Self-organizing energy storage agent cluster system based on ac strong electric interface and control method

Abstract

The application discloses an AC strong current interface-based self-organizing energy storage intelligent agent cluster system and a control method, relates to the technical field of intelligent agent cluster control, and comprises an intelligent agent module, a grid-connected module, a networking module, a communication module, a sampling module and a control module. The intelligent agent module is used for establishing a single-machine AC output structure. The grid-connected module is used for collecting bus synchronization quantities and performing pre-charging and main closing. The networking module is used for connecting multiple energy intelligent agents to a public AC bus through AC strong current interfaces and forming cluster electrical coupling. The communication module is used for multiplexing a public AC bus to superimpose a power line carrier signal and generating a power disturbance identifier through power adjustment. The sampling module is used for collecting operating states and triggering parameter optimization according to an abnormal threshold. The control module is used for asynchronously calling multiple intelligent agent reinforcement learning intelligent agents, and combining a potential field rule engine to correct bidirectional power conversion unit control parameters in the intelligent agent module.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent cluster control technology, specifically to a self-organizing energy storage intelligent agent cluster system and control method based on an AC high-voltage interface. Background Technology

[0002] With the increasing proportion of new energy power generation and changes in electricity demand patterns, electrochemical energy storage systems are gradually evolving from single backup power sources to functions such as frequency regulation, peak shaving and valley filling, microgrid support, and coordinated control of distributed energy resources. Existing energy storage systems typically consist of battery clusters, power conversion units, energy management systems, and communication networks. Some solutions further incorporate modular energy storage, virtual synchronous machine control, droop control, and intelligent optimization algorithms to improve the power regulation capability of energy storage units and the stability of parallel systems. Against the backdrop of software-defined energy storage and edge intelligent control, transforming energy storage units from passive execution components into energy intelligent agents with local sensing, decision-making, and AC output capabilities has become an important research direction for the flexible expansion and autonomous operation of energy storage systems.

[0003] Existing modular energy storage technologies still largely rely on DC-side parallel connection, centralized power conversion units, or centralized energy management units. The connection, disconnection, and power distribution of energy storage units often require unified coordination by a higher-level controller. In this architecture, electrical and communication interfaces are usually independent, and adding new energy storage units requires simultaneous handling of high-voltage wiring, communication access, and control parameter configuration, limiting the flexibility of system expansion. On the other hand, although existing AC parallel systems can improve power distribution through droop control or virtual impedance, problems such as parallel circulating current, uneven output, and parameter adjustment lag can still easily occur when multiple energy storage units are hot-swapped for grid connection, communication failures, controller delays, and inconsistent states of charge. While power line communication can reduce wiring, relying solely on carrier communication is insufficient to express the electrical coupling information of a single unit's operating state. Although multi-agent reinforcement learning can be used for distributed optimization, it is difficult to directly adapt to real-time energy storage control on high-voltage AC buses without device boundary constraints, inference timeout handling, and register-level write-back mechanisms. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing modular energy storage cluster control methods rely on DC parallel connection and centralized power conversion units for energy storage unit interfaces, rely on dedicated communication buses and centralized controllers for cluster collaboration, lack a unified self-organizing control mechanism for hot-swappable access of energy storage units on the AC side, power sharing, and state of charge coordination processes, and how to achieve self-organizing networking, low-speed information interaction, abnormal triggering, and safe closed-loop control of multiple energy storage intelligent agents under the condition of relying only on AC high-voltage interfaces and common AC buses.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a self-organizing energy storage intelligent agent cluster system based on an AC high-voltage interface, comprising an intelligent agent module, a grid-connected module, a networking module, a communication module, a sampling module, and a control module; the intelligent agent module is used to establish a single-unit AC output structure; the grid-connected module is used to collect bus synchronization quantities and perform pre-charging and main closing; the networking module is used to connect multiple energy intelligent agents to a common AC bus via an AC high-voltage interface and form a cluster electrical coupling; the communication module is used to multiplex the common AC bus superimposed with power line carrier signals and generate power disturbance identifiers through power regulation; the sampling module is used to collect operating status and trigger parameter optimization according to anomaly thresholds; the control module is used to asynchronously call multi-agent reinforcement learning intelligent agents and, in conjunction with a latent field rule engine, correct the control parameters of the bidirectional power conversion unit in the intelligent agent module.

[0007] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the intelligent agent module includes a battery pack, a battery management unit, a bidirectional power conversion unit, an edge artificial intelligence module, and an AC high-voltage interface; the DC output terminal of the battery pack is connected to the DC input terminal of the bidirectional power conversion unit, and the AC output terminal of the bidirectional power conversion unit is connected to the AC high-voltage interface; the battery management unit is electrically connected to the battery pack, and collects battery status parameters and generates state of charge, alarm signals, and power current limiting commands; the state of charge, alarm signals, and power current limiting commands are transmitted to the edge artificial intelligence module and the bidirectional power conversion unit via control lines; the bidirectional power conversion unit converts the DC energy output from the battery pack into three-phase AC output according to a four-quadrant operation mode and a grid formation control mode; the AC high-voltage interface is used as the only main electrical interface of the intelligent agent module, and an AC contactor, circuit breaker, synchronous sampling unit, and grid connection protection unit are installed in the AC high-voltage interface; the edge artificial intelligence module is connected to the battery management unit and the bidirectional power conversion unit through control lines, and a channel is established for the edge artificial intelligence module to write control parameters to the bidirectional power conversion unit.

[0008] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the following steps are included: The acquisition of bus synchronization quantities and the execution of pre-charging and main closing include a synchronization sampling unit acquiring the voltage signal on the common AC bus side and the voltage signal on the output side of the local bidirectional power conversion unit, and extracting the voltage amplitude, frequency, and phase; the grid-connected state machine compares the voltage amplitude difference, frequency difference, and phase difference between the common AC bus side and the output side of the local bidirectional power conversion unit with verification thresholds; when the voltage amplitude difference, frequency difference, and phase difference simultaneously meet the verification thresholds and remain continuously and stably for 500ms, grid connection is achieved. The state machine outputs a pre-charging closing command; after the pre-charging auxiliary contactor closes, it maintains voltage equalization for 300ms, and the main power contactor closes; after the main power contactor contacts are in position, the pre-charging auxiliary contactor opens after a 100ms delay; when any level contactor fails to provide a contact position signal, the grid-connected state machine re-outputs the corresponding closing command every 200ms, and the number of retries for the same grid-connected process does not exceed 3; when any of the following conditions are met: synchronization index exceeds limit, closing timeout, multiple closing failures, or AC / DC hardware fault, the grid-connected state machine cancels the closing command, disconnects the pre-charging auxiliary contactor and the main power contactor, and records the fault status.

[0009] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the step of connecting multiple energy intelligent agents to a common AC bus via AC high-voltage interface includes each energy intelligent agent connecting to the common AC bus via a corresponding AC high-voltage interface to form an energy intelligent agent cluster; each energy intelligent agent operates collaboratively by combining electrical coupling with the common AC bus, distributed intelligent rules, and power line communication information interaction; before a new energy intelligent agent is connected to the common AC bus, synchronous verification and closing are performed; when an already connected energy intelligent agent leaves the common AC bus, it performs off-grid exit through the corresponding AC high-voltage interface.

[0010] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the communication module includes a power line carrier communication unit and a power signature communication unit; it is coupled to a common AC bus, and a 2kHz to 5kHz narrowband frequency shift keying carrier signal is superimposed on the common AC bus; the narrowband frequency shift keying carrier signal is injected intermittently with short pulses, and the instantaneous injection peak value is 1% to 3% of the rated voltage; the power line carrier communication unit sends communication frames according to a 200ms communication cycle; the communication frame includes an address field, an instruction field, and a sixteen-bit cyclic redundancy check field; the power signature communication unit adjusts active power and reactive power through the energy intelligent agent to generate a corresponding power disturbance identifier.

