A scheduling method based on energy storage converter group cooperative control

Abstract

The application discloses a kind of based on energy storage converter group cooperative control scheduling method, it is related to energy storage system control technical field, including each energy storage unit is through with its adjoining unit and carries out state information interaction, jointly maintains a distributed system state information library;Based on the state information interaction, the topological structure change of the energy storage unit cluster is detected;In response to detecting the topological structure change or receiving total power scheduling instruction, according to the distributed system state information library generates an adaptive power distribution strategy selection signal;Based on the selection signal, from multiple alternative power distribution strategies, the corresponding strategy is selected and executed;Solved the delay and single point failure problem of centralized scheduling, let energy storage converter group run stably under dynamic topology and complex scheduling instruction, and reduce the total loss of cluster, improve energy utilization efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of energy storage system control, and specifically relates to a scheduling method based on energy storage converter group cooperative control. BACKGROUND

[0002] The existing energy storage power station group control mostly adopts centralized or master-slave architecture and highly depends on preset master control units or upper computers. The central node is responsible for power distribution, state synchronization and operation strategy formulation of the whole group and is the core hub of system coordinated operation.

[0003] This centralized architecture has significant defects: on the one hand, it faces the risk of single point failure, and the failure of the master control unit or the core communication link will directly cause the interruption of system coordination function; on the other hand, it is difficult to adapt to dynamic topology changes, and when new units are connected or old units are disconnected, the central node needs to detect, authenticate and configure again, which is time-consuming and dependent on the availability of the central node, and cannot realize plug and play, and the response may be slow or even interrupted during topology change.

[0004] In summary, the existing technology lacks a distributed intelligent cooperative mechanism without the intervention of the central node, and cannot independently perceive the change of members and quickly reconstruct the control architecture, resulting in poor self-adaptability and slow reconstruction speed of the system when facing random topology changes, which seriously restricts the robustness and flexibility of large-scale energy storage systems. SUMMARY

[0005] The purpose of the present application is to provide a scheduling method based on energy storage converter group cooperative control to solve the problems mentioned in the background.

[0006] A scheduling method based on energy storage converter group cooperative control, the method is executed in parallel by each energy storage unit in the energy storage unit cluster, wherein the adjacent unit refers to other energy storage units that have direct and bidirectional communication links with the current unit in the current communication topology graph of the system;

[0007] The energy storage unit refers to the smallest control entity containing an energy storage battery and a converter and capable of independently executing the processes described in the method;

[0008] The distributed system state information library refers to a structured data set stored locally in each energy storage unit and guaranteed to be consistent by the cooperative maintenance mechanism described below;

[0009] The method comprises:

[0010] Each energy storage unit interacts with its adjacent units to maintain a distributed system state information library;

[0011] Based on the state information interaction, the topology structure change of the energy storage unit cluster is detected;

[0012] in response to detecting the topology change or receiving the total power scheduling instruction, generating an adaptive power allocation strategy selection signal according to the distributed system state information base;

[0013] based on the selection signal, selecting and executing a corresponding strategy from a plurality of alternative power allocation strategies;

[0014] In the process of executing power allocation, the power allocation results of each energy storage unit are verified for consistency by a distributed negotiation mechanism, and the final power allocation instruction is executed after verification.

[0015] Further, the generation of an adaptive power allocation strategy selection signal includes:

[0016] According to the distributed system state information base, the dynamic balance coefficient representing the uniformity of system power distribution and the cooperative power response threshold required by the system are calculated;

[0017] According to the comparison result of the dynamic balance coefficient and the cooperative power response threshold, and the type of the topology change, the adaptive power allocation strategy selection signal is generated.

[0018] Further, the plurality of alternative power allocation strategies at least includes an incremental power allocation strategy and a fast power allocation strategy;

[0019] Wherein, the incremental power allocation strategy focuses on the smoothness of power change, and the fast power allocation strategy focuses on the timeliness of power response.

[0020] Further, the common maintenance of a distributed system state information base includes:

[0021] Each energy storage unit preprocesses the locally collected raw operation data to generate a standardized local data set;

[0022] Exchange statistical characteristics of the local data set with adjacent units to construct a regional data feature map;

[0023] Based on the regional data feature map, the local data set is cleaned to eliminate abnormal data;

[0024] Cross-validation of the cleaned data between adjacent units;

[0025] The verified data is stored in the distributed system state information base.

[0026] Further, detecting the topology change of the energy storage unit cluster includes:

[0027] When a communication link state change with an adjacent unit is detected, a topology change confirmation process is triggered;

[0028] Send a confirmation request to adjacent units with normal communication, and collect topology observation feedback from these adjacent units;

[0029] Compare the topology observation feedback of each unit, and confirm the effective topology change event using the majority decision principle;

[0030] Update the system topology map based on the confirmation result.

[0031] Further, the consistency verification through the distributed negotiation mechanism includes:

[0032] Each energy storage unit collects local power distribution execution data;

[0033] Identify abnormalities in the execution data and mark abnormal data points;

[0034] Correlate and compare the marked abnormal data points with the execution data of adjacent units;

[0035] According to the correlation comparison result, distinguish between system-level abnormalities and unit-level abnormalities;

[0036] Correct the unit-level abnormal execution data, and trigger a power redistribution process for system-level abnormalities;

[0037] The corrected data can only be used for final consistency verification after being confirmed by adjacent units.

[0038] Further, the method further includes a step of calculating a dynamic balance coefficient representing the uniformity of system power distribution, which includes:

[0039] Each energy storage unit monitors the local power regulation capability and operating condition in real time;

[0040] Through state information interaction between adjacent units, a system operating state panoramic view is constructed;

[0041] Based on the panoramic view, the power support relationship between each unit is analyzed to identify system key nodes;

[0042] Evaluate the influence weight of the operating state of the system key nodes on the overall balance of the system;

[0043] According to the influence weight, the dynamic balance coefficient is dynamically calculated.

