Distributed energy storage power station autonomous coordination system and method based on cloud edge agent

Abstract

The application discloses a cloud-edge-end intelligent agent-based distributed energy storage power station autonomous cooperation system and method, relates to the technical field of energy storage control of power systems, and constructs a cloud-edge-end three-layer intelligent agent architecture, cloud-end global optimization and federal learning, edge real-time decision and PBFT consensus, terminal millisecond-level protection and data acquisition, realizes distributed autonomous and continuous evolution of energy storage power stations through multi-intelligent agent reinforcement learning and time scale layered cooperation, and solves the technical problems of high response delay, single-point fault risk, poor expansibility, data island and privacy leakage of the existing centralized control architecture.

Description

Technical Field

[0001] This invention relates to the field of power system energy storage control technology, and in particular to a distributed energy storage power station autonomous collaborative system and method based on cloud-edge-device intelligent agents. Background Technology

[0002] With the accelerated construction of new power systems and the gradual implementation of large-scale grid-side energy storage power stations, stringent requirements are being placed on energy storage control systems across multiple dimensions. These requirements include millisecond-level response capabilities to adapt to grid frequency regulation, the ability to achieve full-site panoramic perception at millions of monitoring points, multi-objective collaborative optimization to improve power station operational profitability, and the ability to adapt to various complex operating conditions to ensure safe and stable equipment operation. Currently, mainstream energy storage control systems still adopt a centralized architecture, which has several inherent shortcomings: data is uniformly aggregated to a central controller for computation before control commands are issued, resulting in control delays of hundreds of milliseconds to seconds due to communication and computing bottlenecks, failing to meet the millisecond-level response standard for grid frequency regulation; a failure of the central controller can directly lead to a complete loss of control of the entire station, indicating weak system fault tolerance; limited architecture expansion makes it difficult to support the smooth expansion of gigawatt-level energy storage clusters and is prone to communication storms; data from different energy storage sites are isolated, forming data silos and hindering global collaborative optimization; and all original operating data from the sites must be uploaded to the cloud, posing a data privacy and security risk.

[0003] Emerging digital technologies such as multi-agent systems, federated learning, and edge computing have been gradually applied to the power industry. Among them, multi-agent systems rely on distributed collaborative decision-making to achieve task decomposition and have the advantages of self-organization and high robustness. Federated learning supports multiple parties to complete joint model training on the premise of retaining the original data locally, which can avoid the problem of data privacy leakage. Edge computing pushes computing power down to the edge nodes of the field, which can significantly shorten the control response latency.

[0004] At present, most of the relevant technology implementation solutions only adopt one of the above technologies, lacking a complete control architecture that integrates multiple technologies, and cannot simultaneously meet the comprehensive business needs of large-scale energy storage power stations in terms of real-time response, operational reliability, and intelligent collaboration. Summary of the Invention

[0005] The purpose of this invention is to provide an autonomous collaborative system and method for distributed energy storage power stations based on cloud-edge-device intelligent agents, which solves the technical problems of high response latency, single point of failure risk, poor scalability, data silos and privacy leakage in the existing centralized control architecture.

[0006] To achieve the above objectives, this invention provides a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents, comprising: The cloud-based intelligent hub, deployed in a data center or cloud platform, is used for global optimization, long-term prediction, strategy generation, and AI model training. Edge computing nodes, deployed at power plant sites or regional aggregation points, are responsible for real-time data processing, local optimization decision-making, rapid response control, and localized execution of cloud strategies, compressing control latency to within 50ms. Terminal execution unit, embedded in BMS, PCS, and sensor field devices, is used for millisecond-level data acquisition, device-level protection control, and status information reporting; The cloud-based intelligent hub, edge computing nodes, and terminal execution units constitute a three-layer intelligent agent architecture. The agents make collaborative decisions through a multi-agent reinforcement learning framework. The agents are trained under the guidance of cloud models, and the edge devices make distributed decisions based on local information. The global optimal power allocation is achieved through communication.

[0007] Preferably, the cloud-based intelligent hub includes a federated learning engine. The federated learning engine is used to maintain a global prediction and optimization model, with copies of the global prediction and optimization model deployed locally at each site. Each site uses local running data to train the model copy through forward inference and backpropagation. The original data does not leave the local site; only the updated model weight parameters are uploaded to the cloud-based intelligent hub. The federated learning engine uses the FedAvg aggregation algorithm to perform a weighted average of the weights uploaded by each site, generating an updated global model and redistributing it to each site. This process is repeated iteratively until the global model converges on the validation set, collaboratively training a better global prediction and optimization model, thereby solving the problems of data silos and privacy. The expression for weighted averaging of uploads from each site using the FedAvg aggregation algorithm is as follows: ; In the formula, Indicates the first Round global model weights; Indicates the total number of stations participating in the aggregation; Indicates station The number of local samples; This represents the total number of samples from all participating parties. ; Indicates station In the Round global model weights Based on this, the updated local weights are obtained after multiple rounds of local training.

[0008] Preferably, edge computing nodes use the Practical Byzantine Fault Tolerance (PBFT) protocol for consensus negotiation. The PBFT protocol includes five phases: Request, Pre-Prepare, Prepare, Commit, and Reply. Edge computing nodes conduct consensus voting through a three-phase voting mechanism formed by Pre-Prepare, Prepare, and Commit, ensuring that the consensus is not greater than [a certain threshold]. Even if a fault-tolerant node fails or communication is interrupted, the system can still reach a consensus on critical security instructions; the total number of edge computing nodes participating in consensus negotiation satisfy .

[0009] Preferably, the terminal execution unit includes a BMS and a PCS with an embedded AI inference chip. The BMS is used to evaluate the health status and charge / discharge constraints of the battery module it monitors in real time, and the PCS is used to evaluate the health status and conversion efficiency preference of its own power conversion device in real time.