[0011] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the sampling module includes a sampling unit, an anomaly determination unit, and a trigger locking unit; the sampling unit collects grid connection point frequency, grid connection point voltage, state of charge, output power, parallel branch circulating current, and bus load fluctuation according to the sampling period; the anomaly determination unit compares the grid connection point frequency with the rated frequency to obtain the frequency deviation, compares the grid connection point voltage with the rated voltage to obtain the voltage deviation, and subtracts the minimum state of charge from the maximum state of charge in the cluster to obtain the state of charge range; the anomaly determination unit performs threshold judgment on the frequency deviation, voltage deviation, state of charge, state of charge range, parallel branch circulating current, bus load fluctuation, and multi-machine power sharing error; threshold judgment The conditions include: absolute frequency deviation greater than 0.2Hz, absolute voltage deviation greater than 5% of rated voltage, state of charge (SOC) below 20%, SOC above 85%, SOC range greater than 15%, parallel branch circulating current greater than 3% of rated phase current, bus load step fluctuation amplitude greater than 10% of rated output, and multi-machine power sharing error exceeding the standard. The trigger locking unit outputs a parameter optimization trigger signal after the same threshold judgment condition is met for 3 consecutive sampling cycles. After the parameter optimization trigger signal is output, the trigger locking unit sets a minimum hold time of 2 seconds and locks the new parameter optimization trigger signal within the minimum hold time. When the minimum hold time ends and all abnormal indicators do not meet the threshold judgment condition for 5 consecutive sampling cycles, the trigger locking unit outputs parameter rollback permission.

[0012] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the control module includes a latent field rule engine, a multi-agent reinforcement learning agent, a parameter pruning unit, and a register write-back unit; the latent field rule engine stores grid connection specifications, equipment limits, and parameter physical boundaries, and establishes legal value ranges for droop correction coefficients, virtual impedance parameters, and power signature parameters; the multi-agent reinforcement learning agent receives the output of the sampling module, asynchronously performs inference after the parameter optimization trigger signal is output, and generates initial control parameters; when the inference time does not exceed the sampling period, the parameter pruning unit performs amplitude limiting and rate of change limitation on the initial control parameters; When the processing time exceeds the sampling period, the parameter trimming unit calls the previous version of stable control parameters and executes safety boundary constraints; the register write-back unit writes the processed droop correction coefficient, virtual impedance parameter, and power control parameter into the register to be activated of the bidirectional power conversion unit, and switches the control parameters in the register to be activated to the currently activated control parameters at the sampling period boundary through parameter version verification; after the control parameters take effect, the control module performs post-activation verification based on the abnormal indicators that triggered this round of parameter optimization; when the post-activation verification fails, the control module calls the previous version of stable control parameters for version rollback; when the post-activation verification passes, the control module records the current round of control parameters as stable control parameters.

[0013] As a preferred embodiment of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface described in this invention, the method includes: establishing a single-unit AC output structure; collecting bus synchronization quantities and performing pre-charging and main closing; connecting multiple energy intelligent agents to a common AC bus via AC high-voltage interface to form cluster electrical coupling; multiplexing the common AC bus and superimposing power line carrier signals, and generating power disturbance indicators through power regulation; collecting operating status and triggering parameter optimization according to abnormal thresholds; asynchronously calling multi-agent reinforcement learning intelligent agents, and combining the latent field rule engine to correct the control parameters of the bidirectional power conversion unit in the intelligent agent module.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for a self-organizing energy storage intelligent agent cluster based on an AC high-voltage interface.

[0015] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for a self-organizing energy storage intelligent agent cluster based on an AC high-voltage interface.

[0016] The beneficial effects of this invention are as follows: The self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface provided by this invention integrates battery packs, battery management units, bidirectional power conversion units, edge artificial intelligence modules, and AC high-voltage interfaces into a single energy intelligent agent. The energy storage unit transforms from a controlled component dependent on external DC interfaces and centralized power conversion units into an autonomous unit with local AC output, status acquisition, and control parameter writing capabilities. Furthermore, a cluster electrical coupling without a centralized controller is formed through a common AC bus, enabling member access, exit, and online computing objects to correspond to the actual contactor status. Power line carrier communication and power disturbance identification reuse the same common AC bus, respectively carrying cluster commands and individual unit status interactions, reducing the constraints of dedicated communication buses on cluster expansion. On the control side, through abnormal threshold triggering, asynchronous multi-agent reinforcement learning inference, inference timeout stabilization parameter invocation, and latent field rule engine safety boundary constraints, control parameters are written to the bidirectional power conversion unit register under the condition of satisfying device boundaries, reducing the impact of inference delay, transient anomalies, and parameter exceeding limits on the parallel operation of the AC bus. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the architecture of the self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface provided by the present invention.

[0019] Figure 2 This is a parameter write-back diagram of the self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface provided by the present invention.

[0020] Figure 3 The overall flowchart of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface provided by the present invention is shown below.

[0021] Figure 4 A schematic diagram of a computer device provided by the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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 protection scope of the present invention.

[0023] Reference Figures 1-2 As one embodiment of the present invention, a self-organizing energy storage intelligent agent cluster system based on an AC high-voltage interface is provided, comprising:

[0024] Intelligent agent module, grid connection module, network topology module, communication module, sampling module, control module;

[0025] The intelligent agent module is used to establish a standalone AC output structure.

[0026] Furthermore, refer to Figure 1The intelligent agent module includes a battery pack, a battery management unit, a bidirectional power conversion unit, an edge artificial intelligence module, and an AC high-voltage interface. The DC output of the battery pack is connected to the DC input of the bidirectional power conversion unit, and the AC output of the bidirectional power conversion unit is connected to the AC high-voltage interface. The battery management unit is electrically connected to the battery pack, collects battery status parameters, and generates state of charge, alarm signals, and power current limiting commands. The state of charge, alarm signals, and power current limiting commands are transmitted to the edge artificial intelligence module and the bidirectional power conversion unit via control lines. The bidirectional power conversion unit converts the DC energy output from the battery pack into three-phase AC output according to a four-quadrant operation mode and a grid formation control mode. The AC high-voltage interface is used as the only main electrical interface of the intelligent agent module, and it is equipped with an AC contactor, circuit breaker, synchronous sampling unit, and grid-connected protection unit. The edge artificial intelligence module is connected to the battery management unit and the bidirectional power conversion unit via control lines, and establishes a channel for the edge artificial intelligence module to write control parameters to the bidirectional power conversion unit.

[0027] It should be noted that one specific scheme for establishing a single-unit AC output structure includes an intelligent module comprising a battery pack, a battery management unit, a bidirectional power conversion unit, an edge AI module, and an AC high-voltage interface. The DC output terminal of the battery pack is connected to the DC input terminal of the bidirectional power conversion unit, and the AC output terminal of the bidirectional power conversion unit is connected to the AC high-voltage interface. The battery pack provides DC energy to the bidirectional power conversion unit, which, operating in a four-quadrant mode and under grid formation control, converts the DC energy output from the battery pack into a three-phase AC output. The three-phase AC output is then output via the AC high-voltage interface.

[0028] The battery management unit (BMU) is electrically connected to the battery pack. The BMU collects battery status parameters and generates state of charge (SOC), alarm signals, and power current limiting commands. SOC parameters include battery voltage, charge / discharge current, and battery temperature. SOC, alarm signals, and power current limiting commands are transmitted via control lines to the edge AI module and the bidirectional power conversion unit (PCU). SOC serves as input to both the sampling and control modules, expressed as a value between 0 and 1, where 0 represents the low charge boundary and 1 represents the full charge boundary. Alarm signals are generated by the BMU based on preset protection boundaries corresponding to battery voltage, charge / discharge current, and battery temperature. Power current limiting commands are generated by the BMU based on the battery pack's current allowable charge / discharge capacity and transmitted to the PCU, ensuring that the PCU executes power commands within the battery pack's allowable power range.

