Battery fast charging strategy multi-objective optimization method based on meta-learning weight

Abstract

The invention discloses a multi-objective optimization method for a quick battery charging strategy based on meta-learning weight, and aims to realize high-efficiency, low-loss, safe and reliable quick charging. The method comprises the following steps: firstly, constructing a comprehensive input basic data set by fusing the real-time state of a battery, environmental parameters and user constraints; establishing a battery model, introducing a residual jump module to improve the battery model, inputting a training set, and training to obtain an optimal model; and subsequently, for three contradictory targets of charging time, temperature rise and capacity attenuation, an improved NSGA-II algorithm is proposed, a temperature sensitive coefficient and a dynamic constraint mechanism are introduced, and a Pareto leading edge is solved by utilizing an optimal model, so that an optimal charging curve is formed. The environment adaptability and the multi-target balance performance of the battery fast charging strategy are remarkably improved, and the method is suitable for intelligent charging scenes of various power batteries under different working conditions.

Description

Technical Field

[0001] The present invention relates to the field of battery charging technology, and in particular to a multi-objective optimization method for a battery fast charging strategy based on meta-learning weights. Background Art

[0002] With the rapid development of electric vehicles and renewable energy storage systems, batteries, as core energy storage units, have become a key bottleneck in terms of charging efficiency and safety, which restricts the development of the industry. Traditional charging strategies (such as constant current-constant voltage charging, pulse charging, etc.) are usually designed based on empirical models. Their core flaw is that they cannot establish a dynamic balance mechanism between charging time, temperature rise control, and capacity attenuation. For example, although constant current fast charging can shorten the charging time, it will cause the local temperature of the battery to rise by more than 50°C, accelerate the collapse of the electrode material structure, and increase the battery capacity attenuation rate by more than 30%; while conservative constant voltage charging can reduce the temperature rise, it requires an additional 2-3 hours of charging time, which is difficult to meet user needs.

[0003] As a complex electrochemical system, the battery's charging process is influenced by the interactions of multiple physical and chemical factors, including temperature, current, and voltage. There is an inherent contradiction between the three key indicators of charging time, temperature rise, and capacity decay. Fast charging is inevitably accompanied by higher currents, which cause significant temperature rise and exacerbate side reactions, leading to accelerated capacity decay. While reducing current can alleviate temperature rise and capacity loss, it significantly prolongs charging time. Currently, the industry lacks an effective methodology that can dynamically balance these mutually constrained objectives under different operating conditions and user preferences to achieve an intelligent charging decision-making mechanism that balances efficiency and safety. Summary of the Invention

[0004] In response to the technical problem in the above-mentioned existing technologies that traditional charging strategies cannot dynamically balance charging time, temperature rise control, and capacity attenuation, this technical solution provides a multi-objective optimization method for battery fast charging strategies based on meta-learning weights. The weights are initialized and dynamically adapted through meta-learning, enabling the model to quickly adapt to new task scenarios such as different battery types, aging levels, and temperature ranges. The backbone network adopts a Transformer or TCN architecture, and a residual module is added to the output of each layer of the network to directly pass the original SOC target value to the deep nodes of the network, avoiding gradient vanishing and enabling the model to better focus on the incremental optimization of polarization voltage reduction and temperature rise control; it can effectively solve the above-mentioned problems.

[0005] The present invention is achieved through the following technical solutions:

[0006] A multi-objective optimization method for battery fast charging strategy based on meta-learning weights, comprising the following steps:

[0007] Step 1: Build a data set: Collect battery status parameters and environmental parameters in real time to build a multi-dimensional input data set;

[0008] Step 2: Build the MAML framework model and train and test the dataset: Based on the multi-dimensional input dataset constructed in Step 1, construct the framework structure and initial parameters for meta-learning. Define the task pool by battery type, aging level, and temperature range. The input features of each task include SOC, SOH, and temperature, and the output is the optimization target of the charging strategy. Then, initialize the meta-weight matrix with a normal random distribution and a dimension of [number of tasks × number of features].

[0009] Step 3: Build a battery charging model: Based on the electrochemical reaction mechanism and the principle of energy conservation, combined with voltage, current, and temperature parameters, differential equations and data-driven methods are used to describe the dynamic characteristics of charging and discharging, and an electrochemical-thermal coupling model is established;

[0010] Step 4: Design a residual jump module: Add a residual jump module to the battery charging model constructed in Step 3. This module addresses the training bottleneck of deep neural networks by stabilizing gradients, preserving features, and implementing incremental learning mechanisms. The battery charging model includes a backbone network using a Transformer or TCN architecture, taking historical charging sequences as input. A residual module is added to the output of each network layer, forcing the model to focus on incremental optimization of polarization voltage reduction and temperature rise control.

[0011] Step 5: Multi-objective optimization: Input the improved NSGA-II algorithm into the trained model, define the multi-objective optimization problem, and achieve the Pareto optimal solution. The specific operation is as follows:

[0012] Step 5.1: Define the objective function: Charging time refers to the total time from the start of charging to the cutoff voltage, which is quantified by the following method; temperature rise refers to the difference between the battery surface temperature and the ambient temperature; capacity decay refers to the irreversible capacity loss of the battery due to side reactions during charging, which is quantified by the following model:

[0013]

[0014] f2=min(max(T cell )-T env )

[0015]

[0016] Among them, f1 represents minimizing charging time, f2 represents minimizing temperature rise, and f3 represents minimizing capacity attenuation; t charge is the charging time, T is the total number of charging stages, Δt is the duration of each stage, T env is the ambient temperature sensor, T cell Built-in temperature sensor for the battery, I rmsis the RMS value of the current, V over is the overvoltage peak, ΔQ loss is the capacity loss, α is the current loss coefficient, β is the overvoltage loss coefficient, To find the integral;

[0017] Design dynamic constraint conditions, voltage safety margin, expression is:

