Modular energy storage bms collaborative management method and system

Abstract

The application discloses a kind of modular energy storage BMS collaborative management method and system, belong to battery control technical field, this method includes: obtaining the state data of battery module and each physical device in energy storage system, physical device is as the node in heterogeneous graph network, and the edge of connecting node is constructed based on the physical coupling relationship between physical device;State feature vector corresponding to each node is extracted based on state data, electrical weight matrix, thermal weight matrix and control weight matrix are constructed based on the state data between nodes;State feature vector and weight matrix are input into neural network containing multiple heterogeneous graph convolution layers, and the power weight prediction value of each battery module, the temperature prediction sequence of each physical device and the thermal risk score are output, and power distribution instruction and rotating speed control instruction are generated, by the application, new and old battery module can be on-line switching and power is fully released.

Description

Technical Field

[0001] This invention belongs to the field of battery control technology, specifically relating to a modular energy storage BMS collaborative management method and system. Background Technology

[0002] Existing modular energy storage systems typically employ a centralized BMS architecture. The battery capacity is fixed at the factory, and all battery modules operate simultaneously, charging and discharging synchronously. The difference in state of charge between modules is minimal, and the BMS only needs to issue unified charge and discharge commands to the entire battery group to meet management requirements. However, when users need to add battery modules in the field to expand capacity, the new modules must be connected while the system is completely shut down. The main control unit requires manual reconfiguration of parameters, and the expansion operation cannot be completed while the system is under load. For industrial scenarios requiring continuous power supply, such as outdoor welding and cutting, shutdown for capacity expansion means work interruption, reducing equipment utilization efficiency.

[0003] To address the aforementioned issues, some existing technical solutions employ a distributed BMS architecture to support hot-swappable modules. In this approach, each battery module has an independent built-in BMS unit, connected to the main control unit via a communication bus such as RS-485. The main control unit can identify newly connected modules and read their status information during system operation. For charge / discharge management, the main control unit collects parameters such as SOC, voltage, and temperature from each module and then issues the same charge / discharge current command to all online modules. Each module relies on its internal passive balancing circuit to perform small-current balancing on individual cells at the end of charging. This solution solves the problem of hot-swappable module identification. However, because the modules are directly connected to the system bus in parallel, the actual output current of each module is distributed according to Ohm's law based on its terminal voltage and internal resistance. When there is a significant SOC difference between new and old modules, the module with higher SOC will bear a larger discharge current due to its higher terminal voltage, causing the SOC difference between modules to continue to widen during operation. When any module reaches the discharge cutoff voltage first, the main controller must stop the discharge of all modules, resulting in a large amount of usable energy remaining in the other modules being unusable. The actual usable capacity of the system is far lower than the sum of the capacities of each module, and it cannot play an effective role under continuous operation conditions. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a modular energy storage BMS collaborative management method and system to resolve the issues present in the background art.

[0005] To achieve the aforementioned objectives, this invention proposes a modular energy storage BMS collaborative management method, comprising: After the new battery module is inserted into the energy storage system, the status data of the battery module and each physical device in the energy storage system are obtained. The physical devices are treated as nodes in the heterogeneous graph network, and the edges connecting the nodes are constructed based on the physical coupling relationship between the physical devices. The edges include electrical coupling edges, thermal coupling edges and control coupling edges. State feature vectors corresponding to each node are extracted based on state data, edge weights of each edge are calculated based on state data between nodes, and electrical weight matrix, thermal weight matrix and control weight matrix are constructed based on edge weights respectively. The state feature vector and weight matrix are input into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score. Power allocation commands and speed control commands are generated based on power weight prediction values, temperature prediction sequences, and thermal risk scores. Power allocation commands are used to control the access, disconnection, or output power of battery modules, while speed control commands are used to regulate the heat dissipation power of the heat dissipation system in the physical device.

[0006] This invention also provides a modular energy storage BMS collaborative management system, which is used to implement the above-described method. The system includes: The graph construction module acquires the state data of the battery module and various physical devices in the energy storage system after the new battery module is inserted into the energy storage system. The physical devices are treated as nodes in the heterogeneous graph network, and the edges connecting the nodes are constructed based on the physical coupling relationship between the physical devices. The edges include electrical coupling edges, thermal coupling edges, and control coupling edges. The weight generation module extracts the state feature vectors corresponding to each node based on the state data, calculates the edge weights of each edge based on the state data between nodes, and constructs the electrical weight matrix, thermal weight matrix and control weight matrix based on the edge weights respectively. The heterogeneous graph prediction module inputs the state feature vector and weight matrix into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score. The collaborative control module generates power allocation instructions and speed control instructions based on power weight prediction values, temperature prediction sequences, and thermal risk scores. It controls the access, disconnection, or output power of the battery module based on the power allocation instructions, and regulates the heat dissipation power of the heat dissipation system in the physical device based on the speed control instructions.

