Edge cloud collaborative battery equalization method and system based on dynamic model migration

Abstract

The invention provides an edge cloud collaborative battery equalization system and method based on dynamic model migration. An edge end monitors a battery state in real time and collects battery parameters; when an abnormal event is detected, extracting feature vectors of battery parameters; the edge end uploads the extracted feature vector to a cloud end, a reinforcement learning algorithm model of the cloud end reads the feature vector, and an optimal phase shift angle delta phi, duty ratio D and resonant capacitance Cr microinstruction are output; the reinforcement learning algorithm model refers to a depth deterministic policy gradient (DDPG) algorithm model; according to a phase shift angle delta phi, a duty ratio D and a resonant capacitor Cr micro instruction generated by a cloud end, an edge end updates a control parameter of the phase shift full bridge converter, a capacitor array switch combination is synchronously switched, and a soft switching time sequence method is used to realize battery equalization. According to the invention, energy can be transferred from the battery module with high voltage to the battery module with low voltage, and cross-battery equalization is realized.

Description

Technical Field

[0001] This invention relates to the field of battery management, and in particular to an edge-cloud collaborative battery balancing system and method based on dynamic model migration. Background Technology

[0002] With the rapid development of electric vehicles and energy storage systems, battery balancing technology, as a key technology to ensure battery pack safety and extend its service life, has received widespread attention. Battery balancing refers to using specific circuit topologies and control strategies to make the parameters such as charge and temperature of each individual cell in a battery pack more consistent, thereby avoiding capacity loss and safety hazards caused by inconsistencies.

[0003] Currently, battery balancing technologies are mainly divided into two categories: passive balancing and active balancing. Passive balancing achieves balancing by consuming excess power through resistors, resulting in low energy utilization. Active balancing, on the other hand, achieves efficient balancing through power transfer. Among these, the phase-shifted full-bridge topology has become one of the mainstream active balancing solutions due to its advantages such as high efficiency and low noise.

[0004] In terms of edge-cloud collaboration technology, CN112101767B proposed a method and system for edge-cloud fusion diagnosis of equipment operating status. This method uses the initial diagnosis model in the cloud for adjustment and processing to obtain the cloud diagnosis model of the target equipment. The cloud diagnosis model is then adjusted at the edge to construct an edge diagnosis model that is more suitable for the target equipment, thus realizing the efficient and accurate reuse of the cloud model [2]. CN112101532B discloses an adaptive multi-model driven equipment fault diagnosis method based on edge-cloud collaboration. Through the time tolerance factor on-demand multi-model branch selection method, the diagnosis model is divided between the edge and the cloud at the layer level, which effectively reduces the latency and realizes cross-operating condition diagnosis.

[0005] In terms of lightweight network applications, CN118965083A introduces a method, system and device for edge fault diagnosis based on lightweight network. This method realizes real-time troubleshooting of mechanical equipment faults by constructing a fault diagnosis model and training it with personalized working condition training samples, which meets the needs of scenarios with high real-time requirements. CN119124624A proposes a cloud-edge collaborative bearing fault detection method. The edge generates preliminary fault detection results through CNN and GRU models, and the cloud generates the final detection results through a local binary residual network model, which effectively solves the problem of difficulty in judging bearing fault types [5].

[0006] With the development of artificial intelligence technology, AI-based battery management systems are gradually emerging. For example, patent document CN118381162B discloses a cloud-edge collaborative energy storage system balancing management method and system. This method acquires operational data before and after balancing the energy storage system, evaluates the balancing effect, obtains balancing optimization parameters, and then controls the balancing on and off of the edge-end energy storage system, effectively reducing cell consistency differences and circulating currents between battery clusters. Another example is patent document CN119764626B, which discloses a cloud-edge combined battery management method and energy storage system. This method utilizes target temperature management strategies and target charge / discharge management strategies to manage the target energy storage device, thereby improving the operational stability of the target energy storage device and achieving effective management of the target energy storage device.

[0007] However, existing technologies still have the following shortcomings: First, the resonant capacitor in traditional phase-shifted full-bridge equalization schemes is fixed, which cannot adapt to the dynamic changes in internal resistance during battery aging. As the battery ages, the internal resistance gradually increases, and the fixed resonant parameters will lead to a reduction in the zero-voltage switching (ZVS) range, increased switching losses, decreased equalization efficiency, and may even cause safety issues such as overheating of switching devices.

