Power control method and device for industrial forklift charger

Abstract

The invention discloses a power control method and device for an industrial forklift charger, and the method comprises the steps: obtaining industrial forklift charging demand data and LLC resonant topology characteristic parameters, and constructing an LLC resonant converter; on the basis of the working frequency range and the power density requirement, through a probability learning algorithm and a random agent model, a random agent model for multi-objective optimization of magnetic element parameters is formed, and a high-frequency low-loss magnetic element is obtained; a high-frequency low-loss magnetic element is integrated into an LLC resonant converter, a random neural control barrier function algorithm is realized, and dynamic power control is performed on a charging process; based on the real-time data, a data-driven hybrid prediction control strategy is realized, and a charging parameter optimization instruction is generated; and in combination with the battery charge and discharge characteristic data and the battery aging state evaluation model, battery state evaluation and charging parameter adaptive adjustment are realized. The problems that a traditional charger is insufficient in control precision, low in efficiency, short in battery life and the like in a complex industrial environment are solved.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to a power control method and device for an industrial forklift charger, applicable to a high-efficiency intelligent charging system for industrial electric forklifts. Background Technology

[0002] As a key component in the energy replenishment of electric industrial forklifts, the performance of industrial forklift charging systems directly impacts the forklift's working efficiency, battery life, and overall operating costs. With increasing industrial automation and increasingly stringent environmental requirements, efficient and intelligent charging systems have become an important development direction for the industry.

[0003] Currently, the common industrial forklift charging technologies on the market mainly fall into two categories: traditional linear charging and switching power supply charging. Traditional linear charging technology has a simple structure, but suffers from problems such as large size, low efficiency, and severe heat generation; while switching power supply charging technology has improved efficiency, it still has significant shortcomings in terms of electromagnetic compatibility, charging curve optimization, and adaptability to different battery types.

[0004] More advanced forklift charging technologies employ a soft-switching topology combined with digital control algorithms, using PWM modulation to achieve power conversion and equipped with basic battery type identification functionality. These systems typically use high-frequency transformers and ordinary winding materials, possessing a certain power density and moderate conversion efficiency, sufficient to meet basic charging needs.

[0005] However, existing technologies still face several significant problems: First, the magnetic components suffer significant losses under high-frequency operating conditions, making it difficult to break through the bottleneck in overall efficiency; second, there is a lack of adaptive charging strategies for battery aging, making it impossible to optimize charging parameters based on the actual state of the battery; and third, in terms of power control algorithms, traditional deterministic models are difficult to cope with complex and ever-changing industrial environments and different battery characteristics, especially in the presence of interference, resulting in insufficient control accuracy and robustness. Summary of the Invention

[0006] The purpose of this invention is to provide a power control method and device for an industrial forklift charger, aiming to solve the technical problems existing in the prior art, such as high high-frequency loss of magnetic components, lack of battery aging adaptive strategy, and insufficient robustness of power control algorithm.

[0007] To achieve the above objectives, the present invention provides a power control method for an industrial forklift charger, comprising the following steps:

[0008] Acquire industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including a primary half-bridge circuit, a resonant network and a secondary rectifier circuit.

[0009] Based on the operating frequency range and power density requirements of the LLC resonant converter, a stochastic surrogate model for multi-objective optimization of magnetic component parameters is formed by constructing a probabilistic learning algorithm and a stochastic surrogate model. The stochastic surrogate model for multi-objective optimization of magnetic component parameters is then optimized in terms of core size, air gap length and winding structure to obtain a high-frequency, low-loss magnetic component.

[0010] The high-frequency, low-loss magnetic components are integrated into the LLC resonant converter. Based on the physical characteristics and operating data of the LLC resonant converter, a stochastic neural network structure is designed and a control barrier function is constructed to realize the stochastic neural control barrier function algorithm, which performs dynamic power control during the charging process and generates a dynamic power control strategy.

[0011] Based on the real-time data of voltage, current and temperature during the charging process under the execution of the dynamic power control strategy, the real-time data is processed by noise-resistant data filtering algorithm and prediction model construction to realize a data-driven hybrid predictive control strategy and generate charging parameter optimization instructions.

[0012] Based on the charging parameter optimization instructions, and combined with battery charging and discharging characteristic data and battery aging state assessment model, a charging curve optimization algorithm and parameter mapping processing are performed to achieve battery state assessment and adaptive adjustment of charging parameters.

[0013] Preferably, the step of acquiring industrial forklift charging demand data and LLC resonant topology characteristic parameters, performing resonant parameter optimization calculations and frequency modulation control design on the industrial forklift charging demand data and the LLC resonant topology characteristic parameters, and constructing an LLC resonant converter including a primary half-bridge circuit, a resonant network, and a secondary rectifier circuit includes:

[0014] Circuit topology analysis and component selection were performed on the industrial forklift charging demand data to design an LLC converter topology.

[0015] Based on the LLC converter topology and the LLC resonant topology characteristic parameters, the parameters of resonant inductance, resonant capacitance and magnetizing inductance are determined by resonant frequency calculation and quality factor optimization.

[0016] By selecting control strategies and designing circuit parameters based on the AC input characteristics, a PFC preamplifier circuit with a power factor greater than 0.99 is achieved, resulting in a stable DC input for the PFC preamplifier circuit.

[0017] Based on the parameters of the resonant inductor, resonant capacitor, and magnetizing inductor, and the stable DC input of the PFC pre-amplifier circuit, the operating frequency range is determined and the modulation strategy is designed to achieve zero-voltage switching and zero-current switching operation across the entire load range, thus obtaining the LLC resonant converter.

[0018] Preferably, the stochastic surrogate model for multi-objective optimization of magnetic component parameters, based on the operating frequency range and power density requirements of the LLC resonant converter, is constructed using a probabilistic learning algorithm and a stochastic surrogate model, including:

[0019] Based on the operating frequency range and power density requirements of the LLC resonant converter, the material performance of PC40 ferrite material was compared and the loss mechanism was analyzed to determine the basic material characteristics of the magnetic core.

[0020] Based on the aforementioned core material property data, a stochastic surrogate model capable of predicting core performance parameters is constructed through Gaussian process regression algorithm and kernel function parameter optimization.

[0021] The random surrogate model capable of predicting magnetic core performance parameters is subjected to cross-validation evaluation and generalization ability testing to obtain a random surrogate model for multi-objective optimization of the magnetic component parameters.

[0022] Preferably, the stochastic surrogate model for multi-objective optimization of the magnetic element parameters is used to optimize the core size, air gap length, and winding structure to obtain a high-frequency, low-loss magnetic element, comprising:

[0023] Based on the stochastic surrogate model and design constraints of the multi-objective optimization of the magnetic element parameters, the Pareto optimal solution search process is performed on the multi-objective Bayesian optimization algorithm to obtain the optimal parameter combination of core size, air gap length and winding structure.

[0024] Based on the core size and air gap length in the optimal parameter combination, and combined with the conductor structure design and winding process optimization of the high-frequency current distribution characteristics, a Litz wire winding structure with low AC resistance and low eddy current loss is realized, and the winding design parameters are obtained.

[0025] Based on the core geometry determined by the optimal parameter combination and the winding design parameters, loss analysis and calculation are performed on the core loss and copper loss to obtain the loss calculation results. Based on the loss calculation results, the thermal resistance network is analyzed and the cooling structure is designed to ensure that the temperature rise of the magnetic component under full load conditions is controlled within a safe range, thus obtaining the high-frequency low-loss magnetic component.

[0026] Preferably, the integration of the high-frequency, low-loss magnetic component into the LLC resonant converter, and the design of a stochastic neural network structure and the construction of a control barrier function based on the physical characteristics and operating data of the LLC resonant converter, to realize a stochastic neural control barrier function algorithm, include:

[0027] Based on the physical characteristics and operating data of the LLC resonant converter, a system dynamic model containing parameter uncertainties is established through system identification methods and model construction.

[0028] The power control requirements of the system dynamic model are processed by Bayesian neural network structure design and variational inference method to construct a stochastic neural network model that can handle uncertainty.

[0029] Based on the system safety constraints and operational boundary conditions determined by the system dynamic model, the barrier function is designed and the safety domain is defined to construct a control barrier function that ensures that the system state is always within the safety region.

[0030] The random neural network model and the control barrier function are subjected to control law design and parameter tuning to realize the random neural control barrier function algorithm.

[0031] Preferably, the real-time data of voltage, current, and temperature during the charging process based on the dynamic power control strategy are processed using an anti-noise data filtering algorithm and a prediction model to realize a data-driven hybrid predictive control strategy and generate charging parameter optimization instructions, including:

[0032] Based on a network of voltage, current, and temperature sensors, the signal is conditioned and the data is preprocessed to obtain real-time data of the charging process.

[0033] Statistical analysis and feature extraction are performed on the real-time data to establish a charging system noise model that includes electromagnetic interference noise, power fluctuation noise, and sensor noise.

[0034] Based on the charging system noise model, effective filtering of charging data is achieved through unscented Kalman filter design and adaptive adjustment of the covariance matrix.

[0035] The filtered historical data is processed by a bidirectional LSTM network structure and Monte Carlo Dropout technique to build a prediction model that can predict the evolution of charging state and quantify prediction uncertainty.

[0036] Based on the aforementioned prediction model, nonlinear model predictive control and stochastic model predictive control are combined to achieve a hybrid predictive control strategy that combines deterministic and stochastic control.

[0037] Based on the control decision of the hybrid predictive control strategy, the charging voltage setpoint, charging current setpoint, and switching frequency are optimized and calculated to generate the charging parameter optimization command.

[0038] Preferably, the step of designing a bidirectional LSTM network structure and processing the filtered historical data using Monte Carlo Dropout technology to construct a predictive model capable of predicting the evolution of the charging state and quantifying prediction uncertainty includes:

[0039] Based on the battery charge and discharge characteristic data, a second-order RC equivalent circuit physical model describing the battery dynamic characteristics is established through equivalent circuit parameter identification and model fitting.

[0040] Based on filtered historical data, bidirectional LSTM layer design and fully connected layer configuration are performed on the normalized multivariate time series to establish a neural network structure that captures forward and backward temporal dependencies, and generate a data-driven network architecture.

[0041] Based on the battery parameters of the second-order RC equivalent circuit physical model and the data-driven network architecture, the physical constraint embedding and neural network parameter optimization are fused to form a hybrid prediction model structure that combines physical constraints and data-driven approaches.

[0042] Based on the hybrid prediction model structure, the Monte Carlo Dropout technique is subjected to multiple forward propagation and probability distribution calculations to obtain the prediction model that can predict the evolution of charging state and quantify prediction uncertainty.

[0043] Preferably, the step of combining nonlinear model predictive control and stochastic model predictive control based on the predictive model to realize a hybrid predictive control strategy combining deterministic and stochastic control includes:

[0044] Based on a nonlinear model predictive control framework, optimization and constraint processing are performed on the prediction time domain to realize a deterministic controller for handling nominal operating conditions;

[0045] Based on the stochastic model predictive control framework, the probability constraints and expected values ​​are optimized to realize a stochastic controller that can handle uncertainties and disturbances.

[0046] The sigmoid function mapping and dynamic weight calculation are applied to the measurement of prediction uncertainty to achieve adaptive fusion of deterministic and stochastic controllers.

[0047] Based on the system's real-time requirements and the complexity of the control algorithm, the control task is divided into layers and the execution cycle is allocated, forming a layered execution strategy that includes a fast control layer, a predictive control layer, and an adaptive layer.

[0048] Based on the adaptive fusion and the hierarchical execution strategy, the control weights are dynamically adjusted and the control sequence generation process is optimized to obtain the hybrid predictive control strategy.

[0049] Preferably, the step of performing charging curve optimization algorithms and parameter mapping processing based on the charging parameter optimization instructions, combined with battery charge / discharge characteristic data and battery aging state assessment models, to achieve battery state assessment and adaptive adjustment of charging parameters includes:

[0050] Characteristic analysis and database construction processing were performed on test data of batteries of different types and aging levels to establish a battery characteristic database including lead-acid batteries and lithium batteries.

[0051] Based on the battery characteristic database, features are extracted and model construction is performed to design a battery aging state assessment model that can evaluate the battery health status.

[0052] Based on the evaluation results of the battery aging state assessment model and the charging parameter optimization instructions, multi-objective optimization and constraints are processed to design a charging curve optimization algorithm that can extend battery life.

[0053] The charging curve optimization algorithm is subjected to parameter mapping and dynamic adjustment to achieve real-time adaptive control of charging parameters and obtain the adjusted charging parameters.

[0054] Based on the adjusted charging parameters, comparative experiments and lifespan models are analyzed and processed to verify the system's effect on extending battery lifespan, and to obtain the battery state assessment and adaptive adjustment of charging parameters.

[0055] The present invention also provides a power control device for an industrial forklift charger, comprising:

[0056] LLC resonant converter power conversion module is used to acquire industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including primary half-bridge circuit, resonant network and secondary rectifier circuit.

[0057] The magnetic component optimization design module is used to construct a multi-objective optimization stochastic surrogate model for magnetic component parameters based on the operating frequency range and power density requirements of the LLC resonant converter through probabilistic learning algorithms and stochastic surrogate models. The module then optimizes the core size, air gap length, and winding structure of the stochastic surrogate model to obtain a high-frequency, low-loss magnetic component.

[0058] A stochastic neural control module is used to integrate the high-frequency, low-loss magnetic components into the LLC resonant converter. Based on the physical characteristics and operating data of the LLC resonant converter, a stochastic neural network structure is designed and a control barrier function is constructed to realize the stochastic neural control barrier function algorithm, which performs dynamic power control during the charging process and generates a dynamic power control strategy.

[0059] The data-driven predictive control module is used to process the real-time data of voltage, current and temperature during the charging process based on the dynamic power control strategy, perform noise-resistant data filtering algorithm and predictive model construction on the real-time data, realize the data-driven hybrid predictive control strategy, and generate charging parameter optimization instructions.

[0060] The battery adaptive charging module is used to perform charging curve optimization algorithms and parameter mapping processing based on the charging parameter optimization instructions, combined with battery charging and discharging characteristic data and battery aging state assessment models, so as to realize battery state assessment and adaptive adjustment of charging parameters.