[0044] As a preferred embodiment of the present application, the method further includes a distributed clock synchronization step for improving the timing accuracy of state information interaction and consistency verification:

[0045] Each energy storage unit adds a local timestamp to the interactive state information;

[0046] The receiving unit calculates the clock bias between the sending unit according to the local timestamp;

[0047] The clock synchronization information is transmitted in the system through a multi-hop transmission mode;

[0048] Based on the transmission result, the clock of each unit is dynamically calibrated.

[0049] As another preferred embodiment of the present application, the method further comprises a policy optimization step based on federated learning, for adaptively improving the accuracy of policy selection:

[0050] A state set containing system state, topology feature and load characteristic is constructed;

[0051] A comprehensive evaluation function of policy execution effect is defined;

[0052] Through a privacy-protected distributed data sharing mechanism, policy execution experience is exchanged between units;

[0053] The federated learning method is used to collaboratively update the policy model parameters.

[0054] Compared with the prior art, the present application has the following advantages:

[0055] The present application replaces the traditional centralized control through distributed state information interaction and topology detection, improves the fault tolerance and topology adaptation capability of cluster scheduling, and accurately matches different scheduling requirements and cluster states through an adaptive policy selection mechanism, taking into account response speed and operation economy; distributed negotiation ensures the consistency and safety of power distribution through accurate handling of exceptions. In this way, not only does it solve the delay and single-point failure problem of centralized scheduling, allowing the energy storage converter group to operate stably under dynamic topology and complex scheduling instructions, but also reduces the total loss of the cluster and improves energy utilization efficiency; it can also adapt to more complex application scenarios through flexible expansion of alternative strategies. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 The timing flow chart of the distributed collaborative control of the present application. DETAILED DESCRIPTION

[0057] The technical solutions of the present application will be described below in conjunction with the embodiments, obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0058] Please refer to Figure 1The application provides a scheduling method based on energy storage converter group cooperative control, which is executed in parallel by each energy storage unit in the energy storage unit cluster, and includes:

[0059] Each energy storage unit interacts with its adjacent units to jointly maintain a distributed system state information library, and by designing a set of maintenance processes including five core links, the application aims to ensure data standardization and high reliability. The core idea is to replace the data verification function of the central node with regional cooperation, including:

[0060] Each energy storage unit preprocesses the locally collected raw operation data to generate a standardized local data set. In practice, raw operation data refers to unprocessed data collected by each energy storage unit through its own sensors, including SOC raw sampling values, output power pulse signals, and battery temperature analog signals.

[0061] The preprocessing process includes data format conversion, unit unification, and normalization processing. By intentionally unifying the power unit to kilowatts and the temperature unit to Celsius, and using the min-max normalization method to map the data to the 0-1 interval, the purpose is to eliminate the interference of different magnitudes and types of data and lay the foundation for subsequent cross-unit unambiguous interaction.

[0062] Exchange statistical characteristics of local data sets with adjacent units to construct a regional data feature map.

[0063] In this embodiment, to reduce communication load and protect data privacy, the method of transmitting a large amount of raw data is abandoned, and instead, the statistical characteristics (including mean, variance, maximum, and minimum) of the local data set are exchanged with adjacent units to construct a regional data feature map.

[0064] Specifically, each unit sends its statistical characteristics to all adjacent units after each preprocessing. After receiving the characteristics of the adjacent units, we construct a map with unit identifiers as nodes and feature similarity as edges. The higher the similarity of the units, the greater the weight of the edges. In this way, it not only greatly reduces the data transmission volume and avoids communication congestion, but also quickly locates the abnormal unit cluster through the feature map.

[0065] Based on the regional data feature map, the local data set is cleaned to eliminate abnormal data. In this embodiment, the cleaning logic fully utilizes the unit correlation revealed by the aforementioned feature map:

[0066] If the local data of a unit deviates from the statistical characteristics of the high correlation degree (weight ≥ 0.8) adjacent unit in the atlas by more than the preset range (such as mean deviation > 2 times variance), it is determined that the data is abnormal data. In order to prevent misjudgment, we stipulate that if the same type of abnormality occurs for 3 consecutive periods, it is determined that the sensor is faulty and an alarm is triggered. This method can more effectively combine regional correlation characteristics and accurately distinguish between transient fluctuations and real abnormalities compared to traditional single-unit threshold cleaning.

[0067] Cross-validation is performed between adjacent units for the cleaned data. In this embodiment, for the cleaned data, a distributed cross-validation mechanism is introduced to further improve data reliability. Specifically, each unit sends the cleaned data to 3 high correlation degree adjacent units, which make judgments based on their own data and regional feature atlas.

[0068] By setting that if 2 or more of the 3 adjacent units feedback "validation passed", the data is considered valid. This design takes advantage of the correlation of units within the region to achieve distributed verification, so that even if a unit is not completely cleaned due to local interference, the problem can be discovered through the verification of adjacent units.

[0069] The validated data is stored in the distributed system state information library. In implementation, we index according to the structure of "data type-collected time-unit identifier" for fast query, and use incremental update to reduce storage resource occupation. Each data entry is attached with cross-validation pass identification and verification unit information to ensure traceability.

[0070] In summary, this step not only realizes data standardization and reliability guarantee, but also fundamentally solves the inherent delay and single-point risk problem of traditional centralized data processing through a complete distributed collaborative processing process.

[0071] Based on state information interaction, the topology structure change of the energy storage unit cluster is detected, and the specific detection process includes:

[0072] When the communication link state change with adjacent units is detected, the topology change confirmation process is triggered. In this embodiment, the communication link state change includes link disconnection, connection recovery, and connection quality drop (such as packet loss rate from ≤5% to ≥50%). Each energy storage unit monitors the link state parameters of each adjacent unit through the communication module in real time, including link connectivity, packet loss rate, and transmission delay.

[0073] When the link state parameter is monitored for 2 consecutive interaction cycles (1 second) to be out of the normal range (such as connectivity interruption, packet loss rate ≥ 50%, transmission delay ≥ 100 ms), or the link is monitored to recover from the abnormal state to the normal range, the topology change confirmation process is triggered immediately to avoid false triggering of detection due to transient link interference and improve detection accuracy.

