Energy storage cooperates with whole life cycle intelligent management and control method, device and storage medium of battery cell

Abstract

The application discloses a kind of energy storage coordination battery full life cycle intelligent management and control method, equipment and storage medium, comprising: by heterogeneous computing device, with hardware level dynamic priority scheduling mode parallel running battery diagnosis algorithm set;According to the battery full-dimension health state data, parallel execution safety intervention branch and optimization control branch;The feedback data generated after the hierarchical intervention measure of the safety intervention branch execution branch and the feedback data generated by the optimization control branch execution balanced current instruction and charge-discharge depth constraint are backfilled to the heterogeneous computing device, and the battery diagnosis algorithm set and the deep reinforcement learning network are incrementally learned correction.The application executes safety intervention and optimization control double branch in parallel through battery diagnosis driving, and the feedback data of double branch execution is uniformly backfilled to realize incremental iteration, solves the data through and collaborative linkage problem between links, effectively improves the safety and economy of fire storage combined operation.

Description

Technical Field

[0001] This application relates to the field of fire-storage collaborative control technology, and in particular to a method, device and storage medium for intelligent management and control of the entire life cycle of energy storage collaborative cells. Background Technology

[0002] A thermal power-storage co-operation system refers to a joint operation system consisting of thermal power units and energy storage power stations. The health status of the battery cells throughout their entire life cycle directly determines the safety and economy of the system. Currently, in thermal power-storage co-operation systems, the battery cell health diagnosis, thermal safety management, optimization scheduling, control execution, and operation and maintenance decision-making processes operate independently. The data and decision results output by each process cannot be effectively reused by other processes, resulting in a fragmented operation of the system as a whole, characterized by "diagnosis without intervention, control without feedback, and execution without iteration."

[0003] Specifically, the aging status and consistency assessment results output by the cell health diagnosis module are not deeply coupled with thermal management, equalization control, and black start scheduling, so aging cells cannot be subject to differentiated thermal load limits and output constraints; control commands are issued in an open-loop manner, lacking accuracy compensation based on actual output feedback, and power deviations caused by cell aging continue to accumulate after long-term operation; equalization strategies and operation and maintenance plans are statically set and cannot be adaptively adjusted according to the dynamic changes in frequency regulation benefits and cell losses; the digital twin system is only used for post-event visualization, and data from various links cannot be aggregated to form effective decision support.

[0004] Due to the lack of data sharing and coordination mechanisms among the aforementioned links, the system cannot form a linkage between safety intervention, optimization control, and operation and maintenance decisions based on the real-time health status of the battery cells. This leads to a series of comprehensive problems, such as the lack of pre-intervention for thermal runaway, the difficulty in balancing frequency regulation benefits with battery cell lifespan, the deterioration of control accuracy over operating time, and the lack of flexible scheduling and fault tolerance during black start processes. These issues restrict the safety and economy of combined thermal and energy storage operations.

[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide a method, device and storage medium for intelligent management and control of the entire life cycle of energy storage collaborative cells, which aims to solve the technical problems of isolated data in each link of the existing thermal energy storage collaborative system, lack of closed-loop linkage, and inability to simultaneously take into account the dynamic balance of thermal safety, control accuracy and life benefits.

[0007] To achieve the above objectives, this application proposes a method for intelligent management and control of the entire lifecycle of energy storage collaborative battery cells, the method comprising: By deploying heterogeneous computing devices at the energy storage power station site, a set of cell diagnostic algorithms is run in parallel using a hardware-level dynamic priority scheduling method. The output data of the set of cell diagnostic algorithms includes cell SOH value, consistency safety entropy value, and cell full-dimensional health status data of abnormal risk level. Based on the comprehensive health status data of the battery cells, a safety intervention branch and an optimization control branch are executed in parallel. The safety intervention branch includes: constructing a comprehensive thermal risk safety entropy model, calculating the thermal risk level based on the comprehensive thermal risk safety entropy model, and executing graded intervention measures according to different threshold ranges of the thermal risk level. The graded intervention measures include conventional heat dissipation, enhanced liquid cooling heat dissipation, inert gas pre-charge suppression, and zoned spray power cutoff. The optimization control branch includes: using the comprehensive health status data of the battery cells as the reinforcement learning environment state, using the net value of the thermal energy storage coordinated frequency regulation benefit minus the battery cell aging loss penalty as the reward function, and outputting equalization current commands and charge / discharge depth constraints through a deep reinforcement learning network; issuing the equalization current commands and charge / discharge depth constraints through three independent hardware command channels: microsecond-level channel, millisecond-level channel, and second-level channel; collecting the actual output value of the power conversion system in real time, generating a compensation amount based on the power deviation between the issued command and the actual output value, and superimposing it into the command of the next control cycle, wherein the correction coefficient of the compensation amount is dynamically adjusted according to the SOH value of the battery cells. The feedback data generated after the safety intervention branch executes the graded intervention measures and the feedback data generated after the optimization control branch executes the equalization current command and charge / discharge depth constraint are fed back to the heterogeneous computing device to perform incremental learning correction on the battery cell diagnostic algorithm set and the deep reinforcement learning network.

[0008] In one embodiment, the step of running a set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site includes: The set of battery cell diagnostic algorithms is divided into three priority levels according to the degree of safety criticality. The frequency modulation anomaly early warning algorithm and the thermal runaway prediction algorithm are set as the highest priority, allocated the first computing power ratio, and the inference latency is limited to not exceeding the first latency threshold. The SOH correction algorithm for cell aging status and the multi-scenario consistency security entropy assessment algorithm are set as the second highest priority, allocated the second computing power ratio, and the inference latency is limited to not exceed the second latency threshold. Set the revenue statistics and monthly lifespan simulation algorithm as the lowest priority, allocate the third computing power ratio, and limit the inference latency to no more than the third latency threshold. When the highest priority algorithm is triggered and runs, the NPU scheduling kernel of the heterogeneous computing device performs hardware-level preemption to automatically reduce the operation frequency of the lowest priority algorithm. When the overall computing load of the heterogeneous computing device exceeds the first load threshold, the operation and update of the lowest priority algorithm are automatically paused until the computing load falls back below the first load threshold and then resumed.

[0009] In one embodiment, after the step of running the set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site, the method further includes: The NPU of the heterogeneous computing device locally stores daily collected first-frequency battery cell timing samples and thermal power operating condition samples. During the idle time slots of the NPU, the daily collected samples are used to perform incremental fine-tuning on the model weights of the battery cell diagnostic algorithm set, while keeping the main inference logic of the algorithm unchanged; After the incremental fine-tuning is completed, the local simulation on the NPU is used to verify whether the SOH estimation error and false alarm rate of the fine-tuned model meet the standards. After verification, the fine-tuned model weights are switched to the current running version without disturbance, and the model weights of the previous version are retained as a rollback backup. During the incremental fine-tuning process performed locally by the NPU, the real-time inference and control functions of the heterogeneous computing device continue to operate uninterruptedly.

[0010] In one embodiment, the step of constructing the comprehensive thermal risk safety entropy model includes: Obtain the full-dimensional health status data of the battery cell output by the set of battery cell diagnostic algorithms. The full-dimensional health status data of the battery cell includes at least the consistency safety entropy value and the battery cell SOH value. Acquire real-time data on the temperature rise gradient of the battery cell tabs and the concentration of combustible gas inside the compartment through high-frequency acquisition devices deployed at the energy storage power station site; Based on the current operating conditions of the thermal power unit, the current weight coefficients of the four dimensions of parameters—cell tab temperature rise gradient data, consistency safety entropy value, cell SOH value, and cabin combustible gas concentration data—are determined according to a preset weight adjustment strategy. The current operating conditions include the unit load rate and AGC operation frequency. The four dimensional parameters are weighted and summed with their respective current weight coefficients to obtain the comprehensive thermal risk safety entropy model.

[0011] In one embodiment, the step of implementing graded intervention measures based on different threshold ranges of the heat risk level includes: When the thermal risk level is in the first threshold range, the normal heat dissipation mode is activated to maintain the current operating parameters of the liquid-cooled or air-cooled unit. When the thermal risk level rises to the second threshold range, switch to enhanced liquid cooling mode to increase the coolant circulation flow and the speed of the cooling unit. When the thermal risk level rises to the third threshold range, the inert gas pre-charge suppression mode is activated to pre-charge inert gas into the energy storage chamber to reduce the oxygen concentration, while automatically reducing the charge and discharge rate of high-risk cells. When the thermal risk level rises to the fourth threshold range, the zoned spray mode is activated and the power output of the faulty battery cell cluster is cut off. Simultaneously, the thermal risk area is highlighted on the digital twin interface and a thermal maintenance work order is generated.

[0012] In one embodiment, the step of using the net value of the thermal energy storage coordinated frequency regulation revenue minus the cell aging loss penalty as the reward function, and outputting the equalization current command and charge / discharge depth constraint through a deep reinforcement learning network, includes: The cell SOH value, consistency safety entropy value, AGC frequency modulation command, real-time electricity price, current charge / discharge rate, and the working condition aging acceleration factor output by the cell aging state SOH correction algorithm are combined into a reinforcement learning environment state vector. The reinforcement learning environment state vector is input into the deep reinforcement learning network to calculate the action strategy that maximizes the long-term net benefit, where the long-term net benefit is the cumulative thermal power and energy storage coordinated frequency regulation power benefit minus the cumulative cell aging loss penalty. Based on the aforementioned action strategy, equalization current commands and upper and lower limits of charge / discharge depth constraints are generated for each cell cluster. In the current control cycle, the equalization current command and charge / discharge depth constraint are output. At the same time, the actual frequency regulation revenue and cell aging loss data collected daily are used as training samples to incrementally learn the deep reinforcement learning network and update the network weights. The equalization operation and maintenance cycle of subsequent control cycles is adjusted according to the updated network weights.

[0013] In one embodiment, the step of issuing the equalization current command and charge / discharge depth constraint through three independent hardware command channels at the microsecond, millisecond, and second levels includes: The equalization current command and the charge / discharge depth constraint command are divided into three levels of sub-commands according to the response speed requirements; A frequency modulation instantaneous power adjustment sub-command is issued through a microsecond-level channel composed of field-programmable gate array hard logic, and the channel refresh cycle does not exceed the first cycle threshold. AGC power scheduling sub-instructions are issued through a millisecond-level channel formed by a hard real-time controller, with the channel refresh cycle being the second cycle threshold. The power plant operating condition switching and start / stop control sub-instructions are issued through a second-level channel composed of the ARM main core, and the channel refresh cycle is the third cycle threshold. Among them, the microsecond-level channel, millisecond-level channel and second-level channel share a unified clock reference based on a precise time protocol, and the timing deviation between the sub-instructions issued by each level of channel does not exceed the first deviation threshold.

[0014] In one embodiment, after the step of feeding back the feedback data generated by the safety intervention branch after executing the graded intervention measures and the feedback data generated by the optimization control branch after executing the equalization current command and charge / discharge depth constraint to the heterogeneous computing device, the method further includes: Enter candidate operation and maintenance solutions into the digital twin interface. The candidate operation and maintenance solutions include balancing cycle parameters, cell replacement threshold parameters, and black start backup capacity parameters. The digital twin simulation engine is used to call the cell aging model and the multi-scenario consistency model to simulate the operating effect of the candidate operation and maintenance scheme under the thermal power operating conditions throughout the year. Output the cell attenuation prediction curve, frequency regulation benefit calculation value and total operation and maintenance cost for each candidate operation and maintenance solution; Automatically recommends the operation and maintenance solution with the highest net benefit, and simultaneously generates virtual maintenance simulation animations; Based on the aforementioned operation and maintenance plan, actual operating data is collected and fed back into the digital twin simulation engine to iteratively correct the simulation model parameters.