[0008] Secondly, existing AI-driven battery management solutions typically require continuous uploading of all data to the cloud for analysis and processing. This not only consumes significant communication bandwidth but also increases system latency, making it difficult to meet the high real-time requirements of battery balancing systems. Especially in environments with poor communication conditions, data transmission delays can lead to delayed balancing decisions and an inability to respond promptly to changes in battery status.

[0009] Finally, with the promotion of battery reuse, the same battery pack may contain batteries with different aging levels, and the difference in internal resistance may exceed 30%. In this case, the efficiency of the fixed topology balancing circuit will decrease significantly, making it difficult to meet the balancing requirements of reused battery packs, thus affecting the overall performance and lifespan of the battery pack.

[0010] Therefore, there is an urgent need for a new type of battery balancing system that can adapt to dynamic changes in battery internal resistance, reduce data transmission pressure, and is suitable for secondary battery packs, so as to improve balancing efficiency and extend battery life. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides an inter-module balancing system and method for energy storage batteries, which addresses the technical problems that traditional balancing circuits cannot solve, such as dynamic changes in battery internal resistance, communication delays, and differences in battery aging. The system achieves balancing between high- and low-voltage batteries by using an array selection switch, transferring energy from high-voltage battery modules to low-voltage battery modules, thus realizing cross-battery balancing.

[0012] The present invention adopts the following technical solution.

[0013] This invention proposes an edge-cloud collaborative battery equalization method based on dynamic model migration, comprising: S1: Real-time monitoring of battery status and collection of battery parameters at the edge; extraction of feature vectors of battery parameters when an abnormal event is detected. S2, the edge device uploads the extracted feature vector to the cloud, and the reinforcement learning algorithm model in the cloud reads the feature vector and outputs the optimal phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr micro-instructions; the reinforcement learning algorithm model refers to the Deep Deterministic Policy Gradient (DDPG) algorithm model. S3 updates the control parameters of the phase-shifting full-bridge converter at the edge based on the micro-instructions of phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr generated in the cloud, and synchronously switches the capacitor array switch combination. The battery balancing is achieved by using a soft switching timing method. The phase-shifting full-bridge converter is isolated by a transformer and connected to the battery pack through a rectifier and filter circuit; the variable resonant capacitor array is connected to the output side of the transformer to form a resonant circuit.

[0014] More preferably, in S1, the edge end continuously monitors the operating status of the battery pack through a lightweight hybrid diagnostic model, and collects battery parameters including voltage, current, temperature, and internal resistance.

[0015] More preferably, in S1, the types of abnormal events include: the absolute value of the SOC change rate is greater than the SOC set threshold; the temperature gradient between monomers is greater than the gradient set threshold; and the internal resistance mutation rate is greater than the mutation set threshold. When any abnormal event occurs, the feature vector of the battery parameters is extracted; key features are immediately extracted from the collected battery parameters using a combination of principal component analysis and wavelet transform to form a feature vector; the dimension of the feature vector is not greater than a set dimension threshold.

[0016] More preferably, in S2, the formula for calculating the value of the resonant capacitor Cr is as follows: Where k is the dynamic scaling factor, For switching frequency, This is the predicted value of the battery's internal resistance.

[0017] More preferably, in S3, the edge end updates the control parameters of the phase-shifting full-bridge converter according to the phase shift angle ΔΦ and the duty cycle D; according to the calculated resonant capacitor Cr value, the edge end controls the MOS transistor matrix to switch the switching combination of the capacitor array.

[0018] More preferably, in S3, the process of performing the soft switching timing includes: pausing the operation of the phase-shifted full-bridge converter; setting up a pre-charging circuit to pre-charge the newly connected capacitor to a set voltage level; updating the switching state of the capacitor array; and restarting the phase-shifted full-bridge converter.

[0019] More preferably, in the collaborative work between the cloud and the edge, the lightweight hybrid diagnostic model at the edge synchronizes its weight parameters with the deep deterministic policy gradient (DDPG) algorithm model in the cloud every 24 hours.