[0074] The confirmation request is sent to the adjacent units in normal communication, and the topology observation feedback of these adjacent units is collected. Specifically, the confirmation request includes the identification of the initiating unit, the specific information of the link state change (such as the interruption of the link of a certain adjacent unit, the time of interruption occurrence), and the feedback time limit (set to 1 interaction cycle 500 ms) when the feedback is requested.

[0075] The initiating unit only sends a request to the adjacent units in normal communication, reducing the invalid communication overhead.

[0076] Each adjacent unit receiving the request generates a topology observation feedback based on the link state data and regional data characteristic map maintained by itself, and the feedback content includes "agree to link state change", "oppose link state change", and the link state data observed by itself (such as whether the link between the adjacent unit and the target unit is normal).

[0077] By comparing the topology observation feedback of each unit, the majority decision principle is used to confirm the effective topology change event. In this embodiment, the number of valid feedback (received within the feedback time limit) collected is counted, and if the number of "agree to link state change" feedback accounts for more than 50%, it is confirmed that there is an effective topology change event, and the specific change type is determined according to the link state change:

[0078] Link interruption or recovery corresponds to "adjacent unit offline / online" topology change, and newly added normal communication unregistered units correspond to "new unit access" topology change;

[0079] If the "agree" feedback accounts for ≤ 50%, it is determined to be transient link interference, and the topology change is not confirmed, and the detection process is terminated. Through the observation feedback of multiple units and the majority decision principle, the problem of false judgment of link state due to communication module failure of a single unit is avoided.

[0080] The system topology map is updated based on the confirmation result. In this embodiment, the system topology map is the cluster topology structure data maintained locally by each energy storage unit, which records the identification and adjacent relationship of all units in the form of node-edge, the node attribute includes the unit running state, and the edge attribute includes the link quality.

[0081] Further, after confirming that there is a topology change event, the initiating unit immediately updates the local system topology map, such as deleting the offline unit node and the corresponding link, adding the online unit node and the link, and updating the link quality attribute;

[0082] The updated topology picture segment is then synchronized to other adjacent units through a normal communication link, each unit is sequentially synchronized and updated, and the full cluster topology graph update completion time is less than or equal to 1 second. The updated topology graph will serve as the basis data for subsequent state information interaction and power distribution strategy selection, ensuring that the scheduling decision is based on the latest cluster topology structure.

[0083] The present application realizes accurate detection and rapid synchronization of topology changes, upgrades the link state monitoring of a single unit to multi-unit cooperative confirmation, solves the pain points of false positives in traditional distributed detection, and at the same time, through the phased feedback and synchronization mechanism, ensures the accuracy of detection while considering the real-time of topology update, so that the cluster can quickly adapt to dynamic scenarios such as new units, exiting units or link faults, and provide a reliable topology basis for subsequent cooperative scheduling.

[0084] In response to detecting a change in the topology structure or receiving a total power scheduling instruction, an adaptive power distribution strategy selection signal is generated based on the distributed system state information library.

[0085] In this embodiment, the total power scheduling instruction is issued by a superior scheduling center (such as a microgrid scheduling platform), and the instruction includes the target total output power, the scheduling response time and the priority (high, medium and low).

[0086] High-priority instructions (such as grid frequency modulation instructions) require a shorter response time, medium-priority instructions (such as regular charge and discharge scheduling) require a moderate response time, and low-priority instructions (such as energy storage capacity balancing scheduling) require a more relaxed response time.

[0087] The process of generating the adaptive power distribution strategy selection signal specifically includes:

[0088] First, based on the distributed system state information library, a dynamic balance coefficient representing the uniformity of system power distribution and a cooperative power response threshold required by the system are calculated. In this embodiment, the calculation method of the cooperative power response threshold T is:

[0089] T = max(|ΔP| / R, M x k); where ΔP: the difference between the target total power and the current total power (unit: kW); R: the response time required by the scheduling instruction (unit: s); M: the sum of the maximum power regulation rates of all available units in the cluster (unit: kW / s); k: safety factor, value 0.8.

[0090] The dynamic balance coefficient representing the uniformity of system power distribution includes:

[0091] Each energy storage unit monitors the local power regulation capability and operating condition in real time. In this embodiment, the power regulation capability refers to the maximum output power that the unit can increase or decrease under the current state, and the monitoring indicators include the difference between the current output power and the rated power, the battery SOC (30%-70% is the optimal interval for regulation capability), the battery temperature (25℃-35℃ is the optimal interval for regulation capability), and the operating load of the energy storage converter;

[0092] The operating condition includes the current load type (resistive load / inductive load), whether it is in the charging and discharging switching process, whether there is a local fault warning, etc. The monitoring data is updated every 500ms to ensure real-time reflection of the unit regulation potential and operating state.

[0093] Through the exchange of state information between adjacent units, a panoramic view of the system operating state is constructed.

[0094] In this embodiment, each unit encapsulates the local monitoring data of power regulation capability and operating condition in the format of "unit identifier-monitoring indicator-value-time stamp", and sends it to all adjacent units;

[0095] After receiving, the adjacent units integrate their own data and adjacent data, and then synchronize to their own adjacent units. Through three rounds of iterative interaction, the data of the whole cluster is synchronized.

[0096] Based on the panoramic view constructed by the synchronized data, a topological graph is used as the basic framework, the unit regulation capability and operating condition data are labeled on the nodes, and the power transmission efficiency between units is labeled on the edges, to intuitively present the whole system operating state.

[0097] It should be noted that during the data synchronization process, if different units have version conflicts (such as different timestamps) for the state data of the same node, the data with the latest timestamp is used to overwrite the old data, to ensure that the panoramic view data maintained by each unit in the whole cluster is consistent.

[0098] Based on the panoramic view, the power support relationship between each unit is analyzed, and the key nodes of the system are identified.