[0015] In addition, to achieve the above objectives, this application also proposes an intelligent management and control device for the entire life cycle of energy storage collaborative battery cells. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described above.

[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the intelligent management and control method for the entire life cycle of energy storage collaborative cells as described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: The technical solution of this application utilizes heterogeneous computing devices deployed at energy storage power stations to run a set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method. The output data of this set of algorithms includes the cell's SOH value, consistency safety entropy value, and comprehensive health status data of the cell, representing its anomaly risk level. Based on this comprehensive health status data, a safety intervention branch and an optimization control branch are executed in parallel. The safety intervention branch includes: constructing a comprehensive thermal risk safety entropy model, calculating the thermal risk level based on the model, and implementing graded intervention measures according to the different threshold ranges of the thermal risk level. These measures include conventional heat dissipation, enhanced liquid cooling, inert gas pre-charging suppression, and zoned spray power cutoff. The optimization control branch uses the comprehensive health status data of the cell as a reinforcement learning environment. The system uses the net value of the thermal power generation and energy storage coordinated frequency regulation revenue minus the cell aging loss penalty as the reward function. A deep reinforcement learning network outputs equalization current commands and charge / discharge depth constraints. These commands are issued through three independent hardware command channels: microsecond-level, millisecond-level, and second-level channels. The actual output value of the power conversion system is collected in real time. A compensation amount is generated based on the power deviation between the issued command and the actual output value and added to the command of the next control cycle. The correction coefficient of the compensation amount is dynamically adjusted according to the cell's state of harmonic equilibrium (SOH). Feedback data generated after the safety intervention branch executes graded intervention measures and feedback data generated after the optimization control branch executes equalization current commands and charge / discharge depth constraints are fed back to the heterogeneous computing device for incremental learning and correction of the cell diagnostic algorithm set and the deep reinforcement learning network.

[0018] The technical solution of this application realizes the real-time driving of cell diagnostic results for safety intervention and optimization control, and the unified iteration of diagnostic and decision models through dual-branch execution feedback. It solves the technical problems of isolated data and lack of closed-loop linkage in the various links of the fire-storage collaborative system, which makes it impossible to take into account thermal safety, control accuracy and dynamic balance of life-benefit. It realizes the technical leap from fragmented open-loop to full-link closed-loop management and control of fire-storage collaborative system. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the first embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative cells in this application; Figure 2 This is a detailed step diagram of step S10 in one embodiment; Figure 3 This is a detailed step diagram of step S20 in one embodiment; Figure 4 This is a schematic diagram of another detailed step of step S20 in one embodiment; Figure 5 This is a flowchart illustrating the second embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative battery cells according to the present invention. Figure 6 This is a flowchart illustrating the third embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative battery cells according to the present invention. Figure 7 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the intelligent management and control method for the entire life cycle of energy storage collaborative cells in the embodiments of this application.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] In a fire-storage collaborative system, the various stages of cell health diagnosis, thermal safety management, optimization scheduling, control execution, and operation and maintenance decision-making follow an independent, sequential, open-loop technical path. This type of solution typically involves the battery management system estimating the cell health status, the energy management system independently issuing power commands, and the thermal management module executing fire-fighting linkages based on a single temperature threshold. Each module operates according to its own logic; diagnostic results do not drive safety intervention or optimization decisions; power commands lack actual output feedback compensation; balancing strategies use statically set fixed thresholds; and digital twins are only used for post-event demonstration.

[0026] The core dilemma faced by the above-mentioned technical approaches is that, although the architecture of diagnosis, control, thermal management and operation and maintenance being independent and open-loop serial is clear, the data between links is not interconnected, the decision-making is not coordinated, and the execution is not iterative. This is fundamentally contradictory to the inherent requirements of the thermal energy storage collaborative system for precise pre-intervention in thermal safety, dynamic balance between frequency regulation benefits and cell lifespan, and long-term maintenance of control precision. It is impossible to form an effective closed-loop linkage between safety intervention, optimized control and operation and maintenance decisions based on the real-time health status of the cells.

[0027] To address the aforementioned shortcomings, this application proposes a method for intelligent management and control of the entire lifecycle of energy storage-coordinated battery cells. This method simultaneously drives two parallel branches—safety intervention and optimization control—using the multi-dimensional health status data output by a set of battery cell diagnostic algorithms. Feedback data from both branches is then uniformly fed back to achieve incremental iteration, constructing a closed-loop system that connects diagnosis, intervention, control, feedback, and iteration. Specifically, a set of battery cell diagnostic algorithms is run in parallel with hardware-level dynamic priority scheduling via a heterogeneous computing device to obtain the cell's SOH value, consistency safety entropy value, and abnormal risk level. The safety intervention branch constructs a comprehensive safety entropy model integrating temperature rise gradient, consistency safety entropy, SOH, and combustible gas concentration, executing graded interventions ranging from conventional heat dissipation to zoned spray power cutoff. The optimization control branch uses frequency regulation revenue minus aging loss penalty as the reward function, outputs an equilibration strategy through reinforcement learning, and distributes it through three independent hardware channels while collecting actual power output for closed-loop accuracy compensation. The compensation coefficient is dynamically adjusted according to SOH. Finally, the feedback data from both branches is fed back to the heterogeneous computing device for incremental learning and correction of the diagnostic algorithms and reinforcement learning network.

[0028] Through the above-mentioned technical means, this application realizes the real-time driving of cell diagnostic results for safety intervention and optimization control, as well as the unified iteration of diagnostic and decision models through dual-branch execution feedback. It solves the technical problems of isolated data and lack of closed-loop linkage in various links of the fire-storage collaborative system, which makes it impossible to take into account thermal safety, control accuracy and dynamic balance of life-benefit. It realizes the technical leap from fragmented open-loop to full-link closed-loop management and control of fire-storage collaborative system.

[0029] Based on this, embodiments of this application provide a method for intelligent management and control of the entire lifecycle of energy storage collaborative cells, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative battery cells according to this application. In this embodiment, the intelligent management and control method for the entire lifecycle of energy storage collaborative battery cells is applied to an access service node, including steps S10~S30: Step S10: Through the heterogeneous computing device deployed at the energy storage power station site, the set of cell diagnostic algorithms is run in parallel in a hardware-level dynamic priority scheduling manner. The output data of the set of cell diagnostic algorithms includes the cell SOH value, consistency safety entropy value and cell full-dimensional health status data of abnormal risk level. In the thermal energy storage co-processing system, a heterogeneous computing device is deployed at the energy storage power station site. This heterogeneous computing device undertakes the parallel operation of a set of cell diagnostic algorithms. The heterogeneous computing device employs a hardware-level dynamic priority scheduling method to schedule and run the set of cell diagnostic algorithms. The set of cell diagnostic algorithms includes a cell aging state (SOH) correction algorithm, a frequency modulation micro-feature anomaly early warning algorithm, and a multi-scenario consistent security entropy assessment algorithm. These algorithms do not compete equally for computing resources on the heterogeneous computing device; instead, they are divided into multiple priority levels according to their safety criticality. The frequency modulation anomaly early warning algorithm and the thermal runaway prediction algorithm, which involve real-time identification of the cell's safety state, are set to the highest priority. When this level of algorithm is triggered, the NPU scheduling kernel of the heterogeneous computing device performs a hardware-level preemption action, automatically reducing the operation frequency of lower-priority algorithms to ensure that the inference latency of the highest-priority algorithm does not exceed twenty milliseconds. The cell aging state (SOH) correction algorithm and the multi-scenario consistent security entropy assessment algorithm, which serve the continuous tracking of the cell's health state, are set to the second highest priority, with an inference latency not exceeding fifty milliseconds. The revenue statistics and monthly lifetime simulation algorithms, which do not involve real-time security control, are set to the lowest priority, with an inference latency of no more than 200 milliseconds. When the overall computing load of the heterogeneous computing device exceeds 90%, the system automatically suspends the operation and update of the lowest priority algorithm until the computing load falls below 90%. Through the above dynamic priority scheduling mechanism, the real-time performance of safety-critical algorithms is effectively guaranteed.

[0030] While running the cell diagnostic algorithm suite on the heterogeneous computing device, the device also undertakes the task of incremental model learning. Daily collected 100Hz-level cell time-series samples and thermal power plant operating condition samples are stored in the NPU's local storage space. When the NPU enters an idle time slot, the system calls upon the aforementioned daily collected samples to perform incremental fine-tuning of the model weights in the cell diagnostic algorithm suite. This incremental fine-tuning process only updates the model weights, keeping the main inference logic of the algorithm unchanged, thus not affecting the basic architecture and inference behavior of the algorithm. After the incremental fine-tuning is completed, the fine-tuned model is simulated and verified locally on the NPU to confirm whether the model's SOH estimation error and false alarm rate meet the preset standards. After verification, the system switches the fine-tuned model weights to the current running version without disturbance, while retaining the previous version's model weights as a rollback backup for rapid recovery in case of anomalies. Throughout the entire process of incremental fine-tuning performed locally on the NPU, the heterogeneous computing device's external real-time inference and control functions remain uninterrupted. The aforementioned incremental learning mechanism enables the cell diagnostic algorithm set to continuously adapt to changes in thermal power operating conditions and drifts in cell aging characteristics during continuous operation, without interrupting system operation for offline updates.

[0031] After the cell diagnostic algorithm suite runs, it outputs comprehensive health status data for the cells. This data includes at least the cell's State of Health (SOH) value, consistency safety entropy value, and anomaly risk level. The SOH value reflects the remaining capacity and aging degree of each cell, the consistency safety entropy value reflects the performance consistency between cells, and the anomaly risk level reflects whether the cell has early abnormal signals at the frequency modulation micro-feature level. These three types of data jointly characterize the cell's health status from different dimensions, providing a data foundation for subsequent safety intervention and optimization control. The above data processing logic ensures that the diagnostic results are not merely status displays, but rather serve as quantitative indicators that can be directly invoked by downstream modules, driving the execution of both the safety intervention and optimization control branches. Through the above, a complete data processing chain is completed from high-frequency raw data acquisition to comprehensive health status data output. The health status data output from this chain will be input into the safety intervention and optimization control branches in subsequent steps.

