the physical constraint projection layer to filter out non-feasible domain commands that violate physical extreme values. The underlying execution module is used to correct the droop coefficient of the underlying controller in real time according to the verified instructions, and to control each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

[0017] This invention discloses a large-model-driven multi-scale collaborative control method and system for multiple types of energy storage. By constructing a spatiotemporal coding strategy to map continuous physical signals into semantic token sequences, and utilizing a fine-tuned large-scale power model to identify grid operating conditions, it generates multi-timescale power decomposition commands based on the remaining cycle life of the energy storage. This effectively solves the problems of insufficient dynamic response and poor operational economy in existing technologies caused by lag in operating condition identification, mismatch in power allocation among multiple types of energy storage, and lack of physical constraints. The feasibility of commands is verified and infeasible commands are filtered out through a physical constraint projection layer. Combined with real-time correction of the droop coefficient of the underlying controller to inject virtual inertia and damping, integrated closed-loop control from data acquisition and large-scale model decision-making to underlying dynamic adjustment is achieved. This significantly improves the collaborative response speed and power allocation accuracy of multiple types of energy storage under complex grid disturbances, enhancing the system's operational reliability and engineering applicability.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a large-model-driven multi-type energy storage multi-scale collaborative control method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a large-scale collaborative control system for multi-type energy storage driven by an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, illustrates a large-model-driven, multi-type energy storage, multi-scale collaborative control method and system according to an embodiment of the present invention. Figure 1 This is a flowchart of a large-model-driven multi-type energy storage multi-scale collaborative control method according to an embodiment of the present invention.

[0023] like Figure 1 As shown, a large-model-driven multi-scale collaborative control method for multiple types of energy storage includes the following steps: S1 collects power grid operation status data and state of charge data of each energy storage unit, and uses a spatiotemporal coding strategy to map continuously changing physical signals into semantic token sequences. S2, input the token sequence into the fine-tuned power big model, identify the current power grid operating conditions and generate multi-timescale power decomposition instructions based on the remaining cycle life of the energy storage unit; S3, based on the power decomposition command, determine the output benchmark and compensation component of each type of energy storage unit, and perform feasibility verification of the command through the physical constraint projection layer to filter out non-feasible domain commands that violate physical extreme values. S4, based on the verified instructions, corrects the droop coefficient of the underlying controller in real time, and controls each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

[0024] Specifically, the core logic of the embodiments of the present invention is implemented through the following three levels: A. Power Grid Situation Awareness and Tokenization Mapping Based on Spatiotemporal Coding. Physical Signal Acquisition: First, acquire power grid frequency deviation. Voltage fluctuations and the state of charge of various energy storage units A multidimensional linear quantization algorithm is used to map continuous physical quantities into semantic token sequences. Mathematical expression: This process transforms physical signals into context vectors that large models can understand. Using a multidimensional linear quantization algorithm, these continuously changing physical signals are converted into semantic token sequences. This spatiotemporal coding strategy is like establishing a dedicated "language encoder" for the power system, enabling large models to "understand" complex electromagnetic transient characteristics and power grid conditions as easily as processing text.

[0025] B. LLM-driven "brain" decision-making and multi-scale power decoupling. Intelligent operating condition identification: The finely tuned power model identifies the current grid state (e.g., load surge, line tripping, or damped oscillation) based on the received semantic sequence; Multi-scale command generation: The large model dynamically generates multi-timescale decomposed commands: at the long scale (hours / minutes), it determines the output benchmark for energy-type energy storage to ensure energy balance; at the short scale (seconds / milliseconds), it determines the compensation component for power-type energy storage to quickly smooth out fluctuations; Lifetime adaptive coordination: The output allocation at the two scales is not fixed. The large model adjusts the allocation coefficient in real time based on the remaining cycle life of different energy storage units to achieve a balance between operational economy and support performance. Includes: Long-scale (hour / minute level): Determining the output benchmark for energy storage systems (ESS-E) Short-scale (second / millisecond level): Determining the compensation components of power-mode energy storage (ESS-P) Collaborative relationships: The decomposition coefficients Based on the large model and the remaining cycle life of the energy storage Adjust in real time.

[0026] C. Adaptive Low-Level Execution Control Incorporating Physical Constraints. To ensure the safety of the large model output, a Physics-Informed Layer is embedded in the low-level controller. An improved droop control is employed, with a droop coefficient... Real-time correction by a large model: ; Virtual inertia is achieved through this equation. With damping Millisecond-level injection.