[0020] This invention also proposes an edge-cloud collaborative battery equalization system based on dynamic model transfer using the above method, including a cloud decision module, an edge execution module, a model fragmentation transfer channel, and a variable resonant capacitor array: The edge execution module includes a lightweight hybrid diagnostic model and a phase-shifting full-bridge converter. The lightweight hybrid diagnostic model combines a random forest algorithm and a Transformer network architecture to read battery operating status data, monitor battery status in real time, and output abnormal events. The phase-shifting full-bridge converter reads the output of the subsequent cloud decision-making module. The model fragmentation migration channel is activated when the edge execution module detects an abnormal event in the battery status. It extracts the feature vectors of key features from the battery operating status data collected by the edge execution module and uploads the feature vectors to the cloud decision module. The cloud-based decision-making module deploys a reinforcement learning algorithm model, reads feature vectors, generates battery lifecycle optimization instructions, and transmits them to the phase-shifting full-bridge converter in the edge execution module; the reinforcement learning algorithm model refers to the Deep Deterministic Policy Gradient (DDPG) algorithm model. The variable resonant capacitor array consists of a capacitor network controlled by a MOS transistor matrix, and reconstructs the resonant parameters according to the battery life cycle optimization instructions issued by the cloud decision module; the variable resonant capacitor array also includes a pre-charging circuit and a soft switching timing controller. By combining the phase-shifted full-bridge converter and the variable resonant capacitor array, and using a soft-switching timing method, battery equalization is achieved.

[0021] More preferably, the edge execution module also includes a phase-shifted full-bridge converter, and the control parameters of the phase-shifted full-bridge converter are adjusted according to the battery life cycle optimization instruction; The phase-shifting full-bridge converter is isolated by a transformer and connected to the battery pack through a rectifier and filter circuit; the variable resonant capacitor array is connected to the output side of the transformer to form a resonant circuit. The types of abnormal events include: the absolute value of the SOC change rate is greater than the SOC set threshold; the temperature gradient between monomers is greater than the gradient set threshold; and the internal resistance mutation rate is greater than the mutation set threshold.

[0022] More preferably, the total capacitance value of the variable resonant capacitor array is calculated as follows: Cr = C0 + ∑(i = 1 to n)(Si·Ci), where Si is the switching state; C0 is the basic capacitance value; and Ci is the capacitance value of the i-th switchable capacitor unit.

[0023] The embodiments of the present invention, by adopting the above technical solutions, have the following beneficial effects: By dynamically reconstructing the resonant capacitor, it can adapt to changes in internal resistance caused by battery aging, reduce switching losses, and expand the ZVS range. A fragmented model migration channel is adopted, uploading compressed feature vectors only when abnormal events occur, reducing communication bandwidth pressure. The feature vector dimension does not exceed 8, and the data volume does not exceed 512 bytes. To address the issue of large differences in internal resistance in secondary battery packs, a variable resonant capacitor array dynamically adjusts the resonant parameters, improving equalization efficiency. The lightweight hybrid diagnostic model has a memory footprint of no more than 50KB and a sampling rate of no less than 10kHz, achieving efficient operation at the edge. The soft switching timing delay does not exceed 5μs, ensuring a smooth capacitor switching process and preventing voltage surges. The cloud uses the DDPG algorithm to optimize long-term lifespan targets, and the edge model is periodically synchronized with the cloud, achieving edge-cloud collaborative full lifecycle management of the battery. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for edge-cloud collaborative battery balancing based on dynamic model migration.

[0025] Figure 2 This is a schematic diagram showing the connection between the phase-shifting full-bridge converter, the variable resonant capacitor array, and the battery pack. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0027] The present invention adopts the following scheme.

[0028] This invention proposes an edge-cloud collaborative battery equalization method based on dynamic model migration, as follows: S1: Real-time monitoring of battery status and collection of battery parameters at the edge; extraction of feature vectors of battery parameters when an abnormal event is detected. In S1, the edge continuously monitors the operating status of the battery pack through a lightweight hybrid diagnostic model, collecting battery parameters including voltage, current, temperature, and internal resistance.