[0032] Step S20: Based on the full-dimensional health status data of the battery cell, execute the safety intervention branch and the optimization control branch in parallel. The safety intervention branch includes: constructing a comprehensive thermal risk safety entropy model, calculating the thermal risk level based on the comprehensive thermal risk safety entropy model, and executing graded intervention measures according to the different threshold ranges of the thermal risk level. The graded intervention measures include conventional heat dissipation, enhanced liquid cooling heat dissipation, inert gas pre-charge suppression, and zoned spray power cutoff. The optimization control branch includes: using the full-dimensional health status data of the battery cell as the reinforcement learning environment state, using the net value of the fire-storage collaborative frequency modulation benefit minus the battery cell aging loss penalty as the reward function, and outputting the equalization current command and charge-discharge depth constraint through a deep reinforcement learning network; issuing the equalization current command and charge-discharge depth constraint through three independent hardware command channels: microsecond-level channel, millisecond-level channel, and second-level channel; collecting the actual output value of the power conversion system in real time, generating a compensation amount based on the power deviation between the issued command and the actual output value, and superimposing it into the command of the next control cycle, wherein the correction coefficient of the compensation amount is dynamically adjusted according to the SOH value of the battery cell. In this embodiment, after obtaining the full-dimensional health status data of the battery cell, two branches are executed in parallel based on this data: a safety intervention branch and an optimization control branch. These two branches share the same set of health status data, but each executes different data processing logic.

[0033] In the safety intervention branch, a comprehensive thermal risk safety entropy model is first constructed. The construction process is as follows: The system acquires full-dimensional health status data of the battery cells from the output of the cell diagnostic algorithm set. This data includes at least the consistency safety entropy value and the cell SOH value. Simultaneously, the system acquires real-time data on the temperature rise gradient of the battery cell tabs and the concentration of combustible gases inside the storage compartment from high-frequency acquisition devices deployed at the energy storage power station site. These four dimensions of data are used as input parameters to the comprehensive thermal risk safety entropy model. Based on the current operating conditions of the thermal power unit, the system determines the current weight coefficients for each of the four dimensions according to a preset weight adjustment strategy. These current operating conditions include the unit load rate and the frequency of AGC actions. The system weights and sums the four dimensions—the temperature rise gradient of the battery cell tabs, the consistency safety entropy value, the cell SOH value, and the concentration of combustible gases inside the storage compartment—with their respective current weight coefficients to obtain the comprehensive thermal risk safety entropy model. The calculation method is as follows: ; in, Represents the temperature rise gradient of the battery cell tabs. Represents consistent security entropy, This represents the health status value of the battery cell. This represents the concentration of flammable gases inside the cabin. w 1. w 2. w 3. w 4 represents the current weighting coefficient for each of the four dimensions mentioned above. This weighting coefficient is determined according to a preset weighting adjustment strategy based on the current operating conditions of the thermal power unit. Specifically, it can be iteratively updated online based on the unit load rate and AGC (Automatic Guided Vehicle) activation frequency. The model outputs a comprehensive safety entropy value. The overall thermal risk level of the battery cell under current operating conditions was quantified.

[0034] Based on the thermal risk level calculated using the comprehensive thermal risk safety entropy model, the system implements tiered intervention measures according to the different threshold intervals in which the thermal risk level falls. When the thermal risk level is in the first threshold interval... When the temperature is below 0.3, the system activates the normal cooling mode, maintaining the current operating parameters of the liquid-cooled or air-cooled unit. When the thermal risk level rises to the second threshold range... When the temperature is greater than or equal to 0.3 and less than 0.6, the system switches to enhanced liquid cooling mode, increasing the coolant circulation flow rate and the radiator speed to improve heat dissipation. When the thermal risk level further rises to the third threshold range... When the thermal risk level is greater than or equal to 0.6 and less than 0.85, the system activates the inert gas pre-charge suppression mode, pre-charging the energy storage chamber with inert gas to reduce the oxygen concentration. This implements pre-suppression from the perspective of oxygen, one of the three elements of combustion, while automatically reducing the charge and discharge rates of high-risk cells to decrease heat generation. When the thermal risk level rises to the fourth threshold range... When the temperature reaches 0.85 or higher, the system activates the zoned spray mode and cuts off the power output of the faulty battery cell cluster. Simultaneously, the thermal risk area is highlighted on the digital twin interface, and a thermal maintenance work order is generated. Through the above four-level gradient intervention, the system adopts intervention measures of different intensities at different stages of thermal runaway, realizing the transformation from passively waiting for threshold triggering to actively suppressing it in a graded manner.

[0035] In the optimized control branch, the full-dimensional health status data of the battery cells is used as the reinforcement learning environment state. The reinforcement learning environment state vector is composed of real-time collected cell SOH values, consistency safety entropy values, AGC frequency regulation commands, real-time electricity prices, current charge / discharge rates, and the operating condition aging acceleration factor output by the cell aging state SOH correction algorithm. The system inputs this environment state vector into a deep reinforcement learning network, which uses the net value of the thermal power-storage coordinated frequency regulation revenue minus the cell aging loss penalty as its reward function. The reward function is calculated as follows: ; in Represents frequency modulation power. Represents real-time electricity price. To accumulate the power benefits of coordinated frequency regulation between thermal power and energy storage, This is the aging penalty coefficient. This is the aging acceleration factor output by the SOH correction algorithm for cell aging status. As the baseline aging rate, The cumulative cell aging loss penalty is applied. The deep reinforcement learning network calculates the action strategy that maximizes the long-term net benefit using the above reward function. This long-term net benefit is the cumulative power gain from thermal-energy storage coordinated frequency regulation minus the cumulative cell aging loss penalty. Based on the above environmental state vector and reward function, the deep reinforcement learning network outputs the equalization current command and the upper and lower limits of charge / discharge depth constraints for each cell cluster. The system issues the above equalization current command and charge / discharge depth constraints through three independent hardware command channels: microsecond-level channel, millisecond-level channel, and second-level channel. The microsecond-level channel is composed of field-programmable gate array hard logic and is used to issue instantaneous power adjustment commands for primary frequency regulation, with a channel refresh period of no more than one millisecond; the millisecond-level channel is composed of a hard real-time controller and is used to issue AGC power scheduling commands, with a channel refresh period of ten milliseconds; the second-level channel is composed of an ARM main core and is used to issue power plant operating condition switching and start / stop control commands, with a channel refresh period of one hundred milliseconds. The three channels mentioned above share a common precise time protocol and a unified clock reference. The timing deviation between sub-instructions issued by each channel does not exceed 0.5 microseconds, ensuring that control tasks at different time scales are executed collaboratively under a unified time coordinate.

[0036] The system acquires the actual output value of the power conversion system in real time and generates a compensation amount based on the power deviation between the issued command and the actual output value. This compensation amount is calculated by taking the set value corresponding to the issued command... Actual output value of the power conversion system The difference between them is multiplied by a correction factor. k The calculation method is as follows: ; in k This is a correction factor, dynamically adjusted based on the cell's SOH value. The more severe the cell aging, the larger the correction factor, to compensate for power response deviations caused by aging. The system will use the above compensation amount... The error is superimposed on the instruction of the next control cycle, so that the instruction issued in the next cycle includes a correction for the execution deviation of the current cycle. Through the above closed-loop accuracy compensation mechanism, the steady-state deviation of power control is reduced from the traditional ±3% to ±0.8%, and this level of accuracy can still be maintained after long-term operation.

[0037] Specifically, the step of implementing graded intervention measures based on different threshold ranges of the heat risk level includes: When the thermal risk level is in the first threshold range, the normal heat dissipation mode is activated to maintain the current operating parameters of the liquid-cooled or air-cooled unit. When the thermal risk level rises to the second threshold range, switch to enhanced liquid cooling mode to increase the coolant circulation flow and the speed of the cooling unit. When the thermal risk level rises to the third threshold range, the inert gas pre-charge suppression mode is activated to pre-charge inert gas into the energy storage chamber to reduce the oxygen concentration, while automatically reducing the charge and discharge rate of high-risk cells. When the thermal risk level rises to the fourth threshold range, the zoned spray mode is activated and the power output of the faulty battery cell cluster is cut off. Simultaneously, the thermal risk area is highlighted on the digital twin interface and a thermal maintenance work order is generated.

[0038] In the safety intervention branch, after calculating the thermal risk level based on the comprehensive thermal risk safety entropy model, graded intervention measures are implemented according to the different threshold intervals in which the thermal risk level falls. Specifically, the thermal risk level is divided into four threshold intervals, each corresponding to an intervention mode of different intensities. When the thermal risk level is in the first threshold interval, the system activates the normal heat dissipation mode to maintain the current operating parameters of the liquid-cooled or air-cooled unit. The comprehensive thermal risk safety entropy value corresponding to the first threshold interval is less than 0.3, indicating that the current thermal risk of the battery cell is at a low level, requiring no additional intervention measures; simply maintaining the current operating state of the heat dissipation unit is sufficient to keep the battery cell temperature within the normal range.

[0039] When the thermal risk level rises to the second threshold range, the system switches to enhanced liquid cooling mode, increasing the coolant circulation flow and the radiator speed. The comprehensive safety entropy value for thermal risk corresponding to the second threshold range is greater than or equal to 0.3 and less than 0.6, indicating that the cell thermal risk has risen to a moderate level, and conventional cooling capacity is insufficient to suppress the temperature rise. At this time, an enhanced cooling command is sent to the liquid cooling unit control unit. This command includes two parameters: an increase in coolant flow and an increase in radiator speed. Upon receiving the enhanced cooling command, the liquid cooling unit increases the coolant circulation flow by adjusting the inverter output frequency of the coolant circulation pump, and simultaneously increases the radiator speed by adjusting the inverter output frequency of the cooling fan, thereby improving the cooling capacity. Through these enhanced liquid cooling measures, the cooling capacity is proactively increased before the thermal risk reaches a dangerous level, preventing the thermal risk from escalating to a higher level.

[0040] When the thermal risk level rises to the third threshold range, the system activates the inert gas pre-charge suppression mode, pre-charging inert gas into the energy storage chamber to reduce the oxygen concentration, while automatically reducing the charge and discharge rates of high-risk battery cells. The comprehensive safety entropy value of thermal risk corresponding to the third threshold range is greater than or equal to 0.6 and less than 0.85, indicating that the thermal risk of the battery cells has risen to a relatively high level. Simply strengthening liquid cooling is insufficient to effectively suppress the thermal risk; pre-emptive suppression from the oxygen dimension of the three elements of combustion is necessary. The specific implementation of inert gas pre-charge suppression is as follows: an opening command is sent to the solenoid valve of the inert gas storage tank located at the top of the energy storage chamber, allowing inert gas to be evenly released into the energy storage chamber through pipelines and nozzles, reducing the probability of combustion by diluting the oxygen concentration inside the chamber. The essential difference between inert gas pre-charge suppression and conventional fire sprinklers is that its pre-charge action is initiated before thermal runaway occurs, representing a preventative pre-emptive intervention rather than a reactive fire suppression measure. Meanwhile, the system identifies high-risk cells based on the abnormal risk levels output by the cell diagnostic algorithm set and sends a charge / discharge rate limiting command to the energy storage converter to reduce the charge / discharge rate of the high-risk cells to a safe range, thereby reducing their heat generation. Through the dual measures of inert gas pre-charging and current limiting for high-risk cells, the system proactively reduces the risk of thermal runaway before the thermal risk worsens.