[0061] The beneficial effects of this invention are:

[0062] 1. By combining a stochastic neural network with a safety control barrier function algorithm, a control algorithm that can ensure the safe operation of the system under conditions of parameter uncertainty and external disturbance is constructed, which solves the problem of insufficient robustness of traditional deterministic control algorithms in complex industrial environments;

[0063] 2. The optimal parameter design of magnetic components is based on a probabilistic learning-based stochastic surrogate model. By constructing a stochastic surrogate model of magnetic materials and component performance, multi-objective optimization design of high-frequency magnetic component parameters is achieved, effectively balancing multiple key indicators such as loss, volume, and temperature rise, and breaking through the limitations of traditional deterministic design methods.

[0064] 3. A noise-resistant hybrid method for data-driven predictive control is introduced, which combines adaptive filtering algorithm with hybrid predictive control strategy to achieve accurate prediction and optimized control of the charging process under noise interference conditions, thereby improving the control accuracy and stability of the system in complex industrial environments.

[0065] 4. An adaptive charging parameter adjustment technology based on battery aging status was developed. By establishing a battery aging assessment model and a charging curve optimization algorithm, intelligent identification and adaptive parameter adjustment of batteries of different types and aging levels were achieved, effectively extending battery life.

[0066] 5. By adopting high-efficiency soft-switching technology based on LLC resonant topology, and through resonant parameter optimization and frequency modulation control, zero-voltage switching and zero-current switching operations are achieved across the entire load range, significantly reducing switching losses and improving the overall system efficiency. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart of a power control method for an industrial forklift charger according to the present invention;

[0069] Figure 2 This is a flowchart illustrating the design of the noise-resistant hybrid method for data-driven predictive control according to the present invention.

[0070] Figure 3 This is a schematic diagram of the power control device for an industrial forklift charger according to the present invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0072] like Figure 1 As shown, the present invention provides a power control method for an industrial forklift charger, comprising the following steps:

[0073] Step S1: Obtain industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including a primary half-bridge circuit, a resonant network and a secondary rectifier circuit.

[0074] Step S2: Based on the operating frequency range and power density requirements of the LLC resonant converter, a random surrogate model for multi-objective optimization of magnetic component parameters is constructed through a probabilistic learning algorithm and a random surrogate model. The core size, air gap length and winding structure of the random surrogate model for multi-objective optimization of magnetic component parameters are then optimized to obtain a high-frequency, low-loss magnetic component.

[0075] Step S3: Integrate the high-frequency, low-loss magnetic components into the LLC resonant converter, and perform stochastic neural network structure design and control barrier function construction based on the physical characteristics and operating data of the LLC resonant converter to realize the stochastic neural control barrier function algorithm, perform dynamic power control on the charging process, and generate a dynamic power control strategy.

[0076] Step S4: Based on the real-time data of voltage, current and temperature during the charging process under the execution of the dynamic power control strategy, the real-time data is processed by noise-resistant data filtering algorithm and prediction model construction to realize the data-driven hybrid predictive control strategy and generate charging parameter optimization instructions;

[0077] Step S5: Based on the charging parameter optimization instructions, combine the battery charging and discharging characteristic data and the battery aging state assessment model to perform charging curve optimization algorithm and parameter mapping processing, so as to realize battery state assessment and adaptive adjustment of charging parameters.

[0078] In one embodiment, the step of acquiring industrial forklift charging demand data and LLC resonant topology characteristic parameters, performing resonant parameter optimization calculations and frequency modulation control design on the industrial forklift charging demand data and the LLC resonant topology characteristic parameters, and constructing an LLC resonant converter including a primary half-bridge circuit, a resonant network, and a secondary rectifier circuit includes:

[0079] Step S1.1: Perform circuit topology analysis and component selection processing on the industrial forklift charging demand data, and design the LLC converter topology.

[0080] The design of an industrial forklift charger first requires analyzing its specific power and voltage requirements. Based on a 10kW power rating and input / output voltage requirements, a half-bridge LLC resonant converter was selected as the primary topology after comparing the advantages and disadvantages of various topologies. Compared to traditional structures, this topology offers higher efficiency, lower electromagnetic interference, and better soft-switching characteristics, making it particularly suitable for industrial forklift charging applications. The half-bridge structure reduces the number of components and simplifies the drive circuitry compared to the full-bridge structure, while maintaining high efficiency over a medium power range.

[0081] In terms of component selection, a 650V rated silicon carbide MOSFET is chosen as the primary switching transistor, featuring low on-resistance and excellent high-frequency switching characteristics, enabling efficient operation at frequencies above 100kHz. The secondary rectification employs a combination of Schottky diodes and synchronous rectifier MOSFETs to reduce rectification losses. High-frequency ceramic capacitors are used in the resonant network, exhibiting low ESR characteristics and good temperature stability.

[0082] The LLC converter topology comprises three main parts: a primary half-bridge circuit consisting of two power MOSFETs and a voltage divider capacitor to convert DC voltage to AC voltage; a resonant network consisting of a resonant inductor, a resonant capacitor, and a magnetizing inductor to enable soft-switching operation and energy transfer; and a secondary rectifier circuit that rectifies the AC voltage to DC output, including rectifier components and an output filter capacitor. This topology design ensures high efficiency and good stability across the entire load range, meeting the stringent requirements of industrial forklift charging.

[0083] Step S1.2: Based on the LLC converter topology and the LLC resonant topology characteristic parameters, determine the parameters of resonant inductor, resonant capacitor and magnetizing inductor through resonant frequency calculation and quality factor optimization.

[0084] The performance of an LLC resonant converter largely depends on the selection of resonant parameters. Based on a 10kW power rating and voltage requirements, the resonant frequency range is first determined, with 100kHz chosen as the nominal resonant frequency to balance the efficiency and size of the magnetic components. The selection of the resonant frequency needs to consider factors such as switching losses, magnetic component losses, and size, avoiding situations where too low a frequency leads to increased magnetic component size or too high a frequency leads to increased switching losses.

[0085] The quality factor Q of a resonant network is a key parameter affecting dynamic response and efficiency. Through simulation analysis of the gain and phase characteristics under different Q values, a Q value of 3.5 was determined to achieve a good balance between load regulation capability and dynamic response. Higher Q values ​​provide better load regulation capability but reduce response speed, while lower Q values ​​improve response speed but reduce load regulation capability.

[0086] Based on the determined resonant frequency and quality factor, the values ​​of the resonant inductance Lr and resonant capacitance Cr are calculated. For a 10kW power rating, the resonant inductance is calculated to be 12μH and the resonant capacitance to be 0.21μF. The magnetizing inductance Lm is chosen considering the transformer turns ratio and voltage gain requirements, and is determined to be 60μH, approximately five times the resonant inductance. This parameter configuration ensures sufficient voltage gain across the entire load range from light to full load, while achieving zero-voltage switching and zero-current switching operation, minimizing switching losses.

[0087] Through time-domain and frequency-domain simulation analysis, the stability and robustness of the selected parameters under various input voltage and load conditions were verified, ensuring reliable operation under various working conditions and meeting the charging needs of industrial forklifts.

[0088] Step S1.3: Select control strategies and design circuit parameters for the AC input characteristics to achieve a PFC preamplifier circuit with a power factor greater than 0.99, thus obtaining a stable DC input for the PFC preamplifier circuit.

[0089] To meet power grid harmonic standards and improve overall efficiency, a high-performance power factor correction (PFC) preamplifier circuit was designed. By comparing various PFC topologies, a continuous-on-mode Boost-type PFC topology was selected. This topology is simple, reliable, and has a mature control strategy, enabling it to achieve a high power factor over a wide input voltage range.

[0090] The PFC control strategy employs a current averaging control method, comprising outer-loop voltage control and inner-loop current control. The outer-loop voltage control generates a current command signal by comparing the output voltage with a reference voltage; the inner-loop current control ensures that the input current follows the waveform of the input voltage, achieving a high power factor. The voltage control loop bandwidth is designed to be 10Hz, and the current control loop bandwidth is designed to be 10kHz, ensuring stability under various operating conditions.

[0091] Key parameters of the PFC circuit include the boost inductor, output filter capacitor, and power switching devices. The boost inductor is designed with a value of 500μH, and the current ripple is controlled within 20% of the rated current. The output filter capacitor is designed with 1200μF / 450V, using low-ESR electrolytic capacitors to ensure that the output voltage ripple is less than 5V. The power switch uses a 650V / 60A silicon carbide MOSFET, which has low on-resistance and excellent switching characteristics, enabling efficient operation at high frequencies.

[0092] Through carefully designed circuit parameters and control strategies, the PFC preamplifier circuit achieves a power factor greater than 0.99 and a total harmonic distortion of less than 5% within an input voltage range of 90-264VAC, meeting international standard requirements. Simultaneously, the PFC output provides a stable 380V DC bus voltage, offering a reliable input power supply to the LLC resonant converter and ensuring efficient and stable operation throughout the charging process.

[0093] Step S1.4: Based on the parameters of the resonant inductor, resonant capacitor, and magnetizing inductor, and the stable DC input of the PFC pre-amplifier circuit, the operating frequency range is determined and the modulation strategy is designed to achieve zero-voltage switching and zero-current switching operation across the entire load range, thus obtaining the LLC resonant converter.

[0094] The control of the LLC resonant converter mainly achieves output voltage stability by adjusting the switching frequency. Based on the previously determined resonant parameters and PFC output characteristics, the operating frequency range of the LLC converter is determined to be 70-150kHz, with the nominal resonant frequency being 100kHz. Under light load conditions, the controller increases the operating frequency and decreases the gain; under heavy load conditions, it decreases the operating frequency and increases the gain, thereby maintaining output voltage stability.

[0095] The frequency modulation control strategy employs a nonlinear control algorithm based on output voltage error. The controller first compares the actual output voltage with the reference voltage to generate an error signal; then, a PI regulator processes the error signal to generate a frequency control signal; finally, a voltage-controlled oscillator converts the control signal into the actual switching frequency. To improve dynamic response performance, the PI regulator parameters are dynamically adjusted using a load adaptive algorithm to optimize control performance under different load conditions.

[0096] To ensure soft-switching operation across the entire load range, an adaptive dead-time control strategy was designed. Under different load conditions, the charging and discharging times of the MOSFET's parasitic capacitance vary, requiring corresponding adjustments to the dead-time. By detecting the rate of change of the switching voltage, the dead-time is dynamically adjusted: increasing the dead-time under light load conditions and decreasing it under heavy load conditions, ensuring zero-voltage switching under various load conditions and minimizing switching losses.

[0097] In one embodiment, the stochastic surrogate model for multi-objective optimization of magnetic component parameters, constructed based on the operating frequency range and power density requirements of the LLC resonant converter through a probabilistic learning algorithm and a stochastic surrogate model, includes:

[0098] Step S2.1: Based on the operating frequency range and power density requirements of the LLC resonant converter, perform material performance comparison and loss mechanism analysis on the PC40 ferrite material data to determine the core basic material characteristic data;

[0099] In the design of magnetic components for LLC resonant converters, the first step is to thoroughly analyze the matching relationship between operating parameters and material properties. The 10kW power rating and LLC topology of an industrial forklift charger dictate that its operating frequency is typically in the 50-200kHz range, which requires the magnetic material to possess excellent high-frequency characteristics.

[0100] The analysis begins with the loss mechanisms of different ferrite materials under high-frequency conditions, mainly including hysteresis loss and eddy current loss. Hysteresis loss is related to the coercivity and maximum magnetic flux density of the material, while eddy current loss is proportional to the resistivity of the material and the square of the operating frequency. By comparing the high-frequency characteristic data of MnZn series (such as PC40 and PC44) and NiZn series ferrite materials, including key parameters such as initial permeability (μi), saturation magnetic flux density (Bs), Curie temperature (Tc), and loss factor (tanδ / μi), it can be found that PC40 ferrite material has the best performance balance point around 100kHz.

[0101] PC40 material has an initial permeability of approximately 2300, a saturation magnetic induction of about 0.5 T, a Curie temperature exceeding 220℃, and a low loss factor (<10×10⁻⁶) in the frequency range of 50-150 kHz. -6 This means that in a 10kW power-level LLC converter, PC40 material can provide sufficient flux carrying capacity while maintaining low core losses.

[0102] In addition, temperature characteristics must be considered. The permeability of PC40 material changes relatively slowly with temperature. Within the operating temperature range (usually 20-100℃), the rate of change of permeability is controlled within ±10%, which is crucial for maintaining the stable operation of LLC resonant converters.

[0103] Step S2.2: Based on the characteristic data of the magnetic core basic material, a stochastic surrogate model capable of predicting the performance parameters of the magnetic core is constructed through Gaussian process regression algorithm and kernel function parameter optimization;

[0104] The construction of the stochastic surrogate model is based on the known properties of magnetic materials, but considering the parameter uncertainties in practical applications, Gaussian Process Regression (GPR) is adopted as the core algorithm of the stochastic surrogate model. It can naturally handle uncertainties and provide the probability distribution of the prediction results.

[0105] First, test data for key parameters such as permeability, loss density, and saturation magnetic flux density at different frequencies, temperatures, and magnetic flux densities were extracted from the PC40 material datasheet to construct an initial training dataset. To enhance the model's accuracy in the actual operating range, the dataset was supplemented through specially designed test experiments, particularly data near the actual operating point of the LLC resonant converter.

[0106] After constructing the training dataset, the mathematical definition of the randomized surrogate model begins. Taking core loss prediction as an example, the model input variables include frequency f, magnetic flux density B, temperature T, and core geometric parameters G (such as effective magnetic path length le, effective cross-sectional area Ae, etc.); the output variable is the loss density per unit volume Pv. The functional relationship is established using the GPR algorithm:

[0107] Pv=f(f,B,T,G)+ε

[0108] Where ε represents Gaussian noise with a specific covariance structure, which can capture the uncertainty of model predictions.

[0109] During model training, the maximum likelihood estimation method was used to optimize the hyperparameters of GPR, including kernel function parameters (such as length scale and signal variance) and noise variance. Considering the characteristics of magnetic materials, a combination of kernel functions was selected, including the RBF (radial basis function) kernel to capture smooth variations and the Matérn kernel to handle potential non-smooth characteristics.

