Dynamic topology charging equalization control method for riding battery changing cabinet

Abstract

The invention discloses a dynamic topology charging equalization control method for a battery changing cabinet, and the method comprises the following steps: S1, collecting multi-source time sequence data of a battery in the battery changing cabinet, and carrying out the preprocessing of the data to generate a data set; s2, constructing a variational auto-encoder model, extracting potential charging characteristics of the battery, and obtaining a health state and charging behavior representation; s3, predicting by using a variational auto-encoder model, and generating battery state prediction data; s4, optimizing a charging topological structure and charging parameters by adopting an improved moth optimization algorithm; s5, the actual charging state of the battery is monitored in real time, and real-time optimization adjustment is carried out; and S6, according to the real-time feedback data, continuously updating the potential space of the variational auto-encoder model. According to the invention, through intelligent prediction and dynamic optimization, topological equalization regulation and control of the battery charging process in the battery changing cabinet is realized, so that the charging efficiency is improved and the service life of the battery is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and in particular to a dynamic topology charging balancing control method for a rider battery swap cabinet. Background Art

[0002] With the rapid development of scenarios such as urban logistics, instant delivery, and shared travel, riders, as important service executors, have put forward higher requirements for the endurance and replacement efficiency of vehicles. Traditional electric bicycles or electric motorcycles mostly rely on manual charging, but under high-frequency delivery tasks, manual charging is inefficient and has long downtime, which can no longer meet the current high-intensity operation needs. In order to improve the efficiency of battery replacement, in recent years, battery swap cabinet systems based on standard battery modules have begun to be widely deployed and have gradually become an important infrastructure for new smart travel and urban logistics. By centrally managing and distributing batteries, battery swap cabinets realize multiple functions such as rapid battery replacement, unified charging, and status monitoring, effectively alleviating the problems of individual riders' difficulty in using electricity, slow charging, and poor endurance.

[0003] However, in actual operation, the large number of batteries in a battery swap cabinet do not operate synchronously, and their states are highly heterogeneous, with complexities such as varying health states, inconsistent charge and discharge cycles, and significant differences in temperature environments. Furthermore, traditional battery swap cabinet systems employ a fixed or preset topology during battery charging, charging batteries sequentially in a set order and lacking the ability to intelligently and dynamically adjust based on the actual battery state. This charging strategy, when faced with a heterogeneous distribution of multiple batteries, can easily lead to overcharging or undercharging some batteries, reducing the overall system's charging efficiency and lifespan utilization, and even posing safety risks.

[0004] Furthermore, while existing technologies have introduced some optimization schemes based on rules or simple predictive models in battery management systems (BMS), such as those prioritizing maximum remaining capacity and timing rotation charging, they generally lack the ability to globally perceive and dynamically adjust the charging topology. In scenarios where riders frequently swap batteries, where battery status changes frequently and health status varies significantly, these static or semi-dynamic control methods struggle to achieve an efficient, balanced, and safe charging process. They suffer from limitations such as delayed response, low resource utilization, and a lack of topology adjustment mechanisms.

[0005] From an algorithmic perspective, in recent years, swarm intelligence optimization algorithms, such as particle swarm optimization, genetic algorithms, and gray wolf optimization algorithms, have gradually been introduced into tasks such as battery scheduling, path optimization, and parameter control. Among them, the moth-flame optimization algorithm (MFO) has gradually attracted the attention of researchers due to its strong global search capabilities and structural flexibility in high-dimensional continuous spaces. However, existing work has mostly used the moth-flame optimization algorithm for general parameter tuning or mathematical function optimization, and has not yet been deeply customized for dynamic structural adjustment problems in complex physical systems. In particular, there is a lack of deep integration mechanisms with data-driven models (such as neural networks and variational models).

[0006] On the other hand, deep learning technologies, particularly variational autoencoders (VAEs), demonstrate excellent generalization and compression capabilities in modeling and predicting multi-source heterogeneous time series data. By probabilistically modeling and abstracting the latent space of raw data, they can effectively extract the underlying health characteristics and charging behavior patterns of batteries during the charging process, providing a foundation for subsequent prediction and optimization. However, existing applications often remain at the level of static prediction or classification judgment, lacking structural linkage with swarm intelligence optimization methods and failing to truly achieve a closed-loop control system of "perception-prediction-decision-making."

[0007] Based on the above analysis, the existing technology has obvious deficiencies in the following aspects: First, there is a lack of real-time state modeling mechanism for the dynamic differences of multiple batteries in the battery swap cabinet, and the potential behavioral correlations between the batteries cannot be fully explored; second, a dynamic optimization strategy tightly coupled with the charging topology structure has not been constructed. Most of the existing methods are static queuing or threshold control, which cannot adapt to complex operating environments; third, there is a lack of effective coordination between swarm intelligence algorithms and data-driven models, and the model prediction and optimization strategies are separated, making it difficult to achieve linkage updates and control responses throughout the entire process; fourth, the system lacks continuous self-learning and online adjustment capabilities, and lacks stability and adaptability when faced with battery state drift, environmental interference or load mutations during long-term operation.

[0008] Therefore, how to provide a dynamic topology charging balancing control method for a rider battery swap cabinet is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] One purpose of the present invention is to propose a dynamic topology charging balancing control method for a rider battery swap cabinet. The present invention combines a variational autoencoder model with an improved moth optimization algorithm. The variational autoencoder is used to model and predict the multi-source time series data of the battery, extract the potential health status and behavioral characteristics of the battery during the charging process, and on this basis, dynamically optimize the charging topology structure and charging parameter configuration through the improved moth optimization algorithm to achieve intelligent adjustment and balanced control of the charging sequence, current, and voltage between batteries. This method has the advantages of structural optimization and adaptation, intelligent control of the charging process, accurate state prediction, and efficient resource allocation. It significantly improves the charging balance, safety, and energy utilization efficiency of the battery swap cabinet in complex operating environments.