[0029] The AC high-voltage interface serves as the sole main electrical interface for the intelligent agent module. It houses an AC contactor, circuit breaker, synchronous sampling unit, and grid-connected protection unit. Through this interface, a single energy intelligent agent can achieve AC-side access capability without an external DC parallel interface. The edge AI module connects to the battery management unit and the bidirectional power conversion unit via control lines, establishing a channel for the edge AI module to write control parameters to the bidirectional power conversion unit. These control parameters include droop coefficient, virtual impedance parameters, and power control parameters. The intelligent agent module's own three-phase AC power output enters the AC high-voltage interface, its output voltage signal enters the grid-connected module, and its state of charge, alarm signals, and power current limiting commands enter the sampling and control modules for status judgment and parameter correction.

[0030] It should also be noted that the intelligent agent module establishes a single-machine AC output structure, transforming the energy storage unit from a traditional battery cluster or DC-side controlled component into an energy intelligent agent with its own AC output capability. In existing energy storage systems, battery packs typically rely on external centralized power conversion units to achieve grid-connected output. Adding or replacing energy storage units involves not only DC-side connections but also reconfiguration of power conversion and control links. This invention integrates the battery pack, battery management unit, bidirectional power conversion unit, edge AI module, and AC high-voltage interface into the same intelligent agent module. The battery management unit collects battery pack voltage, current, and temperature and generates state of charge, alarm signals, and power current limiting commands. The bidirectional power conversion unit converts DC energy into three-phase AC output, and the AC high-voltage interface serves as the sole main electrical interface for AC-side access. Each energy intelligent agent completes power conversion, state acquisition, and parameter writing channel construction before connecting to the common AC bus, reducing reliance on external DC parallel interfaces and centralized power conversion units.

[0031] The grid connection module is used to collect bus synchronization data and perform pre-charging and main closing.

[0032] Furthermore, the process of acquiring bus synchronization signals and executing pre-charging and main closing includes the synchronization sampling unit acquiring the voltage signal on the common AC bus side and the output voltage signal on the local bidirectional power conversion unit, and extracting the voltage amplitude, frequency, and phase. The grid-connected state machine compares the voltage amplitude difference, frequency difference, and phase difference between the common AC bus side and the output voltage of the local bidirectional power conversion unit with the verification thresholds. When the voltage amplitude difference, frequency difference, and phase difference simultaneously meet the verification thresholds and remain stable for 500ms, the grid-connected state machine outputs a pre-charging closing command. Pre-charging auxiliary connection... After the contactor closes, it maintains voltage equalization for 300ms, and the main power contactor closes. After the main power contactor contacts are in position, the pre-charging auxiliary contactor opens after a 100ms delay. When any primary contactor fails to provide a contact position signal, the grid connection state machine re-outputs the corresponding closing command every 200ms. The number of retries for the same grid connection process does not exceed 3. When any of the following conditions are met: synchronization index exceeds limit, closing timeout, multiple closing failures, or AC / DC hardware fault, the grid connection state machine cancels the closing command, disconnects the pre-charging auxiliary contactor and the main power contactor, and records the fault status.

[0033] It should be noted that one specific scheme for collecting bus synchronization quantities and performing pre-charging and main closing includes the grid-connected module receiving the local AC output formed by the intelligent agent module. Since the common AC bus may already be connected to other energy intelligent agents, if a newly added energy intelligent agent is directly closed when the voltage amplitude, frequency, and phase are not aligned, it is prone to generating inrush current and parallel circulating current. Therefore, this embodiment sets up a synchronization sampling, synchronization threshold comparison, pre-charging voltage equalization, and main closing confirmation process before the main power contactor closes, ensuring a clear electrical boundary when the AC high-voltage interface is connected to the common AC bus.

[0034] The synchronous sampling unit acquires the voltage signal from the common AC bus side and the output voltage signal from the local bidirectional power conversion unit, respectively, and extracts the voltage amplitude, frequency, and phase from the common AC bus side, as well as the voltage amplitude, frequency, and phase from the output voltage of the local bidirectional power conversion unit. The grid-connected state machine compares the voltage amplitude difference, frequency difference, and phase difference between the common AC bus side and the output voltage of the local bidirectional power conversion unit with the verification thresholds.

[0035] The voltage amplitude difference is determined by the voltage amplitude on the common AC bus side and the voltage amplitude on the second bus side. The absolute value is obtained by subtracting the output voltage amplitude of the bidirectional power conversion unit of the energy intelligent agent; the frequency difference is obtained by the frequency of the common AC bus and the frequency of the first... The absolute value is obtained by subtracting the output frequency of the bidirectional power conversion unit of the energy intelligent agent; the phase difference is obtained by the phase difference between the common AC bus side and the first... The absolute value is obtained by subtracting the phase values ​​from the output phase of the bidirectional power conversion unit of the energy intelligent agent. This indicates the energy agent's ID index.

[0036] The verification thresholds used by the grid-connected state machine are: voltage amplitude difference not exceeding 2% of the rated voltage, frequency difference not exceeding 0.1Hz, and phase difference not exceeding 5°. When the voltage amplitude difference, frequency difference, and phase difference simultaneously meet the verification thresholds and are continuously and stably maintained for 500ms, the grid-connected state machine outputs a pre-charge closing command. The 500ms holding time is used to filter out instantaneous sampling fluctuations and avoid erroneous closing when the synchronization quantity meets the threshold for a short time but is not stable.

[0037] After receiving the pre-charge closing command, the pre-charge auxiliary contactor closes and maintains voltage equalization for 300ms. After the 300ms voltage equalization is completed, the main power contactor closes. After the main power contactor contacts are in position, the pre-charge auxiliary contactor opens after a 100ms delay, allowing the main power circuit to stably receive the grid-connected current before exiting the pre-charge path. If any level contactor fails to provide a contact position signal, the grid-connected state machine re-outputs the corresponding closing command every 200ms. Specifically, if the grid-connected state machine does not receive the corresponding contact position signal within 200ms after outputting the pre-charge auxiliary contactor closing command or the main power contactor closing command, a closing timeout is recorded, and the number of retries for the same grid-connected process does not exceed 3.

[0038] When any of the following conditions are met: synchronization index exceeding limits, closing timeout, multiple closing failures, or AC / DC hardware fault, the grid-connected state machine cancels the closing command, disconnects the pre-charging auxiliary contactor and the main power contactor, and records the fault status. The grid-connected module outputs the main power contactor contact status, the pre-charging auxiliary contactor contact status, and the fault status. When the main power contactor contact feedback is in place and no fault status is recorded, the energy agent enters a state where it can connect to the common AC bus; when the contactor is disconnected or a fault status is recorded, the energy agent does not participate in the network as an online unit on the common AC bus.

[0039] The networking module is used to connect multiple energy intelligent agents to a common AC bus via AC high-voltage interfaces and form a cluster electrical coupling.

[0040] Furthermore, multiple energy intelligent agents are connected to the public AC bus via AC high-voltage interfaces, forming an energy intelligent agent cluster. Each energy intelligent agent operates collaboratively by combining the electrical coupling of the public AC bus, distributed intelligent rules, and power line communication information interaction. Before a new energy intelligent agent is connected to the public AC bus, synchronous verification and closing are performed. When an energy intelligent agent that has been connected is disconnected from the public AC bus, it is disconnected from the grid through the corresponding AC high-voltage interface.

[0041] It should be noted that one scheme for connecting multiple energy intelligent agents to a common AC bus via AC high-voltage interfaces specifically includes a networking module using the common AC bus as the high-voltage interconnection medium for multiple energy intelligent agents. Each energy intelligent agent is connected to the common AC bus via a corresponding AC high-voltage interface, forming an energy intelligent agent cluster. Since each energy intelligent agent already has a single-unit AC output structure in the intelligent agent module, and synchronization verification and closing control are completed in the grid connection module, the networking module does not need to set up a centralized power conversion unit, nor does it need to set up a central controller to uniformly complete the grid connection.