[0041] When the thermal risk level rises to the fourth threshold range, the zoned sprinkler mode is activated and the power output of the faulty battery cell cluster is cut off. Simultaneously, the thermal risk area is highlighted on the digital twin interface, and a thermal maintenance work order is generated. A comprehensive safety entropy value for thermal risk corresponding to the fourth threshold range is greater than or equal to 0.85, indicating that the battery cell thermal risk has reached its highest level, and thermal runaway may occur at any time. At this point, an opening command is sent to the solenoid valves of the zoned sprinkler system located within the energy storage compartment, allowing fire water or extinguishing agent to be released through the nozzles to the corresponding risk area for fire suppression. The zoned sprinkler mode differs from full-compartment sprinkler; it only sprays the area with the highest thermal risk level, avoiding unnecessary damage to equipment in non-faulty areas caused by full-compartment sprinkler. At the same time, a power cut-off command for the faulty cluster is sent to the energy storage converter, disconnecting the electrical connection of the faulty battery cell cluster to prevent the continued charging and discharging of the faulty cell from exacerbating thermal runaway. Simultaneously, the location and extent of the thermal risk area are highlighted on the digital twin interface for remote viewing by maintenance personnel, and a thermal maintenance work order containing the fault location, fault time, and suggested handling measures is automatically generated and sent to the maintenance management system. Through the above four-level gradient intervention, the system adopts intervention measures of different intensities at different stages of thermal risk development, realizing a complete intervention chain from conventional heat dissipation to enhanced liquid cooling, and from inert gas pre-charging to zoned spray power cut-off.

[0042] Furthermore, the step of issuing the equalization current command and charge / discharge depth constraint through three independent hardware command channels at the microsecond, millisecond, and second levels includes: The equalization current command and the charge / discharge depth constraint command are divided into three levels of sub-commands according to the response speed requirements; A frequency modulation instantaneous power adjustment sub-command is issued through a microsecond-level channel composed of field-programmable gate array hard logic, and the channel refresh cycle does not exceed the first cycle threshold. AGC power scheduling sub-instructions are issued through a millisecond-level channel formed by a hard real-time controller, with the channel refresh cycle being the second cycle threshold. The power plant operating condition switching and start / stop control sub-instructions are issued through a second-level channel composed of the ARM main core, and the channel refresh cycle is the third cycle threshold. Among them, the microsecond-level channel, millisecond-level channel and second-level channel share a unified clock reference based on a precise time protocol, and the timing deviation between the sub-instructions issued by each level of channel does not exceed the first deviation threshold.

[0043] In this embodiment, within the optimized control branch, equalization current commands and charge / discharge depth constraints are issued through three independent hardware command channels: microsecond-level channel, millisecond-level channel, and second-level channel. The specific implementation process of issuing these three command channels is as follows.

[0044] The equalization current command and charge / discharge depth constraint command output by the deep reinforcement learning network are decomposed into three levels of sub-commands based on response speed requirements. The equalization current command and charge / discharge depth constraint command contain multiple control requirements for different control time scales. The instantaneous power regulation requirement of primary frequency regulation has the fastest response speed, immediately adjusting the output power of the energy storage system to support the grid frequency when the grid frequency fluctuates; this requirement has a response time in the millisecond range. The AGC power dispatch requirement has a medium response speed, completing power regulation within a certain time according to the automatic generation control command issued by the grid dispatch center; this requirement has a response time in the tens of milliseconds range. The power plant operating mode switching and start / stop control requirements have a slow response speed, involving energy storage system operating mode switching and equipment start / stop operations; this requirement has a response time in the hundreds of milliseconds range. Based on these differences in response speed, the original integrated control command is decomposed into three independent sub-command sets, each carrying control content for a corresponding time scale, and each sub-command set is routed to the corresponding hardware command channel for execution.

[0045] A frequency modulation instantaneous power adjustment sub-command is issued via a microsecond-level channel constructed from field-programmable gate array (FPGA) hard logic. The refresh period of this channel is no more than one millisecond. The FPGA hard logic does not run an operating system; it directly executes the command issuance logic through hardware circuitry, thus achieving microsecond-level command response. After receiving the frequency modulation instantaneous power adjustment sub-command, the channel outputs the command as a hardware signal directly to the instantaneous power adjustment port of the energy storage converter. Upon receiving this signal, the energy storage converter completes the power adjustment action within a microsecond.

[0046] AGC power scheduling sub-commands are issued via a millisecond-level channel comprised of a hard real-time controller, with a refresh period of ten milliseconds. The hard real-time controller runs a real-time operating system, ensuring that the execution time of the control task has a deterministic upper bound. After receiving the AGC power scheduling sub-command, the channel sends the power scheduling command to the AGC power scheduling port of the energy storage converter via the industrial Ethernet protocol within the ten-millisecond refresh period. Upon receiving the command, the energy storage converter completes the power adjustment action within tens of milliseconds.

[0047] Sub-instructions for power plant operating condition switching and start / stop control are issued via a second-level channel comprised of an ARM main core. This channel has a refresh period of one hundred milliseconds. The ARM main core runs a non-real-time operating system and handles management control tasks that are not time-sensitive. After receiving the sub-instructions, the channel transmits them to the operating condition management port of the energy storage converter via the industrial Ethernet protocol within the one hundred millisecond refresh period. Upon receiving the instruction, the energy storage converter completes the operating condition switching or start / stop control action within several hundred milliseconds.

[0048] The aforementioned microsecond-level, millisecond-level, and second-level channels share a unified clock reference based on a precise time protocol, with timing deviations between sub-instructions issued by each channel not exceeding 0.5 microseconds. The system uses a precise time protocol clock synchronization module deployed in heterogeneous computing devices to synchronize the local clocks of the field-programmable gate array (FPGA), hard real-time controller, and ARM core with a unified master clock, based on the IEEE 1588 precise time protocol standard. The clock synchronization module periodically sends synchronization messages to the three hardware units, and each hardware unit corrects its local clock frequency and phase according to the synchronization messages to ensure consistency in timing references. Under this unified clock reference, sub-instructions issued by each channel carry precise timestamp information. After receiving these sub-instructions, the energy storage converter aligns the execution timing of each instruction according to the timestamp. Through the coordinated implementation of these three independent hardware instruction channels, the system can execute control tasks at different time scales (microsecond, millisecond, and second) on the same time coordinate, achieving hierarchical issuance and precise timing alignment of multi-level instructions from instantaneous power regulation to operating condition switching.

[0049] Step S30: Feedback data generated after the safety intervention branch executes graded intervention measures and feedback data generated after the optimization control branch executes equalization current command and charge / discharge depth constraint are fed back to the heterogeneous computing device to perform incremental learning correction on the cell diagnostic algorithm set and the deep reinforcement learning network.

[0050] After the safety intervention branch executes tiered intervention measures, feedback data related to the execution process of these measures is generated. This feedback data includes the actual trigger time, execution duration, and changes in thermal risk level after each intervention level. Similarly, after the optimization control branch executes balancing current commands and charge / discharge depth constraints, feedback data related to the execution process of these balancing commands is generated. This feedback data includes the actual balancing current value achieved by each cell cluster, the actual charge / discharge rate executed, and the changes in cell SOH value and consistency safety entropy value after executing the commands. The feedback data from both the safety intervention branch and the optimization control branch are fed back into the heterogeneous computing device as a data source for subsequent incremental learning and correction.

[0051] After receiving the aforementioned dual-branch feedback data, the heterogeneous computing device uses this data to incrementally learn and correct the cell diagnostic algorithm set. The SOH (State of Health) correction algorithm within the cell diagnostic algorithm set receives the actual charge / discharge rate and cell degradation data from the optimization control branch feedback, compares it with the algorithm's original predictions, and fine-tunes the weight parameters of the SOH correction model based on the comparison results, continuously improving the SOH estimation accuracy over long-term operation. The anomaly warning algorithm within the cell diagnostic algorithm set receives the tiered intervention measure execution data from the safety intervention branch feedback, performs correlation analysis with the micro-features triggered by the algorithm's warning, and adjusts the warning threshold or fine-tunes the weight parameters of the anomaly feature extraction layer based on the analysis results to reduce the false alarm rate and increase the warning lead time. The multi-scenario consistency safety entropy assessment algorithm within the cell diagnostic algorithm set receives data from the dual-branch feedback involving differences in cell output and temperature, compares it with the consistency safety entropy value output by the algorithm, and updates the consistency assessment parameters based on the comparison results, making the consistency assessment more accurately reflect the actual performance differences between cells.

[0052] The heterogeneous computing device also uses the aforementioned dual-branch feedback data to incrementally learn and correct the deep reinforcement learning network. The system daily compiles actual frequency modulation revenue and cell aging loss data, and inputs these statistical results as training samples into the deep reinforcement learning network for incremental learning. The network updates its weights based on the incremental learning results, making the reward function: ; The balance between the calculated benefit items and loss penalty items is optimized. After the network weights are updated, the system adjusts the balance operation and maintenance cycle parameters of subsequent control cycles according to the updated network weights, so that the balance strategy is adapted to the actual degradation rate of the battery cells, avoiding over-balancing or under-balancing.

[0053] The unified feedback and incremental learning correction process of the aforementioned dual-branch feedback data constitutes a closed-loop iterative mechanism throughout the system's entire lifecycle. The execution feedback data from the safety intervention branch and the optimization control branch are no longer merely stored as operational logs, but are reused as data fuel to drive the continuous evolution of the diagnostic and decision-making models. Through this mechanism, the estimation accuracy of the cell health status by the cell diagnostic algorithm set continuously improves with the extension of operating time, and the optimization effect of the deep reinforcement learning network on the equilibrium strategy continuously strengthens with the accumulation of operating data. After six months of continuous operation, the SOH estimation error decreased from the initial 2.7% to 1.5%, the false alarm rate of anomaly warning decreased from the initial 3.2% to 2.4%, the annualized comprehensive benefit of the project was further improved compared to traditional solutions, and the average annual degradation rate of the cells decreased.

[0054] Further, see Figure 2 , Figure 2This is a detailed step diagram of step S10 in one embodiment, based on the shown... Figure 2 The step of running the set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site includes S11~S16: Step S11: Divide the set of cell diagnostic algorithms into three priority levels according to their safety criticality. Step S12: Set the frequency modulation anomaly early warning algorithm and thermal runaway prediction algorithm to the highest priority, allocate the first computing power ratio, and limit the inference delay to no more than the first delay threshold. Step S13: Set the cell aging state SOH correction algorithm and the multi-scenario consistency security entropy evaluation algorithm as the second highest priority, allocate the second computing power ratio, and limit the inference latency to not exceed the second latency threshold. Step S14: Set the revenue statistics and monthly lifespan simulation algorithm to the lowest priority, allocate the third computing power ratio, and limit the inference latency to not exceed the third latency threshold. Step S15: When the highest priority algorithm is triggered and runs, the NPU scheduling kernel of the heterogeneous computing device performs hardware-level preemption to automatically reduce the operation frequency of the lowest priority algorithm. Step S16: When the overall computing power load of the heterogeneous computing device exceeds the first load threshold, the operation and update of the lowest priority algorithm is automatically paused until the computing power load falls back below the first load threshold and then resumes.