[0007] The beneficial effects of this invention are as follows: This invention first abstracts physical devices such as battery modules, power devices, and heat dissipation systems as heterogeneous graph nodes and constructs electrical, thermal, and control coupling edges, thereby achieving unified modeling of multi-physics coupling relationships. Then, it extracts state features and calculates edge weights based on state data to construct a weight matrix. After iterative propagation using a multi-layer heterogeneous graph convolutional network, multiple prediction heads collaboratively output power weight prediction values, temperature prediction sequences, and thermal risk scores, accurately identifying the available power capacity of each module and predicting thermal runaway trends. Subsequently, based on the prediction values, it actively generates target power commands and heat dissipation control commands allocated according to weights, enabling online switching between old and new modules with full power release, while simultaneously achieving predictive active heat dissipation. Ultimately, it solves the problem in existing technologies where large amounts of residual capacity are caused by SOC differences, and the system's usable capacity is far lower than the sum of the capacities of each module. This maximizes the actual usable capacity and operational safety of the energy storage system without interrupting operation. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of a modular energy storage BMS collaborative management method according to the present invention. Figure 2 This is a comparison diagram of the discharge process and capacity release effect between the method of the present invention and the prior art; Figure 3 This is a schematic diagram of the structure of a modular energy storage BMS collaborative management system according to the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0010] like Figure 1 As shown, a modular energy storage BMS collaborative management method includes: After a new battery module is inserted into the energy storage system, the state data of the battery module and each physical device in the energy storage system are obtained. The physical devices are treated as nodes in a heterogeneous graph network, and edges connecting the nodes are constructed based on the physical coupling relationship between the physical devices. The edges include electrical coupling edges, thermal coupling edges, and control coupling edges.

[0011] When a new battery module is connected to the system via a quick-connect interface, the BMS unit sends a broadcast message to the main control unit. The broadcast message includes the battery module's rated capacity, current state of charge (SOC), current terminal voltage, and current casing temperature. Upon receiving the broadcast message, the main control unit selects an unused free address, encapsulates this free address into a response message, and replies to the BMS unit. In this embodiment, the main control unit is an edge computing device. In other embodiments, the main control unit can also be a cloud computing device, in which case the BMS unit uploads the broadcast message to the cloud computing device via a cellular network. After receiving the response message containing the free address, the BMS unit uses the free address as the unique communication identifier for the new battery module in the system. Subsequently, the BMS unit uses the free address as the source address to send a second message to the main control unit. The second message includes the battery module's status data, including the number of cells connected in series, the rated voltage of each cell, the cumulative cycle count, and the current equivalent internal resistance.

[0012] After receiving the second message, the main control unit creates a record for the new battery module in the local database. The record includes data such as the new battery module's address, capacity, cell configuration, and historical cycle count, thereby completing the registration of the new battery module.

[0013] In this embodiment, after the new battery module is inserted into the energy storage system, the following is also included: Calculate the difference between the current voltage of the new battery module and the current voltage of the energy storage system bus. Define the absolute value of this difference as the voltage difference. If the voltage difference is greater than the preset voltage threshold, close the pre-charge relay to allow the new battery module to form a pre-charge circuit with the energy storage system bus through the current limiting resistor. In the pre-charge circuit, the current is limited by the current limiting resistor and the voltage difference is reduced. When the voltage difference drops below the preset safety threshold, open the pre-charge relay and close the main relay to allow the new battery module to be connected to the system bus.

[0014] After the new battery module is registered, the main control unit obtains the current terminal voltage of the new battery module and reads the current voltage value of the system bus. It calculates the difference between the current terminal voltage of the new battery module and the current voltage value of the bus, and uses the absolute value of the two as the voltage difference. If the voltage difference is greater than the preset voltage threshold, it indicates that there is a large potential difference between the new battery module and the system bus. If the main relay is closed directly to connect the new battery module to the bus at this time, due to the extremely small initial impedance of the circuit, this voltage difference will generate a transient current in the parallel circuit. Since the transient current is calculated based on the voltage difference divided by the total impedance of the circuit, if the voltage difference is large, the transient current may exceed the rated connecting capacity of the relay contacts, leading to contact erosion. To avoid the above situation, if the voltage difference exceeds the preset voltage threshold, the main control unit sends a pre-charge command to the BMS unit. Based on the pre-charge command, the BMS unit closes the pre-charge relay. A current-limiting resistor is connected in series in the circuit where the pre-charge relay is located. The new battery module forms a circuit with the system bus through this current-limiting resistor. The current-limiting resistor limits the transient current to a safe range. As the pre-charge current flows through the circuit, it slowly charges or discharges the filter capacitor on the system bus side, causing the bus voltage to gradually approach the terminal voltage of the new battery module. At the same time, the polarization effect and internal resistance voltage drop inside the battery module change dynamically under the action of current, causing the terminal potentials of the two to approach each other. Thus, the voltage difference between the new battery module and the bus gradually decreases under the action of the current-limiting resistor.

[0015] During the pre-charging process, the main control unit samples the terminal voltage of the new battery module and the system bus voltage at fixed intervals and recalculates the voltage difference. When the voltage difference drops below the preset voltage threshold, the BMS unit disconnects the pre-charging relay and simultaneously closes the main relay, and the new battery module is officially connected to the system bus and begins to participate in the system's power distribution.