[0110] Step S2.3: Perform cross-validation evaluation and generalization ability test on the random surrogate model that can predict the magnetic core performance parameters to obtain the random surrogate model for multi-objective optimization of the magnetic component parameters.

[0111] After model training, cross-validation is performed to evaluate model performance and ensure good generalization ability on data outside the training set. The final stochastic surrogate model can not only predict core performance parameters (such as losses) under given conditions, but also provide a measure of prediction uncertainty, which is crucial for subsequent robust optimization design.

[0112] In one embodiment, the stochastic surrogate model for multi-objective optimization of the magnetic element parameters is used to optimize the core size, air gap length, and winding structure to obtain a high-frequency, low-loss magnetic element, including:

[0113] Step S2.4: Based on the stochastic surrogate model and design constraints of the multi-objective optimization of the magnetic element parameters, perform Pareto optimal solution search processing on the multi-objective Bayesian optimization algorithm to obtain the optimal parameter combination of magnetic core size, air gap length and winding structure;

[0114] Based on the stochastic surrogate model, the multi-objective optimization design of magnetic components aims to balance multiple objectives such as efficiency, volume, cost, and reliability. Here, a Multi-Objective Bayesian Optimization (MOBO) algorithm based on the stochastic surrogate model is employed, which can efficiently explore Pareto optimal solutions in complex design spaces.

[0115] First, define the design variables, including: core shape and size (e.g., EE, EC, PQ type), effective core cross-sectional area Ae, magnetic circuit length le, air gap length lg, number of winding turns N, and winding structure parameters. Simultaneously, define multiple optimization objectives: minimize total loss (including core loss and copper loss), minimize volume and weight, minimize temperature rise, and maximize flux utilization.

[0116] Design constraints include: magnetic flux density B ≤ 0.3T (to avoid saturation); temperature rise ΔT ≤ 40℃ (to ensure reliability); window area utilization ku ≤ 0.4 (to consider manufacturing feasibility); and resonant characteristic requirements (matching LLC resonant converter).

[0117] The optimization process employs Expected Improvement (EI) and Expected Hypervolume Improvement (EHVI) functions to balance exploration and exploitation, effectively addressing the trade-offs in multi-objective optimization. Because the stochastic surrogate model provides predictive uncertainty, the optimization algorithm can prioritize exploring regions of high uncertainty, improving search efficiency.

[0118] Through iterative optimization, a Pareto optimal solution set was obtained, from which the most balanced design scheme was selected: a PQ-type magnetic core structure with an effective cross-sectional area of ​​540 mm² was adopted.2 The magnetic circuit length is 78mm, the primary winding of the main transformer has 12 turns, and the air gap length of the resonant inductor is 0.8mm. With these parameters, at a power level of 10KW, the core loss is controlled within 22W, the temperature rise does not exceed 35℃, and the size and magnetic flux utilization are maintained at a small size.

[0119] Step S2.5: Based on the core size and air gap length in the optimal parameter combination, and combined with the conductor structure design and winding process optimization of the high-frequency current distribution characteristics, a Litz wire winding structure with low AC resistance and low eddy current loss is realized, and the winding design parameters are obtained.

[0120] In high-frequency LLC resonant converters, the skin effect and proximity effect of the conductor significantly increase AC resistance, leading to increased copper losses. Litz wires, composed of multiple strands of insulated fine wires connected in parallel, can effectively reduce these high-frequency effects, but their design requires meticulous optimization to achieve optimal performance.

[0121] First, based on the operating frequency and current requirements of the LLC converter, the skin depth δ is calculated:

[0122]

[0123] Where ρ is the resistivity of copper, f is the operating frequency, and μ is the permeability of copper. At a frequency of 100kHz, the skin depth of copper is approximately 0.21mm. According to the skin effect theory, the diameter of a single conductor should be less than twice the skin depth; therefore, enameled wire with a diameter of 0.1mm is chosen as the basic unit of the Litz wire.

[0124] Next, determine the total cross-sectional area of ​​the Litz line. Based on the 10kW power rating and the voltage and current characteristics of the LLC converter, the maximum primary current is approximately 40A, and the maximum secondary current is approximately 100A. Considering 4A / mm²... 2 According to the current density design criteria, the total cross-sectional area of ​​the primary Litz wire needs to be 10 mm². 2 The secondary side needs to be 25mm. 2 .

[0125] The key to the structural design of Litz wire lies in the number of strands and the braiding method. By comparing the equivalent resistance at high frequencies of different braiding structures (such as concentric circles and hexagons), a seven-bundle, three-layer structure was chosen: the primary stage uses a 7×7×7 = 343-strand structure (each strand with a diameter of 0.1 mm), and the secondary stage uses a 7×7×10 = 490-strand structure. This structure, while ensuring sufficient cross-sectional area, allows each conductor to be distributed as evenly as possible across the conductor's cross-section, minimizing high-frequency effects.

[0126] The winding process of Litz wire is also crucial. By optimizing the interlayer arrangement and winding tension control, stray capacitance between layers is minimized, reducing the no-load loss of the resonant transformer. Simultaneously, a flattened winding technique is employed to increase the heat dissipation area and improve heat distribution. Finite element electromagnetic field analysis verifies that the optimized Litz wire winding, at a frequency of 100kHz, has an AC resistance increase rate (compared to DC resistance) controlled to within 1.2 times, far lower than the 4-5 times increase of ordinary windings, significantly reducing copper losses at high frequencies and improving overall efficiency.

[0127] Step S2.6: Based on the core geometry determined by the optimal parameter combination and the winding design parameters, perform loss analysis and calculation on the core loss and copper loss to obtain the loss calculation results. Based on the loss calculation results, analyze the thermal resistance network and design the cooling structure to ensure that the temperature rise of the magnetic component under full load conditions is controlled within a safe range, thereby obtaining the high-frequency low-loss magnetic component.

[0128] Efficient thermal management is crucial for ensuring the reliability and long lifespan of magnetic components. In a 10kW industrial forklift charger operating continuously, the thermal management of the magnetic components requires optimized design to ensure temperature rise is kept within safe limits.

[0129] First, a detailed thermal model of the magnetic component is established, including the distribution of heat sources and the heat transfer path. The main heat sources include: core losses (hysteresis losses and eddy current losses), winding copper losses (DC losses and additional AC losses), and dielectric losses of the insulating material. The temperature distribution under different operating conditions is simulated using finite element thermal analysis software.

[0130] Thermal resistance network analysis is a core part of the design process. The thermal resistance network of a magnetic component includes: the thermal resistance Rth (core-amb) from the core to the environment, the thermal resistance Rth (winding-core) from the winding to the core, and the thermal resistance Rth (winding-amb) directly from the winding to the environment. By optimizing these thermal resistance values, heat dissipation efficiency can be improved.

[0131] Key optimization measures include: Core surface treatment: While ensuring insulation performance, the core surface is treated with high thermal conductivity epoxy resin to reduce Rth (core-amb); Winding structure optimization: A flat winding design is adopted to increase the contact area with the core and reduce Rth (winding-core); At the same time, the interlayer spacing of the windings is optimized to allow natural air convection and reduce Rth (winding-amb); Application of thermally conductive materials: Thermally conductive silicone with a thermal conductivity ≥3W / (m·K) is filled between the windings and the core to significantly reduce contact thermal resistance; Cooling structure design: Based on the overall structure of the charger, a directional airflow channel is designed to enhance forced air cooling of the magnetic components; For applications with particularly high power density, a heat dissipation base plate is considered to be added to the bottom of the core to achieve additional heat dissipation by conducting heat to the chassis. Thermal simulation analysis verified that, under the most severe operating conditions (ambient temperature 40℃, continuous full-load operation), the optimized magnetic component's maximum temperature is controlled below 85℃, with a temperature rise not exceeding 45℃. This is far below the maximum operating temperature limit of PC40 material (120℃) and the temperature resistance rating of the insulation material (typically Class F, 155℃), ensuring the long-term reliable operation of the magnetic component. Furthermore, an NTC temperature sensor is integrated into the surface of the magnetic component to monitor temperature changes in real time and implement temperature protection functions through control, further improving safety and reliability.

[0132] In one embodiment, integrating the high-frequency, low-loss magnetic component into the LLC resonant converter, and performing stochastic neural network structure design and control barrier function construction based on the physical characteristics and operating data of the LLC resonant converter to realize a stochastic neural control barrier function algorithm, includes:

[0133] Step S3.1: Based on the physical characteristics and operating data of the LLC resonant converter, a system dynamic model containing parameter uncertainties is established through system identification methods and model construction;

[0134] Modeling the power control system of an LLC resonant converter is fundamental to achieving efficient control. Compared to traditional topologies, the dynamic characteristics of an LLC resonant converter are more complex, especially considering parameter uncertainties and external disturbances, requiring a more accurate system model.

[0135] First, a small-signal model of the LLC resonant converter is established based on physical principles. The LLC resonant converter comprises three key energy storage components: resonant inductor Lr, resonant capacitor Cr, and magnetizing inductor Lm. Its state variables are typically chosen as the resonant capacitor voltage vc, the resonant inductor current iLr, and the magnetizing inductor current iLm. Based on the state-to-state averaging method, the small-signal model of the system can be obtained:

[0136] dx / dt=Ax+Bu

[0137] y = Cx + Du

[0138] Where x = [vc, iLr, iLm]T is the state variable, u is the control input (usually the offset of the switching frequency fs), y is the output variable (such as the output voltage vo), and A, B, C, and D are system matrices, which are related to the system parameters and the operating point.

[0139] However, the physical model suffers from two main problems: first, parameter uncertainties (such as the variation of magnetic component parameters with temperature and operating point); and second, model errors caused by simplifying assumptions. Therefore, it is necessary to revise the model by incorporating system identification methods.

[0140] System identification employs parameter estimation methods based on experimental data. First, an excitation signal (such as a pseudo-random binary sequence PRBS) is injected into the system, and then the system's input and output data are recorded. The system's parameter matrix is ​​estimated using algorithms such as least squares, maximum likelihood estimation, or subspace identification (N4SID).

[0141] To capture the uncertainty of the system, a probabilistic description of the parameters is introduced. For example, the magnetizing inductance Lm can be expressed as:

[0142] Lm=Lm nominal +ΔLm

[0143] Where Lm nominal ΔLm is the nominal value, and it is a random variable following a specific distribution (such as a truncated Gaussian distribution), whose mean and variance are estimated from experimental data. Ultimately, the established system dynamic model contains deterministic and stochastic uncertainties, and can be represented as a state-space model and a corresponding description of parameter uncertainties. This model will serve as the basis for subsequent stochastic neural controller design, accurately reflecting the dynamic characteristics and uncertainties of the system, and providing a reliable basis for robust control.

[0144] Step S3.2: The power control requirements of the system dynamic model are processed by Bayesian neural network structure design and variational inference method to construct a stochastic neural network model that can handle uncertainty;

[0145] Stochastic neural networks are the core component for achieving robust power control, and their structural design needs to consider the nonlinear characteristics, parameter uncertainties, and real-time computing capabilities of the system simultaneously.

[0146] First, the basic architecture of the stochastic neural network is determined. Considering the dynamic characteristics and control requirements of the LLC resonant converter, a structure based on a Bayesian Neural Network (BNN) is chosen, which can naturally handle uncertainty and provide a probability distribution for prediction.

[0147] The network's input layer includes: current state variables (resonant capacitor voltage vc, resonant inductor current iLr, magnetizing inductor current iLm), reference output (such as target output voltage Vref), historical control input (switching frequency fs of the previous few moments), and possible external conditions (such as ambient temperature, load current, etc.).

[0148] The hidden layers employ a multi-layered structure, specifically three hidden layers with 64, 32, and 16 nodes per layer, respectively. The activation function chosen is ReLU (Rectified Linear Unit), which maintains computational simplicity while handling non-linear relationships. Unlike traditional neural networks, the weights and biases of a BNN are not deterministic values ​​but rather probability distributions. For example, the weight w connecting the input x and the hidden node h can be expressed as:

[0149] w~N(μ w ,σ w 2 )

[0150] Where μ w It is the mean of the weights, σ w 2 Variance represents the uncertainty of the weights.

[0151] The network's output layer provides the probability distribution of the control variable (switching frequency fs), typically represented as the mean μfs and variance σfs. 2 , representing the optimal control value and the control uncertainty, respectively. To improve the computational efficiency of BNN in real-time systems, variational inference is used for both training and inference.

[0152] Specifically, the Flipout method is used to achieve efficient variational Monte Carlo sampling, reducing the correlation between weighted samples and improving sampling efficiency. The training of the stochastic neural network employs maximization of the evidence lower bound (ELBO) within a variational inference framework, combined with the backpropagation algorithm, and uses the Adam optimizer for parameter updates. Training data comes from system identification experiments and actual operating data, including various operating conditions and disturbances. The final designed stochastic neural network structure not only handles the nonlinear control requirements of the LLC resonant converter but also quantifies the uncertainty of control decisions, providing necessary probabilistic information for the subsequent control barrier function, thus achieving safe and reliable power control.

[0153] Step S3.3: Based on the system safety constraints and operating boundary conditions determined by the system dynamic model, design and define the safety domain for the barrier function to construct a control barrier function that ensures the system state is always within the safety domain;

[0154] A control barrier function (CBF) is a mathematical tool that ensures the system state always meets safety constraints. In industrial forklift charger systems, safety constraints include electrical limits (such as maximum current and voltage limits), thermal limits (such as maximum device temperature), and power boundaries.

[0155] First, define the system's set of safe states C:

[0156] C = {x∈X|h(x)≥0}

[0157] h(x) is called the barrier function. When h(x) ≥ 0, the system state x is considered safe. The design of the barrier function needs to consider various safety constraints of the system. For LLC resonant converters, key safety constraints include: resonant current limit: |iLr| ≤ Imax, to prevent device overcurrent; output voltage range: Vmin ≤ vo ≤ Vmax, to protect the load device; switching frequency range: fs_min ≤ fs ≤ fs_max, to ensure soft switching operation; magnetic flux density limit of magnetic components: B ≤ Bsat, to prevent magnetic saturation.