[0099] In this embodiment, when the output power of a unit is insufficient, if its adjacent units can quickly supplement the corresponding power and the regulation capability matching degree is ≥80%, we determine that there is a "positive support relationship" between the two.

[0100] Subsequently, the number of positive support relationships of each unit (i.e. the number of adjacent units that can effectively support the node) and the number of times of being supported (i.e. the number of adjacent units that can effectively support the node) are counted, and then a comprehensive judgment is made in combination with the importance of its position in the topological graph.

[0101] And in order to quantitatively "position importance", we introduce the "betweenness centrality" index in graph theory, the top 20% of the unit is located in the core connectivity area.

[0102] The identification rules of the system key nodes are as follows:

[0103] At the same time, the following three conditions are met:

[0104] 1. The betweenness centrality ranking of the node in the graph is ≤20%.

[0105] 2. The number of positive support relationships is ≥N×10%. Wherein, N is the total number of units of the cluster.

[0106] 3. The number of supported times is ≥5.

[0107] The influence weight of the running state of the system key node on the overall balance of the system is evaluated. In this embodiment, the influence weight is determined by the analytic hierarchy process, and a three-level evaluation model of "target layer-criterion layer-index layer" is constructed:

[0108] The target layer is the influence weight of the key node on the system balance, the criterion layer includes the adjustment ability, the support range, and the running stability, and the index layer corresponds to the specific monitoring index (such as the adjustable power ratio corresponding to the adjustable power, the number of positive support units corresponding to the support range, and the working condition fluctuation amplitude within 1 minute corresponding to the running stability).

[0109] The weights of each index are determined by expert scoring, the comprehensive score of each key node is calculated, and the score is normalized to obtain the influence weight ω_i (0<ω_i≤1). The higher the weight, the greater the influence of the node on the system balance, and the sum of the weights of all key nodes is ≥0.6, ensuring that the core influencing factors are considered.

[0110] The dynamic balance coefficient is dynamically calculated according to the influence weight. The dynamic balance coefficient (K) is used to quantitatively represent the overall power distribution uniformity and running stability of the system considering the influence weight of the key node, and its value range is 0 to 1. The closer to 1, the more balanced the system power distribution, and the higher the stability.

[0111] Specifically, the traditional coefficient only focuses on the power value distribution and cannot reflect the influence of the core unit state on the system balance. Therefore, the dynamic balance coefficient K is designed, which innovatively combines the running state of the key node and the power distribution uniformity of the non-key node. The calculation method of the dynamic balance coefficient K is as follows:

[0112] 1. Calculation of the key node running state score S:

[0113] S is a comprehensive score, which is calculated by the following formula:

[0114] S=A×B×C;

[0115] A (State of Charge Factor): A = 1 - |D - 0.5| / 0.5. Wherein, D is the battery state of charge (SOC) of the key node;

[0116] B (Power Regulation Capability Factor): B = E / F. Wherein, E is the output power currently available for the node (unit: kW), and F is the rated power of the node (unit: kW);

[0117] C (Health Status Factor): The initial value is 1. When the battery temperature G exceeds the maximum allowable temperature G_max, C is attenuated as (G_max / G) ^ 2;

[0118] 2. Calculation of power distribution uniformity σ of non-key nodes:

[0119] σ is the standard deviation coefficient of the output power of the non-key node group, σ = Std(H) / Mean(H). Wherein, H is the output power set of the non-key node group;

[0120] Comprehensive calculation:

[0121] The calculation formula of the dynamic balance coefficient K is: K = (Σω × S) × 0.7 + (1 - σ) × 0.3. Wherein, ω is the influence weight of each key node;

[0122] The value range of the coefficient K is 0-1. In practice, we agree that: K ≥ 0.8 is judged as balanced power distribution of the system, 0.5 ≤ K < 0.8 is judged as basically balanced, and K < 0.5 is judged as unbalanced. Thus, the problem that the traditional coefficient cannot reflect the influence of the core unit state on the system balance is solved, and the coefficient is more suitable for the actual operation scene.

[0123] Subsequently, according to the comparison result of the dynamic balance coefficient K and the cooperative power response threshold T, and the type of the topology structure change, an adaptive power distribution strategy selection signal is generated.

[0124] Specifically, the type of topology structure change is divided into three types:

[0125] Type A is a slight change (the proportion of the number of added / offline units to the total number of units is small, and there is no key link interruption);

[0126] Type B is a moderate change (the proportion of the number of added / offline units to the total number of units is moderate, or a single non-key link is interrupted);

[0127] Type C is a severe change (the proportion of the number of added / offline units to the total number of units is large, or multiple key links are interrupted, and the key link refers to the link connecting the core unit of the cluster).

[0128] Specifically, the judgment rules for high, medium and low T values are:

[0129] A high response threshold T_high and a low response threshold T_low are set, and the calculation method is T_high=0.7*M, T_low=0.3*M, wherein M is the sum of the maximum power adjustment rates of all available units in the cluster in the current state (unit: kW / s) calculated as described above;

[0130] If T≥T_high, it is a high response demand; if T_low≤T<T_high, it is a medium response demand; and if T<T_low, it is a low response demand.

[0131] The selection signal is in a coded form, corresponding to the preset alternative power distribution strategy, and the specific matching rules are as follows:

[0132] When K<0.5 (power imbalance) and T<T_low (low response demand), the code corresponding to the gradual power distribution strategy is generated regardless of the topology change type;

[0133] When T≥T_high (high response demand) and the topology change type is A or B, the code corresponding to the fast power distribution strategy is generated regardless of the value of K;

[0134] When 0.5≤K<0.8 (basic balance), T_low≤T<T_high (medium response demand), and the topology change type is A, the code corresponding to the SOC balancing strategy or the minimum loss strategy can be selected.

[0135] This step generates a selection signal through dynamic parameter calculation and multi-scene matching, which not only realizes adaptive selection of strategies, but also brings non-obvious advantages: compared with traditional fixed strategy scheduling, the strategy can be flexibly switched according to the real-time state of the cluster and the scheduling demand, which not only guarantees the scheduling effect, but also prolongs the service life of the energy storage unit, for example, the minimum loss strategy is used when the power distribution is balanced and the response demand is low, which reduces unnecessary power adjustment loss and solves the problem that traditional fixed strategies cannot meet multiple dimensional demands.