[0016] In the first comparison after registration, if the voltage difference is less than or equal to the preset voltage threshold, the main control unit skips the pre-charging stage and directly sends a main access command to the BMS unit. The BMS unit then closes the main relay to complete the access.

[0017] In this embodiment, the edges connecting nodes are constructed based on the physical coupling relationship between physical devices, including: The physical equipment includes battery modules, power devices, heat dissipation systems, and load interfaces, and electrical coupling edges are established between battery module nodes, between battery module nodes and power device nodes, and between power device nodes and load interface nodes.

[0018] After the battery module is connected, status data of power devices, heat dissipation system, load interface and environmental sensors are collected. The status data of power devices includes junction temperature, switching frequency, drive voltage and heat sink temperature of inverter IGBT module and DC-DC converter MOSFET module. The status data of heat dissipation system includes real-time speed of each cooling fan, inlet and outlet temperature difference, power consumption and vibration frequency. The status data of load interface includes real-time power and current waveform of each load port, and the harmonic content and power factor of the current waveform are extracted by fast Fourier transform. The status data of environmental sensors includes ambient temperature and wind speed.

[0019] Battery modules, power devices, cooling systems, load interfaces, and environmental sensors are abstracted as nodes in a heterogeneous graph network. For example, each battery module corresponds to a battery module node, each inverter IGBT module or DC-DC converter MOSFET module corresponds to a power device node, each cooling fan corresponds to a cooling system node, each load output port corresponds to a load interface node, and each environmental temperature and humidity sensor corresponds to an environmental sensor node. Edge types include electrically coupled edges, thermally coupled edges, and control coupled edges.

[0020] Thermal coupling edges are established between battery module nodes whose physical spatial distance is less than a preset thermal coupling distance threshold, between battery module nodes and adjacent power device nodes, between power device nodes and heat dissipation system nodes responsible for cooling the power device, and between heat dissipation system nodes and environmental sensor nodes.

[0021] Establish control coupling edges between the heat dissipation system node and the power device node cooled by the heat dissipation system, and between the heat dissipation system node and the battery module node cooled by the heat dissipation system.

[0022] Electrical coupling edges exist between nodes with current transmission relationships, such as between battery modules, between battery modules and power devices, and between power devices and load interfaces. Thermal coupling edges exist between nodes with heat conduction or convection heat transfer relationships. For example, a thermal coupling edge can be established between battery module nodes whose physical spatial distance is less than a preset thermal coupling distance threshold. This is because the heat generated by adjacent battery modules during charging and discharging will be conducted and radiated through air gaps, metal mounting brackets, or cabinet structures, creating mutual thermal influence. This means that a temperature rise in one module will exacerbate the heat accumulation in adjacent modules. Furthermore, thermal coupling edges can be established between battery module nodes and adjacent power device nodes, between power device nodes and the heat dissipation system nodes responsible for cooling those power devices, and between heat dissipation system nodes and environmental sensor nodes. Control coupling edges exist between nodes with active control logic relationships, such as the cooling control of power devices by a cooling fan.

[0023] State feature vectors corresponding to each node are extracted based on state data, edge weights of each edge are calculated based on state data between nodes, and electrical weight matrix, thermal weight matrix and control weight matrix are constructed based on edge weights.

[0024] In this embodiment, electrical weight matrix, thermal weight matrix, and control weight matrix are constructed based on edge weights, including: The self-admittance is calculated based on the mutual admittance between a node and all its neighboring nodes connected by electrical coupling edges. The physical admittance matrix is ​​constructed based on the self-admittance of the node and the mutual admittance between the nodes. The physical admittance matrix is ​​then normalized to obtain the electrical weight matrix.

[0025] For battery module nodes, extract the state feature vector corresponding to the battery module. The state feature vector includes the current state of charge, current terminal voltage, current casing temperature, current equivalent internal resistance, and the rate of change of current between the previous sampling period and the current sampling period. For power device nodes, extract state feature vectors including junction temperature, switching frequency, and heat sink temperature. For cooling system nodes, extract state feature vectors including rotational speed, inlet / outlet temperature difference, and power consumption. For load interface nodes, extract state feature vectors including real-time power and harmonic content. For environmental sensor nodes, extract state feature vectors including ambient temperature and wind speed.

[0026] Based on the physical coupling relationships between nodes, electrical weight matrices, thermal weight matrices, and control weight matrices are constructed respectively. For the construction of the electrical weight matrix, the current terminal voltage and equivalent DC internal resistance of each online battery module, as well as the equivalent input impedance of the power device and load interface, are first collected. For the mutual admittance between node i and node j connected by an electrical coupling edge, different calculation methods are used depending on the node type.

[0027] For battery module node i and battery module node j connected in parallel to the same system bus, the mutual admittance is calculated based on the following formula: ,in, , This represents the equivalent internal resistance of the corresponding battery module. Indicates the terminal voltage difference. This is a nonlinear limiting function. During system topology updates (when battery modules are plugged in or unplugged), this limiting function blocks the current transfer path of uncharged modules. During steady-state online operation, its value is always 1. When the system topology is updated and there are uncharged battery modules, the function sets its value to 0 to block the current transfer path between the uncharged battery module and its neighboring nodes. For paths between other online battery modules, the function's value remains 1.