[0158] Based on these constraints, construct a combined barrier function:

[0159] h(x) = min{h1(x),h2(x),...,h n (x)}

[0160] Where h i (x) is the barrier function corresponding to the i-th constraint. To achieve smooth control, logarithmic or exponential barriers are typically used to approximate the min function. Considering the uncertainty of the system model, it is necessary to extend the traditional CBF to a stochastic control barrier function (Stochastic CBF). The stochastic CBF expresses the safety requirements through probabilistic constraints:

[0161] Pr[h(x)≥0]≥1-δ

[0162] Where δ is a small probability of violation (e.g., 0.001), representing the upper limit of the probability that the system state violates the safety constraints.

[0163] Based on the system dynamic model and the output of the stochastic neural network, the probability distribution of the future state x under a given control input u can be calculated, thereby assessing the probability of satisfying the safety constraints. Specifically, for the control input u and the current state x, if the following conditions are met:

[0164]

[0165] Where f(x,u) is the system dynamic equation, and α is a Class K function (usually chosen as a linear function α(h) = γh, γ>0), then the control input u is considered safe. The finally constructed stochastic control barrier function will serve as a safety guarantee mechanism for subsequent stochastic neural control algorithms, ensuring that the system does not violate safety constraints while pursuing performance goals.

[0166] Step S3.4: Perform control law design and parameter tuning on the stochastic neural network model and the control barrier function to implement the stochastic neural control barrier function algorithm.

[0167] The stochastic neural control algorithm combines the learning capabilities of stochastic neural networks with the safety guarantees of control barrier functions, forming a power control system that is both adaptive and safe. The core idea of ​​the algorithm is to generate candidate control values ​​through a stochastic neural network, then use a control barrier function to filter out unsafe control inputs, and finally select the optimal control strategy while ensuring safety.

[0168] The specific implementation steps are as follows: State observation and preprocessing: In each control cycle (usually 10-100μs), system state variables (such as resonant capacitor voltage vc, resonant inductor current iLr, etc.) are collected and processed by signal filtering and normalization. Control value generation: The preprocessed state variables are input into a random neural network, and the network outputs the probability distribution of the control variable (switching frequency fs), expressed as the mean μfs and variance σfs. 2 Safety Verification: Sample multiple candidate control values ​​from the control variable distribution (typically 5-10 points). For each candidate value ui, calculate the corresponding control barrier function value hi. Only retain safe control values ​​that satisfy hi ≥ 0. Performance Optimization: From the set of safe control values, select the control value that minimizes the performance cost function J(x,u) as the final control output. The performance cost function typically includes factors such as steady-state error, dynamic response speed, and control energy consumption. Execution and Learning: Apply the final determined control value to the system, and record the control results for online learning and updating of the stochastic neural network.

[0169] The mathematical expression of the algorithm is as follows:

[0170] u*=argmin {u∈Usafe} J(x,u)

[0171] in

[0172] To address the time-varying characteristics and parameter drift in the system, an online parameter update mechanism was implemented. An incremental learning strategy was adopted, with new data receiving higher weights, ensuring the network can adapt to changes in the system.

[0173] In practical applications, considering the limitations of computing resources, several optimizations were made: a fast approximation method based on variational inference was adopted to reduce the sampling computation burden; sparse updates based on importance were implemented, updating network parameters only when the state changes significantly; and a look-up table was used to accelerate some computation processes.

[0174] The final stochastic neural control algorithm can make decisions within a 10μs control cycle, meeting the real-time control requirements of the LLC resonant converter, while ensuring the safety and optimal performance of the system under various operating conditions.

[0175] like Figure 2 As shown, in one embodiment, the real-time data of voltage, current, and temperature during the charging process based on the dynamic power control strategy are processed using an anti-noise data filtering algorithm and a prediction model to realize a data-driven hybrid predictive control strategy and generate charging parameter optimization instructions, including:

[0176] Step S4.1: Based on the voltage, current and temperature sensor network, condition the signal and preprocess the data to obtain real-time data of the charging process.

[0177] Efficient control of industrial forklift chargers relies on high-quality real-time data acquisition. To this end, a comprehensive sensor network was designed to monitor electrical parameters, temperature parameters, and battery status parameters. Electrical parameter monitoring includes measurements of key points such as input voltage, input current, PFC bus voltage, LLC resonant current, output voltage, and output current. Voltage measurements employ a high-precision resistor divider network and isolation amplifier to ensure measurement accuracy better than 0.5%; current measurements utilize a closed-loop Hall effect sensor with high bandwidth (DC-200kHz) and good linearity (nonlinearity <0.1%), capable of accurately capturing rapidly changing current waveforms.

[0178] The temperature monitoring system employs a multi-point distributed structure, placing temperature sensors at key locations such as MOSFETs, magnetic components, heat sinks, and the environment. Power device temperature is monitored using NTC thermistors directly attached to the device surface, with a response time of less than 1 second. Magnetic component temperature monitoring utilizes PT100 platinum resistance thermometers, exhibiting good linearity and long-term stability. Ambient temperature monitoring employs digital temperature sensors with an accuracy of ±0.5℃. All temperature signals are processed by conditioning circuitry before being sent to the control system for over-temperature protection and heat dissipation control. Battery status monitoring includes battery voltage, current, temperature, and internal resistance estimation. Battery temperature monitoring uses multi-point temperature measurement technology, placing temperature sensors at different locations within the battery pack to capture temperature distribution. Battery internal resistance estimation uses the AC impedance method, injecting a small signal perturbation and measuring the voltage response to calculate the battery's dynamic internal resistance, providing crucial parameters for subsequent battery status assessment. The signal conditioning circuitry employs a multi-stage processing architecture to ensure signal quality. The first stage consists of pre-filtering and protection circuitry to prevent overvoltage and transient interference. The second stage comprises differential amplification and isolation circuitry to improve the common-mode rejection ratio (CMRR>80dB) and achieve safe isolation. The third stage is a precision conditioning circuit, including gain adjustment, bias compensation, and anti-aliasing filtering, ensuring the signal is compatible with the ADC's input range. All analog signal paths employ shielding and star grounding topologies to minimize external interference.

[0179] The data acquisition system adopts a hierarchical architecture, setting different sampling rates according to signal characteristics and control requirements. High-speed signals (such as resonant current) are sampled at 100kHz using a 12-bit SAR ADC; medium-speed signals (such as output voltage and current) are sampled at 10kHz using a 16-bit Sigma-Delta ADC; and low-speed signals (such as temperature) are sampled at 10Hz using a 24-bit high-precision ADC. All ADCs employ synchronous sampling technology to ensure time consistency of multi-channel data.

[0180] The data preprocessing algorithm is executed in real time on an FPGA or DSP, including outlier detection and processing, digital filtering, signal calibration, data compression, and feature extraction. Outlier detection employs a combination of statistical methods and model prediction to effectively identify abnormal data caused by sensor faults or interference. Digital filtering uses a configurable combination of IIR and FIR filters, automatically selecting the optimal filtering parameters based on different signal characteristics. Signal calibration considers the sensor's nonlinear characteristics and temperature drift, achieving high-precision calibration through lookup table interpolation and polynomial fitting. Through this complete data acquisition and preprocessing system, the charger can acquire high-quality real-time data, providing a reliable foundation for subsequent noise analysis, predictive control, and battery state assessment.

[0181] Step S4.2: Perform statistical analysis and feature extraction on the real-time data to establish a charging system noise model that includes electromagnetic interference noise, power supply fluctuation noise, and sensor noise.

[0182] Noise sources in industrial environments are complex and diverse, significantly impacting charging control accuracy. Systematic analysis of collected real-time data allows for the identification and modeling of various noise sources, providing a scientific basis for subsequent noise mitigation algorithm design. The noise sources are first identified and classified into four main categories: electromagnetic interference (EMI) noise, power supply fluctuation noise, sensor noise, and load fluctuation noise.

[0183] Electromagnetic interference (EMI) noise primarily originates from the high-frequency switching process of power switching devices and radiated interference from surrounding industrial equipment. Spectral analysis revealed that EMI noise is mainly concentrated at the switching frequency (100kHz) and its harmonic frequencies, with amplitude varying with load, and is most significant under full-load conditions. Furthermore, intermediate-frequency interference (1-10kHz) from industrial frequency converters and random pulse interference from welding equipment were also observed. Statistical features of these interferences were extracted, including frequency distribution, amplitude distribution, and time correlation, establishing an EMI noise description based on a Gaussian mixture model.

[0184] Power supply fluctuation noise originates from grid quality fluctuations, primarily manifested as amplitude and frequency fluctuations in input voltage. Through long-term data recording and analysis, it was found that grid voltage fluctuations range from ±10% of the nominal value, mainly consisting of low-frequency fluctuations of 0.1-10Hz and harmonic components of the grid frequency (50 / 60Hz). Time-domain and frequency-domain analyses of these fluctuations were performed, establishing a power supply fluctuation noise description based on an autoregressive moving average (ARMA) model. This model accurately captures the time-varying and frequency characteristics of power supply fluctuations.

[0185] Sensor noise comprises various components, including thermal noise, quantization noise, and conversion noise. Through repeated measurements and statistical analysis under steady-state conditions, the noise characteristics of various sensors were determined. Voltage sensor noise exhibits a Gaussian distribution with a standard deviation of approximately 0.05% of full scale; current sensor noise, in addition to the Gaussian component, also includes proportional noise related to current magnitude; temperature sensor noise exhibits a combination of low-frequency drift and high-frequency random fluctuations. Based on these analyses, noise models for various sensors were established, including noise power spectral density and probability distribution characteristics.

[0186] Load fluctuation noise originates from changes in battery internal resistance and fluctuations in electrochemical reactions during charging. Analysis of charging data from batteries of different types and aging stages reveals that load fluctuations exhibit significant non-stationary characteristics, with the most pronounced fluctuations occurring at transition points during the charging process (such as the transition from constant current to constant voltage). Time-frequency analysis of these fluctuations was performed, and wavelet transform was used to capture the non-stationary characteristics, establishing a load fluctuation noise description based on a hidden Markov model. This model divides the charging process into multiple states, each with different noise characteristics.

[0187] Based on the above analysis, an overall noise model of the charging system was established, and the noise characteristics of the system were represented by a mixture of Gaussian hidden Markov models (MGHMM).

[0188]

[0189] Where w i The weight of the i-th Gaussian component is N(x|μ). i ,Σ i The noise level is a multidimensional Gaussian distribution. The model parameters are estimated from real-world data using the Expectation-Maximization (EM) algorithm, accurately describing the statistical and time-varying characteristics of the noise. Furthermore, a correlation model between noise characteristics and system state is established, describing the variation of noise under different charging stages (pre-charging, constant current charging, constant voltage charging) and different operating conditions (light load, full load, ambient temperature variation). This correlation model provides a basis for the adaptive adjustment of the noise model, ensuring accurate description of system noise characteristics under various operating conditions.

[0190] Step S4.3: Based on the charging system noise model, effective filtering of charging data is achieved through unscented Kalman filter design and adaptive adjustment of the covariance matrix.

[0191] To address the complex noise environment in industrial forklift charging systems, a multi-level adaptive filtering architecture was designed to effectively suppress the impact of various noises on charging data. The filtering system consists of three layers: a pre-filter layer, a main filter layer, and a post-processing layer, each handling noise with different characteristics. The pre-filter layer primarily handles obvious outliers and high-frequency noise, employing a combination of median filtering and amplitude limiting filtering to effectively remove pulse interference and sudden outliers while preserving the rapid change characteristics of the signal.

[0192] The main filtering layer is the core of the system, employing an adaptive filtering algorithm based on the unscented Kalman filter (UKF). Compared to traditional Kalman filters, the UKF can better handle nonlinear systems, eliminating the need for linearization approximations and maintaining higher estimation accuracy. The filter's state equations are based on the dynamic model of the charging system, including state variables such as voltage, current, and temperature, and their derivatives; the observation equations are based on sensor characteristics and the measurement model. To handle parameter uncertainties and variations in operating conditions within the system, an adaptive adjustment mechanism for the covariance matrix is ​​implemented.

[0193] Adaptive adjustment of the covariance matrix is ​​a key technique for improving filtering performance. The observation noise covariance matrix R_k and the process noise covariance matrix Q_k are dynamically updated recursively. Where λ R and λ Q It is the forgetting factor (range 0.95-0.99), K k It is the Kalman gain, v k This refers to the observation residual. This adaptive mechanism can dynamically adjust the filter parameters based on actual observation data to adapt to changes in noise characteristics under different operating conditions. For example, during the constant current phase of battery charging, the system is relatively stable, and the noise mainly comes from the sensor. At this time, R... k The value is relatively small; however, it changes significantly dynamically during charging mode switching or sudden load changes, at which point Q... k The corresponding increase improves the filter's ability to track state changes.

[0194] To handle potential non-Gaussian noise, the filtering system incorporates a particle filter (PF) element, forming a UKF-PF hybrid filter. The statistical characteristics of the observed residuals are monitored in real time. When a significant deviation from the Gaussian assumption is detected in the noise distribution (e.g., through kurtosis and skewness tests), the system automatically switches to particle filter mode, using a ensemble of particles to represent the state distribution, thus improving filtering accuracy in non-Gaussian environments. The number of particles is dynamically adjusted based on computational resources and accuracy requirements, typically between 100 and 500, controlling the computational burden while maintaining accuracy.

[0195] To address sudden interference and sensor failures, an anomaly detection and reset mechanism was designed. When the observation residual exceeds a dynamic threshold (typically set to 3-5 times the predicted observation covariance), a state estimation reset process is triggered to prevent filter divergence. The reset process employs a multi-model approach, maintaining multiple alternative state estimates and switching to the most reliable estimate when an anomaly is detected, ensuring the stability and continuity of the filter.

[0196] For periodic noise present in the system (such as interference related to the power grid frequency or switching frequency), the filtering system combines a notch filter and an adaptive linear predictor for targeted suppression. The center frequency of the notch filter is automatically locked to the detected characteristic frequency, and its bandwidth is dynamically adjusted according to the signal characteristics; the adaptive linear predictor estimates the amplitude and phase of the periodic interference in real time using the recursive least squares (RLS) algorithm, and subtracts the predicted interference component from the original signal.