[0055] In this embodiment, the heterogeneous computing device runs the set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method. The specific implementation process of this dynamic priority scheduling is as follows: The set of cell diagnostic algorithms is divided into three priority levels based on their safety criticality. The set includes a frequency modulation anomaly early warning algorithm, a thermal runaway prediction algorithm, a cell aging state SOH correction algorithm, a multi-scenario consistent safety entropy assessment algorithm, a revenue statistics algorithm, and a monthly lifespan simulation algorithm. These algorithms differ in their importance to system safety. The frequency modulation anomaly early warning algorithm and the thermal runaway prediction algorithm are directly related to the timely detection of dangerous states such as cell thermal runaway and frequency modulation anomalies; their inference latency directly affects the system's response speed to safety risks. The cell aging state SOH correction algorithm and the multi-scenario consistent safety entropy assessment algorithm serve the continuous tracking and assessment of cell health status; their inference latency affects the timeliness of diagnostic results but does not directly affect safety response. The revenue statistics algorithm and the monthly lifespan simulation algorithm serve economic analysis and long-term operation and maintenance planning; their inference latency does not involve real-time control or safety assurance. Based on these differences in safety criticality, the system divides these algorithms into three priority levels, providing a basis for subsequent differentiated computing power allocation.

[0056] The frequency modulation anomaly early warning algorithm and the thermal runaway prediction algorithm are set as the highest priority, allocated the first computing power percentage, and have their inference latency limited to no more than a first latency threshold. The first computing power percentage is 60%, and the first latency threshold is 20 milliseconds. These two algorithms are set as the highest priority because they play a core role in real-time safety monitoring during the coordinated operation of thermal power and energy storage. The frequency modulation anomaly early warning algorithm analyzes the micro-feature changes of the battery cell during frequency modulation in real time, enabling it to issue early warning signals when early signs of performance degradation appear. The thermal runaway prediction algorithm, through comprehensive analysis of the battery cell's temperature change trend and internal state parameters, can identify potential risks before thermal runaway occurs. If the inference latency of these two algorithms exceeds 20 milliseconds due to insufficient computing power resources, it may delay the issuance of safety warnings and increase the risk of thermal runaway. Therefore, the system allocates the highest computing power percentage and the strictest latency constraints to these algorithms.

[0057] The cell aging state SOH correction algorithm and the multi-scenario consistency security entropy assessment algorithm are set as the second highest priority, allocated a second share of computing power, and their inference time is limited to a strict second latency threshold. The second computing power share is 30%, and the second latency threshold is 50 milliseconds. The cell aging state SOH correction algorithm outputs the SOH value and aging acceleration factor of each cell by coupling and analyzing the cell's charge / discharge data and operating condition data. The multi-scenario consistency security entropy assessment algorithm outputs the consistency security entropy value and outlier degree by analyzing the performance consistency of the cell under different scenarios. The output results of the above two algorithms are important input parameters for the thermal risk comprehensive security entropy model and the reinforcement learning environment state. Although their inference latency does not directly determine the real-time performance of the safety response, it affects the timeliness of the data on which the safety intervention and optimization control branches are based. Limiting the inference latency of the above algorithms to within 50 milliseconds can ensure the timely update of diagnostic data while prioritizing the computing power of safety-critical algorithms.

[0058] The revenue statistics and monthly lifespan simulation algorithms are set to the lowest priority, allocated a third share of computing power, and their inference latency is limited to no more than a third latency threshold. The third computing power share is 10%, and the third latency threshold is 200 milliseconds. The revenue statistics and monthly lifespan simulation algorithms serve the system's economic analysis and long-term operation and maintenance planning; their operation cycle is daily or monthly, and they do not involve real-time control or security assurance. Therefore, these algorithms are allocated the lowest computing power share and the most lenient latency constraints. When computing resources are scarce, the execution of these algorithms can be appropriately delayed without affecting the system's safe operation and control performance.

[0059] Building upon the prioritization and computing power allocation described above, a hardware-level preemption mechanism is further introduced. When the highest-priority algorithm is triggered, the NPU scheduling kernel of the heterogeneous computing device performs hardware-level preemption, automatically reducing the computation frequency of the lowest-priority algorithm. This hardware-level preemption mechanism differs from software-level priority queuing or load balancing; it is executed directly by the NPU scheduling kernel without going through the operating system scheduling layer, thus the preemption latency is on the microsecond level. When the highest-priority algorithm is triggered, the NPU scheduling kernel immediately suspends the current computation task of the lowest-priority algorithm and reallocates computing resources to the highest-priority algorithm, ensuring that the highest-priority algorithm can obtain the necessary inference resources immediately.

[0060] The overall computing load of heterogeneous computing devices is monitored in real time. When the overall computing load exceeds a first load threshold, the operation and update of the lowest priority algorithm are automatically paused until the computing load falls below the first load threshold. The first load threshold is 90%. When the overall computing load exceeds 90%, it indicates that the system is close to full-load operation. Continuing to maintain the concurrent operation of all algorithms may lead to a general increase in overall inference latency, thereby affecting the real-time performance of safety-critical algorithms. At this time, the operation and update of the lowest priority algorithm are automatically paused to release some computing resources to alleviate computing pressure. When the computing load falls below 90%, the operation and update of the lowest priority algorithm are resumed to ensure that the economic analysis task can be executed normally when computing power is sufficient. Through the above collaborative implementation, heterogeneous computing devices can simultaneously ensure the real-time performance of safety-critical algorithms, the timeliness of diagnostic algorithms, and the availability of economic algorithms under limited computing resources, realizing differentiated and dynamic scheduling and management of computing resources.

[0061] Further, see Figure 3 , Figure 3 This is a detailed step diagram of step S20 in one embodiment, based on the shown... Figure 3 The steps for constructing the comprehensive thermal risk safety entropy model include S21~S24: Step S21: Obtain the full-dimensional health status data of the battery cell output by the battery cell diagnostic algorithm set. The full-dimensional health status data of the battery cell includes at least the consistency safety entropy value and the battery cell SOH value. Step S22: Obtain real-time data on the temperature rise gradient of the battery cell tabs and the concentration of combustible gas in the compartment through a high-frequency acquisition device deployed at the energy storage power station site. Step S23: Based on the current operating conditions of the thermal power unit, determine the current weight coefficients for the four dimensions of parameters, namely, the temperature rise gradient data of the battery cell tabs, the consistency safety entropy value, the SOH value of the battery cell, and the concentration of combustible gas in the cabin, according to the preset weight adjustment strategy. The current operating conditions include the unit load rate and the frequency of AGC actions. Step S24: The four dimensional parameters are weighted and summed with their respective current weight coefficients to obtain the thermal risk comprehensive safety entropy model.

[0062] In this embodiment, within the safety intervention branch, a comprehensive thermal risk safety entropy model is first constructed. The core of this construction process lies in fusing multiple dimensions of cell state information to comprehensively reflect the degree of thermal risk of the cell.

[0063] The system acquires comprehensive health status data of battery cells from a set of cell diagnostic algorithms. These algorithms run in parallel on heterogeneous computing devices, and their output includes at least a consistency safety entropy value and a cell state of equilibrium (SOH) value. The consistency safety entropy value, output by a multi-scenario consistency safety entropy assessment algorithm, quantifies the degree of consistency among cells in parameters such as voltage, temperature, and internal resistance. A lower consistency safety entropy value indicates greater performance differences between cells, making them more prone to localized overcharging, over-discharging, or heat concentration. The cell state of equilibrium (SOH) value, output by a cell aging state SOH correction algorithm, reflects the percentage of the cell's current remaining capacity relative to its initial capacity. A lower SOH value indicates more severe aging, higher internal resistance, and higher heat generation under the same charge and discharge conditions. These two data points serve as two input dimensions to the comprehensive thermal risk safety entropy model, characterizing the causative factors of thermal risk from the perspective of the cell's internal state.

[0064] Data on the temperature rise gradient of the battery cell tabs and the concentration of combustible gases inside the storage compartment are acquired in real time by a high-frequency acquisition device deployed at the energy storage power station site. The high-frequency acquisition device collects the battery cell tab temperature and surface temperature at a sampling rate of 1 kHz, and calculates the temperature rise gradient of the battery cell tabs based on this temperature data. This temperature rise gradient reflects the rate of temperature increase of the battery cell per unit time; a larger temperature rise gradient indicates a faster rate of heat generation and a higher risk of heat accumulation. The concentration of combustible gases inside the storage compartment is collected by a gas sensor array, which monitors the concentration of carbon monoxide and hydrogen inside the storage compartment. These gases are typical precipitation products in the early stages of thermal runaway of the battery cell; an increase in their concentration indicates that abnormal reactions such as electrolyte decomposition or diaphragm damage may have occurred inside the battery cell, serving as important precursor signals for thermal runaway. These two types of data serve as two additional input dimensions to the comprehensive safety entropy model of thermal risk, characterizing the thermal risk from the perspectives of the external thermal behavior of the battery cell and gas precipitation.

[0065] Based on the current operating conditions of the thermal power units, the current weight coefficients for each of the four dimensions are determined according to a preset weight adjustment strategy. The current operating conditions of the thermal power units include the unit load rate and the frequency of AGC (Automatic Generator Control) actions. The unit load rate reflects the current output level of the thermal power unit. A higher load rate typically leads to a greater charging and discharging frequency and depth of the energy storage system, resulting in a faster rate of temperature rise and aging of the battery cells. Therefore, the weight coefficients for the temperature rise gradient dimension and the battery cell SOH (State of Health) value dimension should be increased accordingly. The frequency of AGC actions reflects the grid's requirements for the thermal power units' regulation capabilities. A higher AGC frequency requires the energy storage system to respond to frequency regulation commands more frequently, thus increasing the weight coefficient for the battery cell consistency safety entropy dimension. The preset weight adjustment strategy is stored in the system, containing the mapping relationship between the weight coefficients of the four dimensions under different combinations of unit load rates and AGC action frequencies. The system queries this mapping relationship based on the real-time monitored unit load rate and AGC action frequency to determine the weight coefficients for each of the four dimensions at the current moment.

[0066] The comprehensive thermal risk safety entropy model is obtained by weighting and summing four parameters—cell tab temperature rise gradient data, consistency safety entropy value, cell SOH value, and cabin combustible gas concentration data—with their respective current weighting coefficients. The calculation method is as follows: ; in This represents the temperature rise gradient data of the battery cell tabs obtained in step S22. This represents the consistency security entropy value obtained in step S21. SOH This represents the SOH value of the battery cell obtained in step S21. This represents the acquired data on the concentration of combustible gases inside the cabin. w 1. w 2. w 3. w 4 represents the current weighting coefficient for each cell. The cell's SOH value is expressed in the formula as 1. SOH The reason for including the form in the calculation is that a higher SOH value indicates a healthier cell and a lower thermal risk, while 1 SOH A higher value indicates a more severe degree of cell aging and a higher thermal risk. After the above transformation, all four dimensions are unified as positive indicators where a higher value signifies a higher thermal risk. The comprehensive safety entropy value is output by multiplying the four dimensions by their respective weighting coefficients and then summing the results. This is the calculation result of the comprehensive thermal risk safety entropy model. This result quantifies the comprehensive thermal risk level of the battery cell under the current operating conditions and is used as the basis for implementing graded intervention measures. Through the above implementation, a complete thermal risk modeling process was completed, from acquiring multi-source heterogeneous data, determining the adaptive weights for operating conditions, to weighted summation and fusion, providing a quantitative decision-making basis for the subsequent implementation of graded intervention measures.