[0018] V min +0.1·SOH<V cutoff <V max -0.05·(100-SOH);

[0019] In the above formula, V min 、V max The minimum / maximum allowable voltage specified by the battery manufacturer, SOH is the battery health status (0-100%);

[0020] Current mutation rate: |dI / dt|<0.2C / s, to prevent lithium deposition;

[0021] Step 5.2: Initialize the population: Each individual is encoded as a three-dimensional vector [I max ,V end ,t charge ], the population size is adaptively adjusted according to the solution space dimension (base value 200);

[0022] x=[I max ,V end ,t charge ];

[0023] Among them, I max is the maximum charging current, V end is the cut-off voltage, t charge is the charging time;

[0024] Step 5.3: Compare the individuals in the population pairwise, define the dominance relationship, use the ε-dominance improvement strategy, set a loose threshold (±1.5°C) for the temperature rise target to avoid premature loss of feasible solutions;

[0025] Step 5.4: Calculate the crowding degree and calculate the crowding distance of each individual on the objective function to measure the density of its neighboring solutions. Perform weighted processing on minimizing charging time, minimizing temperature rise, and minimizing capacity decay, and introduce weight coefficients w1, w2, and w3.

[0026]

[0027] d i The crowding distance of the individual on the objective function, f i (x j ),fi (x j+1 ) is the evaluation value of the jth individual and the j+1th individual on the i-th objective function, max(f i ) and min(f i ) are the maximum and minimum values ​​of all individuals under the i-th objective function, n is the total number of individuals in the population (the number of solutions), is the sum of the differences in the evaluation values ​​of all adjacent individuals;

[0028] Dynamically adjust the target weight based on the battery health status:

[0029] W crowd =Softmax([w1,w2+δ·(1-SOH),w3]);

[0030] W crowd is the final generated dynamic weight vector, δ is the aging sensitivity coefficient, and when SOH < 80%, the capacity attenuation weight w3 increases by 20%;

[0031] Step 5.5: Introduce the crowding distance correction to avoid the temperature rise target dominating the solution set distribution in high temperature scenarios. The correction coefficient is expressed as:

[0032]

[0033] Among them, T sensitivity is the temperature sensitivity coefficient, which is 1.2 in high temperature scenario and 1.0 in normal temperature scenario; d i new is the crowding distance of the corrected individual on the objective function;

[0034] Step 5.6: Crossover and mutation: simulate binary crossover, cross the parent individuals to generate offspring, and control the crossover probability P c and distribution index η c ;

[0035]

[0036] x1, x2 are input variables, m is the result of dynamic calculation, which is used for optimizing weights and resource allocation ratios in subsequent operations; η c is the exponential adjustment coefficient, usually a positive real number greater than 1;

[0037] Mutate the offspring and introduce mutation probability P m and distribution index η m :

[0038]

[0039] In the above formula, Δx represents the parameter correction value, r is the correction factor, and x ub is the preset target value, xib For the polynomial mutation of the benchmark value, a temperature sensitivity coefficient is introduced, and the mutation probability is increased to 0.25 at high temperature to enhance the local search P m ;

[0040] Step 5.7: Filter user preferences and obtain user preference value λ∈[0,1] through past behavior analysis 3 , generate personalized weighted targets:

[0041] J user =λ1f1+λ2f2+λ3f3;

[0042] In the above formula, J user is the user's comprehensive utility function;

[0043] Combining the TOPSIS multi-criteria decision-making method, the strategy parameters closest to the ideal solution are selected. The polarization voltage prediction value is updated every 5% SOC interval. When the actual temperature rise deviates from the predicted value by more than 15%, online re-optimization is triggered. The capacity decay model parameters are updated after each charge, and the meta-weight matrix is ​​reconstructed every 50 cycles.

[0044] Step 5.8: Termination and output: When the maximum number of iterations is reached or the solution set converges, output the Pareto frontier Q_pareto, that is, the curve that minimizes charging time, minimizes temperature rise, and minimizes capacity decay. Otherwise, return to step 6.

[0045] Furthermore, the specific steps of constructing the data set in step 1 include:

[0046] Step 1.1: Data collection: Experimental data is collected in real time through the charge and discharge steps. The experimental data includes battery state parameters including voltage, current, temperature, SOC, SOH, and internal resistance, and environmental parameters including temperature, humidity, and historical cycle records.

[0047] Step 1.2: Design operating condition coverage: Design an orthogonal experimental matrix covering different charge rates (0.5C-3C), temperature ranges (-10°C to 50°C), battery aging levels (SOH 100% to 60%), and simulate extreme operating conditions;

[0048] Step 1.3: Data cleaning and alignment: Eliminate data that exceeds the battery's nominal parameters, process outliers, and use cubic spline interpolation to unify the time base for asynchronous data to ensure time synchronization;

[0049] Step 1.4: Data enhancement: Use FGSM (Fast Gradient Sign Method, a classic adversarial sample generation algorithm that uses the gradient information of the neural network to quickly generate perturbation samples that can cause model misclassification) to eliminate power frequency interference and improve the robustness of the dataset.