[0010] A method for controlling dynamic topology charging balance of a rider battery swap cabinet according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect multi-source time series data of batteries in the rider's battery swap cabinet, pre-process the multi-source time series data, and generate a processed data set;

[0012] S2. Construct a variational autoencoder model. Using the trained dataset, extract the potential features of the battery charging process in the rider’s battery swap cabinet, and obtain the potential representation of the health status and charging behavior of the battery in the rider’s battery swap cabinet.

[0013] S3. Use the variational autoencoder model to predict the future charging state of the battery in the rider's battery swap cabinet and generate battery state prediction data;

[0014] S4. Based on the generated battery status prediction data, the health status of the batteries in the rider battery swap cabinet, and the potential representation of the charging behavior, an improved moth optimization algorithm is used to optimize the charging topology between the batteries in the rider battery swap cabinet, and the charging sequence, charging voltage, and current of the batteries in the rider battery swap cabinet are dynamically adjusted;

[0015] S5. During the charging process, the actual charging status of the battery in the rider's battery swap cabinet is monitored in real time, and real-time optimization and adjustment are performed;

[0016] S6. Based on the real-time feedback data during the charging process, the latent space of the variational autoencoder model is continuously updated.

[0017] Optionally, the multi-source time series data specifically includes the voltage, current, temperature, remaining capacity and charging time of the battery in the rider battery swap cabinet at multiple time steps, which is used to comprehensively characterize the state evolution characteristics of the battery in the rider battery swap cabinet during the charging process.

[0018] Optionally, the multi-source time series data is preprocessed, specifically including data cleaning, missing value filling, normalization and time alignment, to improve data quality and ensure the consistency of data input to the variational autoencoder and the effectiveness of modeling.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Design a variational autoencoder model structure for learning the potential characteristics of the battery charging process. The variational autoencoder model structure includes an encoder, a decoder, and a latent space module. The encoder includes a multi-layer neural network, each layer contains a certain number of neurons, and the activation function is ReLU. The latent space module is set to a fixed dimension Z to represent the potential health status and charging behavior characteristics of the battery in the rider's battery swap cabinet;

[0021] S22, inputting the preprocessed data set into the encoder, which extracts features from the data set and outputs the mean vector and standard deviation vector of the latent space;

[0022] S23. In the latent space, according to the mean vector and standard deviation vector output by the encoder, a reparameterization method is used to perform random sampling to generate latent variables;

[0023] S24, inputting the latent variables into a decoder, which is a symmetrically structured neural network and is used to restore the features of the latent variables, including time-series reconstruction values of multi-dimensional features of voltage, current, and temperature;

[0024] S25. Determine the reconstruction error between the decoder output and the original input data, and combine the KL divergence between the latent space distribution and the standard normal distribution as a loss function to measure the reconstruction accuracy of the variational autoencoder model and the consistency of the latent space distribution;

[0025] S26. Based on the loss function, the variational autoencoder model is iteratively trained through backpropagation and gradient descent methods, and the encoder outputs a stable latent space representation for accurately reflecting the health status and charging behavior characteristics of the battery.

[0026] Optionally, the decoder structure is symmetrical to the encoder and is used to receive latent variables and reconstruct the original input data set.

[0027] Optionally, the latent variable is a low-dimensional feature vector that obeys a standard normal distribution, which is used to represent the evolution characteristics of the charging state of the battery in the rider's battery swap cabinet.

[0028] Optionally, the S3 specifically includes:

[0029] S31. Select a recently processed data set as input window data, which contains the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider's battery swap cabinet at multiple consecutive time steps;

[0030] S32. Input the input window data into the encoder part of the trained variational autoencoder, extract the latent space representation within the corresponding time period, and obtain the compressed features of the current state of the battery in the rider's battery swap cabinet;

[0031] S33. Perform temporal splicing of the current potential representation and the historical potential representation to construct a temporal evolution trajectory of the battery characteristics in the rider's battery swap cabinet to describe the charging state change trend;

[0032] S34. Based on the time series evolution trajectory, a sliding window method is used to construct a prediction input for future time steps, and the input is input to the decoder module to deduce the future state;

[0033] S35. Obtain battery status prediction data for multiple future time steps through the decoder output, where the battery status prediction data includes characteristic values of voltage, current, temperature, and remaining capacity corresponding to the prediction time point;

[0034] S36: Output the battery state prediction data as a battery state prediction result.

[0035] Optionally, the input window data specifically includes the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider battery swap cabinet in multiple consecutive time steps, which is used to capture the state change characteristics of the battery in the rider battery swap cabinet during the continuous time evolution process.

[0036] Optionally, the S4 specifically includes:

[0037] S41. Based on the battery status prediction data P and the potential representation H of the health status and charging behavior of the battery in the rider battery swap cabinet, construct an initial fitness function F = f(P, H, T) for charging topology optimization, where T represents the battery charging topology in the rider battery swap cabinet;

[0038] S42. Initialize the moth individual set in the moth optimization algorithm Each moth individual M i represents a charging topology configuration and the corresponding charging parameter combination, and N represents the total number of individuals initialized and generated in the moth optimization algorithm;

[0039] S43, for each moth individual M i , calculate the fitness value according to the fitness function:

[0040]

[0041] in, Indicates that based on the current moth individual M i The charging topology and parameters represented by Indicates that based on the current moth individual M iThe estimated result of the battery health status in the rider battery swap cabinet generated by the configuration, α and β are weight coefficients, F(M i ) represents the i-th moth individual M i The fitness value of

[0042] S44. Based on the deviation between the battery state prediction data generated by the variational autoencoder and the real-time monitoring data, as well as the deviation between the battery health state prediction and the actual measurement value, dynamically adjust the control parameter b in the moth spiral search to control the position update step size of the individual moth:

[0043]