[0067] Further, see Figure 4 , Figure 4 This is a schematic diagram illustrating another detailed step of step S20 in one embodiment, based on the shown... Figure 4 The step of using the net value of the thermal energy storage coordinated frequency regulation revenue minus the cell aging loss penalty as the reward function, and outputting the equalization current command and charge / discharge depth constraint through a deep reinforcement learning network, includes S25~S28: Step S25: Combine the cell SOH value, consistency safety entropy value, AGC frequency modulation command, real-time electricity price, current charge / discharge rate, and the working condition aging acceleration factor output by the cell aging state SOH correction algorithm into a reinforcement learning environment state vector; Step S26: Input the reinforcement learning environment state vector into the deep reinforcement learning network and calculate the action strategy that maximizes the long-term net benefit. The long-term net benefit is the cumulative power gain of the thermal power and energy storage coordinated frequency regulation minus the cumulative cell aging loss penalty. Step S27: Generate equalization current commands and upper and lower limits of charge / discharge depth constraints for each cell cluster according to the action strategy. Step S28: Output the equalization current command and charge / discharge depth constraint in the current control cycle. At the same time, use the actual frequency regulation revenue and cell aging loss data statistically collected daily as training samples to incrementally learn the deep reinforcement learning network and update the network weights. Adjust the equalization operation and maintenance cycle of subsequent control cycles according to the updated network weights.

[0068] In the optimized control branch, the net value of the thermal-storage coordinated frequency regulation revenue after deducting the cell aging loss penalty is used as the reward function. A deep reinforcement learning network outputs a balanced current command and charge / discharge depth constraints. In the specific implementation of the balanced command output by this deep reinforcement learning network, the cell SOH value, consistency safety entropy value, AGC frequency regulation command, real-time electricity price, current charge / discharge rate, and the operating condition aging acceleration factor output by the cell aging state SOH correction algorithm are combined into a reinforcement learning environment state vector. The cell SOH value and consistency safety entropy value are derived from the output data of the cell diagnostic algorithm set, reflecting the remaining lifespan percentage of the cell and the degree of performance consistency between cells, respectively. The AGC frequency regulation command originates from the automatic generation control command issued by the power grid dispatch center, which includes the target power value that the energy storage system needs to respond to at the current moment. The real-time electricity price comes from the electricity market trading system, reflecting the unit revenue of the energy storage system participating in frequency regulation services during the current period. The current charge / discharge rate comes from the real-time operating parameters of the energy storage converter, reflecting the current output intensity of the cell. The aging acceleration factor originates from the cell aging state SOH correction algorithm. This factor quantifies the accelerating effect of the current thermal power operating conditions on the cell aging rate; a larger value indicates a faster cell aging rate under the current operating conditions. The above six data points describe the current state of the system from different dimensions, including cell health status, grid dispatch demand, market profitability, current operating intensity, and the impact of aging conditions. The system arranges these six data points into a six-dimensional vector in a preset order. This vector is the reinforcement learning environment state vector, which fully contains all the state information required for the deep reinforcement learning network to make decisions.

[0069] The combined reinforcement learning environment state vector is input into a deep reinforcement learning network to calculate the action policy that maximizes the long-term net gain. This long-term net gain is the cumulative power gain from coordinated frequency regulation of thermal power and energy storage minus the cumulative cell aging loss penalty. The deep reinforcement learning network adopts a deep Q-learning network architecture, which includes an input layer, multiple fully connected hidden layers, and an output layer. The input layer receives the aforementioned six-dimensional environment state vector, the hidden layers extract state features through nonlinear transformations, and the output layer outputs the Q-values ​​corresponding to each candidate action. During the training phase, the network parameters are trained using historical running data, enabling the network to accurately estimate the long-term net gain obtainable by executing each candidate action under a given environment state. The calculation method for this long-term net gain is as follows: ; in Represents frequency modulation power. Represents real-time electricity price. To accumulate the power benefits of coordinated frequency regulation between thermal power and energy storage, This is the aging penalty coefficient. This is the aging acceleration factor output by the SOH correction algorithm for cell aging status. As the baseline aging rate, To mitigate the cumulative cell aging loss penalty, the network inputs the current environmental state vector into the network during decision-making, calculates the Q-value corresponding to each candidate action, and then selects the candidate action with the largest Q-value as the action strategy that maximizes long-term net benefits. This action strategy selects a decision scheme that achieves the optimal balance between long-term frequency regulation benefits and cell lifespan loss under the current cell health state, grid dispatch demand, and market revenue conditions.

[0070] The system generates equalization current commands and upper and lower limits for charge / discharge depth for each cell cluster based on the generated action strategy. The action strategy, presented as a high-dimensional vector, contains the target equalization current value and the allowable charge / discharge depth range for each cell cluster in the next control cycle. This action strategy vector is then parsed into independent commands for each cell cluster. For each cell cluster in the energy storage power station, the system extracts the target equalization current value corresponding to that cluster as the equalization current command. This command guides the cluster to perform charge / discharge operations at that current value in the next control cycle to achieve inter-cluster equalization. Simultaneously, the system extracts the upper and lower limits for charge / discharge depth as constraints for that cluster. The upper limit specifies the highest allowable state of charge (SOC) value for the cluster during charging, and the lower limit specifies the lowest allowable SOC value during discharging. By setting different charge / discharge depth constraints for cell clusters in different health states, the system can implement differentiated protection for cell clusters with higher aging levels and poor consistency during charge / discharge scheduling, preventing these cell clusters from accelerating degradation due to deep charge / discharge.

[0071] The equalization current command and charge / discharge depth constraints generated during the current control cycle are sent to the power conversion system via a three-level independent hardware command channel for execution. Simultaneously, using daily statistical data on actual frequency regulation gains and cell aging losses as training samples, the deep reinforcement learning network undergoes incremental learning and its weights are updated. The equalization maintenance cycle for subsequent control cycles is adjusted based on the updated network weights. Within the current control cycle, the equalization current command and charge / discharge depth constraints are sent to the power conversion system via the aforementioned three-level command channel, and the power conversion system executes the charge / discharge operations according to these commands. During execution, the actual frequency regulation power gains obtained within the control cycle and the aging losses generated by the cells during the actual execution are recorded in real time. At the end of each daily statistical cycle, the sum of the actual frequency regulation gains and the sum of the actual cell aging losses for all control cycles of the day are summarized, and this statistical data is used as a set of training samples. When the deep reinforcement learning network enters the incremental learning phase, the aforementioned training samples are used to train the network, and the network weights are fine-tuned using the backpropagation algorithm to make the network's estimation of long-term net gains more accurate. After the network weights are updated, the system evaluates whether the current balanced maintenance cycle is still optimal based on the updated network weights. If the evaluation results indicate that extending or shortening the balanced maintenance cycle can further improve long-term net benefits, the system adjusts the balanced maintenance cycle parameters for subsequent control cycles accordingly. Through the above incremental learning mechanism, the deep reinforcement learning network can continuously optimize its decision-making ability during continuous operation, and the balancing strategy can adapt to the aging process of battery cells and changes in operating conditions, realizing a technological leap from static fixed to dynamic adaptive balancing strategies.

[0072] Furthermore, you can also view Figure 5 , Figure 5 This is a flowchart illustrating the second embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative cells according to the present invention. Figure 5 After the step of running the set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site, the method further includes steps S40~S80: Step S40: The NPU of the heterogeneous computing device locally stores the daily collected first frequency cell timing samples and thermal power operating condition samples. Step S50: During the idle time slot of the NPU, the daily collected samples are used to perform incremental fine-tuning on the model weights of the battery cell diagnostic algorithm set, while keeping the main inference logic of the algorithm unchanged. Step S60: After the incremental fine-tuning is completed, verify whether the SOH estimation error and false alarm rate of the fine-tuned model meet the standards in the local simulation of the NPU. Step S70: After verification, switch the fine-tuned model weights to the current running version without disturbance, and retain the model weights of the previous version as a rollback backup. In step S80, during the incremental fine-tuning process performed locally by the NPU, the real-time inference and control functions of the heterogeneous computing device continue to operate uninterruptedly.

[0073] In this embodiment, the heterogeneous computing device runs the cell diagnostic algorithm set in parallel using a hardware-level dynamic priority scheduling method. Then, it further utilizes the NPU's local storage and idle computing power to perform an incremental learning task, enabling the model weights of the cell diagnostic algorithm set to be continuously optimized during continuous operation. The specific implementation process of this incremental learning task is as follows: The NPU of the heterogeneous computing device locally stores daily-collected cell time-series samples and thermal power plant operating condition samples at the first frequency. During the real-time execution of the cell diagnostic algorithm set on the heterogeneous computing device, the system continuously collects cell time-series data and thermal power plant operating condition data at a preset first frequency. The cell time-series data includes the voltage, current, tab temperature, and surface temperature of each cell at each sampling moment. The thermal power plant operating condition data includes the real-time load rate, AGC activation frequency, and peak shaving depth of the thermal power unit. After collection, the above data is indexed according to the sampling time to form a time-series sample sequence. Due to the large volume and high sampling frequency of the daily sampled data, it is difficult to store all of the high-frequency data. Therefore, the system performs equal-interval thinning of the original sampled data at the preset first frequency and stores it in the NPU's local solid-state storage unit. The daily stored time-series samples and thermal power plant operating condition samples constitute the daily incremental training dataset. This training dataset retains key information on the cell status change trend and thermal power plant operating condition change pattern from the original data, while keeping the data volume within the acceptable range of the NPU's local storage capacity.

[0074] During the NPU's idle time slots, the stored daily collected samples are used to incrementally fine-tune the model weights of the cell diagnostic algorithm set, while maintaining the main inference logic of the algorithms unchanged. The NPU is not always operating at full capacity during daily operations; its computational load is relatively low during nighttime or periods with sparse frequency modulation commands. These periods are referred to as idle time slots. The system initiates an incremental fine-tuning task during these idle time slots, using the stored daily incremental samples to perform incremental fine-tuning on each algorithm model in the cell diagnostic algorithm set. This incremental fine-tuning algorithm is based on parameter fine-tuning techniques from transfer learning. It iteratively updates the model weights using new training samples, based on the original model weights, allowing the model to adapt to the cell characteristics and operating condition changes reflected by the new samples while retaining its original knowledge. During incremental fine-tuning, only the numerical values ​​of the model weights are updated; the architecture and computational flow of the main inference logic of the algorithms remain unchanged. The main inference logic of the algorithms includes structural algorithmic architectures such as the number of network layers, activation function types, and loss function definitions. These architectures remain unchanged, ensuring that the incrementally fine-tuned model is fully compatible with the original model in terms of interface and behavior. Through the aforementioned incremental fine-tuning mechanism, the cell diagnostic algorithm set can gradually approach the optimal model parameters after changes in cell characteristics and operating conditions by continuously collecting new samples, without changing the basic algorithm architecture.

[0075] After incremental fine-tuning, the system verifies whether the SOH estimation error and false alarm rate of the fine-tuned model meet the preset standards through local simulation on the NPU. During incremental fine-tuning, the model weights change; whether this change leads to improved model performance needs to be verified. A portion of incremental samples is allocated locally on the NPU as a validation set, which is not used in incremental fine-tuning training. After fine-tuning, the validation set is input into the fine-tuned model for inference, and the error between the model's output SOH estimate and the actual value, as well as the false alarm rate of the model's output warning signal, are calculated. The SOH estimation error and false alarm rate are compared with preset standards, which can be quantifiable indicators such as the SOH estimation error not exceeding a certain percentage and the false alarm rate not exceeding a certain percentage. If both the error and false alarm rate do not exceed the corresponding preset standards, the fine-tuned model is deemed to have passed verification. Through this local simulation verification mechanism, the system can confirm whether the performance of the fine-tuned model meets expectations before deploying it in actual operation, preventing the incorrect deployment of models that do not meet performance standards.