[0010] Preferably, the three-layer intelligent agents make collaborative decisions through a multi-agent reinforcement learning framework, execute collaborative responses according to a time-scale hierarchical mechanism, and achieve interaction through an intelligent agent collaboration mechanism; Multi-agent reinforcement learning frameworks include: Actor networks output action strategies based on the current state; Critic network, evaluating the state-action value function; An experience replay buffer stores historical interaction data (including state, action, reward, next state, and termination flag) for training. A consistency coordination mechanism controls the power difference between agents within a preset threshold. The time-scale stratification mechanism is as follows: Millisecond level, <10ms, protection actions and PWM modulation are handled by the terminal execution unit; Power tracking is completed in milliseconds, from 10ms to 100ms, by the collaboration between edge computing nodes and terminal execution units. The timeframe is within seconds, ranging from 1 to 10 seconds, with regional coordination completed by edge computing nodes. For timeframes of minutes and above, strategy optimization is completed collaboratively by the cloud-based intelligent hub and edge computing nodes. Intelligent agent cooperation mechanisms include: Task decomposition and allocation: The Contract Network protocol is adopted. The task initiator publishes a tender notice, and each intelligent agent bids after assessing its own capabilities and costs. The initiator selects the successful bidder through comprehensive evaluation and forms a task allocation contract. Information interaction and negotiation: Data distribution is based on a publish-subscribe mechanism. The agent sends a subscription request carrying a specified data topic identifier. The data publisher pushes topic data in a targeted manner according to the subscription matching result, reducing the number of broadcast messages transmitted across the network and avoiding global message broadcasting. Conflict resolution and consistency maintenance: The power configuration parameters of each agent are adjusted through a consistency coordination mechanism to control the power difference between agents within a preset threshold.

[0011] This invention also provides a method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents, comprising the following steps: S1. The terminal execution unit periodically collects energy storage data and performs local protection and periodic status reporting; The terminal execution unit collects battery electrical quantities, status quantities, and environmental quantity data at a period of ≤1ms and performs data preprocessing. When any parameter of voltage, current, or temperature exceeds its respective preset threshold, a three-level protection mechanism is triggered, and the processed status information is reported to the corresponding edge computing node at a period of 10ms to 100ms. S2, edge computing nodes complete data verification, power decision-making, and regional consistency collaborative allocation; Edge computing nodes aggregate the status data reported by the terminal execution units under their jurisdiction and fuse them to construct local status vectors; receive global power targets issued by the cloud intelligent hub or calculate frequency modulation requirements based on local frequency deviations; call the local Actor network to make power allocation decisions, and negotiate with adjacent edge computing nodes through a consensus algorithm to complete the power command calculation within 50ms and send it to the terminal PCS for execution; S3. Edge computing nodes achieve fault tolerance and critical instruction confirmation through the PBFT consensus protocol; When a preset condition is triggered, adjacent edge computing nodes initiate the PBFT consensus protocol, sequentially voting through five stages: Request, Pre-Prepare, Prepare, Commit, and Reply, to ensure that the consensus is no greater than [a certain threshold]. In the event of a fault-tolerant node failure or communication interruption, key instructions are agreed upon. S4. The cloud solves the global charging and discharging strategy with a 15-minute granularity and issues baseline commands. The cloud-based intelligent hub is based on new energy output forecast, load forecast, electricity price forecast data and energy storage SOC evolution forecast. With a 15-minute granularity and the goal of maximizing the benefits throughout the entire life cycle, it uses MILP or dynamic programming algorithms to solve the global charging and discharging strategy for the next 24 hours and distributes it to each edge computing node as a decision reference baseline. S5. Use the federated averaging algorithm to complete cross-site model collaborative training; Each site trains its local model using local historical data, uploading only the model weights (excluding the original data) to the cloud. The cloud uses the FedAvg algorithm to aggregate the weights from each site to generate a global model, which is then redistributed to achieve cross-site collaborative training with "data not leaving the local area". S6. The Actor-Critic network is iteratively optimized through online fine-tuning and periodic global retraining. Edge computing nodes continuously collect experience data during actual operation and fine-tune strategies through online gradient updates of the Actor-Critic network. The cloud periodically (weekly / monthly) aggregates the de-identified experience statistics of each site for global model retraining and distributes the updated model to the edge computing nodes, forming a continuous optimization closed loop of "offline pre-training → online fine-tuning → periodic retraining".

[0012] Preferably, the protection levels of the three-level protection mechanism in S1 are: Level 1 warning, Level 2 protection, and Level 3 emergency. Level 1 Warning: If the voltage / temperature exceeds the warning threshold, the relevant edge computing node will be notified and the charging / discharging power will be reduced. The response time is <50ms. Secondary protection: If the voltage / temperature exceeds the protection threshold, the output of the PCS cluster will be actively cut off, with a response time of <20ms; Level 3 Emergency: If a short circuit or insulation fault occurs, a hardware-level hard trip will be executed with a response time of <5ms.

[0013] Preferably, the specific process of S2 is as follows: S21. The edge computing node receives status data reported by all terminal execution units within its jurisdiction, performs data fusion and consistency verification (cross-validation of multi-source data at the same measurement point), and constructs the local operating state vector at the current moment. The expression is: ; In the formula, This indicates the state of charge of the first battery cluster; This indicates the health status of the first battery cluster; This indicates the maximum allowable charge / discharge power of the first battery cluster; This indicates the minimum permissible charge / discharge power of the first battery cluster; This indicates the temperature of the first battery cluster; S22, Receive global power command issued by cloud-based intelligent hub. (e.g., AGC scheduling instructions) or calculate the power required for a single frequency modulation based on the local frequency deviation. Determine the total power target to be undertaken in this region. ; The expression is: ; In the formula, This represents the power-frequency regulation coefficient for primary frequency modulation, and is the droop coefficient. This indicates the power grid frequency deviation, which is the difference between the actual frequency of the power grid and the rated frequency. S23. Call the local Actor network to make power allocation decisions, inputting the state vector. Output the power allocation ratio of each PCS The allocation objectives are determined, and a consensus algorithm is used to exchange allocation schemes with adjacent edge computing nodes. Regional consensus is achieved through 2-3 iterations. The allocation objectives include balanced allocation requirements and constraints. Constraints include SOC upper and lower limits, power ramp-up rate constraints, and rated capacity constraints. The balanced allocation requirements are... , Indicates the first Power allocated to each PCS; S24. Complete the power command calculation within 50ms and send it to the terminal PCS for execution. Monitor the execution deviation in real time. If the deviation is greater than the threshold, trigger closed-loop fine-tuning.