[0044] Among them, b0 is the initial spiral control parameter, k is the adjustment coefficient, tanh is the hyperbolic tangent function, is the i-th eigenvalue in the battery state vector predicted by the variational autoencoder model, is the corresponding ith eigenvalue in the battery state vector obtained by real-time monitoring, n is the dimension of the battery state eigenvector, is the jth eigenvalue in the battery health status vector predicted by the variational autoencoder model, is the jth eigenvalue corresponding to the actual measured battery health state, m is the dimension of the battery health state eigenvector, λ is the weighting coefficient, and Δ0 is the normalization threshold;

[0045] S45. Based on the updated control parameter b in the moth spiral search, the position of the moth is updated using a spiral function according to the relative position between the moth and the corresponding flame. The moth spirals toward the current guiding flame in the solution space:

[0046]

[0047] in, is the updated position of the i-th moth individual in the current iteration, i.e., the current candidate solution, represents the position of the i-th moth individual before the update in the current iteration, D is the Euclidean distance between the i-th moth and the current corresponding flame, t is a random variable, exp is the exponential function, and cos(2πt) represents the periodic component that forms the spiral path;

[0048] S46. Divide the moth individuals into K subgroups according to the battery groups or status characteristics in the rider's battery swap cabinet, and determine the local optimal moth individual for each subgroup Update the global best moth position:

[0049]

[0050] in, represents the global best moth position in the current iteration, Indicates traversing between j = 1 and K, returning the moth individual corresponding to the index j that minimizes the objective function value, λ ′ is the balance coefficient, C j represents the centroid coordinates of the jth subgroup, represents the fitness value of the local optimal moth individual in the j-th subgroup;

[0051] S47, based on the updated global optimal moth position Re-evaluate and rank the fitness of individual moths to determine the flame set for the next round of spiral renewal;

[0052] S48. After the spiral update, the current candidate solution for each moth individual A hybrid local search strategy combining simulated annealing and perturbation vector is introduced for local fine-tuning:

[0053]

[0054] in, represents the solution of the i-th moth individual after fine-tuning by hybrid local search, η is the local search step length adjustment coefficient, For the current moth individual The fitness value, T o is the simulated annealing temperature parameter, ξ is the normal perturbation vector;

[0055] S49. Comparison of solutions after hybrid local search fine-tuning With the current candidate solution The fitness value of the moth is selected as the individual moth state in the next generation population;

[0056] S410, repeat steps S43 to S49 until the preset convergence condition is met Or the maximum number of iterations is reached, where ε represents the fitness convergence threshold, and the final optimized battery charging topology T in the rider battery swap cabinet is output. (opt) , Charging sequence S (opt) , charging voltage V (opt) With the charging current I (opt) .

[0057] Optionally, the S6 specifically includes:

[0058] S61. Real-time collection of feedback data from the battery in the rider's battery swap cabinet during the charging process, including voltage, current, temperature, charging time, and remaining capacity;

[0059] S62. Standardize the collected real-time feedback data and use it as new input data for the variational autoencoder model;

[0060] S63, inputting the real-time normalized data into the currently trained encoder module to obtain the real-time potential representation, and performing comparative analysis with the potential space distribution generated during the original training;

[0061] S64, determining the degree of deviation of the latent space representation, and triggering the latent space update process when the distribution of the new input data in the latent space deviates from the original distribution structure by more than a set threshold;

[0062] S65. Perform incremental training or fine-tuning using the processed data set and new input data to update encoder and decoder parameters of the variational autoencoder and optimize the adaptability of the latent space to the current battery state distribution;

[0063] S66. Using the updated latent space for battery state prediction and state estimation input in the moth optimization algorithm process.

[0064] The beneficial effects of the present invention are:

[0065] The present invention achieves accurate modeling, dynamic prediction and intelligent optimization of the battery charging status in the rider battery swap cabinet by introducing a variational autoencoder model and an improved moth optimization algorithm, overcoming the static, unbalanced and delayed response defects of the charging strategy in the prior art. In the heterogeneous state of multiple batteries, the variational autoencoder can effectively extract the potential characteristics of the battery at different charging stages, construct a temporal latent space reflecting the health status and charging behavior, and improve the system's perception and prediction accuracy of battery status changes. By linking with real-time monitoring data during the charging process, the model has the ability to adaptively update, so that it maintains high sensitivity and accuracy to the system status during long-term operation, and enhances the stability and usability of the model in the actual battery swap environment.

[0066] In terms of optimization control, the present invention constructs a fitness function that integrates state prediction information and introduces a multi-flame collaborative search mechanism and a hybrid local search strategy, which enables the improved moth optimization algorithm to have stronger global search capabilities and local fine-tuning capabilities. Through the adaptive dynamic adjustment of the spiral search parameters, the optimization strategy can quickly respond to battery state fluctuations and improve the control accuracy, avoiding the shortcomings of traditional swarm intelligence algorithms in complex systems, such as slow convergence speed and many local optimal traps. Furthermore, through real-time topology structure updates and charging parameter fine-tuning, the present invention effectively realizes the dynamic allocation of control parameters such as charging sequence, current, and voltage between batteries, fundamentally improving the balance and energy utilization efficiency of the charging process.

[0067] Therefore, the present invention not only improves the intelligent control capability of the rider battery swap cabinet system under complex and dynamic working conditions, but also realizes the deep integration of data-driven prediction and optimization algorithms, extending the battery life while ensuring charging safety, and has broad application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flow chart of a dynamic topology charging balancing control method for a rider battery swap cabinet proposed by the present invention;

[0070] Figure 2 This is a schematic diagram of the multi-flame collaborative search and local fine-tuning mechanism in the moth optimization algorithm of the dynamic topology charging balancing control method for the rider battery swap cabinet proposed in the present invention;

[0071] Figure 3 This is a schematic diagram of battery state modeling and future state prediction based on a variational autoencoder for a dynamic topology charging balancing control method for a rider battery swap cabinet proposed in the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1 、 Figure 2 and Figure 3 , a dynamic topology charging balancing control method for a rider battery swap cabinet, comprising the following steps:

[0074] S1. Collect multi-source time series data of batteries in the rider's battery swap cabinet, pre-process the multi-source time series data, and generate a processed data set;

[0075] S2. Construct a variational autoencoder model. Using the trained dataset, extract the potential features of the battery charging process in the rider’s battery swap cabinet, and obtain the potential representation of the health status and charging behavior of the battery in the rider’s battery swap cabinet.