[0076] After successful verification, the fine-tuned model weights are seamlessly switched to the current running version, while the previous version's model weights are retained as a rollback backup. Seamless switching means that the switching process does not affect the real-time inference and control tasks being performed on the heterogeneous computing device. To achieve seamless switching, a dual-weight storage area mechanism is adopted, dividing the NPU's storage space into an active area and a spare area. The active area stores the currently running model weights, while the spare area stores the newly fine-tuned model weights. Once verification is successful, the fine-tuned model weights in the spare area are marked as pending activation, and the switching action is performed after the current control cycle ends. During the switch, the spare area is switched to the active area, and the original active area is switched to the spare area. The entire switching process involves only changes to pointers or flags, without loading or unloading any data. After the switch is complete, the new model weights take effect immediately. Simultaneously, the previous version's model weights (i.e., the model weights before the switch) are retained in the spare area as a rollback backup. When the fine-tuned model encounters anomalies during actual operation, it can be quickly rolled back to the previous version through the same seamless switching mechanism, ensuring that the system operation is not interrupted due to model anomalies.

[0077] Throughout the entire process of incremental tuning performed locally on the NPU, the real-time inference and control functions of the heterogeneous computing device operate uninterruptedly. The incremental tuning process is executed during the NPU's idle time slots, but even during idle time slots, the NPU may still receive real-time inference and control tasks, such as frequency modulation anomaly warning requests or AGC power scheduling instructions. To ensure that these real-time tasks are not affected by incremental tuning tasks, a task priority management mechanism is implemented in the NPU. Real-time inference and control tasks are configured with the highest priority, while incremental tuning tasks are configured with a lower priority. When a real-time task arrives, the NPU pauses the incremental tuning task, switches computing resources to the real-time task, and resumes execution of the incremental tuning task after the real-time task is completed. Therefore, from the outside perspective, the real-time inference and control functions of the heterogeneous computing device always maintain an uninterrupted operating state. Through implementation, the set of cell diagnostic algorithms was continuously optimized without stopping the machine or affecting real-time control. The model accuracy of the set of cell diagnostic algorithms improved continuously with the extension of running time. The SOH estimation error and the false alarm rate of the warning both showed a significant decrease after six months of continuous operation, effectively avoiding the accuracy drift problem caused by cell aging and changes in operating conditions in the static deployment model.

[0078] Furthermore, you can also view Figure 6 , Figure 6 This is a flowchart illustrating the third embodiment of the intelligent management and control method for the entire lifecycle of energy storage collaborative cells according to the present invention. Figure 6After the step of feeding back the feedback data generated by the safety intervention branch after implementing graded intervention measures and the feedback data generated by the optimization control branch after implementing equalization current commands and charge / discharge depth constraints to the heterogeneous computing device, the method further includes steps S90 to S130: Step S90: Enter candidate operation and maintenance schemes in the digital twin interface. The candidate operation and maintenance schemes include balancing cycle parameters, cell replacement threshold parameters, and black start backup capacity parameters. Step S100: The cell aging model and multi-scenario consistency model are called through the digital twin simulation engine to simulate the operation effect of the candidate operation and maintenance scheme under the thermal power operating conditions throughout the year. Step S110: Output the cell attenuation prediction curve, frequency regulation benefit calculation value and total operation and maintenance cost corresponding to each candidate operation and maintenance solution. Step S120: Automatically recommend the operation and maintenance plan with the highest net benefit and simultaneously generate a virtual maintenance simulation animation; Step S130: Based on the operation and maintenance plan, collect actual operation data and feed it back into the digital twin simulation engine to iteratively correct the simulation model parameters.

[0079] In this embodiment, feedback data generated by the safety intervention branch after implementing tiered intervention measures and feedback data generated by the optimization control branch after implementing balancing current commands and charge / discharge depth constraints are fed back into the heterogeneous computing device. Furthermore, the fed-back data is used to drive a digital twin simulation engine to simulate and optimize the operation and maintenance plan.

[0080] Candidate maintenance and operation (O&M) plans are entered into the digital twin interface. These plans include balancing cycle parameters, cell replacement threshold parameters, and black-start reserve capacity parameters. The digital twin interface serves as the human-machine interface between O&M personnel and the digital twin simulation engine. O&M personnel select the parameters included in the candidate O&M plan within the interface. The balancing cycle parameter specifies the time interval for the energy storage system to periodically perform cell balancing operations, determining the frequency of these operations. The cell replacement threshold parameter specifies the level at which the cell's state of harmonic equilibrium (SOH) value must be replaced, directly determining the cell's actual lifespan and replacement cost. The black-start reserve capacity parameter specifies the proportion of reserve capacity reserved by the energy storage system to cope with black-start scenarios. This parameter does not participate in frequency regulation services during normal system operation; it is only used as an emergency power source during black starts. O&M personnel input the value ranges of the above three parameters through interface controls. Multiple candidate values ​​can be entered for each parameter, and these values ​​can be freely combined to form multiple candidate O&M plans. The system stores the candidate operation and maintenance solutions entered by the operation and maintenance personnel in the input buffer of the digital twin simulation engine, waiting for the simulation to be executed.

[0081] The digital twin simulation engine invokes the cell aging model and the multi-scenario consistency model to simulate the operational effects of the aforementioned candidate maintenance solutions under year-round thermal power plant operating conditions. The digital twin simulation engine is a simulation computing platform deployed on a plant-level server or in the cloud, which incorporates the cell aging model and the multi-scenario consistency model. The cell aging model reuses the calculation logic of the SOH correction algorithm for cell aging state from the cell diagnostic algorithm set. This model can simulate and calculate the SOH decay curve of the cell over a specific time period based on inputs such as the cell's charge / discharge rate, charge / discharge depth, cumulative cycle count, and thermal power plant operating parameters. The multi-scenario consistency model reuses the calculation logic of the multi-scenario consistency safety entropy assessment algorithm from the cell diagnostic algorithm set. This model can simulate and calculate the evolution trend of the consistency safety entropy value between cells over time based on differences in voltage, temperature, and state of charge between cells. The simulation engine takes the balancing cycle parameters, cell replacement threshold parameters, and black start reserve capacity parameters from the candidate operation and maintenance schemes as input conditions, and uses the thermal power plant operating condition time series contained in the historical operation data of the reinjection as the operating condition input. It simulates the system operation process under the above candidate operation and maintenance schemes on a year-round time scale. During the simulation, the cell aging model updates the SOH value of each cell according to the balancing operation and the change of charge and discharge rate, while the multi-scenario consistency model updates the consistency safety entropy value according to the balancing effect and the aging differences of each cell. The two models are coupled to reflect the performance evolution of the system during the year-round operation.

[0082] The simulation engine outputs the cell attenuation prediction curve, the calculated frequency regulation revenue for each scenario, and the total operation and maintenance cost for each candidate operation and maintenance scheme. For each candidate operation and maintenance scheme, the simulation engine generates three quantitative outputs after completing the year-long simulation. The cell attenuation prediction curve, with time on the horizontal axis and the average cell SOH (State of Health) on the vertical axis, reflects the average aging trend of all cells in the energy storage power station during the year's operation. This curve can be used to assess the impact of different schemes on cell lifespan. The calculated frequency regulation revenue for each scenario divides the year into typical and atypical frequency regulation scenarios, and calculates the frequency regulation power revenue for each scenario. This value can be used to assess the impact of different schemes on frequency regulation service revenue. The total operation and maintenance cost is calculated by summarizing the balanced operation manpower cost, cell replacement material cost, and opportunity cost of black start reserve capacity occupancy. This total cost can be used to assess the impact of different schemes on operation and maintenance investment. The simulation engine stores the above three quantitative outputs in the simulation results database according to the candidate operation and maintenance scheme number.

[0083] Based on the cell attenuation prediction curves, scenario-specific frequency regulation revenue calculations, and total operation and maintenance costs corresponding to each candidate operation and maintenance scheme, the system automatically recommends the operation and maintenance scheme with the highest net benefit and simultaneously generates a virtual maintenance simulation animation. The net benefit is calculated by subtracting the total operation and maintenance cost from the scenario-specific frequency regulation revenue calculation, while also considering deductions after converting the lifespan loss reflected in the cell attenuation prediction curve into monetary value. The simulation engine iterates through all candidate operation and maintenance schemes, calculates the net benefit value for each group of schemes, and marks the scheme with the highest net benefit value as the recommended scheme. This recommended scheme and its corresponding quantitative indicators are highlighted in the digital twin interface for operation and maintenance personnel to view and confirm. Simultaneously, the simulation engine calls the 3D visualization module to generate a virtual maintenance simulation animation based on the annual simulation process data of the recommended scheme. This animation, in the form of a 3D virtual scene, demonstrates the changes in cell aging distribution during the annual operation of the energy storage power station under the recommended scheme, as well as the execution location and execution time of each balancing operation and cell replacement operation in the 3D scene. By watching the animation, operation and maintenance personnel can intuitively understand the specific implementation process of the recommended scheme throughout the year.

[0084] Based on the recommended operation and maintenance (O&M) scheme, actual operational data is collected and fed back into the digital twin simulation engine to iteratively correct the simulation model parameters. After the O&M personnel confirm the adoption of the recommended scheme, the system deploys the various parameters (balancing cycle, cell replacement threshold, black start backup capacity) issued by the scheme to the actual operating system. During actual operation, actual cell attenuation data and actual frequency regulation benefit data are continuously collected, and these actual data are compared item by item with the predicted data output by the simulation engine for the scheme. If there is a deviation between the actual data and the predicted data, the deviation is used as a correction signal input to the digital twin simulation engine to iteratively correct the parameters in the cell aging model and the multi-scenario consistency model, so that the prediction accuracy of the simulation engine is continuously improved in the feedback of actual operation. Through the above implementation, a complete O&M simulation closed loop is realized, from scheme input to simulation, from quantitative output to automatic recommendation, from scheme implementation to data feedback correction. This transforms O&M decision-making from reactive post-event response to proactive pre-event planning. The accuracy rate of predicting the effect of the O&M scheme reaches 92% after multiple iterations, and the total O&M cost is reduced by 22% compared with the traditional method.

[0085] This application provides an intelligent management and control device for the entire life cycle of energy storage collaborative battery cells. The intelligent management and control device for the entire life cycle of energy storage collaborative battery cells includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the intelligent management and control method for the entire life cycle of energy storage collaborative battery cells in the above embodiment 1.

[0086] The following is for reference. Figure 7The diagram illustrates a structural schematic suitable for implementing the intelligent management and control device for the entire lifecycle of energy storage collaborative battery cells in the embodiments of this application. The intelligent management and control device for the entire lifecycle of energy storage collaborative battery cells in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The energy storage collaborative battery cell full life cycle intelligent management and control device shown is merely an example and should not impose any limitations on the function and scope of use of the embodiments of this application.