[0014] Preferably, the preset conditions in S3 include four types of triggering scenarios, namely: Scenario 1: Receiving critical safety instructions: Critical safety instructions include fire-fighting linkage trip instructions, power grid fault emergency trip instructions, and station shutdown instructions; perform PBFT three-phase consensus verification on critical safety instructions, and execute the critical safety instructions after successful verification. Scenario 2: Receiving edge computing node failure instructions: When the heartbeat of any edge computing node times out for more than 3 seconds, or the data reported by the edge computing node deviates from the data of the neighboring edge computing nodes by more than 30%, the remaining normal edge computing nodes, excluding the failed edge computing node, initiate a consensus process to confirm the running status of the edge computing node. Scenario 3: Receiving monitoring power allocation adjustment instructions: When the power target change rate is greater than 30%·Pn / s, negotiate and confirm the power adjustment plan with other edge computing nodes to complete the collaborative adjustment of power allocation; Pn represents the rated power of the energy storage power station; Scenario 4: Receiving node change instructions: Responding to events such as the addition of a new edge computing node or the active offline maintenance of an edge computing node, updating the cluster node list and view number through a consensus mechanism.

[0015] Therefore, the above-mentioned autonomous collaborative system and method for distributed energy storage power stations based on cloud-edge-device intelligent agents has the following beneficial effects: (1) Improved response speed and real-time performance: Through a time-scale hierarchical mechanism of millisecond-level local processing by terminal execution units, real-time decision-making within 50ms by edge computing nodes, and minute-level global optimization in the cloud, the system response latency has been reduced from hundreds of milliseconds to seconds in the traditional architecture to less than 50 milliseconds, meeting the stringent real-time requirements of primary frequency regulation of the power grid. Experimental data show that the regulation efficiency has been improved by more than 80%, and the frequency regulation performance K value has reached more than 2.5.

[0016] (2) Improved control accuracy and operating efficiency: Distributed power allocation is carried out by edge computing nodes based on a multi-agent reinforcement learning framework, reducing the power tracking error from ±2%~±5% of the traditional architecture to within ±0.5%; SOC balance is achieved through a multi-agent consistency coordination mechanism, reducing the SOC standard deviation from 0.03~0.05 to below 0.01, making the charging and discharging states of each battery cluster more consistent, effectively extending the overall service life of the battery pack by 2~3 years, increasing the number of annual charging and discharging cycles from 250 to 340, and reducing the cost per kilowatt-hour of the entire life cycle by 0.05~0.10 yuan / kWh.

[0017] (3) Enhanced system reliability and resilience: By adopting the PBFT Byzantine Fault Tolerant consensus protocol between edge computing nodes, the reliability and resilience of the system can be improved within a specified timeframe. Even if a node fails or communication is interrupted, the system can still reach a consensus on critical safety instructions; the terminal execution unit has an independent three-level protection mechanism (alarm → protective power reduction → hardware tripping), and can still complete device-level protection locally even if communication with the upper layer is completely interrupted; single-point failures do not spread to the whole system, and the system availability rate is increased from 99.5% to over 99.9%.

[0018] (4) Reduced operation and maintenance costs: The intelligent autonomous collaborative platform reduces operation and maintenance costs by 30% to 50% and reduces the demand for on-site operation and maintenance personnel by 50% to 70%, significantly reducing the total life cycle operation cost of energy storage power stations.

[0019] (5) Data privacy and security: Through the federated learning mechanism, the original operating data (time-series sampled values ​​such as voltage, current, SOC, and temperature) of each energy storage power station does not leave the local area. Only the model weight parameters (high-dimensional mathematical matrix, which cannot be used to deduce the original data) are uploaded, which fundamentally solves the risk of data privacy leakage in cross-site collaborative training, meets the compliance requirements of the power grid industry that data does not leave the park, and realizes cross-site knowledge sharing and collaborative modeling.

[0020] (6) System scalability and deployment flexibility: This invention adopts a layered distributed architecture. New energy storage power stations or edge computing nodes only need to load the global model and connect to the consensus network to integrate into the system without modifying the existing architecture. It supports smooth expansion from the megawatt level to the gigawatt level. The containerized microservice architecture further supports rapid deployment, elastic expansion and fault isolation of functions, adapting to the needs of phased construction and rolling development.

[0021] (7) Continuous evolution and adaptive capability: Through reinforcement learning, the closed loop is continuously optimized by “cloud offline pre-training → edge online fine-tuning → cloud periodic retraining”. The system strategy evolves continuously with the accumulation of operating data and can adapt to time-varying factors such as battery aging, electricity price fluctuations, and power grid characteristic evolution, without the need for frequent manual adjustment of control parameters.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the structure of the distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to the present invention; Figure 2 This is a diagram of the multi-agent reinforcement learning framework of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the federated learning engine of the present invention. Figure 4 This is a flowchart of the PBFT consensus algorithm of the present invention; Figure 5 This is a flowchart of the intelligent agent collaboration mechanism of the present invention; Figure 6 This is a hardware deployment topology diagram of Embodiment 1 of the present invention. Detailed Implementation

[0024] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] Please see Figure 1 A distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents, including a cloud intelligent hub, edge computing nodes and terminal execution units; The cloud-based intelligent hub is deployed in a data center or cloud platform for global optimization, long-term prediction, policy generation, and AI model training; the cloud-based intelligent hub includes a federated learning engine; such as Figure 3 As shown, the federated learning engine maintains the global prediction and optimization model, with copies of the model deployed locally at each site. Each site uses its local data to train the model copy through forward inference and backpropagation, keeping the original data on-premises and only uploading the updated model weights to the cloud-based intelligent hub. The federated learning engine uses the FedAvg aggregation algorithm to perform a weighted average of the weights uploaded by each site, generating an updated global model, which is then redistributed to each site. This process is iterated until the global model converges on the validation set, collaboratively training a better global prediction and optimization model, thereby solving the problems of data silos and privacy. A site refers to each independent energy storage power station participating in the federated collaborative training, i.e., each energy storage power station unit that deploys the edge computing nodes of this invention.