[0076] S3. Use the variational autoencoder model to predict the future charging state of the battery in the rider's battery swap cabinet and generate battery state prediction data;

[0077] S4. Based on the generated battery status prediction data, the health status of the batteries in the rider battery swap cabinet, and the potential representation of the charging behavior, an improved moth optimization algorithm is used to optimize the charging topology between the batteries in the rider battery swap cabinet, and the charging sequence, charging voltage, and current of the batteries in the rider battery swap cabinet are dynamically adjusted;

[0078] S5. During the charging process, the actual charging status of the battery in the rider's battery swap cabinet is monitored in real time, and real-time optimization and adjustment are performed;

[0079] S6. Based on the real-time feedback data during the charging process, the latent space of the variational autoencoder model is continuously updated.

[0080] The present invention combines a variational autoencoder with an improved moth optimization algorithm to construct an integrated charging management method with state perception, behavior modeling, predictive decision-making, and dynamic control capabilities. Compared with the traditional fixed topology charging strategy, this method can accurately extract the potential health status and charging behavior characteristics of the battery from multi-source time series data, and predict the future state through the model, providing a highly reliable basis for optimization control. By dynamically optimizing the charging topology structure and key parameters such as voltage and current, intelligent scheduling and balanced control of the charging process are achieved, which improves the overall system energy efficiency and battery life. At the same time, the introduction of real-time feedback mechanism and model adaptive update capability enables the system to continuously optimize the control strategy according to the changes in the state during operation, has good environmental adaptability and long-term stability, and significantly improves the charging safety, balance and automation level.

[0081] In this embodiment, the multi-source time series data specifically includes the voltage, current, temperature, remaining capacity and charging time of the battery in the rider battery swap cabinet at multiple time steps, which is used to comprehensively characterize the state evolution characteristics of the battery in the rider battery swap cabinet during the charging process.

[0082] In this embodiment, the preprocessing of the multi-source time series data specifically includes data cleaning, missing value filling, normalization and time alignment, which are used to improve data quality and ensure the consistency of data input to the variational autoencoder and the effectiveness of modeling.

[0083] In this embodiment, S2 specifically includes:

[0084] S21. Design a variational autoencoder model structure for learning the potential characteristics of the battery charging process. The variational autoencoder model structure includes an encoder, a decoder, and a latent space module. The encoder includes a multi-layer neural network, each layer contains a certain number of neurons, and the activation function is ReLU. The latent space module is set to a fixed dimension Z to represent the potential health status and charging behavior characteristics of the battery in the rider's battery swap cabinet;

[0085] S22, inputting the preprocessed data set into the encoder, which extracts features from the data set and outputs the mean vector and standard deviation vector of the latent space;

[0086] S23. In the latent space, according to the mean vector and standard deviation vector output by the encoder, a reparameterization method is used to perform random sampling to generate latent variables;

[0087] S24, inputting the latent variables into a decoder, which is a symmetrically structured neural network and is used to restore the features of the latent variables, including time-series reconstruction values of multi-dimensional features of voltage, current, and temperature;

[0088] S25. Determine the reconstruction error between the decoder output and the original input data, and combine the KL divergence between the latent space distribution and the standard normal distribution as a loss function to measure the reconstruction accuracy of the variational autoencoder model and the consistency of the latent space distribution;

[0089] S26. Based on the loss function, the variational autoencoder model is iteratively trained through backpropagation and gradient descent methods, and the encoder outputs a stable latent space representation for accurately reflecting the health status and charging behavior characteristics of the battery.

[0090] The present invention realizes the deep modeling and efficient feature abstraction of the multi-dimensional time series data of the battery charging process in the rider battery swap cabinet by constructing a variational autoencoder model with a reasonable structure and clear functions. The model adopts a multi-layer neural network to construct the encoder structure, which can effectively extract the nonlinear behavior characteristics of the battery in different charging stages, and represent the health status and charging behavior of the battery through a latent space of fixed dimensions, thereby enhancing the consistency and interpretability of the state modeling. Through re-parameterized sampling and symmetric decoder structure, the model can not only reconstruct the time series values of key features such as voltage, current, and temperature, but also has strong generalization ability. The loss function constructed by combining reconstruction error and KL divergence ensures that the model maintains the structural stability of the latent space while learning the data distribution. Finally, through backpropagation and gradient optimization training, the potential representation of the encoder output has a high expressive ability for battery health and charging behavior. This model provides a reliable characterization basis for subsequent state prediction and charging topology optimization, and effectively improves the system's intelligent analysis and control capabilities under complex and changeable battery operating conditions.

[0091] In this embodiment, the decoder structure is symmetrical to the encoder and is used to receive latent variables and reconstruct the original input data set.

[0092] In this embodiment, the latent variable is a low-dimensional feature vector that obeys a standard normal distribution, which is used to represent the evolution characteristics of the charging state of the battery in the rider's battery swap cabinet.

[0093] In this embodiment, S3 specifically includes:

[0094] S31. Select a recently processed data set as input window data, which contains the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider's battery swap cabinet at multiple consecutive time steps;

[0095] S32. Input the input window data into the encoder part of the trained variational autoencoder, extract the latent space representation within the corresponding time period, and obtain the compressed features of the current state of the battery in the rider's battery swap cabinet;

[0096] S33. Perform temporal splicing of the current potential representation and the historical potential representation to construct a temporal evolution trajectory of the battery characteristics in the rider's battery swap cabinet to describe the charging state change trend;

[0097] S34. Based on the time series evolution trajectory, a sliding window method is used to construct a prediction input for future time steps, and the input is input to the decoder module to deduce the future state;

[0098] S35. Obtain battery status prediction data for multiple future time steps through the decoder output, where the battery status prediction data includes characteristic values of voltage, current, temperature, and remaining capacity corresponding to the prediction time point;

[0099] S36: Output the battery state prediction data as a battery state prediction result.