[0028] For the mutual admittance between battery module nodes and power device nodes, and between power device nodes and load interface nodes, the formula is as follows: (The formula is missing from the original text.) ,in, and This is the equivalent input impedance of the corresponding node (for battery modules, it is their equivalent DC internal resistance). This refers to the impedance of the connecting lines between nodes.

[0029] The edge computing gateway calculates the self-admittance based on the mutual admittance between a node and all its neighboring nodes connected by electrical coupling edges. Specifically, the self-admittance of node i is calculated as the sum of the mutual admittances of all electrical coupling edges directly connected to node i. Then, the self-admittance of each node is used as the diagonal element and the mutual admittance between nodes is used as the off-diagonal element to construct a physical admittance matrix. The physical admittance matrix is ​​then row-normalized to obtain the electrical weight matrix. The elements in the electrical weight matrix represent the coupling strength between nodes (including battery modules, power devices and load interfaces) when current is transferred and distributed through the electrical network.

[0030] The thermally coupled edge weights are calculated using the temperature difference, spatial distance, and thermal resistance parameters between nodes connected by the thermally coupled edge. An initial weight matrix is ​​constructed based on all thermally coupled edge weights, and the thermal weight matrix is ​​obtained by row normalization of the initial weight matrix.

[0031] To construct the thermal weight matrix, the thermal coupling weights are first calculated based on the physical distance and thermal resistance coefficient between nodes. The calculation formula for the thermal coupling edge weight between node i and node j is as follows: ,in, For heat conduction area, Let be the spatial distance between the physical devices corresponding to node i and node j. The thermal resistivity of the material is expressed in K·m / W. This is the temperature difference correction factor, with dimensions 1 / K. The value represents the temperature difference between nodes. An initial weight matrix is ​​then constructed, where the element in the i-th row and j-th column represents the weight of the thermal coupling edge between node i and node j. If there is no thermal coupling edge between the two nodes, the corresponding element is set to 0. This initial edge weight matrix is ​​then row-normalized to obtain the thermal weight matrix. The control coupling edge weights are calculated based on the current rotation speed of the heat dissipation system nodes, the temperature deviation between the current temperature and the target temperature of the controlled node, and the temperature sensitivity parameters. A control coupling edge weight matrix is ​​constructed based on all the control coupling edge weights, and the control weight matrix is ​​obtained by row normalization of the control coupling edge weight matrix.

[0032] For constructing the control weight matrix, the edge computing gateway sets the edge weights based on the control logic relationship. Taking the control coupling from the cooling fan node k to the power device node j as an example, the formula for calculating the control coupling edge weights is: ,in, This is the current fan speed. This is the fan's maximum speed. This is the current junction temperature of the power device. This refers to the target operating temperature of the power device. This is the temperature sensitivity parameter. Then, following the same construction method as the thermal weight matrix described above, the control coupling edge weights are constructed into a control weight matrix.

[0033] A neural network with L heterogeneous graph convolutional layers is constructed. For the l-th layer, the hidden state update of node i consists of two parts: a diffusion term and a reaction term. The diffusion term represents the information propagation between nodes through heterogeneous edges. For electrically coupled edges, node i aggregates the hidden states of all its neighboring nodes connected by electrically coupled edges in the (l-1)-th layer. During aggregation, the corresponding edge weights in the electrical weight matrix are used for weighted summation, and a linear transformation is performed through a trainable weight matrix. For thermally coupled edges and control-coupled edges, node i also aggregates the information of its corresponding neighboring nodes, performs weighted summation using the edge weights in the thermal weight matrix and control weight matrix respectively, and performs a linear transformation through their respective trainable weight matrices. Finally, the aggregation results of the above three edge types are added together to obtain the diffusion term of node i in the l-th layer.

[0034] The reaction term represents the local evolution of a node's internal state. For node i, the reaction term is obtained by inputting the node's original state feature vector into a multilayer perceptron, where the nonlinear activation function is either ReLU or LeakyReLU. The final hidden state of node i in layer l is the sum of the diffusion term and the reaction term. The edge computing gateway repeats the above process, obtaining the final hidden state vector of each node after propagation through L layers. The final hidden state vector integrates the node's own local state features and the global system state information obtained through multi-hop neighbor propagation.

[0035] The final hidden state vectors of each node are input into the first, second, and third prediction heads, respectively, to achieve parallel prediction of available power weights, future temperature trajectories, and thermal risk scores. Specifically, the first prediction head is the available power weight prediction head. For battery module node i, the available power weight prediction head inputs the node's final hidden state vector into the fully connected layer. The fully connected layer contains trainable weight vectors and bias scalars, and the output is mapped to the 0-1 range through a Sigmoid activation function, serving as the predicted power weight value. The closer the available power weight value is to 1, the higher the power output capability of the module in the short term after comprehensively considering electrical state, thermal state, heat dissipation capacity, and load characteristics; the closer the value is to 0, the more likely the module should stop power output.