[0026] Edge computing nodes are deployed at power plant sites or regional aggregation points, responsible for real-time data processing, local optimization decision-making, rapid response control, and localized execution of cloud strategies, compressing control latency to within 50ms; consensus negotiation between edge computing nodes is achieved using the Practical Byzantine Fault-Tolerant (PBFT) protocol; such as Figure 4 As shown, the PBFT protocol includes five phases: Request, Pre-Prepare, Prepare, Commit, and Reply. Phase 1, Request: The client sends a request message to the master node. , Indicates an operation. Represents a timestamp. Indicates the client identifier; Phase 2, Pre-Prepare: The master node assigns sequence numbers to all replica nodes and broadcasts a pre-prepare message. , Indicates the view number. Indicates the serial number. Indicates a message summary; Phase 3, Prepare: Each replica node broadcasts a prepare message after verification. ,receive After receiving a preparation message, the system enters the preparation complete state. Indicates the number of fault-tolerant nodes to ensure message order consistency; Phase 4, Commit: Each replica node broadcasts a commit message. ,receive After each message is submitted, the request is executed. At this stage, a quorum is reached to ensure cross-viewing... Figure 1 To the point of being compatible; Phase 5, Reply: Each replica node sends a reply message to the client. The client received Confirm the result after receiving a consistent reply; Edge computing nodes conduct consensus voting through a three-stage voting mechanism of pre-preparation, preparation, and submission to ensure that the consensus is no greater than [a certain threshold]. Even if a fault-tolerant node fails or communication is interrupted, the system can still reach a consensus on critical security instructions; the total number of edge computing nodes participating in consensus negotiation satisfy .

[0027] The terminal execution unit is embedded in the BMS, PCS, and sensor field devices for millisecond-level data acquisition, device-level protection control, and status information reporting. The terminal execution unit includes a BMS and a PCS with embedded AI inference chips. The BMS is used to evaluate the health status and charge / discharge constraints of the battery modules it monitors in real time, and the PCS is used to evaluate the health status and conversion efficiency preference of its own power conversion devices in real time.

[0028] The cloud-based intelligent hub, edge computing nodes, and terminal execution units constitute a three-layer intelligent agent architecture. The intelligent agents make collaborative decisions through a multi-agent reinforcement learning framework, execute collaborative responses according to a time-scale hierarchical mechanism, and interact through an intelligent agent collaboration mechanism.

[0029] like Figure 2 As shown, the multi-agent reinforcement learning framework includes an Actor network, a Critic network, an experience replay buffer, and a consensus coordination mechanism. The Actor network outputs action policies based on the current state. The Critic network is used to evaluate the state-action value function. The experience replay buffer stores historical interaction data (including state, action, reward, next state, and termination flag) for training. The consensus coordination mechanism controls the power difference between agents within a preset threshold.

[0030] The time-scale hierarchical mechanism is as follows: millisecond level, <10ms, protection actions and PWM modulation are handled by the terminal execution unit; hundred millisecond level, 10ms~100ms, power tracking is completed by the edge computing node and the terminal execution unit in collaboration; second level, 1s~10s, regional coordination is completed by the edge computing node; and minute level and above, strategy optimization is completed by the cloud intelligent hub and the edge computing node in collaboration.

[0031] like Figure 5 As shown, the intelligent agent collaboration mechanism includes: Task decomposition and allocation: using the Contract Network protocol, the task initiator publishes a tender notice, each intelligent agent bids after assessing its own capabilities and costs, and the initiator selects the winning bidder through comprehensive evaluation, forming a task allocation contract; Information interaction and negotiation: data distribution is based on a publish-subscribe mechanism, intelligent agents send subscription requests carrying specified data topic identifiers, and data publishers push topic data in a targeted manner according to the subscription matching results, reducing the number of broadcast messages transmitted across the network and avoiding global message broadcasting; Conflict resolution and consistency maintenance: the power configuration parameters of each intelligent agent are adjusted through a consistency coordination mechanism to control the power difference between intelligent agents within a preset threshold.

[0032] The method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents includes the following steps: S1. The terminal execution unit periodically collects energy storage data and performs local protection and periodic status reporting; The terminal execution unit collects data and performs moving average filtering and bad value removal on the data. The preprocessed data is written to the local cache and a hardware timestamp (GPS / BeiDou synchronization) is added to ensure the time consistency of the data across the entire station (error <1μs). The collected data includes, but is not limited to, the categories shown in Table 1. Table 1

[0033] Table 2

[0034] When any parameter of voltage, current or temperature exceeds its respective preset threshold, a three-level protection mechanism is triggered. The processed status information (SOC, SOH, maximum allowable charge / discharge power, etc.) is packaged and reported to the edge computing node at a configurable period of 10ms to 100ms. The reporting adopts a publish-subscribe mode and only reports changes (incremental reporting) to save bandwidth. The three-level protection mechanism is a local protection control, as shown in Table 2.