[0050] Furthermore, the specific operation of step 2 includes:

[0051] Step 2.1: Receive the input multi-dimensional data set through the input layer and standardize the data;

[0052] Step 2.2: Construct the meta-weight matrix: Construct a matrix of dimension N×d and initialize it. The initialization method is:

[0053]

[0054] N(0,σ 2 ) represents normal distribution, σ represents standard deviation, and d represents a positive parameter;

[0055] Step 2.3: Meta-train and meta-test the dataset; the meta-training stage is to calculate task-specific parameters on the support set (extract key feature fragments (such as time series data of internal resistance mutation and significant fluctuation of polarization voltage) from the meta-weight matrix) to generate task common features; the meta-testing stage is to evaluate the meta-weight matrix W meta The generalization performance and back-propagation meta-weight matrix W meta The gradient direction of

[0056] When supporting set meta-training, the attention mechanism is introduced to dynamically weight key features (such as internal resistance mutation at low temperature and polarization effect at high SOC) to generate W that adapts to new battery scenarios. metaNEW ;

[0057] Based on the combination of battery type, aging degree, and temperature range, the task urgency score is calculated and the task priority ranking formula is as follows:

[0058] Score = w1·temperature deviation + w2·aging rate + w3·SOC gradient;

[0059] Among them, w1, w2, and w3 represent dynamic weights, which are the initial weights for charging speed, temperature control, and battery life protection, respectively; the temperature deviation is the absolute difference between the current battery temperature and the optimal operating temperature; the aging rate is the percentage of capacity decay per unit cycle number; and the SOC gradient is the absolute value of the difference between the current SOC and the target SOC, determined by the failure rate statistics in the historical cycle records.

[0060] Step 2.3: Generate training and test sets and output them.

[0061] Furthermore, the residual jump module described in step 4 directly transmits the original SOC target value to the deep nodes of the network through cross-layer connections to avoid gradient disappearance; the output of the residual branch is weighted and summed with the output of the gated network to generate a dynamic weight W t :

[0062] W t =Sigmoid(W meta [SOC t ,SOH,T cell ])+δ·Residual(SOC t );

[0063] Where δ is a learnable scaling factor, W meta is the meta-weight matrix with the dimension [number of tasks × number of features]], which is initialized to obey the normal distribution with mean 0 and standard deviation 0.01; Residual (SOC t ) is the original SOC signal, and GRU is introduced to dynamically adjust the jump weight to suppress burst noise;

[0064] The TCN module uses dilated causal convolution to extract local temporal features (such as voltage mutations and short-term current fluctuations) to capture local variation patterns of battery parameters. The Transformer module uses a self-attention mechanism to establish global dependencies (such as the long-term correlation between SOC and temperature, and the impact of aging on polarization voltage). To address the vanishing gradient problem in deep networks, a cross-layer residual jump structure is designed: the output of each layer of the network is weightedly fused with the original input through a learnable parameter z, expressed as:

[0065] h out =TCN(h ist )+z·h ist

[0066] In the formula, h out The mixed output result is a weighted combination of the features processed by the temporal convolutional network TCN and the original historical data; TCN (h ist ) is the temporal convolutional network for historical data h ist The time series feature extraction results of h ist is the input historical data feature vector, which usually contains the time series observation values ​​of voltage, current, and temperature; z is the weighting coefficient, which is a learning parameter used to balance the ratio of new features extracted by TCN and the retention of original historical data;

[0067] Introduce GRU to dynamically adjust jump weights and suppress sudden noise:

[0068] g t =σ(W g ·[hist ,h out ]),

[0069] h filtered =g t ⊙h out +(1-g t )·h ist

[0070] g t is the gate signal, σ is the Sigmoid function, W g is the gating weight matrix, h filtered The final output after gated filtering is the dynamic fusion of TCN processing results and original historical data;

[0071] Predict charging trends: Input capacity attenuation data ΔQ loss , using LSTM network to predict the future capacity attenuation curve Q future (t); Define pseudo health indicators:

[0072]

[0073] γ is the material degradation coefficient, which is 0.12 for NCA and 0.08 for LFP. This indicator comprehensively reflects the available capacity and potential aging risk of the battery. e is the base, which is a number with a value of approximately 2.71828.

[0074] Use the training set to test the model. Use the cleaned and aligned test set to evaluate model performance and calculate key indicators: charging time prediction error, temperature rise control error, and capacity decay fit. If the test error exceeds the threshold, return to step 4 to reconstruct the residual module parameters or adjust the GRU gating weights.

[0075] Beneficial effects

[0076] The multi-objective optimization method for battery fast charging strategy based on meta-learning weights proposed in this invention has the following advantages compared with the existing technology:

[0077] (1) The present invention initializes and dynamically adapts weights through meta-learning, so that the model can quickly adapt to task scenarios with different battery types, aging levels, and temperature ranges. The backbone network adopts the Transformer or TCN architecture, and a residual module is added to the output end of each network layer to directly pass the original SOC target value to the deep nodes of the network to avoid gradient disappearance, so that the model can better focus on the incremental optimization of polarization voltage reduction and temperature rise control. A multi-objective optimization problem is defined, and the three conflicting objectives of charging time, temperature rise, and capacity attenuation are converted into a weighted single objective function. The improved NSGA-II algorithm is combined to solve the Pareto frontier, generate a non-dominated solution set that takes into account both efficiency and safety, and form the optimal charging curve.

[0078] (2) The present invention is based on a dynamic adaptation mechanism of meta-learning weights, which can quickly respond to complex working conditions such as different battery types, aging degrees, and temperature ranges. In addition, the key features of the dynamic weighting of the attention mechanism are introduced, such as accurately capturing internal resistance mutations at low temperatures and focusing on polarization effects at high SOC, to ensure that the charging strategy can be flexibly adjusted with environmental changes. The original charging demand signal is directly transmitted to the deep nodes of the network through the residual jump term, and the weights are dynamically adjusted in combination with the gated network, which effectively suppresses the gradient vanishing problem under sudden working conditions, enables the model to focus on the incremental optimization of polarization voltage reduction and temperature rise control, and significantly improves the dynamic response capability and optimization accuracy of the charging process.

[0079] (3) The present invention transforms the three mutually constrained key indicators of charging time, temperature rise, and capacity decay into a weighted single objective function and combines it with the improved NSGA-II algorithm to solve the Pareto frontier, generating a non-dominated solution set that takes into account both efficiency and safety, and forming an optimal charging curve. This method can effectively control temperature rise and slow down capacity decay while ensuring that the charging time is as short as possible. For example, in emergency charging scenarios, a solution with a slightly longer charging time but extremely low temperature rise and ultra-small capacity decay can be provided; and in daily fast charging scenarios, a strategy with the shortest charging time and temperature rise and capacity decay within a reasonable range can be provided.