[0027] Specifically, the system incorporates several key features: **Physical Perception Layer Embedding:** To filter out potentially infeasible commands generated by the large model, a physical perception layer is embedded in the underlying controller. This layer strictly limits commands from violating the physical limits of energy storage (such as power overload or exceeding capacity limits). **Real-Time Parameter Correction:** The system employs improved droop control, where key control coefficients are corrected by the large model based on real-time operating conditions. This allows the system to inject millisecond-level virtual inertia and damping into the grid, significantly enhancing its resilience to dynamic disturbances. **Closed-Loop Feedback Mechanism:** Leveraging the generalization learning capability of the large model to predict future trends, and combined with real-time feedback on energy storage status, the system uses a closed-loop correction mechanism to adjust the power decoupling frequency point, ensuring dynamic optimal resource allocation without requiring precise physical modeling.

[0028] The method of this invention effectively solves the problems of insufficient dynamic response caused by delayed operating condition identification, power allocation mismatch and lack of physical constraints in the prior art. It realizes integrated closed-loop control from situational awareness and large model decision-making to low-level execution, significantly improves the collaborative response speed and power allocation accuracy of various types of energy storage under complex grid disturbances, and enhances the reliability and engineering applicability of system operation.

[0029] To implement the method of the above embodiments, such as Figure 2 This invention proposes a large-scale collaborative control system for multi-type energy storage driven by a large model. The system 10 includes: The data acquisition and spatiotemporal coding module 100 is used to acquire power grid operation status data and state of charge data of each energy storage unit, and to map continuously changing physical signals into semantic token sequences using a spatiotemporal coding strategy.

[0030] The large model decision module 200 is used to input the token sequence into the fine-tuned power large model, identify the current power grid operating conditions, and generate power decomposition instructions at multiple time scales based on the remaining cycle life of the energy storage unit.

[0031] The physical constraint projection module 300 is used to determine the output benchmark and compensation component of each type of energy storage unit based on the power decomposition command, and to perform feasibility verification on the command through the physical constraint projection layer to filter out non-feasible domain commands that violate physical extreme values.

[0032] The underlying execution module 400 is used to correct the droop coefficient of the underlying controller in real time according to the verified instructions, and to control each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

[0033] The system of this invention effectively solves the problems of delayed operating condition identification, power allocation mismatch and lack of physical constraints in the prior art. It realizes integrated collaborative control from data acquisition, intelligent decision-making to dynamic execution, significantly improves the response speed and allocation accuracy of various types of energy storage under complex power grid disturbances, and enhances the reliability and engineering applicability of the system operation.

[0034] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0035] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A large-scale collaborative control method for multiple types of energy storage driven by a large model, characterized in that, Includes the following steps: Collect power grid operation status data and state of charge data of each energy storage unit, and use a spatiotemporal coding strategy to map continuously changing physical signals into semantic token sequences; The token sequence is input into a finely tuned power model to identify the current grid operating conditions and generate power decomposition instructions at multiple time scales based on the remaining cycle life of the energy storage unit. Based on the power decomposition instructions, the output benchmark and compensation components of each type of energy storage unit are determined, and the instructions are verified for feasibility through the physical constraint projection layer to filter out non-feasible domain instructions that violate physical extreme values. Based on the verified instructions, the droop coefficient of the underlying controller is corrected in real time, and each energy storage unit is controlled to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

2. The method according to claim 1, characterized in that, The process involves collecting power grid operation status data and state-of-charge data from each energy storage unit, and using a spatiotemporal coding strategy to map continuously changing physical signals into semantically encoded token sequences, including: Real-time acquisition of grid frequency deviation, voltage fluctuation and state of charge of each energy storage unit as input data; A multidimensional linear quantization algorithm is used to perform a mapping operation from continuous physical quantities to discrete semantic space on the input data, generating a semantic token sequence containing electromagnetic transient features and power grid status information; The semantic token sequence is transformed into a context vector that the large model can understand, and used as input for subsequent condition recognition and instruction generation.

3. The method according to claim 2, characterized in that, A multidimensional linear quantization algorithm is used to map the input data from continuous physical quantities to a discrete semantic space, generating a semantic token sequence containing electromagnetic transient features and power grid status information, including: Based on timestamps, spatiotemporal alignment processing is performed on grid frequency deviation, voltage fluctuation and the state of charge of each energy storage unit to construct a multidimensional input vector; The initial semantic token is obtained by applying a quantization function to perform a linear transformation and discretization on the multidimensional input vector. The initial semantic tokens are assembled into a complete semantic token sequence in chronological order. ,in .