[0042] The networking module receives the main power contactor contact status and fault status output by the grid-connected module and determines the energy intelligent entities actually connected to the common AC bus. Energy intelligent entities whose main power contactor contacts are in place but whose fault status is not recorded are marked as pending confirmation energy intelligent entities; energy intelligent entities whose main power contactors are not in place or whose fault status is recorded do not participate in the online operation of the cluster. For pending confirmation energy intelligent entities, the networking module further receives the power disturbance identifier identification result output by the communication module and jointly confirms the online status of the energy intelligent entity based on the main power contactor contact feedback and the power disturbance identifier identification result. When the main power contactor contacts are in place and the power disturbance identifier corresponding to the energy intelligent entity is identified in at least one of the three consecutive 200ms communication cycles, the networking module adds the energy intelligent entity to the online member set; when the main power contactor contacts are in place but the power disturbance identifier corresponding to the energy intelligent entity is not identified in the three consecutive 200ms communication cycles, the networking module keeps the energy intelligent entity in the pending confirmation state. In the pending confirmation state, this energy agent does not participate in the target power allocation calculation within the online member set. The set of IDs of the online energy agents forms the online member set, which is then entered into the communication module and the sampling module.

[0043] Each energy agent operates collaboratively by combining common AC bus electrical coupling, distributed intelligent rules, and power line communication information interaction. Common AC bus electrical coupling means that after each online energy agent is connected to the same common AC bus, it can collect AC side operating parameters such as grid connection point frequency, grid connection point voltage, output power, and parallel branch circulating current. Distributed intelligent rules mean that the edge AI module of each energy agent generates local control parameters based on its collected data and parameter rules. Power line communication information interaction means that the communication module multiplexes the common AC bus superimposed power line carrier signal and generates power disturbance indicators through power adjustment.

[0044] Before a new energy intelligent agent is connected to the common AC bus, it first enters the grid connection module for synchronization verification and closing. After the grid connection module confirms that the main power contactor contacts are in place and no fault status is recorded, the networking module marks the new energy intelligent agent as an energy intelligent agent to be confirmed. After the communication module identifies the power disturbance identifier corresponding to the energy intelligent agent in at least one of three consecutive 200ms communication cycles, the networking module adds the new energy intelligent agent to the online member set. When an energy intelligent agent that has been connected leaves the common AC bus, it performs off-grid exit through the corresponding AC high-voltage interface. After the main power contactor is disconnected, the networking module removes the corresponding energy intelligent agent from the online member set.

[0045] The communication module is used to multiplex the common AC bus superimposed power line carrier signal and generate power disturbance indicators through power regulation.

[0046] Furthermore, the communication module includes a power line carrier communication unit and a power signature communication unit; it is coupled to a common AC bus and superimposed with a 2kHz to 5kHz narrowband frequency shift keying carrier signal; the narrowband frequency shift keying carrier signal is injected intermittently with short pulses, and the instantaneous injection peak value is 1% to 3% of the rated voltage; the power line carrier communication unit sends communication frames according to a 200ms communication cycle; the communication frame includes an address field, an instruction field, and a sixteen-bit cyclic redundancy check field; the power signature communication unit adjusts active power and reactive power through an energy agent to generate corresponding power disturbance identifiers.

[0047] It should be noted that one specific scheme for generating power disturbance indicators through power regulation includes the communication module utilizing a common AC bus to complete low-speed information exchange. Traditional explicit communication buses require separate communication cables to be laid in addition to the high-voltage connection, while this embodiment has already completed the high-voltage connection of multiple energy agents through a common AC bus. Therefore, the common AC bus is further reused to superimpose power line carrier signals, and power disturbance indicators are generated through the power regulation of the energy agents themselves.

[0048] The communication module includes a power line carrier communication unit and a power signature communication unit. It is coupled to a common AC bus, and a 2kHz to 5kHz narrowband frequency shift keying (FSK) carrier signal is superimposed on the common AC bus. This carrier frequency band avoids the control loop band and protection operation band of the bidirectional power conversion unit, preventing interference with main power control and protection operations. Preferably, the two carrier frequencies of the narrowband FSK carrier signal are 3kHz and 4kHz, respectively. The narrowband FSK carrier signal uses short-pulse intermittent injection, with an instantaneous injection peak value of 1% to 3% of the rated voltage. The injection range is lower than the main voltage amplitude, which meets the needs of low-speed communication identification while reducing the impact on the bus power quality; the instantaneous injection peak value is preferably 2% of the rated voltage.

[0049] The power line carrier communication unit transmits communication frames according to a 200ms communication period. Each communication frame contains an address field, a command field, and a 16-bit cyclic redundancy check (CRC) field. The address field indicates the sending or receiving object, the command field indicates the cluster scheduling command, and the 16-bit CRC field is used to verify the address and command fields. The receiving side determines the integrity of the communication frame based on the 16-bit CRC field; if the verification fails, the communication frame is discarded, and the current control parameters of the receiving unit remain unchanged.

[0050] The power signature communication unit adjusts active and reactive power through an energy agent to generate corresponding power disturbance identifiers. These identifiers are used to express the interactive information of a single machine's operating status. This embodiment uses three states—positive disturbance, negative disturbance, and no disturbance—to represent these states, corresponding to power disturbance identifier values ​​of 1, -1, and 0, respectively.

[0051] The power signature communication unit adjusts the active and reactive power of the device without changing the main power regulation target, causing corresponding power fluctuations on the common AC bus. Specifically, the active power disturbance does not exceed 1% to 3% of the rated active power output of the energy agent, preferably 2% of the rated active power output; the reactive power disturbance does not exceed 1% to 3% of the rated reactive power output of the energy agent, preferably 2% of the rated reactive power output. When the power disturbance flag value is 1, the power signature communication unit superimposes the positive active power disturbance and the positive reactive power disturbance within a 200ms communication cycle; when the power disturbance flag value is -1, the power signature communication unit superimposes the reverse active power disturbance and the reverse reactive power disturbance within a 200ms communication cycle; when the power disturbance flag value is 0, the power signature communication unit does not superimpose the power disturbance within the corresponding communication cycle.

[0052] After the power disturbances are superimposed, the active power does not exceed the upper limit of the active power output of the bidirectional power conversion unit, and the reactive power does not exceed the reactive power capacity boundary of the bidirectional power conversion unit. If the power disturbance calculated by the rated ratio exceeds the current remaining output margin, then the current remaining output margin is used as the actual power disturbance. The current remaining output margin includes the active power remaining output margin and the reactive power remaining output margin; the active power remaining output margin is the difference between the upper limit of the active power output of the bidirectional power conversion unit and the current active power, and the reactive power remaining output margin is the difference between the reactive power capacity boundary of the bidirectional power conversion unit and the absolute value of the current reactive power. If the active power disturbance exceeds the active power remaining output margin, then the active power remaining output margin is used as the actual active power disturbance; if the reactive power disturbance exceeds the reactive power remaining output margin, then the reactive power remaining output margin is used as the actual reactive power disturbance. After one communication cycle, the energy agent restores to the original active and reactive power regulation targets. Each energy agent collects fluctuations from the common AC bus and decodes them according to the direction of the power fluctuations to obtain the corresponding power disturbance identifier.

[0053] The communication module outputs the instruction field parsed by the power line carrier communication unit and the power disturbance identifier generated by the power signature communication unit. The instruction field is sent to the control module as a cluster scheduling instruction, and the power disturbance identifier is sent to the sampling module and the control module as interactive information on the single-machine operating status. The communication module also performs a consistency check between the instruction field parsed by the power line carrier communication unit and the power disturbance identifier identified by the power signature communication unit. When the instruction field indicates that a certain energy agent is in an online participation state, but the power disturbance identifier is not identified within the corresponding communication period, the communication module outputs a pending confirmation status to the networking module. In the pending confirmation status, the energy agent does not participate in the new target allocation power calculation and retains the currently effective control parameters. When the power line carrier communication frame verification passes and the power disturbance identifier is identified within the corresponding communication period, the communication module outputs a confirmation status to the networking module.