[0197] The post-processing layer primarily refines the filtering results and enhances signal characteristics, employing wavelet thresholding techniques. The signal is decomposed into coefficients at different frequency scales using wavelet transform, and an adaptive threshold function is applied to each scale coefficient. The signal is then reconstructed through inverse transform. This method effectively preserves the signal's edge characteristics and rapidly changing features while further suppressing residual noise.

[0198] The filtering system is implemented on a DSP or FPGA platform, employing a pipelined architecture and parallel processing technology to keep processing latency below 1ms, meeting real-time control requirements. System evaluation shows that in typical industrial environments, this filtering scheme can improve the signal-to-noise ratio by 15-20dB, with residual noise after voltage signal filtering less than 0.1% and residual noise after current signal filtering less than 0.2%, providing a high-quality data foundation for subsequent predictive control.

[0199] Step S4.4: Perform bidirectional LSTM network structure design and Monte Carlo Dropout technology processing on the filtered historical data to build a prediction model that can predict the evolution of charging state and quantify prediction uncertainty.

[0200] Building a predictive model based on filtered, high-quality data is a crucial step in achieving intelligent charging control. This predictive model needs to accurately forecast future states, including output voltage, current, battery status, and the temperature of key components, while simultaneously quantifying the uncertainty of the prediction. To address this, a hybrid predictive architecture combining physical models and data-driven methods was designed, fully leveraging prior knowledge and historical data.

[0201] Step S4.5: Based on the prediction model, combine nonlinear model predictive control and stochastic model predictive control to realize a hybrid predictive control strategy that combines deterministic control and stochastic control.

[0202] Designing a hybrid predictive control strategy based on a high-precision predictive model is the core of realizing intelligent charging control. Hybrid predictive control combines the advantages of deterministic and stochastic control, enabling it to handle uncertainties and external disturbances while ensuring control performance. First, the control objective system is defined, prioritized as follows: safety objective (keeping within the safety boundary), efficiency objective (maximizing charging efficiency), lifespan objective (extending battery life), and time objective (minimizing charging time).

[0203] The hybrid predictive control framework comprises four key components: a prediction module, a deterministic controller, a stochastic controller, and a decision fusion unit. The prediction module, based on the prediction model built in the previous step, provides the probability distribution of future states, including the predicted mean and covariance. The deterministic controller employs a nonlinear model predictive control (NMPC) framework, primarily handling nominal operating conditions; the stochastic controller, based on stochastic model predictive control (SMPC) theory, specifically handles uncertainties and disturbances; and the decision fusion unit dynamically adjusts the weights of the two control strategies according to the current state to optimize overall performance.

[0204] The deterministic control part (NMPC) calculates the optimal control sequence by solving a finite-time optimization problem. Optimization objectives include output tracking error (such as deviations of voltage and current from reference values) and control input variations (such as the adjustment magnitude of switching frequency). A weight matrix balances the importance of different objectives. Constraints include dynamic equations, input / output limitations, and state change rate constraints. To improve solution efficiency, a real-time iterative method (RTI) is employed, decomposing the optimization problem for each control cycle into a preparation phase and a feedback phase. Most calculations are completed in the preparation phase, while the feedback phase requires only a small amount of computation to obtain the control output, significantly reducing computational latency.

[0205] The Stochastic Control Part (SMPC) considers uncertainties and external disturbances, improving control robustness through probabilistic constraints and expected value optimization. Unlike deterministic control, the objective function of SMPC is in expected value form, taking into account the overall characteristics of the state distribution; the constraints are also probabilistic, allowing for constraint violations with low probability, thus increasing control flexibility. To handle probabilistic constraints, a sampling-based approximation method is used, extracting multiple sample points from the predicted state distribution to transform the probabilistic constraints into a set of deterministic constraints, which are then solved using standard optimization methods. This method is computationally efficient and suitable for real-time control applications.

[0206] The fusion of the two control strategies employs an adaptive weighting mechanism, dynamically adjusting the weights based on the uncertainty and disturbance level of the current state. The fusion formula is: u final =(1-γ)·u NMPC +γ·u SMPCWhere γ∈[0,1] are adaptive weights, calculated using the sigmoid function mapping: γ=sigmoid(α·σ pred -β), where σ pred It is a measure of prediction uncertainty (such as the norm of prediction variance), and α and β are adjustment parameters. When in a high-uncertainty or high-disturbance environment (such as rapid changes in battery state or large fluctuations in external temperature), the weight of the stochastic controller is increased; conversely, the weight of the deterministic controller is biased to improve control accuracy and response speed.

[0207] Step S4.6: Based on the control decision of the hybrid predictive control strategy, perform optimization calculations on the charging voltage setpoint, charging current setpoint, and switching frequency to generate the charging parameter optimization command.

[0208] The final output of the hybrid predictive control strategy needs to be translated into specific charging parameter optimization instructions to directly control the charger's operation. This step converts the control decisions into actual control signals, including key parameters such as charging voltage setpoint, charging current setpoint, and switching frequency. The optimization calculation process considers multiple factors, including control performance, efficiency, battery protection, and hardware limitations.

[0209] First, based on the output of the hybrid predictive control strategy, the optimal charging voltage and current setpoints are calculated. During the constant current charging phase, the current setpoint is the primary control objective. The control algorithm determines the optimal charging current based on the battery type, capacity, and health status, typically 0.2C-0.5C of the rated capacity (C being a multiple of the battery capacity). During the constant voltage charging phase, the voltage setpoint becomes the primary control objective, determined based on the battery's chemical characteristics; for example, it is typically 4.2V / cell for lithium batteries and 2.4V / cell for lead-acid batteries. These basic setpoints are dynamically adjusted based on battery temperature, internal resistance, and aging status to ensure a safe and efficient charging process.

[0210] Secondly, the optimal switching frequency control sequence is calculated. The LLC resonant converter controls the output voltage and current by adjusting the switching frequency. Based on the prediction model and control objectives, the optimal frequency trajectory in the future time domain is calculated, and the frequency value at the current moment is extracted as the control output. The frequency calculation considers soft-switching conditions, efficiency optimization, and dynamic response requirements. While ensuring zero-voltage switching (ZVS), the operating frequency is made as close as possible to the resonant frequency to improve efficiency. At the same time, the rate of frequency change is limited to prevent overshoot and oscillation caused by excessively fast response.

[0211] To improve overall efficiency, the control algorithm also includes an efficiency optimization module. By establishing a loss model, it predicts the efficiency at different operating points (combinations of voltage, current, and frequency) and selects the operating point with the highest efficiency while meeting the control objectives. The loss model considers switching losses, conduction losses, magnetic component losses, and other auxiliary losses, accurately predicting efficiency performance under various operating conditions. Under light load conditions, the control algorithm may select a frequency slightly higher than the optimal control point, sacrificing a small amount of control accuracy for higher efficiency.

[0212] The control command generation also considers hardware protection and safety boundaries. Key parameters (such as resonant current, switching temperature, and output current) are monitored in real time. When these parameters are predicted to approach safety limits, the control algorithm automatically adjusts the commands to ensure operation within the safety boundaries. For example, when the resonant current is detected to be close to its limit, the control algorithm limits the lower frequency limit to prevent excessive resonant current from damaging the device; when the switching temperature is detected to be too high, the output power is appropriately reduced to decrease heat generation.

[0213] The final generated charging parameter optimization instructions consist of three main parts: charging voltage setpoints: including target voltage values ​​and voltage change rate limits, directly controlling the charger's output voltage; charging current setpoints: including target current values ​​and current change rate limits, controlling the charger's output current; and switching frequency control values: including target frequency values, frequency change rate limits, and soft-switching protection parameters, directly controlling the operating frequency of the LLC resonant converter. These instructions are sent to the corresponding hardware modules for execution via a digital control interface. The voltage and current setpoints are converted into analog reference signals via a digital-to-analog converter (DAC), or directly passed as digital values ​​to the digital control loop; the switching frequency control values ​​are converted into actual switching drive signals via a digital pulse width modulation (PWM) module or a digital controlled oscillator (DCO).

[0214] The update rate of control commands varies depending on the control layer: the update rate for fast control layers (such as the current loop) is 10-50kHz to ensure rapid response; the update rate for predictive control layers is 10-100Hz to balance control performance and computational burden; and the update rate for adaptive layers is 0.1-1Hz to adapt to slowly changing characteristics. Through this complete parameter optimization and command generation mechanism, the hybrid predictive control strategy can be effectively transformed into actual control behavior, achieving efficient, safe, and intelligent charging process control, and providing a reliable foundation for subsequent adaptive battery charging.

[0215] In one embodiment, the step of designing a bidirectional LSTM network structure and processing the filtered historical data using Monte Carlo Dropout technology to construct a predictive model capable of predicting the evolution of the charging state and quantifying prediction uncertainty includes:

[0216] Step S4.4.1: Based on the battery charge and discharge characteristic data, a second-order RC equivalent circuit physical model describing the battery dynamic characteristics is established through equivalent circuit parameter identification and model fitting.

[0217] Accurate modeling of battery dynamic characteristics is fundamental to predicting the evolution of the state of charge. This step employs a second-order RC equivalent circuit model, which balances complexity and accuracy and is suitable for real-time applications. First, voltage, current, and temperature data for different battery types under various charge and discharge conditions are collected, including the performance of lead-acid and lithium-ion batteries under constant current charging, constant voltage charging, pulse charge and discharge, and different temperature environments, ensuring the model has good generalization ability.

[0218] The equivalent circuit model includes five key parameters: open-circuit voltage, ohmic internal resistance, and two parallel RC networks. The open-circuit voltage is directly related to the battery's state of charge (SOC), and their relationship curve is obtained through step current experiments and static measurement methods. For lithium-ion batteries, this relationship curve is typically S-shaped, changing relatively smoothly in the middle SOC region and steeply in the high and low SOC regions; the relationship for lead-acid batteries is closer to linear, but shows a significant drop in the low SOC region. These characteristic curves are stored using piecewise polynomial functions or lookup tables for real-time use by the model.

[0219] The ohmic internal resistance represents the battery's instantaneous response characteristics, calculated from the instantaneous voltage change during a current step experiment. Two RC networks represent the battery's short-time constant response (on the order of seconds) and long-time constant response (on the order of minutes), respectively. These parameters are obtained from raw data obtained through mixed-pulse power characteristic testing, and then identified using nonlinear least squares method. To improve identification accuracy, a multi-starting-point optimization strategy is employed to avoid local optima.

[0220] The model parameters are closely related to the battery's state of charge (state of charge, temperature, and degree of aging). Through analysis of extensive experimental data, a mapping relationship between parameters and states is established. For example, internal resistance increases with decreasing temperature, potentially being 2-3 times that at 25°C at 0°C; it also increases with decreasing state of charge, especially showing significant growth in the low state of charge region; and it increases with increasing battery aging, with the internal resistance of an aged battery potentially being 1.5-2 times that of a new battery. These mapping relationships are stored using multidimensional lookup tables or parameterized functions, enabling dynamic adjustment of the model parameters.

[0221] The final established second-order RC equivalent circuit model accurately describes the battery's voltage response during charging and discharging, including instantaneous response, short-term polarization effect, and long-term polarization effect. The model is solved numerically using the forward Euler method, with a time step set to 0.1-1 seconds according to control requirements, balancing computational efficiency and accuracy. Model validation employs cross-validation, using a test dataset different from that used for parameter identification to evaluate model performance. Validation results show that for lithium-ion batteries, the model's voltage prediction error under dynamic conditions is less than 20mV; for lead-acid batteries, the prediction error is less than 30mV, meeting the accuracy requirements for charging control. Furthermore, the model maintains good predictive ability within a temperature range of 0-45℃, providing a reliable physical basis for subsequent state-of-charge prediction.

[0222] Step S4.4.2: Based on the filtered historical data, perform bidirectional LSTM layer design and fully connected layer configuration on the normalized multivariate time series to establish a neural network structure that captures forward and backward temporal dependencies and generate a data-driven network architecture.

[0223] While physical models possess clear physical meaning, they struggle to fully capture the complex nonlinear dynamic characteristics of batteries and charging. Therefore, this step designs a data-driven network architecture based on deep learning to supplement and enhance the physical model. First, the filtered historical data undergoes preprocessing, including outlier handling, missing value imputation, and data normalization. Normalization employs a min-max scaling method, mapping variable values ​​to the [-1, 1] interval to eliminate dimensional differences and accelerate network training convergence.

[0224] The input features include three categories: historical observation data (such as time series of battery voltage, current, temperature, and charging time), typically selecting data from the past 10-30 time steps; control variables (such as charging mode, set voltage, and set current); and environmental and state variables (such as ambient temperature, initial state of charge of the battery, battery type, and capacity). These features undergo feature engineering to generate more expressive derived features, such as power, voltage change rate, and temperature gradient, enhancing the network's learning ability.

[0225] The core of the network employs a bidirectional long short-term memory (Bi-LSTM) structure, which can simultaneously consider both forward and backward information of the sequence, capturing a more comprehensive temporal dependency. Compared with traditional unidirectional LSTM, Bi-LSTM performs better when processing sequences such as charging processes that are both historically influenced and goal-oriented. The network structure consists of two Bi-LSTM layers: the first layer contains 128 hidden units, and the second layer contains 64 hidden units. Each layer is followed by a batch normalization layer to accelerate training and improve generalization ability; a Dropout layer (dropout rate of 0.2) is also configured to prevent overfitting. The output of the Bi-LSTM layers is further processed by fully connected layers to extract high-level features and map them to the prediction target space. The fully connected layers adopt a three-layer structure: the first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 16 neurons. The activation function is ReLU, which balances non-linear expressiveness and computational efficiency. Batch normalization and Dropout mechanisms are also configured between the fully connected layers to enhance the network's robustness.

[0226] The network output layer is designed based on the prediction target, including key state variables such as battery voltage, current, temperature, and state of charge at future moments. The output layer uses a linear activation function to ensure that the predicted values ​​are not limited by range. For multi-step predictions (such as predicting the state evolution over the next 30 minutes), an encoder-decoder architecture is adopted. The decoder part uses a unidirectional LSTM to generate the prediction sequence, avoiding information leakage issues.