[0029] In S1, the types of abnormal events include: the absolute value of the SOC change rate is greater than the SOC set threshold; the temperature gradient between monomers is greater than the gradient set threshold; and the internal resistance mutation rate is greater than the mutation set threshold. When any abnormal event occurs, the feature vector of the battery parameters is extracted; key features are immediately extracted from the collected battery parameters using a combination of principal component analysis and wavelet transform to form a feature vector; the dimension of the feature vector is not greater than a set dimension threshold.

[0030] S2, the edge device uploads the extracted feature vector to the cloud, and the reinforcement learning algorithm model in the cloud reads the feature vector and outputs the optimal phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr micro-instructions; the reinforcement learning algorithm model refers to the Deep Deterministic Policy Gradient (DDPG) algorithm model. In S2, the formula for calculating the resonant capacitance Cr is: Where k is the dynamic scaling factor, For switching frequency, This is the predicted value of the battery's internal resistance.

[0031] S3 updates the control parameters of the phase-shifting full-bridge converter at the edge based on the micro-instructions of phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr generated in the cloud, and synchronously switches the capacitor array switch combination. The battery balancing is achieved by using a soft switching timing method. The phase-shifting full-bridge converter is isolated by a transformer and connected to the battery pack through a rectifier and filter circuit; the variable resonant capacitor array is connected to the output side of the transformer to form a resonant circuit.

[0032] In S3, the edge terminal updates the control parameters of the phase-shifting full-bridge converter according to the phase shift angle ΔΦ and the duty cycle D; based on the calculated resonant capacitor Cr value, the edge terminal controls the MOS transistor matrix to switch the switching combination of the capacitor array.

[0033] In S3, the process of executing the soft switching timing includes: pausing the operation of the phase-shifted full-bridge converter; setting up the pre-charge circuit to pre-charge the newly connected capacitor to the set voltage level; updating the switching state of the capacitor array; and restarting the phase-shifted full-bridge converter.

[0034] In the collaborative work between the cloud and the edge, the lightweight hybrid diagnostic model in the edge synchronizes its weight parameters with the deep deterministic policy gradient (DDPG) algorithm model in the cloud every 24 hours.

[0035] This invention also proposes an edge-cloud collaborative battery equalization system based on dynamic model transfer using the above method, comprising a cloud decision module, an edge execution module, a model fragmentation transfer channel, and a variable resonant capacitor array: The edge execution module includes a lightweight hybrid diagnostic model and a phase-shifting full-bridge converter. The lightweight hybrid diagnostic model combines a random forest algorithm and a Transformer network architecture to read battery operating status data, monitor battery status in real time, and output abnormal events. The phase-shifting full-bridge converter reads the output of the subsequent cloud decision-making module. The model fragmentation migration channel is activated when the edge execution module detects an abnormal event in the battery status. It extracts the feature vectors of key features from the battery operating status data collected by the edge execution module and uploads the feature vectors to the cloud decision module. The cloud-based decision-making module deploys a reinforcement learning algorithm model, reads feature vectors, generates battery lifecycle optimization instructions, and transmits them to the phase-shifting full-bridge converter in the edge execution module; the reinforcement learning algorithm model refers to the Deep Deterministic Policy Gradient (DDPG) algorithm model. The variable resonant capacitor array consists of a capacitor network controlled by a MOS transistor matrix, and reconstructs the resonant parameters according to the battery life cycle optimization instructions issued by the cloud decision module; the variable resonant capacitor array also includes a pre-charging circuit and a soft switching timing controller. By combining the phase-shifted full-bridge converter and the variable resonant capacitor array, and using a soft-switching timing method, battery equalization is achieved.

[0036] The edge execution module also includes a phase-shifted full-bridge converter, which adjusts the control parameters of the phase-shifted full-bridge converter according to the battery life cycle optimization instructions; The phase-shifting full-bridge converter is isolated by a transformer and connected to the battery pack through a rectifier and filter circuit; the variable resonant capacitor array is connected to the output side of the transformer to form a resonant circuit. The types of abnormal events include: the absolute value of the SOC change rate is greater than the SOC set threshold; the temperature gradient between monomers is greater than the gradient set threshold; and the internal resistance mutation rate is greater than the mutation set threshold.