[0036] The second prediction head is the future temperature trajectory prediction head. For all types of nodes, the future temperature trajectory prediction head inputs the node's final hidden state vector into the fully connected layer and outputs a sequence of predicted temperature values ​​for the node over the next K time steps. Temperature trajectory prediction enables edge computing gateways to identify potential heat accumulation risks in advance.

[0037] The third prediction head is a thermal risk scoring head, used for battery module nodes and power device nodes. The thermal risk scoring head inputs the final hidden state vector of the corresponding node into the fully connected layer, and the output is mapped to the 0-1 range through a Sigmoid activation function, serving as the thermal risk score for that node. The closer the risk score is to 1, the higher the probability of thermal runaway occurring in the short term.

[0038] In this embodiment, the training objective of the heterogeneous graph neural network is to minimize the multi-task joint loss function. The multi-task joint loss function includes the available power weight prediction error, temperature trajectory prediction error, thermal runaway risk classification error, and a physical constraint violation penalty term. The available power weight prediction error is the mean square error between the predicted weights and the actual weights of each battery module node. The temperature trajectory prediction error is the mean square error between the predicted temperature and the simulated actual temperature of all nodes at each future time step. The thermal runaway risk classification error is the cross-entropy loss between the risk scores of battery module nodes and power device nodes and the labeled actual risk tags.

[0039] Specifically, the physical constraint violation penalty terms include energy conservation constraints and thermal balance constraints. The energy conservation constraint is the absolute value of the difference between the sum of the predicted output power of all online battery modules and the total system load power requirement. The thermal balance constraint is the difference between the predicted temperature of any node and the temperature limit that its corresponding physical device material can withstand. Network parameters are updated via backpropagation using the gradient descent algorithm. The electrical weight matrix, thermal weight matrix, and control weight matrix are calculated from the state data and do not participate in gradient backpropagation. Only the multilayer perceptron parameters in the reaction term, the trainable weight matrices for each edge type, and the fully connected layer parameters of the three prediction heads participate in gradient updates.

[0040] The state feature vector and weight matrix are input into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score.

[0041] In this embodiment, a power allocation instruction is generated based on the predicted power weight value, the predicted temperature sequence, and the thermal risk score, including: Under discharge conditions, if the power weight prediction value, thermal risk score, and maximum predicted temperature value extracted from the temperature prediction sequence of the online battery module all meet the corresponding safety threshold, the battery module is marked as connected and the target output power limit is issued according to the power weight prediction value. Otherwise, the output power of the battery module is reduced to zero by a preset step size and then disconnected.

[0042] Under the discharge condition of the energy storage system, if the predicted power weight of a module is higher than or equal to the preset access weight threshold, the thermal risk score is lower than the preset risk threshold, and the maximum value of the future temperature trajectory is lower than the critical temperature threshold, then the battery module is marked as connected. Simultaneously, the predicted power weights of all online battery modules are normalized, and the target output power limit for each module is calculated proportionally and sent to the BMS unit. The BMS unit adjusts the duty cycle of the DC-DC converter to ensure that the actual output power of the battery module actively follows the target output power limit.

[0043] If the predicted power weight of a module is lower than the preset exit weight threshold, or the thermal risk score is higher than or equal to the preset risk threshold, or the maximum value of the future temperature trajectory is higher than or equal to the critical temperature threshold, then the module is marked as exiting the state. In each control cycle, the target output power of the battery module is gradually reduced, with the reduction step size being the preset power gradient step size, so that the actual output current of the battery module decreases smoothly and avoids load shock caused by sudden disconnection.

[0044] Power allocation commands and speed control commands are generated based on power weight prediction values, temperature prediction sequences, and thermal risk scores. Power allocation commands are used to control the access, disconnection, or output power of battery modules, while speed control commands are used to regulate the heat dissipation power of the heat dissipation system in the physical device.

[0045] In this embodiment, during the charging process, the maximum difference in the current state of charge of the online battery modules is calculated. When the maximum difference is greater than the preset equalization threshold, a current-limiting charging command is issued to the N battery modules with the highest state of charge, and a full-load charging command is issued to the remaining battery modules. When the maximum difference is less than or equal to the preset equalization threshold, a full-load charging command is issued to all battery modules.

[0046] During charging, the maximum difference in the current state of charge (SOC) of the battery modules is calculated. When the maximum difference exceeds a preset balancing threshold, the main relays of all battery modules remain closed and connected to the charging circuit. However, a current-limiting charging command is issued to high SOC battery modules to limit their charging current limit, while a full-load charging command is issued to low SOC modules, allowing them to accept charging at their maximum safe allowable current. When the maximum difference is less than or equal to the preset balancing threshold, a full-load charging command is issued to all modules, and all modules are connected in parallel to the charging circuit, jointly receiving the charging current output by the charger. The method for determining the upper limit of the charging current in the current-limiting charging command is as follows: first, the reference limit current is obtained by looking up a table based on the SOC range of the high SOC current module; then, the upper limit of the charging current is calculated based on the following formula: Upper limit of charging current = Reference limit current * Power weight prediction value * (1 - Thermal risk score). This formula ensures that when allocating charging limits, higher charging limits are prioritized for modules with higher power weight prediction values ​​and lower thermal risk scores, avoiding including modules with thermal risks in the high-power charging circuit.