4. The method according to claim 1, characterized in that, The step of inputting the token sequence into a fine-tuned power model to identify the current grid operating conditions and generate multi-timescale power decomposition instructions based on the remaining cycle life of the energy storage units includes: Semantic token sequences are input into a fine-tuned large power model, and the attention mechanism of the large model is used to identify the current power grid conditions, including load surges, line trips, and damped oscillations. Based on the identified current grid operating conditions and the remaining cycle life of each energy storage unit, the decomposition coefficients at multiple time scales are dynamically calculated. Based on the decomposition coefficients, a long-scale energy-type energy storage output benchmark and a short-scale power-type energy storage compensation component are generated, forming a multi-time-scale power decomposition command.

5. The method according to claim 4, characterized in that, Based on the decomposition coefficients, a long-scale energy-type energy storage output benchmark and a short-scale power-type energy storage compensation component are generated, forming a multi-time-scale power decomposition command, including: Determine the output benchmark of energy storage system ESS-E over a long timescale. To ensure grid energy balance and determine the compensation component of power-type energy storage ESS-P within a short timescale. To quickly smooth out power grid fluctuations; According to the formula Verify the total power distribution relationship, where the decomposition coefficient α varies with the remaining cycle life of the energy storage. Adjustments are made in real time to reflect changes; The output includes the power reference. and compensation components Multi-timescale power decomposition instructions are sent to the underlying execution stage.

6. The method according to claim 1, characterized in that, The process of determining the output benchmark and compensation component of each type of energy storage unit based on the power decomposition command, and performing feasibility verification of the command through a physical constraint projection layer to filter out infeasible domain commands that violate physical extremes, includes: The power decomposition command is parsed to obtain the target output reference and compensation component of each type of energy storage unit, and the target output reference and compensation component are input to the physical constraint projection layer; In the physical constraint projection layer, the target command is compared with the power overload threshold and energy limit exceedance threshold of the energy storage unit to identify infeasible domain commands that may violate physical extremes. The identified infeasible domain instructions are subjected to projection correction operations to generate feasible domain instructions that satisfy physical constraints, and non-compliant instructions that cannot be corrected are filtered out.

7. The method according to claim 1, characterized in that, The process of real-time correction of the droop coefficient of the underlying controller based on the verified instructions, and controlling each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment, includes: Extract the control parameters contained in the verified instructions, and use a large model to calculate the corrected droop coefficient based on real-time operating conditions. and virtual inertia parameters ; The corrected droop coefficient and virtual inertia parameters Substituting the improved droop control equation, the real-time output power of each energy storage unit is calculated. ; Based on the real-time output power The system controls the injection of millisecond-level virtual inertia and damping into the grid from each energy storage unit, where... .

8. The method according to claim 1, characterized in that, The method also includes constructing a closed-loop feedback mechanism: The generalization learning capability of large models is used to predict future power grid trends, and a closed-loop feedback mechanism is constructed by combining real-time feedback data on energy storage status. Based on future power grid trends and real-time feedback data, the power decoupling frequency point is dynamically corrected to optimize the power weight allocation logic of various types of energy storage in the dynamic response process, achieving dynamic optimal allocation of resources without the need for precise physical modeling.

9. The method according to claim 4, characterized in that, Based on the identified current grid operating conditions and the remaining cycle life of each energy storage unit, the decomposition coefficients at multiple time scales are dynamically calculated, further including: Collect the remaining cycle life and corresponding degradation cost function of each energy storage unit, and convert the degradation cost function into the constraint conditions in the large model prompt words; The constraints and the semantic token sequence of the current power grid operating conditions are input together into the attention mechanism of the power large model to dynamically calculate the multi-timescale decomposition coefficients that enable the operation economy and power grid support performance to achieve Pareto optimality. The multi-timescale decomposition coefficients are output to the power decomposition instruction generation stage to adaptively adjust the power allocation weights between the long-scale energy storage output benchmark and the short-scale power storage compensation component.

10. A large-scale collaborative control system for multi-type energy storage driven by a large model, characterized in that, include: The data acquisition and spatiotemporal coding module is used to acquire power grid operation status data and state of charge data of each energy storage unit, and to map continuously changing physical signals into semantic token sequences using a spatiotemporal coding strategy. The large model decision module is used to input the token sequence into the fine-tuned power large model, identify the current power grid operating conditions, and generate power decomposition instructions at multiple time scales based on the remaining cycle life of the energy storage unit. The physical constraint projection module is used to determine the output benchmark and compensation component of each type of energy storage unit based on the power decomposition command, and to perform feasibility verification of the command through the physical constraint projection layer to filter out non-feasible domain commands that violate physical extreme values. The underlying execution module is used to correct the droop coefficient of the underlying controller in real time according to the verified instructions, and to control each energy storage unit to inject virtual inertia and damping into the grid to complete dynamic response adjustment.

Images