[0054] It should also be noted that the communication module reuses the common AC bus and superimposes the power line carrier signal, and generates power disturbance identifiers through power regulation, completing the interaction between cluster commands and individual unit status without adding dedicated communication cables. Traditional energy storage clusters usually separate the high-voltage connection and the communication connection. In addition to connecting to the electrical circuit, new energy storage units also need to connect to the communication bus and complete address and control configuration, resulting in high complexity in on-site deployment. This invention uses the common AC bus as both a power transmission channel and a low-speed information interaction medium. The power line carrier communication unit superimposes a 2kHz to 5kHz narrowband frequency shift keying carrier signal on the bus, preferably using 3kHz and 4kHz as the two carrier frequencies, and injects short pulses intermittently at an instantaneous injection peak of 1% to 3%, preferably 2%, of the rated voltage. The communication frame includes an address field, a command field, and a sixteen-bit cyclic redundancy check field. If the check fails, the local control parameters are not changed. The power signature communication unit generates three types of power disturbance identifiers, 1, -1, and 0, through constrained disturbances of active and reactive power, so that the operating status of individual units can be perceived by other energy agents in the form of power fluctuations. Carrier communication handles command field exchange, while power disturbance flags handle status exchange, reducing the impact of single communication method failure or wiring constraints on cluster collaboration.

[0055] The sampling module is used to collect the running status and trigger parameter optimization according to the abnormal threshold.

[0056] Furthermore, the sampling module includes a sampling unit, an anomaly detection unit, and a trigger locking unit. The sampling unit collects grid-connected frequency, grid-connected voltage, state of charge (SOC), output power, parallel branch circulating current, and bus load fluctuation according to a sampling period. The anomaly detection unit compares the grid-connected frequency with the rated frequency to obtain the frequency deviation, compares the grid-connected voltage with the rated voltage to obtain the voltage deviation, and subtracts the minimum SOC from the maximum SOC within the cluster to obtain the SOC range. The anomaly detection unit performs threshold judgments on the frequency deviation, voltage deviation, SOC, SOC range, parallel branch circulating current, bus load fluctuation, and multi-unit power sharing error. The threshold judgment condition includes an absolute frequency deviation greater than 0.2 Hz. The following conditions must be met: absolute voltage deviation greater than 5% of rated voltage; state of charge (SOC) below 20%; SOC above 85%; SOC range greater than 15%; parallel branch circulating current greater than 3% of rated phase current; bus load step fluctuation amplitude greater than 10% of rated output; and multi-machine power sharing error exceeding the standard. The trigger locking unit outputs a parameter optimization trigger signal after the same threshold judgment condition is met for three consecutive sampling cycles. After outputting the parameter optimization trigger signal, the trigger locking unit sets a minimum hold time of 2 seconds, locking the new parameter optimization trigger signal within this minimum hold time. When the minimum hold time ends and all abnormal indicators fail to meet the threshold judgment condition for five consecutive sampling cycles, the trigger locking unit outputs parameter rollback permission.

[0057] It should be noted that one scheme for collecting operational status and triggering parameter optimization based on anomaly thresholds specifically includes a sampling module receiving the online member set output by the networking module and collecting operational status data for the energy agents within the online member set. The sampling module includes a sampling unit, an anomaly detection unit, and a trigger locking unit. The sampling unit collects grid connection point frequency, grid connection point voltage, state of charge, output power, parallel branch circulating current, and bus load fluctuations according to a sampling period. The sampling period is set to 10ms to balance the grid disturbance response speed and the computing power load of the edge AI module.

[0058] The anomaly detection unit compares the grid connection point frequency with the rated frequency to obtain the frequency deviation, which is the first frequency deviation. The difference between the grid connection point frequency and the rated frequency is obtained by the energy intelligent agent; the anomaly detection unit compares the grid connection point voltage with the rated voltage to obtain the voltage deviation, and the voltage deviation is the first... The difference between the grid connection point voltage and the rated voltage of the energy intelligent body is used to obtain the state of charge range. The anomaly detection unit subtracts the minimum state of charge from the maximum state of charge within the online member set to obtain the state of charge range. The online member set is output by the networking module and only includes energy intelligent bodies that have been connected to the common AC bus and have not recorded a fault state. Therefore, the state of charge range reflects the difference in state of charge among the energy intelligent bodies actually participating in parallel operation.

[0059] The anomaly detection unit performs threshold judgments on frequency deviation, voltage deviation, state of charge (SCC), SCC range, parallel branch circulating current, bus load fluctuation, and multi-machine power sharing error. Threshold judgment conditions include: absolute frequency deviation greater than 0.2Hz, absolute voltage deviation greater than 5% of rated voltage, SCC below 20%, SCC above 85%, SCC range greater than 15%, parallel branch circulating current greater than 3% of rated phase current, bus load step fluctuation amplitude greater than 10% of rated output, and multi-machine power sharing error exceeding the limit. Bus load fluctuation is the absolute value of the difference between the bus load power in the current sampling period and the bus load power in the previous sampling period. Rated output is the sum of the rated active power of each energy agent in the online member set.

[0060] The multi-machine power sharing error is the deviation between the actual output power of each energy agent in the online member set and the target allocated power. Specifically, the sampling module adds the actual output active power of each energy agent in the online member set to obtain the current total active power output of the cluster; then it sums the actual output active power of each energy agent in the online member set to obtain the current total active power output of the cluster; and then it sums the actual output active power of each energy agent in the online member set to obtain the current total active power output of the cluster. Dividing the rated active power output of the energy agent by the sum of the rated active power outputs of all energy agents in the online member set yields the result of the first division. The target allocation ratio of the energy agent; finally, multiply the current total active power output of the cluster by the first... The target allocation ratio of the energy intelligent agent is obtained. The target allocation of active power by the energy intelligence agent.

[0061] No. The power distribution deviation of the energy agent is the first The absolute value of the difference between the actual output active power of the energy agent and the corresponding target allocated active power, divided by the first... The normalized deviation of the rated active power output of each energy agent is used. The anomaly detection unit takes the maximum normalized deviation of each energy agent in the online member set as the multi-machine power sharing error. The preset power sharing error limit is preferably 0.05. When the multi-machine power sharing error is greater than 0.05, it is determined that the multi-machine power sharing error exceeds the limit; 0.05 means that the deviation of the actual output active power from the target allocated active power exceeds 5% of the corresponding rated active power output.

[0062] The threshold selection corresponds to different controlled objects. Frequency deviation of 0.2Hz and voltage deviation of 5% are used to identify situations where the electrical quantity of the common AC bus deviates from the normal operating range; state of charge below 20% or above 85% is used to identify excessively low or high state of charge of a single energy agent; state of charge range of 15% is used to identify unbalanced state of charge within the cluster; parallel branch circulating current of 3% of rated phase current is used to identify parallel circulating currents that require virtual impedance adjustment; bus load step fluctuation of 10% of rated output is used to identify load disturbances that require power redistribution; and multi-machine power sharing error exceeding the standard is used to identify situations where the actual output of online energy agents deviates from the target allocated power.

[0063] The trigger-lock unit outputs a parameter optimization trigger signal after the same threshold judgment condition is met for three consecutive sampling cycles. Three consecutive sampling cycles are used to prevent false triggering caused by single-point sampling noise; when the sampling cycle is 10ms, three consecutive sampling cycles correspond to a 30ms anti-jitter confirmation time. After outputting the parameter optimization trigger signal, the trigger-lock unit sets a minimum hold time of 2 seconds. During this minimum hold time, it locks the new parameter optimization trigger signal, ensuring the control parameters remain in place for sufficient execution time after adjustment. After the minimum hold time ends, if all abnormal indicators fail to meet the threshold judgment condition for five consecutive sampling cycles, the trigger-lock unit outputs a parameter rollback permission.

[0064] The control module is used to asynchronously call the multi-agent reinforcement learning agent and, in conjunction with the latent field rule engine, correct the control parameters of the bidirectional power conversion unit in the agent module.