[0047] This embodiment also generates a speed control command for the cooling system based on the predicted future temperature trajectories of power device nodes and battery module nodes, and adjusts the cooling power of the cooling system based on the speed control command. Specifically, for each cooling fan node, the predicted temperature sequence of the power device node and battery module node cooled by that fan is extracted, and the maximum predicted temperature value is taken. The target speed is calculated based on the following formula: Target speed = Minimum speed of cooling fan + (Maximum speed - Minimum speed) * (Maximum predicted temperature - Ambient temperature) / (Critical temperature threshold - Ambient temperature). In particular, if the ratio of (Maximum predicted temperature - Ambient temperature) / (Critical temperature threshold - Ambient temperature) is less than 0, it is truncated to 0; if it is greater than 1, it is truncated to 1. The target speed is issued as a speed control command, so that the cooling capacity is improved before the temperature of the power device and battery module actually rises to the dangerous area, avoiding heat accumulation caused by passive response. Based on the above formula, the target speed can be increased when the ambient temperature rises. At the same time, the maximum value of the predicted temperature sequence is used, and the multi-step prediction capability of the graph neural network is used to offset the thermal conduction inertia, realizing feedforward active cooling before heat actually accumulates.

[0048] In this embodiment, to address the graph topology abrupt change caused by the instantaneous connection or disconnection of the battery module, a state-damped transitional stabilization strategy is also implemented. Specifically, when the battery module is connected to or disconnected from the system bus, the energy storage module is set to enter a topology transitional state with a continuously preset transition period. The set of nodes affected by the battery module node changes is identified, and the hidden state update formulas of the nodes in the node set are adjusted.

[0049] Once it is confirmed that the main relay of any battery module has completed its closing or opening action, the system is marked to enter a topology transition state for a preset number of transition cycles. The set of nodes affected by the topology change is identified. For a battery module exit scenario, the affected node set includes all first-order neighbor nodes of the exited battery module that were connected via electrical coupling edges before exiting. For a new module access scenario, the affected node set includes the newly accessed module node itself.

[0050] During the topological transition state, the electrical weight matrix, thermal weight matrix, and control weight matrix are not modified to maintain the stability of graph convolution propagation. The hidden state update formula affecting nodes in the heterogeneous graph convolutional layer is adjusted. Specifically, for the hidden state update of node i in layer l, the original update formula is modified as follows: ,in, The confidence modulation factor is calculated using the following formula: ,in, The initial confidence level modulation coefficients range from 0.3 to 0.6. The decay time constant is half the preset number of transition periods, and k is the current transition period number. The confidence modulation factor decays exponentially with the increase of the transition period, causing the diffusion term weight to gradually return to its normal value of 1, and the reaction term weight to gradually return to its normal value of 1. The significance of this mechanism is that when a topology change occurs, the diffusion information transmitted by neighboring nodes may contain distorted features learned from the old topology. At this time, the dependence on the diffusion term should be reduced, and more reliance should be placed on the local state information provided by the reaction term through the real-time physical sensor measurements of the node itself.

[0051] For newly connected battery modules, the state feature vector is directly used as the initial hidden state vector of that node at layer 0. During the topology transition state, all other battery module connection or disconnection operations in the system are stopped to avoid cascading topology changes before the transition state is stable.

[0052] At the end of each transition cycle, the change in the predicted power weight of all battery modules compared to the previous cycle is checked. When the change in the available power weight of all modules is less than a preset convergence threshold, such as 0.03 to 0.05, the system is determined to have reached a steady state under the new topology, the transition state is terminated early, and normal graph convolutional inference mode is restored. If the early exit condition is not met after completing a preset K transition cycles, the transition state is forcibly exited at the end of the Kth cycle, and the confidence modulation factor is set to zero.

[0053] In this embodiment, when the maximum predicted temperature value or thermal risk score of a battery module exceeds the corresponding safety threshold, it is marked as a pending exit state and its power weight prediction value is reduced. The total output power of the remaining battery modules in the energy storage system and the total load power demand of the system are calculated. When the total output power is greater than or equal to the total load power demand, a tiered exit command is issued to the battery modules in the pending exit state, and a power increase command is issued to the remaining online modules. When the total output power is less than the total load power demand, a power reduction command is issued to the system inverter, and a disconnect command is issued to the battery modules in the pending exit state.

[0054] When the maximum predicted temperature value or thermal risk score of the battery module exceeds the corresponding safety threshold, it means that the battery module is about to enter a dangerous state of overheating under the current operating conditions. It needs to be marked as a state to be exited in advance, and its power weight prediction value should be reduced simultaneously to limit the subsequent power allocation. This will achieve a smooth gradient decrease in the actual output current of the module and avoid sudden changes in bus voltage and electrical shocks caused by instantaneous relay disconnection.

[0055] While marking the battery module as awaiting exit, extract the predicted power weights of all battery modules in the current system, excluding the marked module. Obtain the total load power demand of the current system, calculate the sum of the predicted power weights of the remaining online modules, multiply this sum by the system's rated output power, and obtain the total output power of the remaining modules.