[0035] S2, edge computing nodes complete data verification, power decision-making, and regional consistency collaborative allocation; S21. The edge computing node receives status data reported by all terminal execution units within its jurisdiction, performs data fusion and consistency verification (cross-validation of multi-source data at the same measurement point), and constructs the local operating state vector at the current moment. The expression is: ; In the formula, This indicates the state of charge of the first battery cluster; This indicates the health status of the first battery cluster; This indicates the maximum allowable charge / discharge power of the first battery cluster; This indicates the minimum permissible charge / discharge power of the first battery cluster; This indicates the temperature of the first battery cluster; S22, Receive global power command issued by cloud-based intelligent hub. (e.g., AGC scheduling instructions) or calculate the power required for a single frequency modulation based on the local frequency deviation. Determine the total power target to be undertaken in this region. ; The expression is: ; In the formula, This represents the power-frequency regulation coefficient for primary frequency modulation, and is the droop coefficient. This indicates the power grid frequency deviation, which is the difference between the actual frequency of the power grid and the rated frequency. S23. Call the local Actor network to make power allocation decisions, inputting the state vector. Output the power allocation ratio of each PCS The allocation objectives are determined, and a consensus algorithm is used to exchange allocation schemes with adjacent edge computing nodes. Regional consensus is achieved through 2-3 iterations. The allocation objectives include balanced allocation requirements and constraints. Constraints include SOC upper and lower limits, power ramp-up rate constraints, and rated capacity constraints. The balanced allocation requirements are... , Indicates the first Power allocated to each PCS; S24. Complete the power command calculation within 50ms and send it to the terminal PCS for execution. Monitor the execution deviation in real time. If the deviation is greater than the threshold, trigger closed-loop fine-tuning.

[0036] S3. Edge computing nodes achieve fault tolerance and critical instruction confirmation through the PBFT consensus protocol; When a preset condition is triggered, adjacent edge computing nodes initiate the PBFT consensus protocol, sequentially voting through five stages: Request, Pre-Prepare, Prepare, Commit, and Reply, to ensure that the consensus is no greater than [a certain threshold]. In the event of a fault-tolerant node failure or communication interruption, key instructions are agreed upon; preset conditions include four types of triggering scenarios as shown in Table 3. Table 3

[0037] S4. The cloud solves the global charging and discharging strategy with a 15-minute granularity and issues baseline commands. The prediction target is: New energy output forecast: Output curves of photovoltaic / wind power for the next 4 hours and 24 hours; Load forecasting: The trend of power grid load changes over the next 4 hours and 24 hours; Electricity price forecast: Time-of-use electricity prices in the spot market for the next 24 hours; Energy Storage SOC Evolution Prediction: Based on new energy output forecast, load forecast, and electricity price forecast, the future range of energy storage SOC variation is estimated; Optimization objective: The objective function is to maximize the total lifecycle revenue, i.e.: ; The specific process is as follows: The latest forecast data is loaded into the cloud, from numerical weather prediction, load forecasting systems, and electricity market interfaces; the global optimization engine is called based on MILP or dynamic programming algorithms to generate charging and discharging strategy curves for the next 24 hours with a time granularity of 15 minutes; the strategy curves include the target power value, SOC upper and lower limits, and charging and discharging number limits for each hour; the optimization results are sent to each edge computing node via the MQTT protocol as a reference baseline for real-time edge decision-making; the cloud updates the optimization strategy at least once a day (or updates are triggered based on changes in market electricity prices). S5. Use the federated averaging algorithm to complete cross-site model collaborative training; Each site trains its local model using local historical data, uploading only the model weights (excluding the original data) to the cloud. The cloud uses the FedAvg algorithm to aggregate the weights from each site to generate a global model, which is then redistributed to achieve cross-site collaborative training with "data not leaving the local area". S51, Initialize global model weights in the cloud. and will Distributed to various energy storage power stations participating in federal learning; S52. Each energy storage power station utilizes local historical data and employs the Adam optimizer and MSE loss function to perform... During local training, the raw data does not leave the local machine. S53, Each energy storage power station will use the weights of the trained local model. Upload to the cloud; S54. Execute FedAvg aggregation in the cloud to generate a new global model with the following expression: ; S55. Repeat steps S51 to S54 until the global model converges on the validation set, i.e., the validation loss decreases by less than 1% for 5 consecutive rounds.

[0038] S6. The Actor-Critic network is iteratively optimized through online fine-tuning and periodic global retraining, including three stages: offline pre-training, online fine-tuning, and periodic retraining. Offline pre-training phase (completed in the cloud): S61. The cloud-based simulation environment is built based on more than one year of historical energy storage operation data. S62. Use the MADDPG algorithm to pre-train the multi-agent policy in a simulation environment, with no less than 10,000 training rounds. S63. When the average reward fluctuation over 100 consecutive rounds is less than 5%, the strategy is determined to be converged, and an initial strategy model is generated. Online fine-tuning phase (edge ​​computing nodes complete, continuous operation): S64. Each edge computing node loads a pre-trained model from the cloud as the initial strategy. S65. In actual operation, the status of each control cycle is collected. Execute actions , This indicates that the Actor network is based on the current state. The output action strategy receives environmental feedback; an immediate reward function is provided. for: ; In the formula, This represents the power tracking error, which is the difference between the actual output power and the target power. This represents the weighting coefficient of the SOC deviation term; Indicates the weighting coefficient of the PCS loss term; This represents the weighting coefficient of the frequency modulation revenue term; This indicates the deviation between the current SOC and the target SOC of each energy storage unit; This represents the power conversion loss during PCS operation; S66, Experience Store the data in a local experience replay buffer with a capacity of 100,000 records, using a first-in-first-out replacement strategy. S67. Every 64 new experiences accumulated, perform a gradient update based on the Actor-Critic network: Actor gradient: Update the Actor network; Critic loss: Update the Critic network; Target network soft update: τ=0.005; In the formula, This represents the policy gradient of the Actor network; Represents the mathematical expectation; Represents the Actor network parameters The gradient; Indicates action The gradient; This represents the state-action value function output by the Critic network; The loss function of the Critic network is represented by this. Indicates an immediate reward; Indicates the discount factor; Represents the target Critic network; Indicates the state at the next moment; Indicates the target Actor network in state The action to be output; These represent the parameters of the target network. Indicates the soft update coefficient; Parameters representing online networks; S68. During the low-load period in the early morning every day, the accumulated experience data of the day is replayed for training in batches, with an additional 200 batches trained to consolidate the learning effect. Periodic retraining phase (completed in the cloud, weekly / monthly): S69. Each energy storage power station uploads the de-identification experience statistics for this period to the cloud. The de-identification experience statistics include average reward, strategy entropy value and task success rate. S610 aggregates the experience statistics uploaded by various energy storage power stations in the cloud and continues to train the global model in the simulation environment through transfer learning. S611. The updated global model is redeployed to each edge computing node to complete the iterative upgrade of the strategy.