[0037] The formula for calculating the total capacitance of the variable resonant capacitor array is: Cr = C0 + ∑(i=1 to n)(Si·Ci), where Si is the switching state; C0 is the basic capacitance value; and Ci is the capacitance value of the i-th switchable capacitor unit.

[0038] Example 1 This invention proposes an edge-cloud collaborative battery equalization system based on dynamic model migration, comprising a cloud decision module, an edge execution module, a model fragmentation migration channel, and a variable resonant capacitor array.

[0039] The cloud-based decision-making module deploys a reinforcement learning strategy engine. This engine continuously optimizes battery management strategies through deep reinforcement learning algorithms, generating optimization instructions for the entire battery lifecycle. This module collects historical battery operating data, constructs a battery aging model based on the reinforcement learning strategy, and dynamically adjusts the balancing strategy parameters according to battery performance under different operating conditions. The cloud-based decision-making module possesses massively parallel computing capabilities, enabling it to process data from multiple battery packs simultaneously and customize personalized balancing schemes for each battery pack.

[0040] The edge execution module comprises a lightweight hybrid diagnostic model and a phase-shifting full-bridge converter. The lightweight hybrid diagnostic model combines a random forest algorithm and a Transformer network architecture to achieve real-time monitoring and anomaly detection of battery status. This hybrid diagnostic model has a memory footprint of no more than 50KB, a sampling rate exceeding 10kHz, and can capture transient changes in battery characteristics. The random forest part handles structured data features, while the Transformer part focuses on temporal feature extraction; their collaborative operation improves diagnostic accuracy. The phase-shifting full-bridge converter is responsible for performing the actual energy conversion and equalization operations, adjusting power flow and equalization rate based on the diagnostic model's output and instructions from the cloud.

[0041] The model fragmentation migration channel is activated when a battery anomaly occurs. It compresses key data collected at the edge into feature vectors and uploads them to the cloud, while also receiving micro-instructions from the cloud. This channel employs a highly efficient compression algorithm, ensuring that the compressed feature vector dimension does not exceed 8 and the data volume does not exceed 512 bytes, significantly reducing communication bandwidth requirements. The triggering conditions for model fragmentation migration include: an absolute value of the SOC change rate greater than 5% / second; an inter-cell temperature gradient greater than 3℃; and an internal resistance mutation rate greater than 20%. When any of these conditions are met, the system immediately initiates the data upload process, ensuring that the cloud can promptly obtain anomaly information and respond.

[0042] The variable resonant capacitor array consists of a capacitor network controlled by a matrix of MOSFETs, dynamically reconstructing the resonant parameters based on the Cr value sent from the cloud. The total capacitance of the array is calculated as: Cr = C0 + ∑(i=1 to n)(Si·Ci), where Si represents the switching state (Si=1 when the MOSFET is on, Si=0 when it is off); C0 is the base capacitance; and Ci is the capacitance of the i-th switchable capacitor unit. By controlling the switching states of different MOSFETs, the system can precisely adjust the resonant capacitance value, thereby optimizing energy transfer efficiency. The variable resonant capacitor array also includes a pre-charging circuit and a soft-switching timing controller to prevent voltage surges during capacitor switching. The pre-charging circuit pre-charges the newly connected capacitor to an appropriate voltage level before switching, while the soft-switching timing controller ensures that the switching operation occurs near the current zero-crossing point, effectively suppressing electromagnetic interference during switching transients.

[0043] A phase-shifted full-bridge converter (PSFB) typically consists of four switching transistors (Q1-Q4) forming a full bridge, isolated by a transformer, and then connected to the battery pack via an output rectifier and filter circuit. A variable resonant capacitor array is usually connected to the secondary side (output side) of the transformer to form a resonant circuit, which helps achieve zero-voltage switching (ZVS) of the switching transistors.

[0044] like Figure 2 As shown, the specific connection relationships are as follows: A phase-shifted full-bridge converter (PSFB) typically consists of four switching transistors (Q1-Q4) forming a full bridge, isolated by a transformer, and then connected to the battery pack via an output rectifier and filter circuit. A variable resonant capacitor array is usually connected to the secondary side (output side) of the transformer to form a resonant circuit, which helps achieve zero-voltage switching (ZVS) of the switching transistors.