[0056] When the total output power is greater than or equal to the total load power demand, a tiered exit command is issued to the module to be exited. The BMS unit gradually reduces the output power of the module according to the power tier step size until the output power drops to zero, at which point the main relay is disconnected. A power increase command is issued to the module with the higher power weight prediction value among the remaining online modules, so that the remaining modules can smoothly take over the power share of the module to be exited, maintaining the stability of the output power on the load side.

[0057] When the total output power is less than the total load power demand, the difference between the total load power demand and the total output power is calculated. This difference represents the power shortfall in the system after the module to be decommissioned is removed. A power reduction command is sent to the inverter, which includes a target output power value. The target output power value is equal to the total output power of the remaining modules minus a preset power margin. The power margin is used to reserve thermal management margin for the remaining modules, preventing them from rapidly triggering over-temperature protection due to excessive power load. After receiving the power reduction command, the inverter reduces its output power to the target output power value. The inverter completes the power reduction, and the BMS unit disconnects the main relay, disconnecting the corresponding battery module to be decommissioned from the system bus.

[0058] As attached Figure 2 As shown, compared to existing technologies where the direct parallel connection of modules causes low SOC modules to reach the discharge cutoff threshold first, forcing the system to shut down prematurely and leaving a large amount of usable power, this application generates power distribution instructions through heterogeneous graph prediction. In the early stage of discharge, the output power of the SOC module is dynamically increased and the power of the low SOC module is decreased, so that the SOC curves of each module smoothly converge and synchronously drop to the cutoff threshold, allowing the power to be fully released and significantly extending the effective load-carrying time of the system.

[0059] like Figure 3 As shown, the present invention also provides a modular energy storage BMS collaborative management system, which is used to implement the above-described method. The system includes: The graph construction module acquires the state data of the battery modules and various physical devices in the energy storage system after a new battery module is inserted into the energy storage system. The physical devices are treated as nodes in the heterogeneous graph network, and edges connecting the nodes are constructed based on the physical coupling relationships between the physical devices. The edges include electrical coupling edges, thermal coupling edges, and control coupling edges.

[0060] The weight generation module extracts the state feature vectors corresponding to each node based on the state data, calculates the edge weights of each edge based on the state data between nodes, and constructs the electrical weight matrix, thermal weight matrix and control weight matrix based on the edge weights.

[0061] The heterogeneous graph prediction module inputs the state feature vector and weight matrix into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score.

[0062] The collaborative control module generates power allocation instructions and speed control instructions based on power weight prediction values, temperature prediction sequences, and thermal risk scores. It controls the access, disconnection, or output power of the battery module based on the power allocation instructions, and regulates the heat dissipation power of the heat dissipation system in the physical device based on the speed control instructions.

[0063] It should be noted that the various preset thresholds, algorithm parameters, and specific normalization methods involved in this specification and claims are all determined by those skilled in the art through conventional experiments or limited debugging, based on the specific energy storage system hardware configuration, actual application scenario requirements, historical experimental data, and system computing power. Those skilled in the art can reasonably select and set the above parameters and normalization methods based on actual operating conditions.

[0064] Furthermore, the calculation formulas shown in this specification and accompanying drawings are intended to clearly express the core logical calculation relationships. In practical engineering applications and algorithm code implementation, those skilled in the art will take conventional numerical stability protection measures to avoid numerical anomalies such as division by zero that may occur in the formulas. To facilitate the simplified presentation of the formulas and the clear understanding of the core features, this specification does not specifically describe the above-mentioned division by zero protection process when describing the formulas; such numerical processing details are conventional technical means in this field.

Claims

1. A modular energy storage BMS collaborative management method, characterized in that, include: After the new battery module is inserted into the energy storage system, the status data of the battery module and each physical device in the energy storage system are obtained. The physical devices are treated as nodes in the heterogeneous graph network, and the edges connecting the nodes are constructed based on the physical coupling relationship between the physical devices. The edges include electrical coupling edges, thermal coupling edges and control coupling edges. State feature vectors corresponding to each node are extracted based on state data, edge weights of each edge are calculated based on state data between nodes, and electrical weight matrix, thermal weight matrix and control weight matrix are constructed based on edge weights respectively. The state feature vector and weight matrix are input into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score. Power allocation commands and speed control commands are generated based on power weight prediction values, temperature prediction sequences, and thermal risk scores. Power allocation commands are used to control the access, disconnection, or output power of battery modules, while speed control commands are used to regulate the heat dissipation power of the heat dissipation system in the physical device.

2. The method according to claim 1, characterized in that, After the new battery module is inserted into the energy storage system, it also includes: Calculate the difference between the current voltage of the new battery module and the current voltage of the energy storage system bus. Define the absolute value of this difference as the voltage difference. If the voltage difference is greater than the preset voltage threshold, close the pre-charge relay to allow the new battery module to form a pre-charge circuit with the energy storage system bus through the current limiting resistor. In the pre-charge circuit, the current is limited by the current limiting resistor and the voltage difference is reduced. When the voltage difference drops below the preset safety threshold, open the pre-charge relay and close the main relay to allow the new battery module to be connected to the system bus.