[0227] The network training employed mini-batch gradient descent with a batch size of 64, using the Adam optimizer with an initial learning rate of 0.001 and a learning rate decay strategy. The loss function combined mean squared error and mean absolute percentage error, weighted at a ratio of 0.7:0.3, balancing the accuracy of large and small value predictions. To prevent overfitting, in addition to Dropout, an early stopping strategy was employed: training was halted when the validation loss showed no improvement for 10 consecutive training epochs.

[0228] The dataset is partitioned using time-series cross-validation to ensure the reliability of the model evaluation. Specifically, the data is divided into multiple consecutive segments in chronological order. The first few segments are used for training, and the next segment is used for validation, gradually rolling forward until the average performance is taken as the model evaluation result. This method is more in line with the practical application scenarios of time-series forecasting than simple random partitioning.

[0229] The final designed bidirectional LSTM network structure performs excellently in battery state-of-charge prediction, accurately capturing the complex dynamic characteristics and temporal dependencies during the charging process, providing a powerful data-driven component for subsequent hybrid prediction models. The network exhibits good generalization ability across various charging modes and battery types, with prediction errors significantly lower than traditional time series models.

[0230] Step S4.4.3: Based on the battery parameters of the second-order RC equivalent circuit physical model and the data-driven network architecture, the physical constraint embedding and neural network parameter optimization are fused to form a hybrid prediction model structure that combines physical constraints and data-driven approaches.

[0231] Physical models and data-driven models each have their advantages: physical models have clear physical meaning and good extrapolation capabilities, but may overlook some complex nonlinear relationships; data-driven models can capture complex patterns, but lack physical interpretability and may perform poorly outside the training data range. This step fuses the two models to form a hybrid prediction model under physical constraints, combining the advantages of both to improve prediction performance and reliability.

[0232] The fusion strategy employs a "physics-guided neural network" architecture, embedding the structure and constraints of the physical model into the neural network. First, the equations of the second-order RC equivalent circuit model are discretized, serving as the backbone structure of the neural network. Specifically, a portion of the network directly simulates the computational process of the physical equations, including state-of-charge updates, RC network voltage calculations, and terminal voltage synthesis. The initial weights of this part of the network are set to the corresponding values ​​of physical parameters, such as internal resistance and capacitance.

[0233] Simultaneously, a residual learning mechanism is introduced, allowing the neural network to learn the residual components that the physical model failed to capture. The residual network branches employ a Bi-LSTM structure, with inputs including the original observation data and intermediate state variables of the physical model, and outputs as corrections to the physical model's predictions. This design enables the network to learn complex nonlinear dynamic characteristics and the influence of environmental factors while maintaining the basic structure of the physical model.

[0234] Physical constraints are embedded in the network training process in several ways. First, there are loss function constraints, which, in addition to the standard prediction error, include a physical consistency loss term to penalize predictions that violate physical laws. For example, energy conservation constraints ensure that the predicted battery energy change is balanced with the input / output energy; monotonicity constraints ensure that the state of charge increases monotonically during charging and decreases monotonically during discharging; and range constraints ensure that predicted parameters such as voltage and temperature are within physically reasonable ranges.

[0235] Secondly, there are parameter constraints. Non-negative constraints are applied to the network weights corresponding to the physical parameters (e.g., the resistance value must be greater than zero), and reasonable ranges of variation are set. These constraints are implemented through parameter reprojection or penalty terms to ensure that the network parameters always have physical rationality. Thirdly, there are structural constraints to maintain the causal relationships and computational flow of the physical model. For example, the state-of-charge calculation module must receive current integrals as input; the terminal voltage calculation must consider the contributions of open-circuit voltage, ohmic voltage drop, and polarization voltage.

[0236] The parameter optimization employs a two-stage strategy: In the first stage, the physical model portion is fixed, and only the residual network branch is trained, allowing the network to learn supplementary information from the physical model. In the second stage, all parameters are jointly optimized, but a large regularization constraint is applied to the physical parameters to prevent excessive deviation from their physically reasonable values. The optimization algorithm uses the Adam optimizer in conjunction with a cosine annealing strategy, gradually reducing the learning rate in the later stages of training to fine-tune the parameters.

[0237] Step S4.4.4: Based on the hybrid prediction model structure, perform multiple forward propagation and probability distribution calculations on the Monte Carlo Dropout technique to obtain the prediction model that can predict the evolution of the charging state and quantify the prediction uncertainty.

[0238] Quantifying prediction uncertainty is crucial for robust control decisions, especially in safety-critical applications such as battery charging. This step introduces Monte Carlo Dropout, extending the deterministic hybrid prediction model into a probabilistic prediction model, providing not only point predictions but also quantifying the uncertainty of the predictions. Monte Carlo Dropout is a simple and effective Bayesian deep learning approximation method that transforms a neural network into an approximation of a Bayesian neural network by maintaining Dropout activation during the inference phase.

[0239] First, Dropout layers are added between key layers of the hybrid prediction model, with a dropout rate set to 0.2. Specifically, Dropout mechanisms are configured between Bi-LSTM layers, between Bi-LSTM and fully connected layers, and between fully connected layers. These Dropout layers are used to prevent overfitting during training and to generate random variants of the model during inference, simulating posterior distribution sampling.

[0240] During the inference phase, the same input is forward-propagated multiple times (typically 50-100 times). In each propagation, the Dropout layer randomly masks different neurons, producing slightly different prediction results. The set of these prediction results forms the probability distribution of the prediction. The mean can be calculated as the final prediction value, and the variance or quantiles can be calculated as a measure of uncertainty. For multi-step prediction, a complete probability distribution is generated at each time step. As the prediction time domain extends, the uncertainty usually increases, which is consistent with the increasing prediction difficulty in reality.

[0241] To improve the efficiency of Monte Carlo sampling, an importance-weighted strategy is adopted. Instead of simply averaging all sampling results, weights are assigned based on the consistency of each sample with historical observations, with samples showing higher consistency receiving greater weights. This method can reduce the impact of outlier sampling and improve the robustness of predictions. Weight calculation is based on sample likelihood; for each sample, a corresponding weight is calculated based on its validation loss, with samples having lower validation losses receiving higher weights.

[0242] Uncertainty can be decomposed into two parts: cognitive uncertainty and stochastic uncertainty. Cognitive uncertainty stems from the uncertainty of model parameters, reflecting insufficient training data or limitations in the model structure, and can usually be reduced by increasing training data. Stochastic uncertainty originates from inherent randomness and measurement noise, and cannot be eliminated by increasing data. This model captures these two types of uncertainty through different mechanisms: Dropout sampling primarily captures cognitive uncertainty; while stochastic uncertainty is captured through the direct output of prediction variance.

[0243] In practice, the model's output layer not only predicts the expected value of the target variable but also its variance (or standard deviation). For example, for battery voltage prediction, the model outputs two values, representing the predicted mean and standard deviation of the voltage, respectively. A negative log-likelihood loss function is used during training. This design allows the model to learn prediction uncertainty and automatically increase the prediction variance in high-noise regions.

[0244] The final prediction uncertainty is obtained by combining cognitive uncertainty and stochastic uncertainty. For the predictor variable, its overall prediction distribution can be approximated as a Gaussian distribution, with the mean being the predicted average of multiple samples, and the variance comprising both average stochastic uncertainty and cognitive uncertainty. The reliability of the prediction uncertainty is assessed through calibration; ideally, the 90% confidence interval should contain approximately 90% of the actual observations. Calibration adjustments are performed as necessary using post-processing methods such as temperature scaling.

[0245] In practical applications, predictive uncertainty provides an important basis for control decisions. In regions of high uncertainty, more conservative control strategies should be adopted to increase the safety margin; while in regions of low uncertainty, more aggressive strategies can be used to optimize performance. Furthermore, uncertainty information can be used for active learning, identifying areas where the model lacks knowledge, collecting new data in a targeted manner, and continuously improving model performance.

[0246] By employing Monte Carlo Dropout technology, the resulting predictive model not only accurately predicts the evolution of the charging state but also quantifies the uncertainty of the prediction, providing comprehensive decision support information for subsequent hybrid predictive control. When implemented on an embedded platform, the model utilizes batch processing and parallel computing techniques to keep the computation time for 50 Monte Carlo samplings within 50ms, meeting real-time control requirements.

[0247] In one embodiment, the process of combining nonlinear model predictive control and stochastic model predictive control based on the predictive model to achieve a hybrid predictive control strategy combining deterministic and stochastic control includes:

[0248] Step S4.5.1: Based on the nonlinear model predictive control framework, optimize and constrain the prediction time domain to realize a deterministic controller that can handle nominal operating conditions.

[0249] Deterministic controllers are a fundamental component of hybrid predictive control strategies, primarily responsible for precise control under nominal operating conditions. This step, based on a nonlinear model predictive control framework, designs a deterministic controller suitable for industrial forklift charging. This controller achieves predictive control by online optimization of future control sequences, minimizing the performance objective function while satisfying constraints.

[0250] First, the control objectives and performance indicators are clearly defined. For charging, the main control objectives include: output voltage / current tracking (ensuring the actual output follows the setpoint), smooth state transitions (avoiding drastic changes in control and state variables), and maximizing efficiency (optimizing efficiency while meeting other objectives). These objectives are represented by a weighted performance indicator function, including output tracking error, control rate of change, and efficiency terms. The weights are dynamically adjusted according to the control phase; for example, the smoothness weight is increased during charging mode switching, and the efficiency weight is increased during steady-state operation.

[0251] The prediction time domain length is a key parameter of the controller, affecting control performance and computational burden. A longer prediction time domain can consider effects further into the future, improving control foresight, but increasing computational complexity; a shorter prediction time domain is computationally efficient but may lead to short-sighted behavior. This paper adopts an adaptive prediction time domain strategy, dynamically adjusting the time domain length according to the state and control stage. Specifically, a longer time domain (e.g., corresponding to a 2-second prediction) is used during steady-state operation to fully optimize efficiency and smoothness; a shorter time domain (e.g., corresponding to a 1-second prediction) is used during transient processes or mode switching to improve response speed and computational efficiency.

[0252] Constraints are a crucial component of controllers, ensuring that control inputs and states remain within safe and reasonable ranges. Key constraints include: input constraints (range limits for control variables), input rate of change constraints (rate limits for the rate of change of control variables), output constraints (range limits for output variables), and state constraints (safety limits for internal state variables). To avoid feasibility issues in constraint handling, a soft constraint mechanism is employed, allowing slight violations of certain constraints when necessary, but with penalties imposed. Soft constraints are implemented by introducing slack variables and corresponding penalty terms, with penalty weights set to relatively large values ​​to ensure that constraints are violated only when there are no other options.

[0253] The optimization problem is solved using a real-time iterative method, dividing the optimization process for each control cycle into a preparation phase and a feedback phase. The preparation phase occurs during the execution of the current control cycle and includes linearization, constraint handling, and quadratic programming problem construction. The feedback phase is executed immediately after new state measurements become available, and the control output can be obtained in just a few iterations. This method significantly reduces feedback latency and improves control real-time performance.

[0254] The specific solution algorithm employs a sequential quadratic programming method combined with a warm-start technique. In each iteration, the nonlinear optimization problem is linearized around the current working point, transforming it into a quadratic programming problem, which is then solved using an efficient solver. The warm-start technique uses the solution from the previous step as the initial value for the current step, accelerating the convergence process. To further improve computational efficiency, an early termination strategy is adopted, stopping optimization when the solution reaches sufficient accuracy or the maximum number of iterations is reached.

[0255] Reference trajectory generation is a crucial step in ensuring control performance. Simply using a step reference signal may lead to overshoot and oscillation. This paper employs model-based trajectory planning to generate a smooth and feasible reference trajectory based on dynamic characteristics and constraints. For charging mode switching (such as from constant current to constant voltage), an S-shaped transition curve is used to ensure smooth state changes; for responding to external commands (such as adjusting charging power), a bandwidth-considered filtered reference generator is used to avoid rapid changes that exceed the response capability.

[0256] To address model mismatch and external disturbances, state estimation and disturbance observers are introduced. State estimation employs an extended Kalman filter to estimate the complete state vector based on the model and measurement data. The disturbance observer estimates unmodeled external disturbances and parameter drift, forming a disturbance compensation term to enhance control robustness. These components are tightly integrated with the controller to form a complete control loop.

[0257] The final deterministic controller performs excellently under nominal operating conditions, achieving voltage control accuracy of ±0.5% and current control accuracy of ±1%. Its dynamic response is rapid and smooth, meeting the high-performance control requirements of industrial forklift charging. Implemented on an embedded platform, the controller employs code optimization and parallel computing techniques to keep the computation time for each control cycle within 5ms, satisfying real-time control requirements.

[0258] Step S4.5.2: Based on the stochastic model predictive control framework, optimize the probability constraints and expected values ​​to realize a stochastic controller that handles uncertainties and disturbances.

[0259] While deterministic controllers perform excellently under nominal conditions, their performance may degrade under uncertainty and external disturbances. This step designs a stochastic controller based on a stochastic model predictive control framework, specifically designed to handle uncertainties and disturbances, thereby improving robustness. The stochastic controller treats the state and prediction as random variables, achieving reliable control in uncertain environments through probabilistic constraints and expected value optimization.

[0260] The core of stochastic controllers lies in extending deterministic optimization problems to stochastic ones. The objective function transforms from a deterministic form to an expected value form, taking into account the probability distribution of state predictions. This form balances average performance and risk, avoiding overly aggressive control decisions. Constraints also shift from a deterministic to a probabilistic form. Traditional hard constraints are transformed into probabilistic constraints, allowing for constraint violations with low probabilities, thus improving control flexibility, especially in regions of high uncertainty.

[0261] For safety-critical constraints (such as maximum current limits), a smaller violation probability (e.g., 0.001) can be set to ensure high safety; for performance-related constraints (such as voltage fluctuation range), a relatively larger violation probability (e.g., 0.05) can be set to increase control degrees of freedom. Handling probabilistic constraints is a technical challenge in stochastic control. This paper employs a sampling-based approximation method, extracting multiple sample points from the predicted state distribution to transform the probabilistic constraints into a set of deterministic constraints.