[0136] Based on the selection signal, a corresponding strategy is selected and executed from a plurality of alternative power distribution strategies. In this embodiment, the plurality of alternative power distribution strategies at least includes a gradual power distribution strategy and a fast power distribution strategy, both of which are preset in the control program of each energy storage unit. Auxiliary strategies such as SOC balancing and minimum loss can be expanded according to the scene, and the diversified design of the strategy can adapt to different working condition demands and improve the scheduling flexibility.

[0137] Further, the gradual power distribution strategy focuses on the smoothness of power change and is suitable for scenes where the power quality requirement is high and power sudden change is not allowed, such as micro-grid energy storage clusters that supply power to precision production equipment.

[0138] The implementation logic is: taking the actual output power of each unit as the reference, the power adjustment step and step interval are calculated according to the difference between the allocated target power and the reference power, and the target power is reached through multi-step gradual adjustment. The step size is related to the type of the unit, and the step size of the lithium battery energy storage unit can be slightly larger, and the step size of the lead-acid battery energy storage unit needs to be smaller;

[0139] The step interval is set to 2-3 times of the regular adjustment period, ensuring that the power change rate is controlled within the smooth range allowed by the equipment. At the same time, the battery voltage and temperature data of each unit are collected in real time during the adjustment process, and if the parameter fluctuation exceeds the smooth threshold, the step size is automatically reduced or the interval is extended.

[0140] The present application avoids the influence of power sudden change on the cycle life of energy storage unit battery through gradual adjustment, reduces the interference of voltage and frequency fluctuation on sensitive loads on the user side, solves the contradiction between slow response and large fluctuation in traditional fixed step adjustment, and balances the smoothness and adjustment efficiency.

[0141] The fast power allocation strategy focuses on the timeliness of power response, and is suitable for scenes with high response speed requirements, such as grid frequency regulation, new energy output fluctuation suppression, etc.

[0142] It should be further pointed out that: the energy storage units with fast response speed and sufficient adjustment margin in the cluster are preferentially selected (such as lithium battery units with fast charge and discharge capability, units with SOC in the optimal adjustment interval of 30%-70%), and most of the target power is allocated to these units, and the remaining small part of the power is supplemented by other units.

[0143] When allocating, a one-time adjustment instruction of "target power-current power" is directly used, and small power deviation (allowing single adjustment deviation within ±5%) is ignored, and subsequent correction is made through short period fine adjustment.

[0144] The present application avoids the response delay caused by synchronous fine adjustment of all units under the premise of ensuring the core response timeliness, reduces the adjustment frequency, reduces the switching loss of energy storage converter, and solves the problem of high loss in traditional fast adjustment.

[0145] It should be understood that the allocation formula of the SOC balancing strategy is:

[0146] P=P_all×(Q_avg / Q)×k;

[0147] P: the allocated power of the unit (unit: kW).

[0148] P_all: the target total power of the system (unit: kW).

[0149] Q_avg: the average battery state of charge (SOC) of all units in the cluster.

[0150] Q: The state of charge (SOC) of this cell.

[0151] k: Unit power limitation factor, set according to the unit's rated power, with a value of 0.8-1.0.

[0152] This strategy can effectively reduce the SOC difference between units and avoid overcharging and over-discharging of some units, bringing non-obvious advantages: the balanced SOC state gives the cluster more room for adjustment in subsequent scheduling, improves the overall response capability of the cluster, and solves the problem in traditional scheduling where some units cannot participate in discharge due to low SOC.

[0153] The minimum loss strategy is suitable for long-term operation scenarios. The allocation logic is based on a simplified loss model of each unit (loss is related to output power), and the power allocation scheme with the minimum total loss is solved by conventional optimization algorithm.

[0154] This strategy can reduce the total loss of the cluster, and the reduction in loss over long-term operation not only reduces energy waste, but also reduces the heat generated by the loss, reduces the burden on the heat dissipation system, and improves the overall operational stability of the cluster.

[0155] After selecting a strategy, each unit calculates its own target output power according to the allocation scheme and performs initial power adjustment.

[0156] During the power allocation process, a distributed negotiation mechanism is used to verify the consistency of the power allocation results for each energy storage unit. Specifically:

[0157] Each energy storage unit collects local power allocation execution data. In this embodiment, power allocation execution data refers to the real-time operating data generated after each unit executes the initial power allocation strategy, including actual output power, power regulation rate, battery SOC change rate, battery temperature change trend, and energy storage converter operating status parameters. The data acquisition cycle is consistent with the status information interaction cycle to ensure data timing synchronization. A data acquisition timestamp and unit identifier are included during collection to provide traceability for subsequent correlation and comparison.

[0158] The execution data is anomaly identified, and abnormal data points are marked. In this embodiment, anomaly identification is based on a preset threshold range and trend rules:

[0159] The threshold range is set according to the unit's rated parameters. If the actual output power exceeds the rated power by ±10% or the power regulation rate exceeds the allowable range, it is judged as an abnormal threshold.

[0160] The trend change rule refers to that the fluctuation amplitude of data in three consecutive periods is greater than or equal to 5% and no reasonable scheduling instruction is triggered, and the trend change is determined to be abnormal. After the identification is completed, the abnormal data points are marked with the abnormal type (threshold abnormality / trend abnormality), the occurrence time and the associated power distribution strategy code, so as to facilitate subsequent targeted analysis.

[0161] The labeled abnormal data points are compared with the execution data of the adjacent units. In this embodiment, the labeled abnormal data points and the corresponding normal period data are sent to three high-correlation adjacent units by each unit;

[0162] After receiving, the adjacent units extract the execution data of the same period and the abnormal labeling information of other adjacent units, construct a data correlation matrix, take time as the horizontal axis and unit identification as the vertical axis, mark the abnormal state of the data of each unit in the same period, and focus on comparing whether the occurrence time, abnormal type and value deviation amplitude of the abnormal data points are correlated.