[0080] (4) The design of the residual jump module in this invention cleverly solves the problem of vanishing gradients in deep networks, enabling the model to accurately and in real time focus on incremental optimization of polarization voltage reduction and temperature rise control. The backbone network adopts a Transformer or TCN architecture, and a residual module is added to the output of each network layer to directly pass the original SOC target value to the deep nodes. At the same time, the residual branch output is weighted and summed with the gated network output to generate dynamic weights. GRU is introduced to dynamically adjust the jump weights to suppress burst noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a schematic diagram of the overall architecture of the present invention.

[0082] Figure 2 It is the algorithm flow chart of the present invention. DETAILED DESCRIPTION

[0083] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Under the premise of not departing from the design concept of the present invention, various modifications and improvements made by ordinary persons in this field to the technical solutions of the present invention should fall within the scope of protection of the present invention.

[0084] Example 1:

[0085] Combine Figures 1 to 2 As shown in FIG, a multi-objective optimization method for a battery fast charging strategy based on meta-learning weights includes the following steps:

[0086] Step 1: Constructing a data set: Real-time collection of battery status parameters and environmental parameters to construct a multi-dimensional input data set. The specific steps for constructing the data set include:

[0087] Step 1.1: Data collection: Experimental data is collected in real time through the charge and discharge steps. The experimental data includes battery state parameters including voltage, current, temperature, SOC, SOH, and internal resistance, and environmental parameters including temperature, humidity, and historical cycle records.

[0088] Step 1.2: Design operating condition coverage: Design an orthogonal experimental matrix covering different charge rates (0.5C-3C), temperature ranges (-10°C to 50°C), battery aging levels (SOH 100% to 60%), and simulate extreme operating conditions;

[0089] Step 1.3: Data cleaning and alignment: Eliminate data that exceeds the battery's nominal parameters, process outliers, and use cubic spline interpolation to unify the time base for asynchronous data to ensure time synchronization;

[0090] Step 1.4: Data enhancement: Use FGSM (Fast Gradient Sign Method, a classic adversarial sample generation algorithm that uses the gradient information of the neural network to quickly generate perturbation samples that can cause model misclassification) to eliminate power frequency interference and improve the robustness of the dataset.

[0091] Step 2: Build the MAML framework model and train and test the dataset: Based on the multi-dimensional input dataset constructed in Step 1, construct the framework structure and initial parameters for meta-learning. Define the task pool by battery type, aging level, and temperature range. The input features of each task include SOC, SOH, and temperature, and the output is the optimization target of the charging strategy. Then initialize the meta-weight matrix with a normal random distribution and a dimension of [number of tasks × number of features]. The specific operation methods include:

[0092] Step 2.1: Receive the input multi-dimensional data set through the input layer and standardize the data;

[0093] Step 2.2: Construct the meta-weight matrix: Construct a matrix of dimension N×d and initialize it. The initialization method is:

[0094]

[0095] N(0,σ 2) represents normal distribution, σ represents standard deviation, and d represents a positive parameter;

[0096] Step 2.3: Meta-train and meta-test the dataset; the meta-training stage is to calculate task-specific parameters on the support set (extract key feature fragments (such as time series data of internal resistance mutation and significant fluctuation of polarization voltage) from the meta-weight matrix) to generate task common features; the meta-testing stage is to evaluate the meta-weight matrix W meta The generalization performance and back-propagation meta-weight matrix W meta The gradient direction of

[0097] When supporting set meta-training, the attention mechanism is introduced to dynamically weight key features (such as internal resistance mutation at low temperature and polarization effect at high SOC) to generate W that adapts to new battery scenarios. metaNEW ;

[0098] Based on the combination of battery type, aging degree, and temperature range, the task urgency score is calculated and the task priority ranking formula is as follows:

[0099] Score = w1·temperature deviation + w2·aging rate + w3·SOC gradient;

[0100] Among them, w1, w2, and w3 represent dynamic weights, which are the initial weights for charging speed, temperature control, and battery life protection, respectively; the temperature deviation is the absolute difference between the current battery temperature and the optimal operating temperature; the aging rate is the percentage of capacity decay per unit cycle number; and the SOC gradient is the absolute value of the difference between the current SOC and the target SOC, determined by the failure rate statistics in the historical cycle records.

[0101] Step 2.3: Generate training and test sets and output them.

[0102] Step 3: Construct a battery charging model: Based on the electrochemical reaction mechanism and the principle of energy conservation, combined with voltage, current, and temperature parameters, differential equations and data-driven methods are used to describe the dynamic characteristics of charging and discharging, and an electrochemical-thermal coupling model is established.

[0103] Step 4: Design a residual jump module: Add a residual jump module to the battery charging model constructed in Step 3. This solves the training bottleneck of deep neural networks through gradient stability, feature retention, and incremental learning mechanisms. The battery charging model includes a backbone network, which adopts a Transformer or TCN architecture and takes the historical charging sequence as input. A residual module is added to the output of each network layer to force the model to focus on incremental optimization of polarization voltage reduction and temperature rise control.