[0065] Furthermore, refer to Figure 2The control module includes a latent field rule engine, a multi-agent reinforcement learning agent, a parameter pruning unit, and a register write-back unit. The latent field rule engine stores grid connection specifications, device limits, and parameter physical boundaries, and establishes the legal value ranges for droop correction coefficients, virtual impedance parameters, and power signature parameters. The multi-agent reinforcement learning agent receives the output of the sampling module, asynchronously executes inference after the parameter optimization trigger signal is output, and generates initial control parameters. When the inference time does not exceed the sampling period, the parameter pruning unit applies amplitude and rate-of-change limits to the initial control parameters. When the inference time exceeds the sampling period, the parameter pruning unit calls the upper... The first version of stable control parameters is implemented and safety boundary constraints are enforced. The register write-back unit writes the processed droop correction coefficient, virtual impedance parameter, and power control parameters into the register to be activated of the bidirectional power conversion unit, and switches the control parameters in the register to be activated to the currently activated control parameters at the sampling period boundary through parameter version verification. After the control parameters take effect, the control module performs post-activation verification based on the abnormal indicators that triggered this round of parameter optimization. If the post-activation verification fails, the control module calls the previous version of stable control parameters for version rollback. If the post-activation verification passes, the control module records the control parameters of this round as stable control parameters.

[0066] It should be noted that one scheme for correcting the control parameters of the bidirectional power conversion unit in the intelligent agent module using a latent field rule engine specifically includes the following: the control module receives frequency deviation, voltage deviation, state of charge, output power, parallel branch circulating current, multi-machine power sharing error, parameter optimization trigger signal, and parameter rollback permission from the sampling module. The control module includes a latent field rule engine, a multi-agent reinforcement learning agent, a parameter pruning unit, and a register write-back unit. Compared to centralized energy storage control where a central controller uniformly calculates and issues control commands, this embodiment arranges the control parameter generation process in the edge artificial intelligence module of each energy agent, enabling each energy agent to complete the correction of its own bidirectional power conversion unit control parameters based on its own data collection and cluster operation goals.

[0067] The latent field rule engine stores grid connection specifications, device limits, and physical boundaries of parameters, and establishes the legal value ranges for droop correction coefficients, virtual impedance parameters, and power signature parameters. Specifically, the legal value range of the droop correction coefficient is 0.9 to 1.1, where 1.0 represents the normal droop state, 0.9 represents a 10% reduction in the droop coefficient compared to the normal value, and 1.1 represents a 10% increase in the droop coefficient compared to the normal value. The droop correction coefficient is used to correct the droop coefficient of the bidirectional power conversion unit. The virtual impedance parameters include direct-axis virtual resistance and quadrature-axis virtual inductance. The legal value range of the direct-axis virtual resistance is 0.02Ω to 0.3Ω, preferably 0.15Ω, and the legal value range of the quadrature-axis virtual inductance is 0.05H to 0.8H, preferably 0.4H. The power signature parameters include active power disturbance ratio, reactive power disturbance ratio, and power disturbance flag value. The legal value range of the active power disturbance ratio and reactive power disturbance ratio is 1% to 3% of the rated power, preferably 2% of the rated power. The power disturbance flag value is 1, -1, or 0. The parameter trimming unit performs amplitude limiting and rate of change limitation on the initial control parameters according to the legal value range output by the latent field rule engine. The amplitude limiting process is used to ensure that the control parameters are within the legal value range, and the rate of change limit is used to avoid the power conversion unit control fluctuation caused by excessive adjustment amplitude in a single operation.

[0068] The multi-agent reinforcement learning agent receives the frequency deviation, voltage deviation, state of charge, output power, parallel branch circulating current, and multi-machine power sharing error output from the sampling module, and normalizes the input quantities to the 0-1 range. The multi-agent reinforcement learning agent generates initial control parameters according to the joint reward function, which is expressed as follows:

[0069]

[0070] in, This represents the total value of the joint reward function. This represents the frequency deviation benefit term. This represents the voltage deviation benefit term. This represents the state-of-charge equilibrium benefit term. This indicates the parallel circulation penalty term. This represents the power sharing error penalty term. This indicates a penalty for exceeding parameter limits. This indicates a penalty for changes in actions.

[0071] The frequency deviation benefit term is calculated from the absolute value of the frequency deviation, specifically: divide the absolute value of the frequency deviation by 0.2Hz to obtain the normalized value of the frequency deviation; when the normalized value of the frequency deviation is greater than 1, it is counted as 1; then subtract the normalized value of the frequency deviation from 1 to obtain the frequency deviation benefit term. When the frequency deviation is 0, the frequency deviation benefit term is 1; when the absolute value of the frequency deviation reaches or exceeds 0.2Hz, the frequency deviation benefit term is 0.

[0072] The voltage deviation benefit is calculated from the absolute value of the voltage deviation. Specifically, the absolute value of the voltage deviation is divided by 5% of the rated voltage to obtain a normalized voltage deviation value. When the normalized voltage deviation value is greater than 1, it is counted as 1. Then, the normalized voltage deviation value is subtracted from 1 to obtain the voltage deviation benefit. When the voltage deviation is 0, the voltage deviation benefit is 1; when the absolute value of the voltage deviation reaches or exceeds 5% of the rated voltage, the voltage deviation benefit is 0.

[0073] The state-of-charge (SOC) balance benefit is derived from the SOC range. Specifically, the difference between the maximum and minimum SOCs within the online member set is divided by 15% to obtain a normalized SOC range value. When the normalized SOC range value is greater than 1, it is counted as 1. Then, 1 is subtracted from the normalized SOC range value to obtain the SOC balance benefit. When all energy agents within the online member set have the same SOC, the SOC balance benefit is 1; when the SOC range exceeds 15%, the SOC balance benefit is 0.

[0074] The parallel circulating current penalty term is calculated from the parallel branch circulating current. Specifically, the parallel branch circulating current is divided by 3% of the rated phase current to obtain the normalized value of the parallel circulating current. When the normalized value of the parallel circulating current is greater than 1, it is counted as 1, and this normalized value is used as the parallel circulating current penalty term. When the parallel branch circulating current is 0, the parallel circulating current penalty term is 0; when the parallel branch circulating current reaches more than 3% of the rated phase current, the parallel circulating current penalty term is 1.

[0075] The power sharing error penalty term is calculated from the multi-machine power sharing error. The multi-machine power sharing error is the maximum normalized deviation between the actual output active power of each energy agent in the online member set and the target allocated active power. The multi-machine power sharing error is divided by the preset power sharing error limit of 0.05 to obtain the normalized power sharing error value. When the normalized power sharing error value is greater than 1, it is counted as 1, and this normalized power sharing error value is used as the power sharing error penalty term. When the multi-machine power sharing error is 0, the power sharing error penalty term is 0; when the multi-machine power sharing error reaches 0.05 or higher, the power sharing error penalty term is 1.

[0076] The parameter over-limit penalty is calculated based on the degree to which the initial control parameter exceeds the legal value range. For each type of initial control parameter, if the parameter value is within the corresponding legal value range, the over-limit ratio for that type of parameter is 0; if the parameter value is below the lower limit of the legal value range, the lower limit value is subtracted from the parameter value, and then divided by the span of the corresponding legal value range to obtain the over-limit ratio for that type of parameter; if the parameter value is above the upper limit of the legal value range, the upper limit value is subtracted from the parameter value, and then divided by the span of the corresponding legal value range to obtain the over-limit ratio for that type of parameter. The span of the legal value range is the difference between the upper limit and the lower limit of the legal value for the corresponding control parameter. The parameter over-limit penalty takes the maximum value among the over-limit ratios of each type of control parameter; when the maximum over-limit ratio is greater than 1, the parameter over-limit penalty is counted as 1.

[0077] The action change penalty term is calculated from the change magnitude of the initial control parameter relative to the control parameter written to the bidirectional power conversion unit register in the previous round. For each type of control parameter, the absolute value of the difference between the initial control parameter in this round and the control parameter in the previous round is divided by the span of the corresponding legal value range to obtain the action change ratio of that type of control parameter; the action change penalty term takes the maximum value among the action change ratios of each type of control parameter; when the maximum action change ratio is greater than 1, the action change penalty term is counted as 1.