[0163] According to the correlation comparison result, the system-level abnormality and the unit-level abnormality are distinguished.

[0164] The distinguishing rule is that if an abnormal data point presents the characteristics of “multi-unit same period same type abnormality” in the correlation matrix, and the abnormal value deviation trend is consistent, such as insufficient power regulation rate in multiple units in the same period, it is determined to be a system-level abnormality, which is usually caused by unsynchronized topology change and insufficient strategy parameter adaptability.

[0165] If only a single unit has an abnormal data point, and the data of the same period of the adjacent units is normal, it is determined to be a unit-level abnormality, which is usually caused by unit sensor error and local regulation mechanism jam. The innovation of this step is to break through the traditional single abnormality determination mode, accurately locate the abnormal source through correlation characteristics, and avoid misjudging the unit-level abnormality as system problem leading to excessive processing.

[0166] The data of the unit-level abnormality is corrected, and the power redistribution process is triggered for the system-level abnormality. In this embodiment, the unit-level abnormality correction adopts the “adjacent data reference + historical baseline correction” mode:

[0167] The mean value of the normal execution data of the three adjacent units in the same period is taken as the reference baseline, the normal data baseline of this unit under the same working condition is combined, the correction value is calculated and the abnormal data point is corrected, and the deviation of the corrected data should be less than or equal to 2%;

[0168] When the system-level abnormality triggers the redistribution process, the redistribution request is initiated by the unit that identifies the abnormality earliest, synchronizes to the whole cluster with the correlation comparison result, and generates a power distribution scheme based on the latest topology graph and state information library.

[0169] The corrected data can be used for final consistency verification only after being confirmed by the adjacent units.

[0170] In this embodiment, after the unit-level exception correction is completed, the corrected data and the correction basis are sent to the original three high-correlation adjacent units, the adjacent units verify the rationality of the correction logic and the accuracy of the correction result, and if two or more adjacent units feed back "confirmation effective", the corrected data takes effect;

[0171] In the final consistency verification, each unit collects its valid execution data (corrected data or normal data) to the adjacent unit, calculates the deviation of the actual total output power of the whole cluster from the target total power, and if the deviation is less than or equal to 3% and the data of each unit has no unprocessed exception, the verification is passed, and the final power distribution instruction is executed.

[0172] The distributed negotiation verification mechanism avoids the overall verification failure caused by local exceptions in traditional verification through the logic of exception processing first and then verification. Meanwhile, the differentiated processing of system-level and unit-level exceptions not only solves small faults of units through local correction, but also copes with system-level problems through redistribution, thereby improving the verification pass rate and scheduling stability and solving the low efficiency problem caused by one-size-fits-all processing of exceptions in traditional verification.

[0173] In summary, the present application replaces the traditional centralized control through distributed state information interaction and topology detection, thereby improving the fault tolerance and topology adaptation capability of cluster scheduling; the adaptive strategy selection mechanism can accurately match different scheduling requirements and cluster states, and takes into account the response speed and operation economy; and the distributed negotiation verification ensures the consistency and safety of power distribution through accurate processing of exceptions. In this way, not only the problems of delay and single-point fault of centralized scheduling are solved, and the energy storage converter group can stably operate under dynamic topology and complex scheduling instructions, but also the total loss of the cluster is reduced, and the energy utilization efficiency is improved; and more application scenarios can be adapted through flexible expansion of alternative strategies, including but not limited to integrated photovoltaic storage and charging power stations, user-side peak-valley arbitrage energy storage clusters, and microgrid frequency modulation auxiliary services.

[0174] In this embodiment, when the energy storage converter group executes the power distribution strategy, the units may have execution deviations due to sensor errors, mechanism jamming, etc., and the system level may have overall imbalance due to unsynchronized topology changes and insufficient strategy adaptability. If the unverified distribution result is directly executed, it may cause unstable power supply, unit overload, and even cluster failure.

[0175] As an embodiment of the present application, the method further comprises:

[0176] Each energy storage unit adds a local timestamp to the interactive state information; in this embodiment, the local timestamp refers to the time identifier generated by the clock module of each energy storage unit, which is in the format of "year-month-day hour: minute: second.millisecond", accurate to 1 millisecond, to ensure that the time sequence differentiation accuracy meets the industrial control requirements.

[0177] The clock module adopts a real-time clock (RTC) module built in the energy storage unit, has a power-off retention function, and an initial clock can be remotely calibrated by a superior dispatching center or manually calibrated locally. When encapsulating state information, a local timestamp is added as a fixed field to the head of a data frame, and the timestamp field, state information (SOC, output power, etc.), and a unit identification field together constitute a complete interactive data frame. The data frame is checked by CRC-16 to ensure that the timestamp field is not tampered with during transmission. The design binds the timestamp and state information to provide a direct basis for subsequent clock deviation calculation, solving the problem of being unable to trace the data generation timing in traditional timestamp-free interaction.

[0178] The receiving unit calculates the clock deviation between the receiving unit and the sending unit according to the local timestamp; in this embodiment, the clock deviation calculation adopts a "two-way interaction method", and the specific process is as follows:

[0179] When the receiving unit receives the data frame carrying the timestamp T1 sent by the sending unit, the current local time T2 of the receiving unit is immediately recorded;

[0180] Further, the receiving unit returns a response data frame to the sending unit, and the response data frame carries the local timestamp T2 of the receiving unit and the timestamp T1 of the data received from the sending unit; after receiving the response frame, the sending unit records the current local time T3 of the sending unit, and feeds back T3 to the receiving unit.

[0181] Further, the calculation formula of the clock deviation Δt is as follows:

[0182] Δt = [(T2-T1) + (T3-T4)] / 2; wherein T1 is the local timestamp of the unit A sending a synchronization request packet, T2 is the local timestamp of the unit B receiving the request packet, T3 is the local timestamp of the unit B sending a response packet, and T4 is the local timestamp of the unit A receiving the response packet.