[0104] The residual jump module passes the original SOC target value directly to the deep nodes of the network through cross-layer connections to avoid gradient disappearance; the output of the residual branch is weighted and summed with the output of the gated network to generate a dynamic weight W t :

[0105] W t =Sigmoid(W meta [SOC t ,SOH,T cell ])+δ·Residual(SOC t );

[0106] Where δ is a learnable scaling factor, W meta is the meta-weight matrix with the dimension [number of tasks × number of features]], which is initialized to obey the normal distribution with mean 0 and standard deviation 0.01; Residual (SOC t ) is the original SOC signal, and GRU is introduced to dynamically adjust the jump weight to suppress burst noise;

[0107] The TCN module uses dilated causal convolution to extract local temporal features (such as voltage mutations and short-term current fluctuations) to capture local variation patterns of battery parameters. The Transformer module uses a self-attention mechanism to establish global dependencies (such as the long-term correlation between SOC and temperature, and the impact of aging on polarization voltage). To address the vanishing gradient problem in deep networks, a cross-layer residual jump structure is designed: the output of each layer of the network is weightedly fused with the original input through a learnable parameter z, expressed as:

[0108] h out =TCN(h ist )+z·h ist

[0109] In the formula, h out The mixed output result is a weighted combination of the features processed by the temporal convolutional network TCN and the original historical data; TCN (h ist ) is the temporal convolutional network for historical data h ist The time series feature extraction results of h ist is the input historical data feature vector, which usually contains the time series observation values ​​of voltage, current, and temperature; z is the weighting coefficient, which is a learning parameter used to balance the ratio of new features extracted by TCN and the retention of original historical data;

[0110] Introduce GRU to dynamically adjust jump weights and suppress sudden noise:

[0111] g t =σ(W g ·[h ist ,h out ]),

[0112] h filtered =g t ⊙h out +(1-g t )·h ist

[0113] g t is the gate signal, σ is the Sigmoid function, W g is the gating weight matrix, h filtered The final output after gated filtering is the dynamic fusion of TCN processing results and original historical data;

[0114] Predict charging trends: Input capacity attenuation data ΔQ loss , using LSTM network to predict the future capacity attenuation curve Q future (t); Define pseudo health indicators:

[0115]

[0116] γ is the material degradation coefficient, which is 0.12 for NCA and 0.08 for LFP. This indicator comprehensively reflects the available capacity and potential aging risk of the battery. e is the base, which is a number with a value of approximately 2.71828.

[0117] Use the training set to test the model. Use the cleaned and aligned test set to evaluate model performance and calculate key indicators: charging time prediction error, temperature rise control error, and capacity decay fit. If the test error exceeds the threshold, return to step 4 to reconstruct the residual module parameters or adjust the GRU gating weights.

[0118] Step 5: Multi-objective optimization: Input the improved NSGA-II algorithm into the trained model, define the multi-objective optimization problem, and achieve the Pareto optimal solution. The specific operation is as follows:

[0119] Step 5.1: Define the objective function: Charging time refers to the total time from the start of charging to the cutoff voltage, which is quantified by the following method; temperature rise refers to the difference between the battery surface temperature and the ambient temperature; capacity decay refers to the irreversible capacity loss of the battery due to side reactions during charging, which is quantified by the following model:

[0120]

[0121] f2=min(max(T cell )-T env )

[0122]

[0123] Among them, f1 represents minimizing charging time, f2 represents minimizing temperature rise, and f3 represents minimizing capacity attenuation; t charge is the charging time, T is the total number of charging stages, Δt is the duration of each stage, T env is the ambient temperature sensor, T cell Built-in temperature sensor for the battery, I rms is the RMS value of the current, V over is the overvoltage peak, ΔQ loss is the capacity loss, α is the current loss coefficient, β is the overvoltage loss coefficient, To find the integral;

[0124] Design dynamic constraint conditions, voltage safety margin, expression is:

[0125] V min +0.1·SOH<V cutoff <V max -0.05·(100-SOH);

[0126] In the above formula, V min 、V max The minimum / maximum allowable voltage specified by the battery manufacturer, SOH is the battery health status (0-100%);

[0127] Current mutation rate: |dI / dt|<0.2C / s, to prevent lithium deposition;

[0128] Step 5.2: Initialize the population: Each individual is encoded as a three-dimensional vector [I max ,V end ,t charge ], the population size is adaptively adjusted according to the solution space dimension (base value 200);

[0129] x=[I max ,V end ,t charge ];

[0130] Among them, I max is the maximum charging current, V end is the cut-off voltage, t charge is the charging time;

[0131] Step 5.3: Compare the individuals in the population pairwise, define the dominance relationship, use the ε-dominance improvement strategy, set a loose threshold (±1.5°C) for the temperature rise target to avoid premature loss of feasible solutions;

[0132] Step 5.4: Calculate the crowding degree and calculate the crowding distance of each individual on the objective function to measure the density of its neighboring solutions. Perform weighted processing on minimizing charging time, minimizing temperature rise, and minimizing capacity decay, and introduce weight coefficients w1, w2, and w3.

[0133]

[0134] d i The crowding distance of the individual on the objective function, f i (x j ),f i (x j+1 ) is the evaluation value of the jth individual and the j+1th individual on the i-th objective function, max(f i ) and min(f i ) are the maximum and minimum values ​​of all individuals under the i-th objective function, n is the total number of individuals in the population (the number of solutions), is the sum of the differences in the evaluation values ​​of all adjacent individuals;

[0135] Dynamically adjust the target weight based on the battery health status:

[0136] W crowd =Softmax([w1,w2+δ·(1-SOH),w3]);

[0137] W crowd is the final generated dynamic weight vector, δ is the aging sensitivity coefficient, and when SOH < 80%, the capacity attenuation weight w3 increases by 20%;

[0138] Step 5.5: Introduce the crowding distance correction to avoid the temperature rise target dominating the solution set distribution in high temperature scenarios. The correction coefficient is expressed as:

[0139]

[0140] Among them, T sensitivity is the temperature sensitivity coefficient, which is 1.2 in high temperature scenario and 1.0 in normal temperature scenario; d i new is the crowding distance of the corrected individual on the objective function;

[0141] Step 5.6: Crossover and mutation: simulate binary crossover, cross the parent individuals to generate offspring, and control the crossover probability P c and distribution index η c ;

[0142]