[0078] To adapt the execution process of the multi-agent reinforcement learning agent to the real-time control requirements of the power conversion unit, this embodiment asynchronously calls the multi-agent reinforcement learning agent to perform inference after the parameter optimization trigger signal is output. The multi-agent reinforcement learning agent first undergoes offline pre-training based on the cluster's historical operating conditions. The inputs to the offline pre-training include frequency deviation, voltage deviation, state of charge, output power, parallel branch circulating current, and multi-machine power sharing error. The outputs include droop coefficient, virtual impedance parameters, and power control parameters. During online execution, the multi-agent reinforcement learning agent performs asynchronous inference on the normalized input quantities output by the sampling module based on the offline pre-training results and outputs initial control parameters.

[0079] When the inference time of the multi-agent reinforcement learning agent does not exceed the sampling period, the parameter pruning unit receives the initial control parameters for this round and applies amplitude and rate-of-change constraints to the initial control parameters according to the legal value range established by the latent field rule engine. When the inference time of the multi-agent reinforcement learning agent exceeds the sampling period, the control module does not wait for the inference result of this round. The parameter pruning unit calls the stable control parameters of the previous version and applies safety boundary constraints according to the legal value range established by the latent field rule engine. The stable control parameters of the previous version are the control parameter group that the sampling module did not output a parameter optimization trigger signal after the most recent write to the bidirectional power conversion unit register. When the control module receives the parameter rollback permission, the control module releases the hold state after the parameter optimization trigger and allows the control parameters of the next round to be readjusted according to the current output of the sampling module.

[0080] The control module adjusts the droop coefficient according to the state of charge (SBC). When the SBC is greater than 85%, the control module lowers the droop coefficient; when the SBC is between 30% and 85%, the control module maintains the normal droop coefficient; when the SBC is less than 30%, the control module raises the droop coefficient. In this embodiment, the bidirectional power conversion unit adopts a control definition where a smaller droop coefficient corresponds to a higher output adjustment weight. Therefore, energy agents with a high SBC take on a higher output weight by lowering the droop coefficient, while energy agents with a low SBC reduce their output weight by raising the droop coefficient. The maximum adjustment range of the droop coefficient for a single energy agent within a single scheduling cycle does not exceed 10% of the rated range, preferably 5% of the rated range, and is subject to the upper and lower limits of the legal value range.

[0081] The control module corrects the virtual impedance parameters based on the circulating current of the parallel branches, voltage deviation, and power averaging error. Every 10ms, the control module corrects the virtual impedance parameters based on the circulating current of the parallel branches, voltage deviation, and power averaging error. The adjustment range of a single adjustment does not exceed 15% of the corresponding parameter range, and the preferred adjustment range is 8% of the corresponding parameter range. Furthermore, the parameters are subject to the upper and lower limits of the legal value range throughout the entire process.

[0082] The register write-back unit writes the droop correction factor, virtual impedance parameters, and power control parameters, processed by the parameter trimming unit, into the bidirectional power conversion unit register. The power control parameters include the active power output limit, reactive power response rate, and power ramp-up slope.

[0083] During the register write-back process, the register write-back unit first writes the control parameters processed by the parameter trimming unit into the register to be activated of the bidirectional power conversion unit, and generates a parameter version number for the control parameters in this round. The bidirectional power conversion unit reads the control parameters in the register to be activated at the sampling period boundary and performs register verification. Register verification includes parameter version number consistency verification and write value readback verification. When both parameter version number consistency verification and write value readback verification pass, the bidirectional power conversion unit switches the control parameters in the register to be activated to the currently activated control parameters; when either parameter version number consistency verification or write value readback verification fails, the bidirectional power conversion unit keeps the previous version of the activated control parameters unchanged and returns a write-back failure status to the control module.

[0084] After executing the currently effective control parameters, the bidirectional power conversion unit refreshes the local droop control, virtual synchronous machine control, virtual impedance control, and power signature parameter control logic, and returns the updated voltage, frequency, power, state of charge, and parallel branch circulating current to the sampling module.

[0085] After the control parameters take effect, the sampling module continues to collect abnormal indicators that trigger this round of parameter optimization. If, after the control parameters take effect, the abnormal indicator that triggers this round of parameter optimization still meets the corresponding threshold judgment condition within 5 consecutive sampling periods, the control module determines that the control parameters of this round have failed the post-implementation verification and rolls back to the stable control parameters of the previous version. If the abnormal indicator that triggers this round of parameter optimization does not meet the corresponding threshold judgment condition for 5 consecutive sampling periods, the control module determines that the control parameters of this round have passed the post-implementation verification and records the control parameters of this round as stable control parameters.

[0086] In this embodiment, the post-implementation verification of control parameters is part of the current round of parameter safety rollback process and not a trigger for new parameter optimization. Therefore, even if the trigger locking unit is within the minimum 2-second holding time during the post-implementation verification process, the control module still allows version rollback for control parameters that fail the post-implementation verification. After the version rollback is completed, the control module rewrites the stable control parameters of the previous version into the pending-implementation register of the bidirectional power conversion unit and performs write-back confirmation according to the parameter version number consistency check and the write value readback check.

[0087] The sampling module proceeds to the next round of sampling and anomaly detection based on the data returned by the bidirectional power conversion unit, and the control module proceeds to the next round of parameter correction based on the new sampling results.

[0088] It should also be noted that after parameter optimization is triggered, the control module asynchronously calls the multi-agent reinforcement learning agent and completes safety boundary constraints and register write-back through the latent field rule engine, enabling intelligent optimization to enter the real-time control link of the high-voltage AC bus. Ordinary scheduling methods typically focus on power allocation results, but power electronic equipment is also subject to parameter boundary constraints such as droop coefficient, virtual impedance, power disturbance ratio, and power ramp-up. Inference time may also be mismatched with the sampling period. This invention uses the latent field rule engine to store grid connection specifications, equipment limits, and parameter physical boundaries, establishing a legal value range for the droop correction coefficient. The multi-agent reinforcement learning agent receives frequency deviation, voltage deviation, state of charge, output power, parallel branch circulating current, and multi-machine power sharing error, and generates initial control parameters through a joint reward function. If the inference time does not exceed the sampling period, the parameter pruning unit limits the amplitude and restricts the rate of change of the initial control parameters; if the inference timeout occurs, the previous version of stable control parameters is called and safety boundary constraints are executed. The register write-back unit writes the processed droop correction coefficient, virtual impedance parameter, and power control parameter into the bidirectional power conversion unit register, so that the agent control is transformed from "calculation result" into register parameters that can be executed by the bidirectional power conversion unit.

[0089] Reference Figure 3 One embodiment of the present invention provides a method for a self-organizing energy storage intelligent agent cluster based on an AC high-voltage interface, comprising:

[0090] S1: Establish a stand-alone AC output structure;

[0091] S2: Collect bus synchronization data and perform pre-charging and main closing;

[0092] S3: Connect multiple energy intelligent agents to the common AC bus via AC high-voltage interfaces and form a cluster electrical coupling;

[0093] S4: Multiplex the common AC bus superimposed with the power line carrier signal, and generate a power disturbance indicator through power regulation;

[0094] S5: Collect operating status and trigger parameter optimization according to abnormal thresholds;

[0095] S6: Asynchronously call the multi-agent reinforcement learning agent and combine it with the latent field rule engine to correct the control parameters of the bidirectional power conversion unit in the agent module.

[0096] Reference Figure 4 This embodiment also provides a computer device applicable to a self-organizing energy storage intelligent agent cluster system based on an AC high-voltage interface, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the self-organizing energy storage intelligent agent cluster method based on an AC high-voltage interface as proposed in the above embodiment.