[0045] The specific connection relationships are as follows: (1) Phase-shifting full-bridge converter section: The positive terminal of the input DC power supply is connected to the drain (or collector, in the case of a MOSFET) of Q1 and Q3; the negative terminal of the input DC power supply is connected to the source of Q2 and Q4; the source of Q1 is connected to the drain of Q2, forming the midpoint A of the bridge arm; the source of Q3 is connected to the drain of Q4, forming the midpoint B of the bridge arm. Midpoints A and B are respectively connected to the two ends of the primary side of the high-frequency transformer.

[0046] (2) Transformer section: The primary side of the transformer receives an AC square wave voltage from the full bridge.

[0047] The secondary side of the transformer outputs AC voltage.

[0048] (3) Variable resonant capacitor array: One end of the transformer's secondary winding is connected to one end of the variable resonant capacitor array, and the other end (P2) is connected to the other end of the variable resonant capacitor array. Therefore, the resonant capacitor array is internally connected in parallel and then in series with the transformer's secondary winding. To create resonance in the transformer's secondary winding, the resonant capacitor is typically connected across the two ends of the transformer's secondary winding (i.e., before rectification). This creates a resonant circuit with the transformer's leakage inductance (or an external resonant inductor), thus helping the primary side achieve ZVS (Zero-Voltage Switching).

[0049] (4) Output rectification section (taking full-wave rectification as an example): Four rectifier diodes are connected to the two ends of the transformer secondary side to convert high-frequency AC power into DC power.

[0050] (5) Output filter circuit and battery pack: The rectified output is filtered by a filter inductor (Lf) and a filter capacitor (Cf) before being connected to the battery pack. The positive terminal of the battery pack is connected to the positive terminal of the rectified output, and the negative terminal is connected to the other end of the secondary winding of the transformer.

[0051] In a preferred embodiment, the lightweight hybrid diagnostic model has a memory footprint of 32KB, a sampling rate of 15kHz, a random forest component containing 25 decision trees with a depth of 8 layers, and a Transformer component employing a 4-layer attention mechanism with a hidden layer dimension of 64. This configuration achieves an anomaly detection accuracy of over 95% while maintaining low computational complexity.

[0052] In another preferred embodiment, the compressed feature vector has a dimension of 6 and a data volume of 256 bytes. Key features are extracted by combining principal component analysis and wavelet transform, which further reduces communication overhead while ensuring information integrity.

[0053] During system operation, the edge execution module continuously monitors the battery pack status. When an abnormal event is detected, it uploads the compressed feature vector to the cloud via the model fragmentation migration channel. The cloud decision module analyzes the abnormal data, combines it with historical operation records, generates optimization instructions, and sends them to the edge. Based on the received instructions, the edge execution module adjusts the parameters of the variable resonant capacitor array and performs battery equalization operations through a phase-shifting full-bridge converter. The entire process achieves edge-cloud collaboration, ensuring real-time response capabilities at the edge while fully utilizing the powerful computing resources of the cloud, significantly improving battery equalization efficiency and the overall lifespan of the battery pack.

[0054] Example 2 This embodiment provides a battery balancing method based on an edge-cloud collaborative battery balancing system, such as... Figure 1 As shown. This method utilizes the edge-cloud collaborative battery balancing system described in Embodiment 1 to achieve efficient balanced management of the battery pack through the collaborative work of the edge and the cloud.

[0055] The battery equalization method includes the following steps: First, the edge device monitors the battery status in real time and extracts feature vectors when an abnormal event is detected. The edge device continuously monitors the battery pack's operating status, including parameters such as voltage, current, temperature, and internal resistance, using the lightweight hybrid diagnostic model (random forest + Transformer) described in Example 1. When the system detects abnormal events such as an absolute value of the SOC change rate exceeding 5% / second, an inter-cell temperature gradient exceeding 3℃, or an internal resistance mutation rate exceeding 20%, the edge device immediately extracts key features from the collected data using a combination of principal component analysis and wavelet transform to form a feature vector.