[0039] To verify this invention, a "simulation test platform + actual operating data benchmark + industry standard benchmarking" method was used. The specific experimental conditions were as follows: Simulation testing platform (main data source): Based on the architecture of this invention, a digital twin test platform was built, comprising one cloud-based intelligent hub, six edge computing nodes, and 100 terminal execution units. The test scenarios cover typical operating conditions such as grid frequency disturbance (±0.1Hz), power command step change (0→100%Pn), and partial edge node network outage (simulated fault). Each test condition was repeated no less than 100 times, and the statistical average was taken to ensure the statistical significance of the data. The simulation platform is based on RT-LAB + Python co-simulation, and the communication delay is set according to the measured delay parameters of industrial Ethernet + 5G private network (≤2ms within the station, ≤10ms between stations).

[0040] Actual operating data of energy storage power stations (comparison benchmark): (1) The comparative data of “traditional centralized” comes from the actual operation logs of three 100MW-200MW energy storage power stations that have been put into operation in the early stage, with a time span of 12 consecutive months; (2) The above power plant adopts a traditional centralized EMS architecture. Its response delay, power tracking error, SOC equalization accuracy and other indicators are taken from actual SCADA historical data, not theoretical estimates.

[0041] Industry publicly available standards and literature data (for auxiliary calibration): (1) Some indicators (such as system availability ≥ 99.9%) are benchmarked against GB / T 36547-2018 "Technical Specifications for Electrochemical Energy Storage Systems Connected to the Power Grid" and relevant requirements of IEEE 2030.2; (2) Some values ​​(such as 250 charge-discharge cycles per year) are based on the industry average in the annual white paper of the China Energy Storage Alliance (CNESA).

[0042] The experimental results are shown in Table 4. The system response latency of this invention is reduced to less than 50 milliseconds, and the regulation efficiency is improved by more than 80%, meeting the stringent real-time requirements of power grid frequency regulation; the power tracking error is controlled within ±0.5%, the standard deviation of SOC equalization accuracy is <0.01, and the frequency regulation performance K value reaches more than 2.5; single-point faults do not propagate, local autonomy maintains operation, and the system availability rate is >99%; the intelligent platform reduces operation and maintenance costs by 30%-50% and reduces the need for on-site personnel by 50%-70%; the annual charge and discharge cycles are increased from 250 to 340, the battery life is extended by 2-3 years, and the cost per kilowatt-hour over the entire life cycle is reduced by 0.05-0.10 yuan / kWh; federated learning enables collaborative training with data not leaving the local area, meeting data security compliance requirements; it supports smooth expansion from the megawatt level to the gigawatt level, and flexible deployment in phases.

[0043] Table 4

[0044] Example 1: Typical Deployment of a 500MW / 2000MWh Standalone Energy Storage Power Station This embodiment describes a typical deployment scheme for a 500MW / 2000MWh stand-alone energy storage power station based on the architecture of this invention.

[0045] like Figure 6 As shown, the hardware configuration is as follows: Cloud layer: Deploy a cluster of more than 10 servers, including an AI training center (4 GPU servers), a big data platform (4 storage servers), and a global optimization engine (2 computing servers).

[0046] Edge layer: 16 core switches are deployed to form a redundant ring network, and 6 coordination controllers are used to coordinate regional power. Each coordination controller manages about 80-100 PCS.

[0047] Terminal layer: Deploys nearly 5 million intelligent sensing and execution devices, including battery management system (BMS), energy storage converter (PCS), temperature sensor, voltage and current sensor, etc.

[0048] Communication Network: A redundant industrial Ethernet + 5G / fiber optic hybrid architecture is adopted. The station control layer uses gigabit / 10-gigabit industrial Ethernet, with a ring network topology to ensure reliability; the process layer uses 100-megabit industrial Ethernet or fiber optic bus, with deterministic transmission meeting real-time requirements; remote communication uses a 5G private network or fiber optic leased line to ensure a secure connection with the cloud. Time synchronization uses GPS / BeiDou time service, achieving sub-microsecond time synchronization through the IEEE 1588 PTP protocol, with end-to-end latency controlled at the hundred-microsecond level and latency jitter <1μs.

[0049] Software Deployment: A containerized microservice architecture is adopted, encapsulating functions into independent containers to support rapid deployment, elastic scaling, and fault isolation. Typical microservices include: data acquisition service, real-time control service, optimization computing service, alarm handling service, human-machine interface service, and external interface service.

[0050] Example 2: Power Grid Frequency Regulation Application Scenarios This embodiment describes the specific workflow of the present invention in the power grid frequency regulation application scenario.

[0051] Step 1: The terminal execution unit collects electrical quantities such as grid frequency, voltage, and power at millisecond intervals. Once the frequency deviation is detected to exceed the dead zone (±0.033Hz), a power adjustment request is immediately triggered. Step 2: The edge computing node receives the frequency modulation request, calculates the power allocation in real time based on the energy storage status of the area, and completes the coordinated control of multiple PCS in seconds; Step 3: The cloud-based intelligent hub optimizes the energy storage SOC status in advance based on load forecasting and new energy output forecasting, reserving sufficient capacity for frequency regulation. Step 4: The system response delay is controlled within 50 milliseconds, and the frequency modulation performance index K value reaches 2.5 or above.