[0183] Only when the absolute value |Δt| of the calculated clock deviation is greater than 5 ms, the clock calibration operation is performed.

[0184] Through a multi-hop transmission mode, clock synchronization information is transmitted in the system; in this embodiment, the clock synchronization information refers to the clock deviation data calculated by each unit, the clock calibration state (not calibrated / calibrated / calibrating) of the unit itself, and the identification of the reference clock unit. The reference clock unit is the unit with the highest initial clock calibration accuracy and the most stable operation in the cluster, which is selected by a voting mechanism during the initialization of the cluster (the unit with a voting rate of ≥60% is selected as the reference).

[0185] The multi-hop transmission adopts a "adjacent forwarding + priority sorting" mechanism: each unit preferentially forwards the synchronization information to the adjacent unit in the direction of the reference clock unit, and marks the forwarding hop number (the initial forwarding unit hop number is 0, and the hop number is increased by 1 each time the forwarding is performed) when forwarding. The forwarding is stopped when the hop number exceeds 5 to avoid invalid transmission.

[0186] The synchronization information transmission period is consistent with the state information interaction period (500 ms), ensuring real-time diffusion of deviation data. When a reference clock unit failure is detected (no synchronization information output for more than 3 periods), a new reference unit election is triggered immediately, ensuring the continuity of synchronization information transmission. This transmission mode reduces the communication network load while ensuring full-cluster synchronization information coverage through directional forwarding and hop limit.

[0187] Based on the transmission result, the clock of each unit is dynamically calibrated.

[0188] In this embodiment, the dynamic calibration adopts a "reference tracing + multi-source verification" strategy, and the specific process is as follows: each unit collects the clock synchronization information transmitted by all adjacent units, and selects valid information with a hop count ≤ 3 (the smaller the hop count, the smaller the deviation accumulation);

[0189] The clock deviation data related to the reference clock unit is extracted from the valid information, and the average deviation Δt_avg is calculated. If the absolute value of the average is ≤ 5 milliseconds (the allowable error range of industrial control clock synchronization), it is determined that the current clock does not need to be calibrated, and only the local deviation cache is updated. If the absolute value of the average is > 5 milliseconds, a calibration operation is performed, and the clock value of the unit is adjusted to "current local time - Δt_avg".

[0190] During the calibration process, the unit suspends the execution of the power adjustment instruction (the state information collection is normal), sends a "calibration complete" notification to all adjacent units after the calibration is completed, and the adjacent units recalculate the clock deviation with the unit after receiving the notification. If the deviation of a unit with the reference clock is still > 5 milliseconds after 3 consecutive calibrations, it is determined that the clock module of the unit is faulty, a local alarm is triggered, and the fault information is reported to the upper dispatch center, and at the same time, the average clock deviation of the adjacent unit is temporarily used to replace the clock of the unit, to ensure that it does not affect the overall synchronization of the cluster. This calibration strategy ensures synchronization accuracy through reference tracing, avoids mis-calibration caused by single deviation data through multi-source verification, and improves the reliability of clock synchronization.

[0191] The coordination logic of the clock synchronization method and the original scheduling process: the clock synchronization is throughout the core links such as state information interaction, topology detection, and consistency verification. During state information interaction, the timestamp ensures the synchronization analysis of data in the same period by each unit; during topology detection, the timestamp can accurately determine the occurrence time of link changes, avoiding misjudgment of topology changes caused by asynchronization; during consistency verification, the synchronized clock ensures the time sequence alignment of the power execution data of each unit, improving the accuracy of abnormal identification and correlation comparison. The whole realizes the deep integration of clock synchronization and scheduling process, and solves the scheduling deviation problem caused by asynchronous clock of distributed cluster.

[0192] As an embodiment of the present application, the method further comprises:

[0193] A state set containing system state, topology feature and load characteristic is constructed; in this embodiment, the state set is a basic data set for strategy model training and effect evaluation, covering three types of core features: the system state features include cluster average SOC, total output power, dynamic balance coefficient and cooperative power response threshold;

[0194] The topology features include topology change type, key node number and link connectivity rate;

[0195] The load characteristics include load type (resistive / inductive / capacitive), load fluctuation amplitude (maximum fluctuation value to average value ratio within 5 minutes) and load duration. Each feature is normalized (mapped to the 0-1 interval) to eliminate dimensional influence, and the feature data is aggregated according to "10 minutes as a time window", each time window corresponds to a state set record, and the record associates the power distribution strategy code executed in the window and the execution effect data. The state set is stored locally in each unit, only sharing feature data but not raw running data, balancing model training needs and data privacy protection.

[0196] A comprehensive evaluation function of strategy execution effect is defined; in this embodiment, the comprehensive evaluation function is used to quantify the execution effect of different strategies under the corresponding state set, and the evaluation indexes include three dimensions: economic index (cluster total loss rate, i.e. total loss to total output power ratio), stability index (power fluctuation standard deviation), and responsiveness index (actual response time to required response time ratio).

[0197] The evaluation function formula is F=a×(1-L)+b×(1-σ)+c×(1-T_ratio);

[0198] Wherein, L is the loss rate, σ is the normalized value of power fluctuation standard deviation, T_ratio is the response time ratio, a, b, c are weight coefficients (a+b+c=1, determined by AHP, economic scenario a=0.4, stability scenario b=0.4, responsiveness scenario c=0.4). The value range of F is 0-1, and the closer the F value is to 1, the better the strategy execution effect is. After each time window ends, each unit calculates the F value based on the local state set and execution data, which is the core evaluation basis of strategy experience.

[0199] Through a privacy-protected distributed data sharing mechanism, strategy execution experience is exchanged among units; in this embodiment, each unit anonymizes the state set features in the experience data, deleting the identification information that can be associated with specific units;

[0200] Then, a small Gaussian noise (noise standard deviation ≤0.01) is added to the evaluation function value F to realize differential privacy protection, avoiding the inference of the original running data from the F value.