[0100] By constructing a battery state prediction mechanism centered around a variational autoencoder, the present invention achieves high-precision, multi-dimensional predictions of the future charging state of batteries in rider battery swap cabinets. By selecting a continuous time-step input window containing voltage, current, temperature, remaining capacity, and charging time, the system can comprehensively capture the dynamic behavior characteristics of the battery during its temporal evolution. The latent space representation processed by the encoder not only retains key features but also significantly compresses redundant information. By concatenating the current and historical latent representations, the temporal evolution trajectory of the battery state is further constructed, providing a basis for the model to understand state trends and change patterns. The prediction process uses a sliding window mechanism, enhancing the ability to simulate short- and medium-term charging trends. Ultimately, the decoder outputs state values for multiple future time steps, including key indicators such as voltage, current, temperature, and remaining capacity. This end-to-end state prediction mechanism not only improves the continuity and stability of the prediction but also provides a reliable, real-time input foundation for subsequent charging topology optimization. This method enhances the system's forward-looking scheduling capabilities, effectively avoids risks such as battery overheating and undercharging, and improves the intelligence and operational efficiency of the battery swap system.

[0101] In this embodiment, the input window data specifically includes the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider battery swap cabinet in multiple continuous time steps, which is used to capture the state change characteristics of the battery in the rider battery swap cabinet during the continuous time evolution process.

[0102] In this embodiment, the S4 specifically includes:

[0103] S41. Based on the battery status prediction data P and the potential representation H of the health status and charging behavior of the battery in the rider battery swap cabinet, construct an initial fitness function F = f(P, H, T) for charging topology optimization, where T represents the battery charging topology in the rider battery swap cabinet;

[0104] S42. Initialize the moth individual set in the moth optimization algorithm Each moth individual M i represents a charging topology configuration and the corresponding charging parameter combination, and N represents the total number of individuals initialized and generated in the moth optimization algorithm;

[0105] S43, for each moth individual M i , calculate the fitness value according to the fitness function:

[0106]

[0107] in, Indicates that based on the current moth individual M i The charging topology and parameters represented by Indicates that based on the current moth individual M i The estimated result of the battery health status in the rider battery swap cabinet generated by the configuration, α and β are weight coefficients, F(M i ) represents the i-th moth individual M i The fitness value of

[0108] S44. Based on the deviation between the battery state prediction data generated by the variational autoencoder and the real-time monitoring data, as well as the deviation between the battery health state prediction and the actual measurement value, dynamically adjust the control parameter b in the moth spiral search to control the position update step size of the individual moth:

[0109]

[0110] Among them, b0 is the initial spiral control parameter, k is the adjustment coefficient, tanh is the hyperbolic tangent function, is the i-th eigenvalue in the battery state vector predicted by the variational autoencoder model, is the corresponding ith eigenvalue in the battery state vector obtained by real-time monitoring, n is the dimension of the battery state eigenvector, is the jth eigenvalue in the battery health status vector predicted by the variational autoencoder model, is the jth eigenvalue corresponding to the actual measured battery health state, m is the dimension of the battery health state eigenvector, λ is the weighting coefficient, and Δ0 is the normalization threshold;

[0111] S45. Based on the updated control parameter b in the moth spiral search, the position of the moth is updated using a spiral function according to the relative position between the moth and the corresponding flame. The moth spirals toward the current guiding flame in the solution space:

[0112]

[0113] in, is the updated position of the i-th moth individual in the current iteration, i.e., the current candidate solution, represents the position of the i-th moth individual before the update in the current iteration, D is the Euclidean distance between the i-th moth and the current corresponding flame, t is a random variable, exp is the exponential function, and cos(2πt) represents the periodic component that forms the spiral path;

[0114] S46. Divide the moth individuals into K subgroups according to the battery groups or status characteristics in the rider's battery swap cabinet, and determine the local optimal moth individual for each subgroup Update the global best moth position:

[0115]

[0116] in, represents the global best moth position in the current iteration, Indicates traversing between j = 1 and K, returning the moth individual corresponding to the index j that minimizes the objective function value, λ ′ is the balance coefficient, C j represents the centroid coordinates of the jth subgroup, represents the fitness value of the local optimal moth individual in the j-th subgroup;

[0117] S47, based on the updated global optimal moth position Re-evaluate and rank the fitness of individual moths to determine the flame set for the next round of spiral renewal;

[0118] S48. After the spiral update, the current candidate solution for each moth individual A hybrid local search strategy combining simulated annealing and perturbation vector is introduced for local fine-tuning:

[0119]

[0120] in, represents the solution of the i-th moth individual after fine-tuning by hybrid local search, η is the local search step length adjustment coefficient, For the current moth individual The fitness value, T o is the simulated annealing temperature parameter, ξ is the normal perturbation vector;

[0121] S49. Comparison of solutions after hybrid local search fine-tuning With the current candidate solution The fitness value of the moth is selected as the individual moth state in the next generation population;

[0122] S410, repeat steps S43 to S49 until the preset convergence condition is met Or the maximum number of iterations is reached, where ε represents the fitness convergence threshold, and the final optimized battery charging topology T in the rider battery swap cabinet is output. (opt) , Charging sequence S (opt) , charging voltage V (opt) With the charging current I (opt) .