[0143] x1, x2 are input variables, m is the result of dynamic calculation, which is used for optimizing weights and resource allocation ratios in subsequent operations; η c is the exponential adjustment coefficient, usually a positive real number greater than 1;

[0144] Mutate the offspring and introduce mutation probability P m and distribution index η m :

[0145]

[0146] In the above formula, Δx represents the parameter correction value, r is the correction factor, and x ub is the preset target value, x ib For the polynomial mutation of the benchmark value, a temperature sensitivity coefficient is introduced, and the mutation probability is increased to 0.25 at high temperature to enhance the local search P m ;

[0147] Step 5.7: Filter user preferences and obtain user preference value λ∈[0,1] through past behavior analysis 3 , generate personalized weighted targets:

[0148] J user =λ1f1+λ2f2+λ3f3;

[0149] In the above formula, J user is the user's comprehensive utility function;

[0150] Combining the TOPSIS multi-criteria decision-making method, the strategy parameters closest to the ideal solution are selected. The polarization voltage prediction value is updated every 5% SOC interval. When the actual temperature rise deviates from the predicted value by more than 15%, online re-optimization is triggered. The capacity decay model parameters are updated after each charge, and the meta-weight matrix is ​​reconstructed every 50 cycles.

[0151] Step 5.8: Termination and output: When the maximum number of iterations is reached or the solution set converges, output the Pareto frontier Q_pareto, that is, the curve that minimizes charging time, minimizes temperature rise, and minimizes capacity decay. Otherwise, return to step 5.6.

[0152] Case Study:

[0153] Case Background:

[0154] A new energy vehicle company plans to improve the intelligence of its fast-charging stations. Battery types vary: its models are equipped with NCA (nickel-cobalt-aluminum) and LFP (lithium iron phosphate) batteries, which exhibit significant differences in charging characteristics. When the battery state of health (SOH) of operating vehicles generally drops to 80%, charging efficiency decreases by 20%, and capacity decay accelerates. In high-temperature areas (35°C+), the temperature rise during charging exceeds the standard by 40%, triggering safety current limiting and extending charging time by 50%. Taxi drivers demand extremely fast charging (charging to 80% in 30 minutes), while private car owners are more concerned about battery life (capacity decay rate <1% / year).

[0155] Method validation:

[0156] 1) Scenario Settings: The test subjects are NCA batteries (SOH 100% & 80%) and LFP batteries (SOH 100% & 80%). The environment settings are high temperature (35°C) and room temperature (25°C). The comparison benchmarks are traditional constant current constant voltage (CC-CV) charging and single-target optimization (only minimizing charging time).

[0157] 2) Implementation process: Collect 10,000 sets of charging cycle data and construct a multidimensional dataset containing voltage, current, temperature, SOC, and SOH.

[0158] Define the task pool:

[0159] Task 1: Fast charging optimization of NCA battery (SOH 100%) at 35°C;

[0160] Task 2: Life-priority charging of LFP batteries (SOH 80%) at 25°C;

[0161] Task 3: Safe charging of NCA batteries (SOH 80%) under dynamic temperature rise scenarios.

[0162] Initialize the meta-weight matrix W meta (Dimension: 3 tasks × 6 features), extract common features through support set training:

[0163] NCA battery: polarization voltage mutation characteristics at high SOC (weight +0.3).

[0164] LFP battery: low-temperature internal resistance gradient characteristics (weight +0.2).

[0165] Aging batteries: Strong correlation between capacity decay and RMS current (weight +0.4).

[0166] 3) Dynamic Optimization of the Residual Module: A TCN-Transformer hybrid backbone network is used, with a residual jump coefficient α of 0.15. For sudden temperature rise scenarios (e.g., a 5°C rise in ambient temperature during charging), the GRU gating is dynamically adjusted from 0.6 to 0.3, suppressing noise while preserving the original SOC signal. This reduces the polarization voltage prediction error from 12% to 4%.

[0167] 4) Multi-objective optimization: Improved NSGA-II parameter settings: population size 200, crossover probability 0.8, mutation probability 0.25 at high temperature, and dynamic adjustment of objective weights:

[0168] Taxi scenario: λ = [0.7, 0.2, 0.1] (time priority)

[0169] Private car scenario: λ = [0.3, 0.4, 0.3] (balanced mode)

[0170] 5) Results comparison:

[0171]

[0172]

[0173] in conclusion:

[0174] Multi-scenario adaptability: Safe and efficient strategies are generated for NCA / LFP, new and old batteries, and within a temperature difference of 15°C. Comparative tests of NCA / LFP battery packs show that in aged battery packs with an SOH of 80%, the charging temperature rise under high temperature conditions (35°C) dropped from 18°C ​​in the traditional method to 12°C (a decrease of 33.3%), and the temperature rise exceeding the standard rate dropped from 40% to 5%. The capacity attenuation rate of LFP batteries at 25°C dropped from 0.015% / cycle to 0.008% / cycle (a decrease of 46.7%). The new method significantly improves the fast charging efficiency of NCA batteries, shortening the charging time by 37.8% (45min→28min).

[0175] Target Balancing: Charging time is shortened by 35%, while temperature rise and capacity decay are reduced by 33% and 47%, respectively. In the taxi ultra-fast charging scenario, charging to 80% SOC is achieved in 28 minutes (37.8% faster than traditional methods), while keeping capacity decay within 0.01% / cycle. In the private car lifespan priority scenario, charging time is 38 minutes (18.8% longer than single-target optimization), but capacity decay is reduced by 68% (0.025% → 0.008%).

[0176] Dynamic response mechanism: Through online meta-weight updates and residual gating, the adjustment delay for sudden operating conditions is less than 10 seconds. In the sudden temperature rise scenario (ambient temperature suddenly rises by 5°C), the GRU gating mechanism reduces the polarization voltage prediction error by 66.7% (12% → 4%).