3. The method according to claim 1, characterized in that, Edges connecting nodes are constructed based on the physical coupling relationships between physical devices, including: The physical equipment includes battery modules, power devices, heat dissipation systems, and load interfaces, and electrical coupling edges are established between battery module nodes, between battery module nodes and power device nodes, and between power device nodes and load interface nodes. Thermal coupling edges are established between battery module nodes whose physical spatial distance is less than a preset thermal coupling distance threshold, between battery module nodes and adjacent power device nodes, between power device nodes and heat dissipation system nodes responsible for cooling the power device, and between heat dissipation system nodes and environmental sensor nodes. Establish control coupling edges between the heat dissipation system node and the power device node cooled by the heat dissipation system, and between the heat dissipation system node and the battery module node cooled by the heat dissipation system.

4. The method according to claim 3, characterized in that, Based on the edge weights, electrical weight matrices, thermal weight matrices, and control weight matrices are constructed respectively, including: The self-admittance is calculated based on the mutual admittance between a node and all its neighboring nodes connected by electrical coupling edges. The physical admittance matrix is ​​constructed based on the self-admittance of the node and the mutual admittance between the nodes. The physical admittance matrix is ​​then normalized to obtain the electrical weight matrix. The thermal coupling edge weights are calculated by considering the temperature difference, spatial distance, and thermal resistance parameters between nodes connected by the thermal coupling edge. An initial weight matrix is ​​constructed based on all thermal coupling edge weights, and the thermal weight matrix is ​​obtained by row normalization of the initial weight matrix. The control coupling edge weights are calculated based on the current rotation speed of the heat dissipation system nodes, the temperature deviation between the current temperature and the target temperature of the controlled node, and the temperature sensitivity parameters. A control coupling edge weight matrix is ​​constructed based on all the control coupling edge weights, and the control weight matrix is ​​obtained by row normalization of the control coupling edge weight matrix.

5. The method according to claim 1, characterized in that, Power allocation instructions are generated based on power weight predictions, temperature prediction sequences, and thermal risk scores, including: Under discharge conditions, if the power weight prediction value, thermal risk score, and maximum predicted temperature value extracted from the temperature prediction sequence of the online battery module all meet the corresponding safety threshold, the battery module is marked as connected and the target output power limit is issued according to the power weight prediction value; otherwise, the output power of the battery module is reduced to zero by a preset step size and then disconnected. During charging, the maximum difference in the current state of charge of the online battery modules is calculated. When the maximum difference is greater than the preset equalization threshold, a current-limiting charging command is issued to the N battery modules with the highest state of charge, and a full-load charging command is issued to the remaining battery modules. When the maximum difference is less than or equal to the preset equalization threshold, a full-load charging command is issued to all battery modules.

6. The method according to claim 1, characterized in that, When the battery module is connected to or disconnected from the system bus, the energy storage module is set to enter a topology transition state with a continuous preset transition period. The set of nodes affected by the changes in the battery module nodes is identified, and the hidden state update formula of the nodes in the set of nodes is adjusted.

7. The method according to claim 5, characterized in that, When the maximum predicted temperature or thermal risk score of a battery module exceeds the corresponding safety threshold, it is marked as a pending exit state and its power weight prediction value is reduced. The total output power of the remaining battery modules in the energy storage system and the total load power demand of the system are calculated. When the total output power is greater than or equal to the total load power demand, a tiered exit command is issued to the battery modules in the pending exit state, and a power increase command is issued to the remaining online modules. When the total output power is less than the total load power demand, a power reduction command is issued to the system inverter, and a disconnect command is issued to the battery modules in the pending exit state.

8. A modular energy storage BMS collaborative management system, used to implement the method as described in any one of claims 1-7, characterized in that, The system includes: The graph construction module acquires the state data of the battery module and various physical devices in the energy storage system after the new battery module is inserted into the energy storage system. The physical devices are treated as nodes in the heterogeneous graph network, and the edges connecting the nodes are constructed based on the physical coupling relationship between the physical devices. The edges include electrical coupling edges, thermal coupling edges, and control coupling edges. The weight generation module extracts the state feature vectors corresponding to each node based on the state data, calculates the edge weights of each edge based on the state data between nodes, and constructs the electrical weight matrix, thermal weight matrix and control weight matrix based on the edge weights respectively. The heterogeneous graph prediction module inputs the state feature vector and weight matrix into a neural network containing multiple heterogeneous graph convolutional layers for iterative propagation to obtain the final hidden state vector of each node. The final hidden state vector of each node is then input into multiple prediction heads to output the power weight prediction value of each battery module, the temperature prediction sequence of each physical device, and the thermal risk score. The collaborative control module generates power allocation instructions and speed control instructions based on power weight prediction values, temperature prediction sequences, and thermal risk scores. It controls the access, disconnection, or output power of the battery module based on the power allocation instructions, and regulates the heat dissipation power of the heat dissipation system in the physical device based on the speed control instructions.

Images