[0052] Example 3: Fault Handling and Self-Healing This embodiment describes the self-healing capability of the present invention in fault handling scenarios.

[0053] When a battery cluster fails, the BMS of the terminal execution unit detects the anomaly in milliseconds and immediately reports it to the edge computing node. The edge computing node locates the fault range in seconds, initiates a dynamic reconstruction algorithm, recalculates available power and energy, and adjusts the power allocation strategy. At the same time, it reaches an agreement with other edge computing nodes on fault isolation instructions through the PBFT consensus algorithm to ensure the consistent execution of critical safety instructions.

[0054] In extreme scenarios where 30% of the capacity suddenly fails, the intelligent reconfiguration system can complete power redistribution within 2 seconds, with frequency tracking error controlled within 0.1Hz, and the system maintains normal operation.

[0055] Therefore, this invention adopts the above-mentioned autonomous collaborative system and method for distributed energy storage power stations based on cloud-edge-device intelligent agents. Relying on the time-scale hierarchical mechanism, distributed reinforcement learning, and consensus coordination algorithm, it significantly improves the real-time performance of regulation and power balance control accuracy, extends battery life, and reduces the cost per kilowatt-hour. It enhances the system's fault tolerance and resilience by relying on PBFT consensus and local multi-level protection at the terminal, reducing maintenance manpower and expenses. It ensures that local raw data does not leak out by relying on federated learning, taking into account data compliance and cross-site collaborative modeling. The hierarchical distributed architecture can be smoothly expanded to adapt to different scales of energy storage. With the reinforcement learning iterative closed loop, it can adapt to various dynamic changes in operating conditions. The overall comprehensive performance is superior to traditional centralized solutions.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents, characterized in that: include: The cloud-based intelligent hub, deployed in a data center or cloud platform, is used for global optimization, long-term prediction, strategy generation, and AI model training. Edge computing nodes, deployed at power plant sites or regional aggregation points, are responsible for real-time data processing, local optimization decision-making, rapid response control, and localized execution of cloud strategies, compressing control latency to within 50ms. The terminal execution unit includes a data acquisition terminal execution unit and a control terminal execution unit. The data acquisition terminal execution unit is embedded in the sensor for millisecond-level data acquisition and status information reporting. The control terminal execution unit is embedded in the BMS and PCS that integrate AI inference chips for device-level protection control and real-time evaluation of battery and power device status. The cloud-based intelligent hub, edge computing nodes, and terminal execution units constitute a three-layer intelligent agent architecture. The agents make collaborative decisions through a multi-agent reinforcement learning framework. The agents are trained under the guidance of cloud models, and the edge devices make distributed decisions based on local information. The global optimal power allocation is achieved through communication.

2. The distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 1, characterized in that: The cloud-based intelligent hub includes a federated learning engine. This engine maintains a global prediction and optimization model, with copies deployed locally at each site. Each site uses its local data to train the model copy through forward inference and backpropagation, ensuring the original data remains on its local machine. The updated model weights are then uploaded to the cloud-based intelligent hub. The federated learning engine uses the FedAvg aggregation algorithm to perform a weighted average of the weights uploaded from each site, generating an updated global model which is then redistributed to each site. This process is iterated until the global model converges on the validation set, collaboratively training a global prediction and optimization model, thereby addressing data silos and privacy issues. The expression for weighted averaging of uploads from each site using the FedAvg aggregation algorithm is as follows: ； In the formula, Indicates the first Round global model weights; This represents the total number of stations participating in the aggregation; Indicates station The number of local samples; This represents the total number of samples from all participating parties. ; Indicates station In the Round global model weights Based on this, the updated local weights are obtained after multiple rounds of local training.

3. The distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 1, characterized in that: The edge computing nodes use the Practical Byzantine Fault-Tolerant (PBFT) protocol for consensus negotiation. The PBFT protocol consists of five phases: Request, Pre-Prepare, Prepare, Commit, and Reply. Edge computing nodes conduct consensus voting through a three-stage voting mechanism of pre-preparation, preparation, and submission to ensure that the consensus is no greater than [a certain threshold]. In the event of a fault-tolerant node failure or communication interruption, key security instructions are agreed upon. Total number of edge computing nodes participating in consensus negotiation satisfy .

4. The distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 1, characterized in that: In the control-type terminal execution unit, the BMS is used to evaluate the health status and charge / discharge constraints of the battery module it monitors in real time, and the PCS is used to evaluate the health status and conversion efficiency preference of its own power conversion device in real time.

5. The distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 1, characterized in that: The three-layered agents make collaborative decisions through a multi-agent reinforcement learning framework, execute collaborative responses according to a time-scale hierarchical mechanism, and interact through an agent collaboration mechanism. Multi-agent reinforcement learning frameworks include: Actor networks output action strategies based on the current state; Critic network, evaluating the state-action value function; An experience replay buffer stores historical interaction data for training. A consistency coordination mechanism controls the power difference between agents within a preset threshold. The time-scale stratification mechanism is as follows: Millisecond level, <10ms, protection actions and PWM modulation are handled by the terminal execution unit; Power tracking is completed in milliseconds, from 10ms to 100ms, by the collaboration between edge computing nodes and terminal execution units. The timeframe is within seconds, ranging from 1 to 10 seconds, with regional coordination completed by edge computing nodes. For timeframes of minutes and above, strategy optimization is completed collaboratively by the cloud-based intelligent hub and edge computing nodes. Intelligent agent cooperation mechanisms include: Task decomposition and allocation: The Contract Network protocol is adopted. The task initiator publishes a tender notice, and each intelligent agent bids after assessing its own capabilities and costs. The initiator selects the successful bidder through comprehensive evaluation and forms a task allocation contract. Information interaction and negotiation: Data distribution is based on a publish-subscribe mechanism. The agent sends a subscription request carrying a specified data topic identifier. The data publisher pushes topic data in a targeted manner according to the subscription matching result, reducing the number of broadcast messages transmitted across the network and avoiding global message broadcasting. Conflict resolution and consistency maintenance: The power configuration parameters of each agent are adjusted through a consistency coordination mechanism to control the power difference between agents within a preset threshold.