[0177] Commercial Value: This technology is expected to increase fast-charging station utilization by 20%, extending battery life to 8 years / 2,000 cycles, meeting automakers' TCO (total cost of ownership) optimization needs. Based on two fast-charging cycles per day, battery life will increase from the traditional 5 years / 1,000 cycles to 8 years / 2,000 cycles. Fast-charging station utilization is also expected to increase by 20% (28 minutes / charge vs. the traditional 45 minutes / charge), resulting in an estimated annual revenue increase of ¥15,600 per charging station (based on a 0.3 / kWh charge).

[0178] The above is a detailed introduction to the deep learning-based long-term prediction and dynamic intervention system for Alzheimer's disease provided by the examples of the present invention. The description of the above embodiments is only intended to help understand the method and core concept of the present invention; at the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concepts of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-objective optimization method for battery fast charging strategy based on meta-learning weights, characterized in that: Including steps: Step 1: Build a data set: Collect battery status parameters and environmental parameters in real time to build a multi-dimensional input data set; Step 2: Build the MAML framework model and train and test the dataset: Based on the multi-dimensional input dataset constructed in Step 1, construct the framework structure and initial parameters for meta-learning. Define the task pool by battery type, aging level, and temperature range. The input features of each task include SOC, SOH, and temperature, and the output is the optimization target of the charging strategy. Then, initialize the meta-weight matrix with a normal random distribution and a dimension of [number of tasks × number of features]. Step 3: Build a battery charging model: Based on the electrochemical reaction mechanism and the principle of energy conservation, combined with voltage, current, and temperature parameters, differential equations and data-driven methods are used to describe the dynamic characteristics of charging and discharging, and an electrochemical-thermal coupling model is established; Step 4: Design a residual jump module: Add a residual jump module to the battery charging model constructed in Step 3. The battery charging model includes a backbone network using a Transformer or TCN architecture, and its input is the historical charging sequence. Adding a residual module to the output of each network layer forces the model to focus on incremental optimization of polarization voltage reduction and temperature rise control. Step 5: Multi-objective optimization: Input the improved NSGA-II algorithm into the trained model, define the multi-objective optimization problem, and achieve the Pareto optimal solution. The specific operation is as follows: Step 5.1: Define the objective function: Charging time refers to the total time from the start of charging to the cutoff voltage, which is quantified by the following method; temperature rise refers to the difference between the battery surface temperature and the ambient temperature; capacity decay refers to the irreversible capacity loss of the battery due to side reactions during charging, which is quantified by the following model: f2=min(max(T cell )-T env ) Among them, f1 represents minimizing charging time, f2 represents minimizing temperature rise, and f3 represents minimizing capacity attenuation; t charge is the charging time, T is the total number of charging stages, Δt is the duration of each stage, T env is the ambient temperature sensor, T cell Built-in temperature sensor for the battery, I rms is the RMS value of the current, V over is the overvoltage peak, ΔQ loss is the capacity loss, α is the current loss coefficient, β is the overvoltage loss coefficient, To find the integral; Design dynamic constraints, voltage safety margin, expression is: In min +0.1·SOH<V cutoff <In max -0.05·(100-SOH); In the above formula, V min 、V max The minimum / maximum allowable voltage specified by the battery manufacturer, SOH is the battery health status (0-100%); Current mutation rate: |dI / dt|<0.2C / s, to prevent lithium deposition; Step 5.2: Initialize the population: Each individual is encoded as a three-dimensional vector [I max ,V end ,t charge ], the population size is adaptively adjusted according to the solution space dimension (base value 200); x=[I max ,V end ,t charge ]; Among them, I max is the maximum charging current, V end is the cut-off voltage, t charge is the charging time; Step 5.3: Compare the individuals in the population pairwise, define the dominance relationship, use the ε-dominance improvement strategy, set a loose threshold for the temperature rise target, and avoid premature loss of feasible solutions; Step 5.4: Calculate the crowding degree and calculate the crowding distance of each individual on the objective function to measure the density of its neighboring solutions. Perform weighted processing on minimizing charging time, minimizing temperature rise, and minimizing capacity decay, and introduce weight coefficients w1, w2, and w3. d i The crowding distance of the individual on the objective function, f i (x j ),f i (x j+1 ) is the evaluation value of the jth individual and the j+1th individual on the i-th objective function, max(f i ) and min(f i ) are the maximum and minimum values ​​of all individuals under the i-th objective function, n is the total number of individuals in the population (the number of solutions), is the sum of the differences in the evaluation values ​​of all adjacent individuals; Dynamically adjust the target weight based on the battery health status: W crowd =Softmax([w1,w2+δ·(1-SOH),w3]); W crowd is the final generated dynamic weight vector, δ is the aging sensitivity coefficient, and when SOH < 80%, the capacity attenuation weight w3 increases by 20%; Step 5.5: Introduce the crowding distance correction to avoid the temperature rise target dominating the solution set distribution in high temperature scenarios. The correction coefficient is expressed as: Among them, T sensitivity is the temperature sensitivity coefficient, which is 1.2 in high temperature scenario and 1.0 in normal temperature scenario; d i new is the crowding distance of the corrected individual on the objective function; Step 5.6: Crossover and mutation: simulate binary crossover, cross the parent individuals to generate offspring, and control the crossover probability P c and distribution index η c ; x1, x2 are input variables, m is the result of dynamic calculation, which is used for optimizing weights and resource allocation ratios in subsequent operations; η c is the exponential adjustment coefficient, usually a positive real number greater than 1; Mutate the offspring and introduce mutation probability P m and distribution index η m : In the above formula, Δx represents the parameter correction value, r is the correction factor, and x ub is the preset target value, x ib For the polynomial mutation of the benchmark value, a temperature sensitivity coefficient is introduced, and the mutation probability is increased to 0.25 at high temperature to enhance the local search P m ; Step 5.7: Filter user preferences and obtain user preference value λ∈[0,1] through past behavior analysis 3 , generate personalized weighted targets: J user =λ1f1+λ2f2+λ3f3; In the above formula, J user is the user's comprehensive utility function; Combining the TOPSIS multi-criteria decision-making method, the strategy parameters closest to the ideal solution are selected. The polarization voltage prediction value is updated every 5% SOC interval. When the actual temperature rise deviates from the predicted value by more than 15%, online re-optimization is triggered. The capacity decay model parameters are updated after each charge, and the meta-weight matrix is ​​reconstructed every 50 cycles. Step 5.8: Termination and output: When the maximum number of iterations is reached or the solution set converges, output the Pareto frontier Q_pareto, that is, the curve that minimizes charging time, minimizes temperature rise, and minimizes capacity decay. Otherwise, return to step 6.