[0087] like Figure 7 As shown, the intelligent management and control device for the entire life cycle of the energy storage collaborative battery cell may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1002 or the program loaded from the storage device 1003 into the random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the intelligent management and control device for the entire life cycle of the energy storage collaborative battery cell. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the energy storage collaborative cell lifecycle intelligent management and control device to exchange data with other devices wirelessly or via wired communication. Although the figure shows an energy storage collaborative cell lifecycle intelligent management and control device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0088] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0089] The intelligent management and control device for the entire lifecycle of energy storage collaborative cells provided in this application adopts the intelligent management and control method for the entire lifecycle of energy storage collaborative cells in the above embodiments. It can solve the technical problems of isolated data and lack of closed-loop linkage in existing thermal energy storage collaborative systems, which cannot simultaneously take into account the dynamic balance between thermal safety, control accuracy and lifecycle benefits. Compared with the prior art, the beneficial effects of the intelligent management and control device for the entire lifecycle of energy storage collaborative cells provided in this application are the same as the beneficial effects of the intelligent management and control method for the entire lifecycle of energy storage collaborative cells provided in the above embodiments. Moreover, other technical features in the intelligent management and control device for the entire lifecycle of energy storage collaborative cells are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0090] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0092] This application provides a storage medium, which is a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon. The computer-readable program instructions are used to execute the intelligent management and control method for the entire life cycle of energy storage collaborative battery cells in the above embodiments.

[0093] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0094] The aforementioned computer-readable storage medium may be included in the intelligent management and control device for the entire life cycle of energy storage collaborative cells; or it may exist independently and not be assembled into the intelligent management and control device for the entire life cycle of energy storage collaborative cells.

[0095] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the intelligent management and control device for the entire life cycle of energy storage collaborative cells, the intelligent management and control device for the entire life cycle of energy storage collaborative cells implements the technical content of the above-described embodiment of the intelligent management and control method for the entire life cycle of energy storage collaborative cells.

[0096] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0098] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0099] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described intelligent management and control method for the entire life cycle of energy storage collaborative cells. This solves the technical problems of isolated data and lack of closed-loop linkage in existing thermal energy storage collaborative systems, which cannot simultaneously balance thermal safety, control accuracy, and lifespan benefits. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the intelligent management and control method for the entire life cycle of energy storage collaborative cells provided in the above embodiments, and will not be repeated here.

Claims

1. A method for intelligent management and control of the entire lifecycle of energy storage collaborative battery cells, characterized in that, Includes the following steps: By deploying heterogeneous computing devices at the energy storage power station site, a set of cell diagnostic algorithms is run in parallel using a hardware-level dynamic priority scheduling method. The output data of the set of cell diagnostic algorithms includes cell SOH value, consistency safety entropy value, and cell full-dimensional health status data of abnormal risk level. Based on the comprehensive health status data of the battery cells, a safety intervention branch and an optimization control branch are executed in parallel. The safety intervention branch includes: constructing a comprehensive thermal risk safety entropy model, calculating the thermal risk level based on the comprehensive thermal risk safety entropy model, and executing graded intervention measures according to different threshold ranges of the thermal risk level. The graded intervention measures include conventional heat dissipation, enhanced liquid cooling heat dissipation, inert gas pre-charge suppression, and zoned spray power cutoff. The optimization control branch includes: using the comprehensive health status data of the battery cells as the reinforcement learning environment state, using the net value of the thermal energy storage coordinated frequency regulation benefit minus the battery cell aging loss penalty as the reward function, and outputting equalization current commands and charge / discharge depth constraints through a deep reinforcement learning network; issuing the equalization current commands and charge / discharge depth constraints through three independent hardware command channels: microsecond-level channel, millisecond-level channel, and second-level channel; collecting the actual output value of the power conversion system in real time, generating a compensation amount based on the power deviation between the issued command and the actual output value, and superimposing it into the command of the next control cycle, wherein the correction coefficient of the compensation amount is dynamically adjusted according to the SOH value of the battery cells. The feedback data generated after the safety intervention branch executes the graded intervention measures and the feedback data generated after the optimization control branch executes the equalization current command and charge / discharge depth constraint are fed back to the heterogeneous computing device to perform incremental learning correction on the battery cell diagnostic algorithm set and the deep reinforcement learning network.

2. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, The step of running a set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site includes: The set of battery cell diagnostic algorithms is divided into three priority levels according to the degree of safety criticality. The frequency modulation anomaly early warning algorithm and the thermal runaway prediction algorithm are set as the highest priority, allocated the first computing power ratio, and the inference latency is limited to not exceeding the first latency threshold. The SOH correction algorithm for cell aging status and the multi-scenario consistency security entropy assessment algorithm are set as the second highest priority, allocated the second computing power ratio, and the inference latency is limited to not exceed the second latency threshold. Set the revenue statistics and monthly lifespan simulation algorithm as the lowest priority, allocate the third computing power ratio, and limit the inference latency to no more than the third latency threshold. When the highest priority algorithm is triggered and runs, the NPU scheduling kernel of the heterogeneous computing device performs hardware-level preemption to automatically reduce the operation frequency of the lowest priority algorithm. When the overall computing load of the heterogeneous computing device exceeds the first load threshold, the operation and update of the lowest priority algorithm are automatically paused until the computing load falls back below the first load threshold and then resumed.

3. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, Following the step of running the set of cell diagnostic algorithms in parallel using a hardware-level dynamic priority scheduling method through heterogeneous computing devices deployed at the energy storage power station site, the method further includes: The NPU of the heterogeneous computing device locally stores daily collected first-frequency battery cell timing samples and thermal power operating condition samples. During the idle time slots of the NPU, the daily collected samples are used to perform incremental fine-tuning on the model weights of the battery cell diagnostic algorithm set, while keeping the main inference logic of the algorithm unchanged; After the incremental fine-tuning is completed, the local simulation on the NPU is used to verify whether the SOH estimation error and false alarm rate of the fine-tuned model meet the standards. After verification, the fine-tuned model weights are switched to the current running version without disturbance, and the model weights of the previous version are retained as a rollback backup. During the incremental fine-tuning process performed locally by the NPU, the real-time inference and control functions of the heterogeneous computing device continue to operate uninterruptedly.

4. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, The steps for constructing the comprehensive thermal risk safety entropy model include: Obtain the full-dimensional health status data of the battery cell output by the set of battery cell diagnostic algorithms. The full-dimensional health status data of the battery cell includes at least the consistency safety entropy value and the battery cell SOH value. Acquire real-time data on the temperature rise gradient of the battery cell tabs and the concentration of combustible gas inside the compartment through high-frequency acquisition devices deployed at the energy storage power station site; Based on the current operating conditions of the thermal power unit, the current weight coefficients of the four dimensions of parameters—cell tab temperature rise gradient data, consistency safety entropy value, cell SOH value, and cabin combustible gas concentration data—are determined according to a preset weight adjustment strategy. The current operating conditions include the unit load rate and AGC operation frequency. The four dimensional parameters are weighted and summed with their respective current weight coefficients to obtain the comprehensive thermal risk safety entropy model.

5. The intelligent management and control method for the entire life cycle of energy storage collaborative cells as described in claim 1, characterized in that, The step of implementing graded intervention measures based on different threshold ranges of the heat risk level includes: When the thermal risk level is in the first threshold range, the normal heat dissipation mode is activated to maintain the current operating parameters of the liquid-cooled or air-cooled unit. When the thermal risk level rises to the second threshold range, switch to enhanced liquid cooling mode to increase the coolant circulation flow and the speed of the cooling unit. When the thermal risk level rises to the third threshold range, the inert gas pre-charge suppression mode is activated to pre-charge inert gas into the energy storage chamber to reduce the oxygen concentration, while automatically reducing the charge and discharge rate of high-risk cells. When the thermal risk level rises to the fourth threshold range, the zoned spray mode is activated and the power output of the faulty battery cell cluster is cut off. Simultaneously, the thermal risk area is highlighted on the digital twin interface and a thermal maintenance work order is generated.

6. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, The step of using the net value of the revenue from coordinated frequency regulation of thermal power and energy storage after deducting the penalty for cell aging loss as the reward function, and outputting equalization current commands and charge / discharge depth constraints through a deep reinforcement learning network, includes: The cell SOH value, consistency safety entropy value, AGC frequency modulation command, real-time electricity price, current charge / discharge rate, and the working condition aging acceleration factor output by the cell aging state SOH correction algorithm are combined into a reinforcement learning environment state vector. The reinforcement learning environment state vector is input into the deep reinforcement learning network to calculate the action strategy that maximizes the long-term net benefit, where the long-term net benefit is the cumulative thermal power and energy storage coordinated frequency regulation power benefit minus the cumulative cell aging loss penalty. Based on the aforementioned action strategy, equalization current commands and upper and lower limits of charge / discharge depth constraints are generated for each cell cluster. In the current control cycle, the equalization current command and charge / discharge depth constraint are output. At the same time, the actual frequency regulation revenue and cell aging loss data collected daily are used as training samples to incrementally learn the deep reinforcement learning network and update the network weights. The equalization operation and maintenance cycle of subsequent control cycles is adjusted according to the updated network weights.

7. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, The step of issuing the equalization current command and charge / discharge depth constraint through three independent hardware command channels at the microsecond, millisecond, and second levels includes: The equalization current command and the charge / discharge depth constraint command are divided into three levels of sub-commands according to the response speed requirements; A frequency modulation instantaneous power adjustment sub-command is issued through a microsecond-level channel composed of field-programmable gate array hard logic, and the channel refresh cycle does not exceed the first cycle threshold. AGC power scheduling sub-instructions are issued through a millisecond-level channel formed by a hard real-time controller, with the channel refresh cycle being the second cycle threshold. The power plant operating condition switching and start / stop control sub-instructions are issued through a second-level channel composed of the ARM main core, and the channel refresh cycle is the third cycle threshold. Among them, the microsecond-level channel, millisecond-level channel and second-level channel share a unified clock reference based on a precise time protocol, and the timing deviation between the sub-instructions issued by each level of channel does not exceed the first deviation threshold.

8. The intelligent management and control method for the entire life cycle of energy storage collaborative battery cells as described in claim 1, characterized in that, After the step of feeding back the feedback data generated by the safety intervention branch after executing the graded intervention measures and the feedback data generated by the optimization control branch after executing the equalization current command and charge / discharge depth constraint to the heterogeneous computing device, the method further includes: Enter candidate operation and maintenance solutions into the digital twin interface. The candidate operation and maintenance solutions include balancing cycle parameters, cell replacement threshold parameters, and black start backup capacity parameters. The digital twin simulation engine is used to call the cell aging model and the multi-scenario consistency model to simulate the operating effect of the candidate operation and maintenance scheme under the thermal power operating conditions throughout the year. Output the cell attenuation prediction curve, frequency regulation benefit calculation value and total operation and maintenance cost for each candidate operation and maintenance solution; Automatically recommends the operation and maintenance solution with the highest net benefit, and simultaneously generates virtual maintenance simulation animations; Based on the aforementioned operation and maintenance plan, actual operating data is collected and fed back into the digital twin simulation engine to iteratively correct the simulation model parameters.

9. A smart management and control device for the entire life cycle of energy storage and collaborative battery cells, characterized in that, The energy storage collaborative battery cell full life cycle intelligent management and control device stores a computer program, which, when executed by a processor, implements the energy storage collaborative battery cell full life cycle intelligent management and control method according to any one of claims 1-8.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the intelligent management and control method for the entire life cycle of energy storage collaborative cells as described in any one of claims 1-8.

Images