[0097] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0098] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for implementing a self-organizing energy storage intelligent agent cluster based on an AC high-voltage interface as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface, characterized in that, include: Intelligent agent module, grid connection module, network topology module, communication module, sampling module, control module; The intelligent agent module is used to establish a stand-alone AC output structure; The grid-connected module is used to collect bus synchronization data and perform pre-charging and main closing. The networking module is used to connect multiple energy intelligent agents to a common AC bus via AC high-voltage interfaces and form a cluster electrical coupling. The communication module is used to multiplex the common AC bus superimposed power line carrier signal and generate a power disturbance identifier through power adjustment; The sampling module is used to collect the operating status and trigger parameter optimization according to the abnormal threshold. The control module is used to asynchronously call the multi-agent reinforcement learning agent and, in conjunction with the latent field rule engine, correct the control parameters of the bidirectional power conversion unit in the agent module.

2. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 1, characterized in that: The intelligent agent module includes a battery pack, a battery management unit, a bidirectional power conversion unit, an edge artificial intelligence module, and an AC high-voltage interface; The DC output terminal of the battery pack is connected to the DC input terminal of the bidirectional power conversion unit, and the AC output terminal of the bidirectional power conversion unit is connected to the AC high-voltage interface. The battery management unit is electrically connected to the battery pack. The battery management unit collects battery status parameters and generates state of charge, alarm signals, and power current limiting commands. The state of charge, alarm signals, and power current limiting commands are transmitted to the edge AI module and bidirectional power conversion unit via control lines; The bidirectional power conversion unit converts the DC energy output from the battery pack into three-phase AC output according to the four-quadrant operation mode and the grid formation control mode; The AC high-voltage interface is used as the only main electrical interface for the intelligent agent module to the outside world. The AC high-voltage interface is equipped with an AC contactor, circuit breaker, synchronous sampling unit and grid connection protection unit. The edge AI module connects to the battery management unit and the bidirectional power conversion unit via a control line, and establishes a channel for the edge AI module to write control parameters to the bidirectional power conversion unit.

3. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 2, characterized in that: The process of acquiring bus synchronization signals and performing pre-charging and main closing includes a synchronization sampling unit acquiring the voltage signal on the common AC bus side and the voltage signal on the output side of the local bidirectional power conversion unit, and extracting the voltage amplitude, frequency, and phase. The grid-connected state machine compares the voltage amplitude difference, frequency difference, and phase difference between the common AC bus side and the output side of the local bidirectional power conversion unit with the verification thresholds respectively. When the voltage amplitude difference, frequency difference, and phase difference simultaneously meet the verification threshold and remain stable for 500ms, the grid-connected state machine outputs a pre-charge closing command. After the pre-charge auxiliary contactor closes, it maintains voltage equalization for 300ms before the main power contactor closes. After the main power contactor contacts are in position, the pre-charge auxiliary contactor disconnects after a 100ms delay. When any primary contactor fails to provide a contact position signal, the grid connection state machine re-outputs the corresponding closing command every 200ms. The number of retries for the same grid connection process shall not exceed 3. When any of the following conditions are met: synchronization index exceeds limit, closing timeout, multiple closing failures, or AC / DC hardware fault, the grid-connected state machine cancels the closing command, disconnects the pre-charging auxiliary contactor and the main power contactor, and records the fault status.

4. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 3, characterized in that: The process of connecting multiple energy intelligent agents to a common AC bus via AC high-voltage interfaces includes each energy intelligent agent connecting to the common AC bus via a corresponding AC high-voltage interface, forming an energy intelligent agent cluster. Each energy agent operates collaboratively by combining electrical coupling with the common AC bus, distributed intelligent rules, and power line communication information interaction. Before a new energy intelligent agent is connected to the public AC bus, it undergoes synchronous verification and closing. When an energy intelligent entity that has been connected to the public AC bus is disconnected, it shall perform the off-grid disconnection through the corresponding AC high-voltage interface.

5. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 4, characterized in that: The communication module includes a power line carrier communication unit and a power signature communication unit; The common AC bus is coupled and connected, and a 2kHz to 5kHz narrowband frequency shift keying carrier signal is superimposed on the common AC bus; The narrowband frequency shift keying carrier signal is injected intermittently with short pulses, and the instantaneous injection peak value is 1% to 3% of the rated voltage; The power line carrier communication unit sends communication frames according to a communication period of 200ms; The communication frame includes an address field, an instruction field, and a 16-bit cyclic redundancy check field. The power signature communication unit adjusts active and reactive power through the energy agent to generate corresponding power disturbance identifiers.

6. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 5, characterized in that: The sampling module includes a sampling unit, an anomaly detection unit, and a trigger locking unit; The sampling unit collects grid connection point frequency, grid connection point voltage, state of charge, output power, parallel branch circulating current, and bus load fluctuation according to the sampling period; The anomaly detection unit compares the grid connection point frequency with the rated frequency to obtain the frequency deviation, compares the grid connection point voltage with the rated voltage to obtain the voltage deviation, and subtracts the minimum state of charge from the maximum state of charge in the cluster to obtain the state of charge range. The anomaly detection unit performs threshold judgments on frequency deviation, voltage deviation, state of charge, state of charge range, parallel branch circulating current, bus load fluctuation, and multi-machine power sharing error. The threshold judgment conditions include: absolute value of frequency deviation greater than 0.2Hz, absolute value of voltage deviation greater than 5% of rated voltage, state of charge less than 20%, state of charge greater than 85%, state of charge range greater than 15%, circulating current of parallel branches greater than 3% of rated phase current, bus load step fluctuation amplitude greater than 10% of rated output, and multi-machine power sharing error exceeding the standard. The trigger locking unit outputs a parameter optimization trigger signal after the same threshold judgment condition is met for 3 consecutive sampling cycles. The trigger locking unit sets a minimum hold time of 2 seconds after the parameter optimization trigger signal is output, and locks the new parameter optimization trigger signal within the minimum hold time. When the minimum hold time ends and all abnormal indicators fail to meet the threshold judgment condition for 5 consecutive sampling periods, the lock unit output parameter rollback permission is triggered.

7. The self-organizing energy storage intelligent agent cluster system based on AC high-voltage interface as described in claim 6, characterized in that: The control module includes a latent field rule engine, a multi-agent reinforcement learning agent, a parameter pruning unit, and a register write-back unit. The latent field rule engine stores grid connection specifications, device limits, and physical boundaries of parameters, and establishes the legal value ranges for droop correction coefficients, virtual impedance parameters, and power signature parameters. The multi-agent reinforcement learning agent receives the output of the sampling module, performs inference asynchronously after the parameter optimization trigger signal is output, and generates initial control parameters; When the inference time does not exceed the sampling period, the parameter trimming unit applies amplitude and rate of change limits to the initial control parameters; When the inference time exceeds the sampling period, the parameter pruning unit calls the stable control parameters of the previous version and executes the safety boundary constraints. The register write-back unit writes the processed droop correction coefficient, virtual impedance parameter, and power control parameter into the register to be activated of the bidirectional power conversion unit, and switches the control parameter in the register to be activated to the currently active control parameter at the sampling period boundary through parameter version verification. After the control parameters take effect, the control module performs post-implementation verification based on the abnormal indicators that triggered this round of parameter optimization; If the verification fails after the effect is activated, the control module calls the stable control parameters of the previous version to roll back the version. Once the verification is successful after the system takes effect, the control module will record the control parameters for this round as stable control parameters.

8. A method for a self-organizing energy storage intelligent agent cluster based on an AC high-voltage interface, employing the self-organizing energy storage intelligent agent cluster system based on an AC high-voltage interface as described in any one of claims 1 to 7, characterized in that: This includes establishing a stand-alone AC output structure; Collect bus synchronization data and perform pre-charging and main closing; Multiple energy intelligent agents are connected to a common AC bus via AC high-voltage interfaces and form a cluster electrical coupling. The common AC bus superimposed power line carrier signal is reused, and a power disturbance indicator is generated through power regulation; Collect operational status and trigger parameter optimization based on abnormal thresholds; Asynchronous invocation of a multi-agent reinforcement learning agent, combined with a latent field rule engine, corrects the control parameters of the bidirectional power conversion unit in the agent module.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the self-organizing energy storage intelligent agent cluster method based on AC high-voltage interface as described in any one of claims 1 to 8.

Images