[0123] The present invention proposes a charging topology optimization method that deeply integrates the variational autoencoder model with the improved moth optimization algorithm, aiming to solve the problems of low search accuracy, slow convergence speed, and easy falling into local optimality in traditional swarm intelligence algorithms in battery charging scenarios. In order to adapt to the complex charging environment with multiple batteries in heterogeneous states and frequent dynamic changes in the rider's battery swap cabinet, the present invention introduces state prediction error and health state deviation as the basis for parameter control, dynamically adjusts the spiral update step size, and improves the moth's individual search process's perception of system state changes. At the same time, the proposed multi-subgroup collaborative search mechanism and local disturbance fine-tuning strategy can enhance the stability and accuracy of local solutions while maintaining global search capabilities, and effectively avoid the problem of falling into local extreme values. The algorithm takes into account both global scheduling and individual optimization in iterative control, so that the generated charging topology structure, sequence, voltage and current strategy can fully consider the predicted trend and actual deviation, significantly improving the system's adaptability, balance and energy utilization efficiency, thereby providing a more robust and accurate optimization control solution for high-frequency and dynamic battery swap applications.

[0124] In this embodiment, S6 specifically includes:

[0125] S61. Real-time collection of feedback data from the battery in the rider's battery swap cabinet during the charging process, including voltage, current, temperature, charging time, and remaining capacity;

[0126] S62. Standardize the collected real-time feedback data and use it as new input data for the variational autoencoder model;

[0127] S63, inputting the real-time normalized data into the currently trained encoder module to obtain the real-time potential representation, and performing comparative analysis with the potential space distribution generated during the original training;

[0128] S64, determining the degree of deviation of the latent space representation, and triggering the latent space update process when the distribution of the new input data in the latent space deviates from the original distribution structure by more than a set threshold;

[0129] S65. Perform incremental training or fine-tuning using the processed data set and new input data to update encoder and decoder parameters of the variational autoencoder and optimize the adaptability of the latent space to the current battery state distribution;

[0130] S66. Using the updated latent space for battery state prediction and state estimation input in the moth optimization algorithm process.

[0131] The present invention solves the problem that traditional static models are difficult to cope with battery state drift, environmental changes and actual feedback errors in long-term operation by introducing a latent space adaptive update mechanism based on real-time feedback data. By continuously collecting real-time data of the battery during the charging process, such as voltage, current, temperature, charging time and remaining capacity, the system can fully perceive the dynamic changes of the battery state. The standardized real-time data is input into the trained variational autoencoder model, and the offset between its latent representation and the initial training distribution is detected. When the offset exceeds the threshold, the model fine-tuning process is triggered, effectively preventing model aging and failure. This mechanism continuously optimizes the encoder and decoder through incremental training or parameter fine-tuning, so that the latent space can adapt to the current system state and maintain accurate abstraction and compression expression of battery behavior characteristics. The final updated latent space not only improves the accuracy of battery state prediction, but also provides a more reliable state estimation input for the moth optimization algorithm, thereby improving the rationality and responsiveness of the charging scheduling strategy and significantly enhancing the long-term stability and adaptive operation capability of the system.

[0132] Example 1:

[0133] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a logistics park, where an average of more than 150 takeout, express and shared delivery riders frequently replace batteries every day. In order to meet the needs of high-intensity operations around the clock, the park deployed a 24-slot intelligent battery exchange cabinet for centralized management, charging and scheduling of battery modules. However, in the actual operation process, it was found that due to the uneven frequency of battery replacement by riders and large differences in the degree of battery aging, the fixed-sequence charging strategy adopted by the original battery exchange cabinet system could not be intelligently adjusted according to the real-time status of the battery, resulting in overcharging of some batteries and excessive use of some batteries in cycles, resulting in the overall proportion of available batteries dropping to about 76.4%, and the average battery temperature rising by 18.6°C, posing a safety hazard.

[0134] In order to solve the above problems, this embodiment deploys a dynamic topology charging balancing control method for a rider battery swap cabinet based on the present invention in the battery swap cabinet system. The system first accesses the multi-source data interface of all batteries, collects multi-dimensional time series data such as voltage, current, temperature, remaining capacity, charging time, etc. of each battery, and standardizes it as input data. It also uses a variational autoencoder model to deeply model and predict the battery status, and extracts the potential characteristics of the charging behavior and the latent vector of the health status. In the early stage of system deployment, the battery charging data of the past 30 days was trained to construct an initial latent space model. The training data contains approximately 115,200 records, covering the behavior of all batteries under different temperature and load conditions.

[0135] In the prediction module, the system updates the input window every five minutes, predicting the voltage, current, temperature rise, and health trends of each battery cell in the current topology over the next hour. Simultaneously, the system dynamically updates the latent space distribution based on the difference between real-time and predicted data (with an average error of approximately 3.2%), ensuring that the modeling results consistently accurately reflect the current operating status.

[0136] In the optimization control link, the system constructs a fitness function including state deviation, health index, and temperature rise risk based on the above prediction results, and uses the improved moth optimization algorithm to jointly optimize the charging topology and parameters. The moth algorithm combines the dynamic spiral parameter adjustment mechanism with the local disturbance correction strategy during the search process, so that the entire optimization process has both global search capabilities and fine-tuning of sensitive battery strategies. In this battery swap cabinet scenario, the algorithm performs 100 optimization iterations per round, and the total calculation time is controlled within 12 seconds, which is much lower than the system control interval time (30 seconds), and can achieve quasi-real-time optimization control.

[0137] During the first operation cycle (7 days) after deploying the method of the present invention, the system recorded various operating data and conducted comparative analysis with the traditional fixed topology control strategy.