6. A method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents as described in any one of claims 1-5, characterized in that, Includes the following steps: S1. The terminal execution unit periodically collects energy storage data and performs local protection and periodic status reporting; The terminal execution unit collects battery electrical quantities, status quantities, and environmental quantity data at a period of ≤1ms and performs data preprocessing. When any parameter of voltage, current, or temperature exceeds its respective preset threshold, a three-level protection mechanism is triggered, and the processed status information is reported to the corresponding edge computing node at a period of 10ms to 100ms. S2, edge computing nodes complete data verification, power decision-making, and regional consistency collaborative allocation; Edge computing nodes aggregate the status data reported by the terminal execution units under their jurisdiction and fuse them to construct local status vectors; receive global power targets issued by the cloud intelligent hub or calculate frequency modulation requirements based on local frequency deviations; call the local Actor network to make power allocation decisions, and negotiate with adjacent edge computing nodes through a consensus algorithm to complete the power command calculation within 50ms and send it to the terminal PCS for execution; S3. Edge computing nodes achieve fault tolerance and critical instruction confirmation through the PBFT consensus protocol; When a preset condition is triggered, adjacent edge computing nodes initiate the PBFT consensus protocol, sequentially voting through five stages: Request, Pre-Prepare, Prepare, Commit, and Reply, to ensure that the consensus is no greater than [a certain threshold]. In the event of a fault-tolerant node failure or communication interruption, key instructions are agreed upon. S4. The cloud solves the global charging and discharging strategy with a 15-minute granularity and issues baseline commands. The cloud-based intelligent hub is based on new energy output forecast, load forecast, electricity price forecast data and energy storage SOC evolution forecast. With a 15-minute granularity and the goal of maximizing the benefits throughout the entire life cycle, it uses MILP or dynamic programming algorithms to solve the global charging and discharging strategy for the next 24 hours and distributes it to each edge computing node as a decision reference baseline. S5. Use the federated averaging algorithm to complete cross-site model collaborative training; Each site trains its local model using local historical data, and only uploads the model weights to the cloud. The cloud uses the FedAvg algorithm to aggregate the weights of each site to generate a global model, which is then redistributed to achieve cross-site collaborative training. S6. The Actor-Critic network is iteratively optimized through online fine-tuning and periodic global retraining. Edge computing nodes continuously collect experience data during actual operation and fine-tune strategies through online gradient updates of the Actor-Critic network; the cloud periodically aggregates the desensitized experience statistics of each site to retrain the global model and distributes the updated model to the edge computing nodes, forming a continuous optimization closed loop.

7. The method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 6, characterized in that, The three protection levels in S1 are: Level 1 Warning, Level 2 Protection, and Level 3 Emergency. Level 1 Warning: If the voltage / temperature exceeds the warning threshold, the warning will be reported to the relevant edge computing node, and the charging / discharging power will be reduced. The response time is <50ms. Secondary protection: If the voltage / temperature exceeds the protection threshold, the output of the PCS cluster will be actively cut off, with a response time of <20ms; Level 3 Emergency: If a short circuit or insulation fault occurs, a hardware-level hard trip will be executed with a response time of <5ms.

8. The method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 6, characterized in that, The specific process of S2 is as follows: S21. The edge computing node receives the status data reported by all terminal execution units within its jurisdiction, performs data fusion and consistency verification, and constructs the local operating state vector at the current moment. The expression is: ； In the formula, This indicates the state of charge of the first battery cluster; This indicates the health status of the first battery cluster; This indicates the maximum allowable charge / discharge power of the first battery cluster; This indicates the minimum permissible charge / discharge power of the first battery cluster; This indicates the temperature of the first battery cluster; S22, Receive global power command issued by cloud-based intelligent hub. Alternatively, calculate the power required for primary frequency modulation based on the local frequency deviation. Determine the total power target to be undertaken in this region. ; The expression is: ； In the formula, This represents the power-frequency regulation coefficient for primary frequency modulation; Indicates the power grid frequency deviation; S23. Call the local Actor network to make power allocation decisions, inputting the state vector. Output the power allocation ratio of each PCS The allocation objectives are determined, and a consensus algorithm is used to exchange allocation schemes with adjacent edge computing nodes. Regional consensus is achieved through 2-3 iterations. The allocation objectives include balanced allocation requirements and constraints. Constraints include SOC upper and lower limits, power ramp-up rate constraints, and rated capacity constraints. The balanced allocation requirements are... , Indicates the first Power allocated to each PCS; S24. Complete the power command calculation within 50ms and send it to the terminal PCS for execution.

9. The method for a distributed energy storage power station autonomous collaborative system based on cloud-edge-device intelligent agents according to claim 6, characterized in that, The preset conditions in S3 include four types of trigger scenarios, namely: Scenario 1: Receiving critical safety instructions: Critical safety instructions include fire-fighting linkage trip instructions, power grid fault emergency trip instructions, and station shutdown instructions; perform PBFT three-phase consensus verification on critical safety instructions, and execute the critical safety instructions after successful verification. Scenario 2: Receiving edge computing node failure instructions: When the heartbeat of any edge computing node times out for more than 3 seconds, or the data reported by the edge computing node deviates from the data of the neighboring edge computing nodes by more than 30%, the remaining normal edge computing nodes, excluding the failed edge computing node, initiate a consensus process to confirm the running status of the edge computing node. Scenario 3: Receiving monitoring power allocation adjustment instructions: When the power target change rate is greater than 30%·Pn / s, negotiate and confirm the power adjustment plan with other edge computing nodes to complete the collaborative adjustment of power allocation; Pn represents the rated power of the energy storage power station; Scenario 4: Receiving node change instructions: Responding to events such as the addition of a new edge computing node or the active offline maintenance of an edge computing node, updating the cluster node list and view number through a consensus mechanism.

Images