2. The multi-objective optimization method for battery fast charging strategy based on meta-learning weights according to claim 1, characterized in that: The specific steps for constructing the dataset in step 1 include: Step 1.1: Data collection: Experimental data is collected in real time through the charge and discharge steps. The experimental data includes battery state parameters including voltage, current, temperature, SOC, SOH, and internal resistance, and environmental parameters including temperature, humidity, and historical cycle records. Step 1.2: Design operating condition coverage: Design an orthogonal experimental matrix to cover different charging rates, temperature ranges, and battery aging levels, and simulate extreme operating conditions; Step 1.3: Data cleaning and alignment: Eliminate data that exceeds the battery's nominal parameters, process outliers, and use cubic spline interpolation to unify the time base for asynchronous data to ensure time synchronization; Step 1.4: Data enhancement: FGSM is used to eliminate power frequency interference and improve the robustness of the dataset.

3. The multi-objective optimization method for battery fast charging strategy based on meta-learning weights according to claim 1, characterized in that: The specific operation method of step 2 includes: Step 2.1: Receive the input multi-dimensional data set through the input layer and standardize the data; Step 2.2: Construct the meta-weight matrix: Construct a matrix of dimension N×d and initialize it. The initialization method is: N(0,σ 2 ) represents normal distribution, σ represents standard deviation, and d represents a positive parameter; Step 2.3: Meta-train and meta-test the dataset; the meta-training phase is to calculate task-specific parameters on the support set and generate task-common features; the meta-test phase is to evaluate the meta-weight matrix W meta The generalization performance and back-propagation meta-weight matrix W meta The gradient direction of When training on the support set, the attention mechanism is introduced to dynamically weight key features to generate W that adapts to new battery scenarios. metaNEW ; Based on the combination of battery type, aging degree, and temperature range, the task urgency score is calculated and the task priority ranking formula is as follows: Score = w1·temperature deviation + w2·aging rate + w3·SOC gradient; Among them, w1, w2, and w3 represent dynamic weights, which are the initial weights for charging speed, temperature control, and battery life protection, respectively; the temperature deviation is the absolute difference between the current battery temperature and the optimal operating temperature; the aging rate is the percentage of capacity decay per unit cycle number; and the SOC gradient is the absolute value of the difference between the current SOC and the target SOC, determined by the failure rate statistics in the historical cycle records. Step 2.3: Generate training and test sets and output them.

4. The multi-objective optimization method for battery fast charging strategy based on meta-learning weights according to claim 1, characterized in that: The residual jump module described in step 4 directly transfers the original SOC target value to the deep nodes of the network through cross-layer connections to avoid gradient disappearance; the output of the residual branch is weighted and summed with the output of the gated network to generate a dynamic weight W t : W t =Sigmoid(W meta [SOC t ,SOH,T cell ])+δ·Residual(SOC t ); Where δ is a learnable scaling factor, W meta is the meta-weight matrix with the dimension [number of tasks × number of features]], which is initialized to obey the normal distribution with mean 0 and standard deviation 0.01; Residual (SOC t ) is the original SOC signal, and GRU is introduced to dynamically adjust the jump weight to suppress burst noise; The TCN module uses dilated causal convolution to extract local temporal features and capture the local change pattern of battery parameters. The Transformer module establishes global dependencies through the self-attention mechanism. To solve the problem of vanishing gradients in deep networks, a cross-layer residual jump structure is designed: the output of each layer of the network is weightedly fused with the original input through a learnable parameter z, which is expressed as: h out =BC(h ist )+z·h ist In the formula, h out The mixed output result is a weighted combination of the features processed by the temporal convolutional network TCN and the original historical data; TCN (h ist ) is the temporal convolutional network for historical data h ist The time series feature extraction results of h ist is the input historical data feature vector, which usually contains the time series observation values ​​of voltage, current, and temperature; z is the weighting coefficient, which is a learning parameter used to balance the ratio of new features extracted by TCN and the retention of original historical data; Introduce GRU to dynamically adjust jump weights and suppress sudden noise: g t =σ(W g ·[h ist ,h out ]), h filtered =g t ⊙h out +(1-g t )·h ist g t is the gate signal, σ is the Sigmoid function, W g is the gating weight matrix, h filtered The final output after gated filtering is the dynamic fusion of TCN processing results and original historical data; Predict charging trends: Input capacity attenuation data ΔQ loss , using LSTM network to predict the future capacity attenuation curve Q future (t); Define pseudo health indicators: γ is the material degradation coefficient, which is 0.12 for NCA and 0.08 for LFP. This indicator comprehensively reflects the available capacity and potential aging risk of the battery. e is the base, which is a number with a value of approximately 2.71828. Use the training set to test the model. Use the cleaned and aligned test set to evaluate model performance and calculate key indicators: charging time prediction error, temperature rise control error, and capacity decay fit. If the test error exceeds the threshold, return to step 4 to reconstruct the residual module parameters or adjust the GRU gating weights.

Images