[0138] Table 1 Comparison of key indicators before and after implementation of the present invention

[0139]

[0140]

[0141] The following is a detailed paragraph-by-paragraph analysis of the "Comparison Table of Optimization Effects of Rider Battery Swapping Cabinets." Combining the various indicators in the table, this systematically summarizes the technical advantages and actual effects of the method of the present invention in actual scenarios:

[0142] After the present invention was deployed and operated for 7 days in a battery swap cabinet system in a rider-dense area in Nanshan District, Shenzhen, the system operating efficiency and battery management level were significantly improved. From the core indicator of "average available battery ratio", before optimization, only about 76.4% of the batteries were replaceable at any time. After adopting the method of the present invention, the proportion rose to 91.3%, an increase of 14.9%. This shows that this method can significantly improve the efficiency of battery resource scheduling, increase the turnover rate and frequency of use of batteries, and provide riders with more stable battery replacement guarantees.

[0143] In terms of the number of failed battery swaps per day, before optimization, the system experienced up to 12 failed swaps per day due to insufficient battery charge or untimely charging. After implementing dynamic topology optimization, this number dropped to 3 per day, a reduction of approximately 75%, significantly improving the user's battery swap success rate and service experience. Furthermore, the percentage of batteries operating at high temperatures dropped from 21.7% before optimization to 8.2%, a decrease of over 62%. This demonstrates that this method can effectively reduce the risk of battery overheating and enhance system safety under high-frequency, high-load operation.

[0144] Regarding battery thermal stability, the average battery temperature rise decreased from 18.6°C to 10.3°C, a 44.6% decrease. This demonstrates that this method's optimization of charging current and voltage scheduling has achieved effective thermal management, significantly extending battery life and reducing failure rates. Furthermore, the standard deviation of battery charge balance decreased from 0.164 to 0.142, a 13.5% improvement. This indicates a more balanced charging process, significantly reduced differences between batteries, and a more stable overall system health.

[0145] In terms of prediction accuracy, the variational autoencoder model used in this paper achieved a prediction error of 3.2%, demonstrating its strong modeling capabilities and excellent generalization performance for multi-source data, providing a highly reliable decision-making basis for subsequent optimization. Regarding battery life, it is estimated that the optimized charging control strategy can extend battery life by approximately 2.1 months, indirectly reducing operation and maintenance costs and the frequency of battery replacements.

[0146] In terms of response speed, traditional systems are unable to implement real-time scheduling and rely on static control rules. However, the optimization calculations of our invention take an average of only 12 seconds, far less than the system control cycle, achieving near-real-time policy updates and execution. Regarding human dependency, the system required ≥5 manual interventions per week before optimization, but this was reduced to ≤1 after optimization, reducing the manual burden by approximately 80% and improving the level of automation and intelligence.

[0147] Overall, this invention achieves significant improvements in multiple key metrics, including charging efficiency, battery life, safety and stability, system response speed, and operation and maintenance costs, demonstrating its practicality and advancement in complex urban battery swapping scenarios. This technical solution not only possesses strong adaptability and scalability, but also provides a solid technical foundation for the intelligent upgrade of future smart battery swapping platforms.

[0148] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A dynamic topology charging balancing control method for a rider battery swap cabinet, characterized in that: The steps include: S1. Collect multi-source time series data of batteries in the rider's battery swap cabinet, pre-process the multi-source time series data, and generate a processed data set; S2. Construct a variational autoencoder model. Using the trained dataset, extract the potential features of the battery charging process in the rider’s battery swap cabinet, and obtain the potential representation of the health status and charging behavior of the battery in the rider’s battery swap cabinet. S3. Use the variational autoencoder model to predict the future charging state of the battery in the rider's battery swap cabinet and generate battery state prediction data; S4. Based on the generated battery status prediction data, the health status of the batteries in the rider battery swap cabinet, and the potential representation of the charging behavior, an improved moth optimization algorithm is used to optimize the charging topology between the batteries in the rider battery swap cabinet, and the charging sequence, charging voltage, and current of the batteries in the rider battery swap cabinet are dynamically adjusted; S5. During the charging process, the actual charging status of the battery in the rider's battery swap cabinet is monitored in real time, and real-time optimization and adjustment are performed; S6. Based on the real-time feedback data during the charging process, the latent space of the variational autoencoder model is continuously updated.

2. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The multi-source time series data specifically includes the voltage, current, temperature, remaining capacity and charging time of the battery in the rider battery swap cabinet at multiple time steps, which is used to comprehensively characterize the state evolution characteristics of the battery in the rider battery swap cabinet during the charging process.

3. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The preprocessing of the multi-source time series data specifically includes data cleaning, missing value filling, normalization and time alignment, which are used to improve data quality and ensure the consistency of data input to the variational autoencoder and the effectiveness of modeling.

4. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The S2 specifically includes: S21. Design a variational autoencoder model structure for learning the potential characteristics of the battery charging process. The variational autoencoder model structure includes an encoder, a decoder, and a latent space module. The encoder includes a multi-layer neural network, each layer contains a certain number of neurons, and the activation function is ReLU. The latent space module is set to a fixed dimension Z to represent the potential health status and charging behavior characteristics of the battery in the rider's battery swap cabinet; S22, inputting the preprocessed data set into the encoder, which extracts features from the data set and outputs the mean vector and standard deviation vector of the latent space; S23. In the latent space, according to the mean vector and standard deviation vector output by the encoder, a reparameterization method is used to perform random sampling to generate latent variables; S24, inputting the latent variables into a decoder, which is a symmetrically structured neural network and is used to restore the features of the latent variables, including time-series reconstruction values of multi-dimensional features of voltage, current, and temperature; S25. Determine the reconstruction error between the decoder output and the original input data, and combine the KL divergence between the latent space distribution and the standard normal distribution as a loss function to measure the reconstruction accuracy of the variational autoencoder model and the consistency of the latent space distribution; S26. Based on the loss function, the variational autoencoder model is iteratively trained through backpropagation and gradient descent methods, and the encoder outputs a stable latent space representation for accurately reflecting the health status and charging behavior characteristics of the battery.

5. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 2, characterized in that: The decoder structure is symmetrical to the encoder and is used to receive latent variables and reconstruct the original input dataset.

6. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 2, characterized in that: The latent variable is a low-dimensional feature vector that obeys a standard normal distribution and is used to represent the evolution characteristics of the charging state of the battery in the rider's battery swap cabinet.

7. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The S3 specifically includes: S31. Select a recently processed data set as input window data, which includes the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider's battery swap cabinet at multiple consecutive time steps; S32. Input the input window data into the encoder part of the trained variational autoencoder, extract the latent space representation within the corresponding time period, and obtain the compressed features of the current state of the battery in the rider's battery swap cabinet; S33. Perform temporal splicing of the current potential representation and the historical potential representation to construct a temporal evolution trajectory of the battery characteristics in the rider's battery swap cabinet to describe the charging state change trend; S34. Based on the time series evolution trajectory, a sliding window method is used to construct a prediction input for future time steps, and the input is input to the decoder module to deduce the future state; S35. Obtain battery status prediction data for multiple future time steps through the decoder output, where the battery status prediction data includes voltage, current, temperature, and remaining capacity characteristic values corresponding to the prediction time points; S36: Output the battery state prediction data as a battery state prediction result.

8. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 7, characterized in that: The input window data specifically includes the voltage, current, temperature, remaining capacity and charging time characteristics of the battery in each rider battery swap cabinet in multiple continuous time steps, which is used to capture the state change characteristics of the battery in the rider battery swap cabinet during the continuous time evolution process.

9. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The S4 specifically includes: S41. Based on the battery status prediction data P and the potential representation H of the health status and charging behavior of the battery in the rider battery swap cabinet, construct an initial fitness function F = f(P, H, T) for charging topology optimization, where T represents the battery charging topology in the rider battery swap cabinet; S42. Initialize the moth individual set in the moth optimization algorithm Each moth individual M i represents a charging topology configuration and the corresponding charging parameter combination, and N represents the total number of individuals initialized and generated in the moth optimization algorithm; S43, for each moth individual M i , calculate the fitness value according to the fitness function: in, Indicates that based on the current moth individual M i The charging topology and parameters represented by Indicates that based on the current moth individual M i The estimated result of the battery health status in the rider battery swap cabinet generated by the configuration, α and β are weight coefficients, F(M i ) represents the i-th moth individual M i The fitness value of S44. Based on the deviation between the battery state prediction data generated by the variational autoencoder and the real-time monitoring data, as well as the deviation between the battery health state prediction and the actual measurement value, dynamically adjust the control parameter b in the moth spiral search to control the position update step size of the moth individual: Among them, b0 is the initial spiral control parameter, k is the adjustment coefficient, tanh is the hyperbolic tangent function, is the i-th eigenvalue in the battery state vector predicted by the variational autoencoder model, is the corresponding ith eigenvalue in the battery state vector obtained by real-time monitoring, n is the dimension of the battery state eigenvector, is the jth eigenvalue in the battery health status vector predicted by the variational autoencoder model, is the jth eigenvalue corresponding to the actual measured battery health state, m is the dimension of the battery health state eigenvector, λ is the weighting coefficient, and Δ0 is the normalization threshold; S45. Based on the updated control parameter b in the moth spiral search, the position of the moth is updated using a spiral function according to the relative position between the moth and the corresponding flame. The moth spirals toward the current guiding flame in the solution space: in, is the updated position of the i-th moth individual in the current iteration, i.e., the current candidate solution, represents the position of the i-th moth individual before the update in the current iteration, D is the Euclidean distance between the i-th moth and the current corresponding flame, t is a random variable, exp is the exponential function, and cos(2πt) represents the periodic component that forms the spiral path; S46. Divide the moth individuals into K subgroups according to the battery groups or status characteristics in the rider's battery swap cabinet, and determine the local optimal moth individual for each subgroup Update the global best moth position: in, represents the global best moth position in the current iteration, Indicates traversing between j = 1 and K, returning the moth individual corresponding to the index j that minimizes the objective function value, λ ′ is the balance coefficient, C j represents the centroid coordinates of the jth subgroup, represents the fitness value of the local optimal moth individual in the j-th subgroup; S47, based on the updated global optimal moth position Re-evaluate and rank the fitness of individual moths to determine the flame set for the next round of spiral renewal; S48. After the spiral update, the current candidate solution for each moth individual A hybrid local search strategy combining simulated annealing and perturbation vector is introduced for local fine-tuning: in, represents the solution of the i-th moth individual after fine-tuning by hybrid local search, η is the local search step length adjustment coefficient, For the current moth individual The fitness value, T o is the simulated annealing temperature parameter, ξ is the normal perturbation vector; S49. Comparison of solutions after hybrid local search fine-tuning With the current candidate solution The fitness value of the moth is selected as the individual moth state in the next generation population; S410, repeat steps S43 to S49 until the preset convergence condition is met Or the maximum number of iterations is reached, where ε represents the fitness convergence threshold, and the final optimized battery charging topology T in the rider battery swap cabinet is output. (opt) , Charging sequence S (opt) , charging voltage V (opt) With the charging current I (opt) .

10. A dynamic topology charging balancing control method for a rider battery swap cabinet according to claim 1, characterized in that: The S6 specifically includes: S61. Real-time collection of feedback data from the battery in the rider's battery swap cabinet during the charging process, including voltage, current, temperature, charging time, and remaining capacity; S62. Standardize the collected real-time feedback data and use it as new input data for the variational autoencoder model; S63, inputting the real-time normalized data into the currently trained encoder module to obtain the real-time potential representation, and performing comparative analysis with the potential space distribution generated during the original training; S64, determining the degree of deviation of the latent space representation, and triggering the latent space update process when the distribution of the new input data in the latent space deviates from the original distribution structure by more than a set threshold; S65. Perform incremental training or fine-tuning using the processed data set and new input data to update encoder and decoder parameters of the variational autoencoder and optimize the adaptability of the latent space to the current battery state distribution; S66. Using the updated latent space for battery state prediction and state estimation input in the moth optimization algorithm process.

Images