[0262] The computational efficiency of this sampling approximation method is directly related to the number of samples. To balance computational burden and approximation accuracy, an adaptive sampling strategy is adopted. In regions of high uncertainty or near constraints, the number of samples is increased to improve approximation accuracy; in regions of low uncertainty or far from constraints, the number of samples is reduced to lower the computational burden. The sample generation employs an importance sampling technique, prioritizing the generation of more samples near constraint boundaries to improve the accuracy and efficiency of constraint processing.

[0263] The solution to the stochastic optimization problem employs a scenario-based approach, treating multiple samples as multiple scenarios to construct a deterministic equivalence problem. To improve solution efficiency, a relaxation approximation method and a distributed optimization strategy are used to decompose the large-scale problem into multiple small-scale subproblems that are solved in parallel. This approach significantly reduces computational complexity, enabling the stochastic controller to be applied in real time.

[0264] The robustness of stochastic controllers can be enhanced through risk-sensitive optimization. Traditional expected value optimization may overlook low-probability extreme cases. Therefore, a conditional risk value metric is introduced to address the risk at the tails of the distribution. The objective function is modified to a weighted combination of expected value and conditional risk value, with the weight parameters dynamically adjusted based on the state. During safety-critical phases (such as when the battery is nearing full charge or the temperature is high), risk sensitivity is increased to improve safety; during normal operation, risk sensitivity is decreased to optimize performance.

[0265] To address the uncertainty propagation problem in multi-step forecasting, a closed-loop forecasting strategy is adopted. Traditional open-loop forecasting assumes that future control inputs will be executed as planned, neglecting the possibility of feedback adjustments, leading to excessive accumulation of uncertainty. Closed-loop forecasting considers the feedback adjustment capability at future moments, models future control decisions through a pre-defined feedback strategy, obtains a more accurate estimate of uncertainty propagation, and avoids overly conservative control decisions.

[0266] The computational efficiency of stochastic controllers is a critical consideration for real-time applications. Solving the entire stochastic optimization problem within each control cycle can be computationally burdensome. Therefore, an approximate dynamic programming approach is employed, decomposing the multi-step optimization problem into a series of single-step problems and connecting the decisions at each step using value function approximation techniques. Specifically, a neural network is used to approximate the future cost function, simplifying the optimization problem, which originally required complete time-domain prediction, into a short-time-domain problem, significantly reducing computational complexity while maintaining long-term optimality.

[0267] The final stochastic controller performs excellently under uncertainties and disturbances, effectively handling variations in model parameters (such as battery internal resistance fluctuations), increased measurement noise, and external disturbances (such as grid fluctuations). Compared to a deterministic controller, the stochastic controller exhibits a more than 50% reduction in control performance degradation under high uncertainty environments, demonstrating significant robustness advantages. The controller is implemented on an embedded platform, employing parallel computing and approximation techniques to keep computation time within 20ms, meeting real-time control requirements.

[0268] Step S4.5.3: Perform sigmoid function mapping and dynamic weight calculation on the prediction uncertainty metric to achieve adaptive fusion of deterministic and stochastic controllers.

[0269] Deterministic and stochastic controllers each have their advantages: the former provides precise control and fast response under nominal conditions, while the latter provides robustness and risk management in uncertain environments. This step designs an adaptive fusion mechanism that dynamically adjusts the weights of the two controllers based on the degree of uncertainty of the current state to achieve "optimal two-boundary" control performance. The core of adaptive fusion is dynamic weight calculation based on predicted uncertainty.

[0270] First, a predictive uncertainty metric is defined to quantify the level of uncertainty in the current state prediction. This metric is based on the output of the prediction model and comprehensively considers the predictive uncertainties of multiple key state variables. The specific calculation uses a weighted norm form, incorporating the prediction variance and covariance of each state variable, and reflects the importance of different state variables through a weight matrix. For charging, the uncertainties in voltage and current predictions typically have higher weights, while the uncertainties in temperature and state of charge predictions have lower weights, reflecting their relative importance in control decisions.

[0271] Forecast uncertainty measurement requires normalization to eliminate the influence of dimensions and limit it to a reasonable range. Normalization employs an adaptive method based on the statistical characteristics of historical observations. Specifically, it maintains a sliding window statistic of forecast uncertainty (such as the mean and standard deviation of the past 100 control periods) and then calculates the normalized uncertainty. This normalization ensures that the normalized values ​​are roughly distributed within a reasonable range under normal operating conditions, facilitating subsequent processing.

[0272] Normalized uncertainty is mapped into fusion weights using the sigmoid function. The sigmoid function, with its smooth S-shaped curve, can map inputs of any range to the 0-1 interval, making it ideal as a weight calculation function. The adjustment parameters in the function control the steepness and midpoint of the response curve, respectively. A larger steepness parameter makes the weights more sensitive to changes; a larger midpoint parameter indicates a larger region where a deterministic controller is preferred.

[0273] This setup ensures that both controllers contribute significantly at normal uncertainty levels; at low uncertainty, the deterministic controller is used almost entirely; and at high uncertainty, the stochastic controller is relied upon primarily. To prevent frequent fluctuations in control weights from causing control instability, a smoothing mechanism is introduced. The fused weights used in practice are the low-pass filtered results of the original weights, ensuring gradual changes in control weights and avoiding control jitter caused by abrupt changes.

[0274] The final control output is a weighted combination of the outputs of the deterministic and stochastic controllers. This combination method favors precise control in low-uncertainty environments and robust control in high-uncertainty environments, automatically adapting to different operating conditions. In addition to adaptive fusion based on predicted uncertainty, an adjustment mechanism based on the control stage is also introduced. Control requirements and uncertainty characteristics differ at different stages of the charging process, necessitating corresponding adjustments to the fusion strategy.

[0275] For example, in the initial stage of charging, battery state assessment may be inaccurate and uncertain, so the weight of the stochastic controller is increased; in the stable stage of constant current charging, behavior is more deterministic, so the weight of the deterministic controller is increased; in the transition stage from constant current to constant voltage, dynamic changes are significant, so the weight of the stochastic controller is appropriately increased to improve control robustness. This stage-based adjustment is achieved by modifying the parameters of the sigmoid function, setting different parameter values ​​for different charging stages.

[0276] The adaptive fusion mechanism also considers the controller consistency problem. When the outputs of two controllers are significantly different (e.g., opposite directions), simple weighting may lead to suboptimal results. To address this, a consistency check mechanism is introduced, calculating the cosine similarity between the two control outputs. When the similarity is less than a threshold, it indicates that the controller outputs are inconsistent. In this case, simple weighting is not used; instead, the output of one controller is selected based on historical performance, or a more conservative compromise strategy is adopted to avoid insufficient control caused by mutual cancellation.

[0277] To evaluate and optimize the fusion strategy, an adaptive performance monitoring mechanism was designed. Performance metrics (such as tracking error, energy consumption, and stability) of the two controllers under different uncertainty levels are continuously recorded, establishing a performance-uncertainty mapping relationship. Based on this data, the sigmoid function parameters are updated periodically to optimize the fusion strategy. This closed-loop optimization ensures that the fusion mechanism can adapt to long-term changes in characteristics, such as battery aging or seasonal environmental changes.

[0278] The final adaptive fusion mechanism intelligently balances control precision and robustness, performing excellently under various operating conditions. Compared to using either controller alone, the fusion strategy improves overall performance by 15-25%, with particularly significant advantages in scenarios where uncertainty levels vary considerably. The computational overhead of the fusion mechanism is minimal, mainly involving uncertainty metric calculation and sigmoid mapping, and its impact on real-time performance is negligible, making it suitable for embedded implementations.

[0279] Step S4.5.4: Based on real-time requirements and control algorithm complexity, the control task is divided into layers and the execution cycle is allocated to form a hierarchical execution strategy that includes a fast control layer, a predictive control layer and an adaptive layer.

[0280] Real-time implementation of hybrid predictive control strategies faces challenges related to computational resources and time constraints. To meet control requirements at different time scales while ensuring efficient execution of computational tasks with limited resources, this step designs a hierarchical execution strategy, dividing control tasks into multiple levels according to time scale and functional characteristics, and rationally allocating computational resources and execution cycles.

[0281] The hierarchical execution strategy comprises three main layers: a fast control layer, a predictive control layer, and an adaptive layer, each responsible for control tasks of different time scales and complexities. This hierarchical design follows the "divide and conquer" principle, breaking down complex control problems into multiple relatively independent but collaborative sub-problems, thereby improving overall efficiency and reliability.

[0282] The fast control layer is the lowest level control loop, interacting directly with the hardware interface to execute basic regulation and control functions. Its main characteristics include: short execution cycle (typically 20-100μs), ensuring timely response to rapid changes; simple control algorithm, mainly implementing basic current and voltage loop control, using simplified models and pre-calculated control tables, resulting in a low computational burden; deterministic execution, employing fixed-cycle scheduling to ensure strict determinism in control timing; and safety features, implementing basic protection functions such as overcurrent and overvoltage protection.

[0283] Fast control layers are typically implemented in hardware on dedicated controllers (such as digital signal processors or field-programmable gate arrays), using assembly language or optimized C code to minimize execution time. Control algorithms employ lookup table interpolation or simplified calculation methods to avoid complex mathematical operations. For example, frequency control of an LLC resonant converter can be achieved through a pre-calculated frequency-gain table, directly looking up the control output based on the current error and operating point, without requiring online optimization calculations.

[0284] The predictive control layer is the core control layer, implementing model-based predictive control functions. Its main characteristics include: moderate execution cycle (typically 10-100ms), balancing control performance and computational burden; high algorithm complexity, implementing complete nonlinear and stochastic model predictive control algorithms, including state prediction, optimization solving, and fusion decision-making; priority scheduling, employing a priority scheduling mechanism to ensure timely completion of critical computational tasks; and forward-looking control, optimizing future control sequences based on predictive models to achieve advanced control objectives.

[0285] Predictive control layers are typically implemented on the main controller (such as a high-performance microprocessor or embedded computing platform), using C / C++ and numerical optimization libraries. To meet real-time requirements, several optimization techniques are employed: computational task decomposition, breaking down large optimization problems into smaller problems for parallel solution; pre-computation, utilizing idle time to pre-compute possible control schemes; approximate solutions, using approximate algorithms when time is tight, sacrificing a small amount of accuracy for faster computation; and early stopping, terminating iterations early when the solution reaches sufficient accuracy or the time budget is exhausted.

[0286] The adaptive layer is the highest-level control layer, responsible for long-term optimization and parameter tuning. Its main characteristics include: long execution cycle (typically 1-10 seconds), handling slowly changing characteristics; variable computational complexity, dynamically adjusting computational depth and accuracy based on resource availability; background execution, running with low priority during idle periods without affecting critical control functions; and adaptive optimization, updating model parameters, control weights, and optimization objectives to adapt to changes.

[0287] The adaptive layer typically shares the computing platform with the predictive control layer, but employs a task scheduling mechanism to ensure that it does not interfere with critical control functions. Its functions include model parameter updates, control performance evaluation, fusion strategy optimization, and long-term goal planning. These tasks are usually computationally intensive but do not require strict real-time performance; they can be completed in stages over multiple control cycles or executed centrally when the load is low.

[0288] The three control layers collaborate through data sharing and command interfaces. The adaptive layer provides the predictive control layer with updated model parameters and control strategies; the predictive control layer provides the fast control layer with optimized control sequences and reference trajectories; and the fast control layer feeds back the actual execution results and status to the upper layers. This bidirectional information flow ensures coordination among the layers, forming a complete control loop.

[0289] To ensure the reliability of hierarchical control, a fault detection and handling mechanism was designed. When the higher-level control cannot provide control output in a timely manner due to computational delays or algorithmic issues, the lower-level control can operate autonomously and maintain basic functions. For example, the fast control layer can continue execution based on the most recently valid control parameters, or switch to a preset safety control mode to ensure stable operation until the higher-level control returns to normal.

[0290] The implementation of the tiered execution strategy takes into account the characteristics of the hardware platform. For single-processor-based systems, priority-based preemptive scheduling is used to ensure that high-priority tasks can be executed on time. For multi-core processors, core isolation technology is used to allocate different control layers to different processor cores to avoid resource contention. For heterogeneous computing platforms, tasks are allocated according to computing characteristics, with control tasks requiring deterministic execution assigned to dedicated hardware and complex optimization tasks assigned to general-purpose processors.

[0291] This hierarchical execution strategy enables the hybrid predictive control algorithm to be implemented efficiently with limited computing resources, meeting control requirements at different time scales. Actual testing shows that this strategy runs stably on embedded platforms, with jitter in the fast control layer less than 5μs, computational latency in the predictive control layer controlled within 50ms, and background tasks in the adaptive layer not affecting foreground control performance. Overall, it meets the real-time control requirements for industrial forklift charging.

[0292] Step S4.5.5: Based on the adaptive fusion and the hierarchical execution strategy, dynamically adjust the control weights and optimize the control sequence generation process to obtain the hybrid predictive control strategy.

[0293] The preceding steps designed a deterministic controller, a stochastic controller, an adaptive fusion mechanism, and a hierarchical execution strategy. This step integrates these components into a complete hybrid predictive control strategy, enabling dynamic adjustment of control weights and generation of optimized control sequences. This is the final integration stage of the hybrid predictive control strategy, ensuring that all components work together to achieve optimal performance.

[0294] First, a dynamic adjustment mechanism for control weights is established. Control weights comprise two levels: one is the fusion weight of the deterministic and stochastic controllers, determined by an adaptive fusion mechanism; the other is the weight coefficients within the control objective, such as tracking error weights, control smoothness weights, and efficiency weights. These internal weights need to be dynamically adjusted based on the charging stage, state, and control performance to adapt to the control requirements of different operating conditions.