[0201] When sharing, the "adjacent unit on-demand sharing" mode is adopted. When a unit encounters a new state set (the similarity with the local history set is <80%), it sends an experience query request to the three adjacent units with high correlation, and the adjacent units only return the anonymized "state set feature-strategy code-F value" data, without transmitting the original data.

[0202] The federated learning method is adopted to collaboratively update the strategy model parameters.

[0203] In this embodiment, by selecting federated learning, the core purpose is to gather the running experience of the whole cluster to improve the intelligence level of the strategy model under the premise of strictly complying with data privacy protection.

[0204] Further, the initial model is constructed using the decision tree algorithm. Among various decision tree algorithms, the CART (Classification and Regression Tree) algorithm is selected, mainly because its model structure is clear, the calculation efficiency is high, and the interpretability is strong, which is very suitable for deployment and operation on edge devices such as energy storage units which have limited computing resources.

[0205] In the initial parameter setting, the maximum depth of the tree is set to 10, and the minimum sample split number of the node is set to 5. These values are a good starting point between preventing overfitting and ensuring model performance according to our engineering experience, and can be continuously optimized through the federated learning process.

[0206] The federated learning update process is as follows: the reference clock unit serves as the federated learning coordinator, and initiates a model update every 24 hours. Each unit trains a local model based on the latest experience data, and calculates the model parameter update amount. To ensure data security, each unit sends the parameter update amount to the coordinator after homomorphic encryption processing. The coordinator performs weighted averaging on the encrypted update amounts of all units, generates a global parameter update amount, and decrypts it to broadcast to each unit. Each unit adjusts the local model parameters accordingly to complete the collaborative update.

[0207] The innovation of the present application lies in that the clock synchronization ensures the consistency of the scheduling timing, and the model optimization and experience sharing improve the strategy adaptation accuracy. Both of them cooperate with the core scheduling process, which not only solves the asynchronous and strategy rigidification problems of the distributed cluster, but also considers data privacy and running efficiency, significantly improving the reliability and economy of the energy storage group scheduling.

[0208] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A scheduling method based on collaborative control of energy storage converter groups, characterized in that, The method is executed in parallel by each energy storage unit in the energy storage unit cluster, including: Each energy storage unit interacts with its neighboring units to exchange status information and jointly maintains a distributed system status information database. Based on the interaction of the state information, the topological changes of the energy storage unit cluster are detected; In response to detecting the topology change or receiving a total power scheduling command, an adaptive power allocation strategy selection signal is generated based on the distributed system state information database; Generate an adaptive power allocation strategy selection signal, including: Based on the distributed system state information database, calculate the dynamic balance coefficient K, which characterizes the uniformity of power distribution in the system, and the required cooperative power response threshold T of the system; Based on the comparison result between the dynamic balance coefficient K and the cooperative power response threshold T, and the type of topology change, the adaptive power allocation strategy selection signal is generated; The calculation method for the coordinated power response threshold T is: T=max(|ΔP| / R,M×k), where ΔP is the difference between the target total power and the current total power, R is the response time required by the scheduling instruction, M is the sum of the maximum power adjustment rates of all available units in the cluster, and k is the safety factor. The dynamic balance coefficient K is calculated as follows: K=(Σω×S)×0.7+(1-σ)×0.3, where ω is the influence weight of each critical node; S is the comprehensive score, S=A×B×C, where A is the state of charge factor, B is the power regulation capability factor, C is the health status factor; and σ is the standard deviation coefficient of the output power of the non-critical node group. Based on the selection signal, a corresponding strategy is selected and executed from a variety of alternative power allocation strategies; During the power allocation process, a distributed negotiation mechanism is used to verify the consistency of the power allocation results of each energy storage unit, and the final power allocation command is executed after the verification is successful.

2. The scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, The multiple alternative power allocation strategies include at least a progressive power allocation strategy and a fast power allocation strategy; The gradual power allocation strategy focuses on the smoothness of power changes, while the fast power allocation strategy focuses on the timeliness of power response.

3. The scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, The jointly maintained distributed system state information database includes: Each energy storage unit preprocesses the raw operating data collected locally to generate a standardized local dataset; The statistical features of the local dataset are exchanged with those of adjacent units to construct a regional data feature map. Based on the regional data feature map, the local dataset is cleaned to remove abnormal data; The cleaned data is cross-validated between adjacent cells; The verified data is stored in the distributed system state information database.

4. The scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, Detecting topology changes in the energy storage unit cluster includes: When a change in the communication link status with a neighboring unit is detected, the topology change confirmation process is triggered. Send acknowledgment requests to neighboring units with normal communication and collect topology observation feedback from these neighboring units; By comparing the topology observation feedback from each unit, the majority decision principle is used to confirm valid topology change events; Update the system topology diagram based on the confirmation results.

5. A scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, The consistency of power allocation results for each energy storage unit is verified through a distributed negotiation mechanism, including: Each energy storage unit collects local power allocation execution data; The execution data is anomaly identified, and abnormal data points are marked. The annotated abnormal data points are correlated and compared with the execution data of adjacent units; Based on the correlation comparison results, distinguish between system-level anomalies and unit-level anomalies; Perform data correction for unit-level anomalies and trigger power redistribution processes for system-level anomalies; The corrected data can only be used for final consistency verification after being confirmed by the adjacent units.

6. The scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, The method further includes: Each energy storage unit adds a local timestamp to the status information it interacts with; The receiving unit calculates the clock deviation between itself and the sending unit based on the local timestamp; Clock synchronization information is transmitted within the system via multi-hop transmission. The clocks of each unit are dynamically calibrated based on the transmission results.

7. A scheduling method based on collaborative control of energy storage converter groups according to claim 1, characterized in that, The method further includes: Construct a state set that includes system state, topology characteristics, and load characteristics; Define a comprehensive evaluation function for the effectiveness of strategy execution; Experience in strategy execution can be shared among different units through a privacy-preserving distributed data sharing mechanism. Federated learning is used to collaboratively update the policy model parameters.

Images