[0056] Next, the edge device uploads the extracted feature vectors to the cloud, where it generates micro-instructions for the phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr. Upon receiving the feature vectors, the cloud analyzes them using the Deep Deterministic Policy Gradient (DDPG) algorithm, which focuses on optimizing the battery's long-term lifespan. Based on the current battery state and historical data, the DDPG algorithm calculates the optimal values ​​for the phase shift angle ΔΦ, duty cycle D, and resonant capacitor Cr, and then sends these parameters as micro-instructions to the edge device.

[0057] The formula for calculating the resonant capacitance Cr is as follows: Where k is the dynamic scaling factor, with a value range of 0.8 ≤ k ≤ 1.2. For switching frequency, This is the predicted internal resistance value of the battery. In practical applications, the k value can be dynamically adjusted according to the battery's aging state, temperature conditions, and load characteristics to obtain the optimal resonance effect. For example, when the battery is in a low-temperature environment, the k value may be close to 0.8; while in a high-temperature environment, the k value may be close to 1.2.

[0058] Subsequently, the edge device updates the phase-shifted full-bridge control parameters and synchronously switches the capacitor array switching combination. After receiving the micro-instructions from the cloud, the edge device immediately updates the control parameters of the phase-shifted full-bridge converter, including the phase shift angle ΔΦ and duty cycle D. Simultaneously, based on the calculated resonant capacitor Cr value, the edge device controls the MOS transistor matrix in the variable resonant capacitor array to switch the switching combination of the capacitor array, achieving dynamic reconstruction of the resonant parameters.

[0059] Finally, the soft switching sequence is executed: pause full-bridge operation → pre-charge capacitors → update switch states → restart full-bridge. To avoid voltage surges and electromagnetic interference during capacitor switching, the system employs soft switching technology. The specific steps are: first, pause the operation of the full-bridge converter; then, pre-charge the newly connected capacitors to an appropriate voltage level using the pre-charging circuit; next, update the switch states of the capacitor array; and finally, restart the full-bridge converter. The delay of the entire soft switching sequence is controlled within 5μs to ensure the continuity and stability of the equalization process.

[0060] In the collaborative operation between the cloud and the edge, the edge model synchronizes its weight parameters with the cloud every 24 hours. This periodic synchronization mechanism ensures that the lightweight model at the edge can be updated in a timely manner to reflect the latest battery characteristics and optimization strategies, while avoiding the network burden caused by frequent communication.

[0061] In a preferred embodiment, the DDPG algorithm employs a four-layer neural network structure, including two Actor networks and two Critic networks, with 128 neurons in the hidden layer, a learning rate of 0.001, and a discount factor γ of 0.95. This configuration ensures algorithm convergence while effectively handling control problems in continuous action spaces, such as battery balancing.

[0062] In another preferred embodiment, the soft handover timing delay is controlled within 3μs. By employing a high-speed optocoupler and optimized drive circuitry, the delay time during the handover process is further reduced, thereby improving the system response speed and equalization efficiency.

[0063] Using the above method, the battery balancing system can dynamically adjust the balancing parameters according to the real-time status of the battery, thereby achieving efficient balancing management of the battery pack, extending the service life of the battery pack, and improving energy utilization efficiency.

[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for edge cloud collaborative battery equalization based on dynamic model migration, characterized in that, Comprise: S1, edge end real-time monitoring battery status, collecting battery parameters; When detecting abnormal events, extracting the feature vector of battery parameters; S2, the edge end uploads the extracted feature vector to the cloud, and the cloud's reinforcement learning algorithm model reads the feature vector and outputs the optimal phase shift angle ΔΦ, duty ratio D, and resonant capacitor Cr micro instruction; the reinforcement learning algorithm model refers to the deep deterministic policy gradient DDPG algorithm model; S3, according to the phase shift angle ΔΦ, duty ratio D, and resonant capacitor Cr micro instruction generated by the cloud, the edge end updates the control parameters of the phase shift full-bridge converter, synchronously switches the capacitor array switch combination, and uses the soft switching timing method to realize battery balancing; The phase shift full-bridge converter is connected to the battery pack through a transformer isolation and a rectifier filter circuit; the variable resonant capacitor array is connected to the output side of the transformer, forming a resonant loop.

2. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In S1, the edge end continuously monitors the running state of the battery pack through the lightweight hybrid diagnostic model, and collects the parameters of the battery including voltage, current, temperature, and internal resistance.

3. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In S1, the types of abnormal events include the absolute value of SOC change rate greater than the SOC set threshold, the temperature gradient between single batteries greater than the gradient set threshold, and the internal resistance mutation rate greater than the mutation set threshold; When any of the abnormal events occurs, the feature vector of the battery parameters is extracted; the key features are immediately extracted from the collected battery parameters using the combination of principal component analysis and wavelet transform to form the feature vector; The dimension of the feature vector is not greater than the dimension set threshold.

4. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In S2, the calculation formula of the resonant capacitor Cr value is where k is a dynamic scaling factor, is the switching frequency, is the battery internal resistance prediction value.

5. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In S3, the edge end updates the control parameters of the phase shift full-bridge converter according to the phase shift angle ΔΦ and the duty ratio D; and according to the calculated resonant capacitor Cr value, the edge end controls the MOS tube matrix to switch the switch combination of the capacitor array.

6. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In S3, the process of executing soft switching timing includes pausing the operation of the phase shift full-bridge converter; Setting a pre-charge circuit to pre-charge the newly connected capacitor to a set voltage level; Updating the switch state of the capacitor array; Restarting the phase shift full-bridge converter.

7. The battery balancing method based on dynamic model migration and edge-cloud collaboration according to claim 1, wherein: In the collaborative work of the cloud and the edge, the lightweight hybrid diagnostic model in the edge synchronizes the weight parameters with the deep deterministic policy gradient DDPG algorithm model in the cloud once every 24 hours.

8. A battery balancing system based on dynamic model migration and edge-cloud collaboration, comprising a cloud decision module, an edge execution module, a model fragmentation migration channel, and a variable resonant capacitor array, wherein: The edge execution module includes a lightweight hybrid diagnostic model and a phase-shifted full-bridge converter; the lightweight hybrid diagnostic model combines a random forest algorithm and a Transformer network architecture, reads battery operating state data, monitors battery state in real time, and outputs abnormal events; The phase-shifted full-bridge converter reads the output of the subsequent cloud decision module; When the edge execution module detects an abnormal event in the battery state, the model fragmentation migration channel is activated, extracts the feature vector of the key features in the battery operating state data collected by the edge execution module, and uploads the feature vector to the cloud decision module; The cloud decision module deploys a reinforcement learning algorithm model, reads the feature vector, generates battery life cycle optimization instructions, and transmits them to the phase-shifted full-bridge converter in the edge execution module; the reinforcement learning algorithm model refers to a deep deterministic policy gradient (DDPG) algorithm model; The variable resonant capacitor array is composed of a capacitor network controlled by a MOS tube matrix, and reconstructs the resonance parameters according to the battery life cycle optimization instructions issued by the cloud decision module; the variable resonant capacitor array also includes a pre-charge circuit and a soft switching timing controller; Combined with the phase-shifted full-bridge converter and the variable resonant capacitor array, the method of soft switching timing is used to complete battery equalization.

9. The edge-cloud collaborative battery equalization system based on dynamic model migration according to claim 8, characterized in that: The edge execution module further includes a phase-shifted full-bridge converter that adjusts the control parameters of the phase-shifted full-bridge converter according to the battery life cycle optimization instructions; The phase-shifted full-bridge converter is connected to the battery pack through a transformer isolation and a rectifier filter circuit; the variable resonant capacitor array is connected to the output side of the transformer, forming a resonant circuit; The types of abnormal events include an absolute value of SOC change rate greater than a SOC set threshold, a temperature gradient between single cells greater than a gradient set threshold, and a sudden change rate of internal resistance greater than a sudden change set threshold.

10. The edge-cloud collaborative battery equalization system based on dynamic model migration according to claim 8, characterized in that: The total capacitance value of the variable resonant capacitor array is calculated by the formula: Cr=C0+∑(i=1 to n)(Si· Ci), where Si is the switch state; C0 is the basic capacitance value; and Ci is the capacitance value of the ith switchable capacitor unit.

Images