[0295] Internal weight adjustments are based on multiple factors: During the charging phase, the tracking error weight is increased in the initial charging stage to ensure rapid attainment of the target operating point; during the constant current charging phase, the efficiency weight is increased to optimize energy conversion efficiency; and during the constant voltage charging phase, the control smoothness weight is increased to avoid the impact of voltage fluctuations on battery life. Regarding battery status, when approaching constraint boundaries, the weight of the corresponding constraints is increased to ensure safety margins; when external disturbances are detected, the robustness-related weight is increased to improve anti-interference capabilities. Control performance is continuously monitored, and when a certain indicator deteriorates significantly, the corresponding weight is increased for targeted optimization. User preferences are also considered, adjusting the weight configuration according to the user-defined charging mode (e.g., standard mode, fast mode, energy-saving mode) to meet the needs of different application scenarios.

[0296] Weight adjustment employs a combination of rule-based and learning-based approaches. Rule-based adjustment uses pre-defined condition-action mappings, such as "increase the temperature constraint weight if the battery temperature is too high." Learning-based adjustment analyzes historical data to identify the correlation patterns between optimal weight configuration and operating conditions, gradually optimizing the adjustment strategy. The combined use of these two methods ensures basic safety and performance through rules, while learning provides long-term optimization and adaptability.

[0297] Secondly, the generation and processing of the optimized control sequence are implemented. Based on the outputs of the deterministic and stochastic controllers, and adaptive fusion weights, the final control sequence is calculated. The control sequence includes control values ​​at multiple time steps, but only the value of the current step is executed immediately; the values ​​of subsequent steps are used as references and updated in the next control cycle. The generation of the control sequence needs to consider the following factors: control consistency, ensuring that the generated control sequence is smooth and coherent, avoiding oscillations caused by drastic changes; constraint satisfaction, ensuring that the control sequence satisfies constraints under various operating conditions, including hard and soft constraints; real-time performance, the control sequence generation process needs to be completed within the allocated time budget to ensure timely control; and robustness guarantee, the control sequence should have sufficient robustness to cope with prediction errors and external disturbances.

[0298] The control sequence generation employs a hierarchical processing approach, consistent with the hierarchical execution strategy. The fast control layer is responsible for the real-time execution of the basic control sequence; the predictive control layer computes the optimized control sequence and decomposes it into commands executable by the fast control layer; the adaptive layer monitors control performance and adjusts control parameters and strategies. This hierarchical processing ensures that the control sequence maintains consistency and optimality across different time scales.

[0299] To improve the quality and robustness of the control sequence, several enhancement techniques are introduced: predictive correction, which corrects the output of the predictive model based on historical prediction error patterns to reduce the impact of deviations; constraint softening, which uses soft constraints for non-critical constraints to increase control flexibility and avoid performance degradation caused by minor constraints; feedback enhancement, which embeds a feedback mechanism into the predictive control framework to compensate for prediction errors and external disturbances in a timely manner, thereby improving control robustness; and multi-model switching, which maintains multiple control models and automatically switches them according to the state and operating conditions to adapt to the control requirements of different operating conditions.

[0300] The final hybrid predictive control strategy is a complete closed-loop control, including state estimation, predictive modeling, optimization calculation, control fusion, and execution supervision. The control process includes: data acquisition and preprocessing (acquiring state data, filtering and preprocessing it to provide high-quality input for subsequent control calculations); state estimation and prediction (estimating the complete state based on current observations and historical data, and predicting future state evolution, including quantifying prediction uncertainty); controller calculation (calculating the control outputs of the deterministic and stochastic controllers, including current control values ​​and future control sequences); adaptive fusion (calculating fusion weights based on prediction uncertainty and state to generate the final control decision); hierarchical execution (decomposing the control decision into execution commands at different time scales and assigning them to the corresponding control layers); and performance monitoring (continuously monitoring control performance, collecting data for adaptive optimization, and triggering control strategy adjustments when necessary).

[0301] This complete control loop executes cyclically at different frequencies across various control layers, forming multi-timescale collaborative control. The fast control layer executes basic control functions at a high frequency; the predictive control layer updates and optimizes the control strategy at a medium frequency; and the adaptive layer adjusts control parameters and strategies at a low frequency.

[0302] Through this multi-level, adaptive hybrid predictive control strategy, industrial forklift charging maintains excellent control performance under various operating conditions. Compared with traditional control methods, this strategy significantly improves control accuracy, response speed, energy efficiency, and robustness, especially when facing uncertainties and external disturbances, where its performance advantages are even more pronounced. Practical application tests show that this control strategy can improve voltage control accuracy to ±0.5% and current control accuracy to ±1%, while maintaining good performance even in the presence of disturbances, providing a strong guarantee for the efficient and reliable operation of industrial forklifts.

[0303] In one embodiment, the step of performing a charging curve optimization algorithm and parameter mapping processing based on the charging parameter optimization instructions, combined with battery charge / discharge characteristic data and a battery aging state assessment model, to achieve battery state assessment and adaptive adjustment of charging parameters includes:

[0304] Step S5.1: Perform characteristic analysis and database construction processing on battery test data of different types and aging levels to establish a battery characteristic database that includes lead-acid batteries and lithium batteries.

[0305] To support battery status assessment and adaptive adjustment of charging parameters, a database of battery characteristics covering different types and aging levels was established through systematic testing, providing data support for subsequent evaluation and optimization.

[0306] The test subjects include lead-acid batteries (AGM, gel) and lithium batteries (lithium iron phosphate, ternary lithium) commonly used in industrial forklifts, covering different capacity levels (such as 200Ah, 300Ah, etc. for 48V systems) and service life (new batteries to over 5 years). The tests include four core categories: capacity testing (measuring actual capacity and retention rate under 0.2C constant current discharge at 25℃); internal resistance testing (combining AC and DC methods to measure internal resistance under different states of charge and temperatures, and constructing a three-dimensional characteristic graph); charge and discharge characteristic testing (recording voltage curves, temperature, and efficiency under different charging modes, currents, and discharge rates, focusing on the impact of aging on charge acceptance); and aging characteristic testing (sulfation testing for lead-acid batteries, SEI film growth testing for lithium batteries, and parameter extraction using EIS, dQ / dV, etc.).

[0307] Data acquisition is conducted using high-precision equipment in a constant-temperature environment, with sampling rates improved at key points. The resulting database includes a basic characteristic library (basic parameters), a charge / discharge characteristic library (curves and efficiency data), and an aging characteristic library (characteristic parameters and trends). It employs a relational structure, supports multi-dimensional queries, and features an automatic update mechanism. This database provides a reliable foundation for battery aging assessment and charging optimization, assists in customizing optimal charging strategies, and provides data resources for future battery management technology research and development.

[0308] Step S5.2: Based on the battery characteristic database, extract features and build models to design a battery aging status assessment model that can evaluate the battery health status.

[0309] Accurately assessing battery health is a prerequisite for achieving adaptive charging. This step, based on a battery characteristic database, uses feature extraction and model building to design an aging state assessment model that can quantify battery health.

[0310] First, feature selection and extraction are performed. Through correlation analysis and other methods, key features such as capacity retention rate, internal resistance growth rate, charge acceptance capability, voltage recovery characteristics, and self-discharge rate are screened for lead-acid batteries and standardized to form feature vectors. For lithium batteries, key features such as capacity retention rate, internal resistance spectrum characteristics, voltage plateau characteristics, coulombic efficiency, and differential capacity curve characteristics are screened and similarly standardized to form feature vectors.

[0311] Based on these characteristics, a multi-model fusion evaluation system is constructed. Considering the complex nonlinearity of battery aging, a fusion strategy is adopted that includes an equivalent circuit model (ECM), a support vector regression model (SVR), and a long short-term memory network model (LSTM).

[0312] The equivalent circuit model, based on electrochemical principles, simulates the battery as a combination of components such as resistors and capacitors. Parameters are extracted by fitting charge-discharge curves to establish a mapping with the degree of aging. While the physical meaning is clear, it has limitations in handling nonlinear aging. The support vector regression model can handle nonlinear relationships in high-dimensional spaces. Using feature vectors as input and health status indicators as output, it maps features to a high-dimensional space using radial basis functions. Hyperparameters are optimized through grid search and cross-validation, balancing fitting and generalization capabilities.

[0313] The Long Short-Term Memory (LSTM) network model can handle time-series characteristics. It takes a battery's historical usage time series as input and outputs predictions of its current and future health status. The network consists of two LSTM layers with 64 hidden units each and two fully connected layers. Dropout is used to prevent overfitting, and the model is trained using the Adam optimizer and mean squared error loss function. To improve accuracy and robustness, a fusion technique combining weighted averaging and stacked ensemble is employed. First, basic weights are determined based on the model's performance on the validation set. Then, a gradient boosting decision tree is trained as a meta-learner, dynamically adjusting the weights to achieve adaptive fusion and fully leverage the strengths of each model. Model evaluation uses multiple metrics, including root mean square error (RMSE). Through hold-out and k-fold cross-validation, the fusion model achieves an RMSE of 2.8% and an R² of [missing data - likely a missing value]. 2 The RSE was 0.94, and the RMSE for lithium batteries was 3.2%. 2 The efficiency score is 0.92, which is better than the single model. This model not only outputs the State of Health (SOH) index, but also provides fine-grained aging characteristic analysis, providing a basis for charging curve optimization. Moreover, when implemented on an embedded platform, the inference time is controlled within 100ms through compression and quantization techniques, meeting the requirements for real-time evaluation.

[0314] Step S5.3: Based on the evaluation results of the battery aging state assessment model and the charging parameter optimization instructions, process the multi-objective optimization and constraints, and design a charging curve optimization algorithm that can extend battery life.

[0315] The battery charging process directly impacts its lifespan and performance, especially for aging batteries, where a suitable charging strategy can significantly slow down further aging. This step, based on the battery aging status assessment results and previously generated charging parameter optimization instructions, designs a charging curve optimization algorithm to achieve precise control of the charging process and maximize battery lifespan. First, the multiple objectives of charging curve optimization are clearly defined, including: extending battery lifespan (primary objective), maintaining charging efficiency (secondary objective), controlling charging time (constraints), and maintaining charging safety (hard constraints). These objectives involve certain conflicts and trade-offs, requiring a multi-objective optimization method to find the optimal balance.

[0316] The core of charging curve optimization is adjusting charging parameters based on battery aging conditions, including the charging current curve, voltage limits, charging termination conditions, and temperature control strategies. For lead-acid batteries, the main optimized parameters include: initial charging current, constant current stage current, constant current to constant voltage transition point voltage, constant voltage stage voltage value, charging termination current, and charging termination time. For lithium-ion batteries, the main optimizations are: constant current stage current, constant voltage stage voltage, constant current to constant voltage transition point, charging termination current, and charging pulse parameters (e.g., when using pulse charging).

[0317] The optimization algorithm employs a model-based multi-objective optimization framework, combining battery aging mechanism models and experimental data to predict the impact of different charging parameters on battery performance and lifespan. For lead-acid batteries, mathematical models of aging mechanisms such as sulfation formation, active material shedding, and grid corrosion are established; for lithium batteries, models of aging mechanisms such as SEI film growth, lithium deposition, and structural stress are established. These models can predict the aging rate under specific charging conditions, providing a theoretical basis for optimization.

[0318] Multi-objective optimization employs an improved non-dominated sorting genetic algorithm (NSGA-III), which effectively handles multi-objective problems and seeks Pareto optimal solutions. The algorithm process includes: initializing the population (each individual represents a set of charging parameters); evaluating multiple objective function values ​​for each individual (such as predicted battery life extension rate, charging efficiency, charging time, etc.); selection based on non-dominated sorting and reference point methods; generating a new population through crossover and mutation operations; and iterative optimization until convergence or the maximum number of iterations is reached. To improve optimization efficiency, an adaptive mutation rate and elitist retention strategy are employed to accelerate the convergence process.

[0319] Constraint handling is crucial in the optimization process. Key constraints include: safety constraints (such as maximum charging current, maximum battery temperature, and maximum charging voltage), time constraints (charging time must not exceed specified limits), and hardware constraints (power and accuracy limitations of the charging equipment). Constraint handling employs a penalty function method, transforming the degree of constraint violation into a penalty term in the objective function to guide the algorithm in searching the feasible solution space. For safety-critical constraints (such as temperature limits), a higher penalty weight is applied to ensure the safety of the solution.

[0320] To accommodate batteries with varying degrees of aging, a parameter mapping mechanism based on aging state was designed. The battery state of health (SOH) was divided into multiple intervals (e.g., 90-100%, 70-90%, 50-70%, <50%), with differentiated optimization strategies designed for each interval. For example, for batteries in good health (SOH>90%), optimization focuses on charging efficiency and time; for moderately aged batteries (SOH 70-90%), a balance is struck between extending lifespan and charging efficiency; and for severely aged batteries (SOH<70%), the primary goal is to extend remaining lifespan while significantly reducing charging stress.

[0321] Specific optimization strategies were designed to address the different characteristics of lead-acid and lithium batteries. For lead-acid batteries, the following aspects were optimized: reducing the initial charging current to decrease sulfation and gas generation; employing a multi-stage charging current that is dynamically adjusted based on the depth of charge; extending the absorption phase time to improve charging completion; adding an equalization charging phase to reduce imbalance between individual battery cells; and controlling the charging temperature to prevent overheating and accelerated aging.

[0322] For lithium-ion batteries, optimization strategies include: reducing the constant current stage current according to the degree of aging to reduce the risk of lithium deposition; slightly reducing the constant voltage stage voltage (e.g., from 4.2V to 4.15V or 4.1V) to significantly extend cycle life; adopting multi-stage constant current charging to replace the traditional constant current-constant voltage mode; implementing temperature-compensated charging voltage to automatically adjust the charging voltage in low and high temperature environments; and for severely aged batteries, using pulse charging technology to reduce polarization effects and lithium deposition.

[0323] The optimization algorithm outputs a complete set of charging parameters and charging curves, including current and voltage setpoints for each stage, stage transition conditions, and temperature compensation coefficients. These parameters are converted into executable control commands for the charger through a charging parameter mapping function, directly guiding the charging process. To verify the optimization effect, an accelerated aging test scheme was designed to compare the battery aging rates under standard charging and optimized charging. Experimental results show that for moderately aged lead-acid batteries, the optimized charging strategy can extend the remaining lifespan by 20-35%; for moderately aged lithium batteries, it can extend the remaining lifespan by 15-25%, while only increasing the charging time by 10-15%, demonstrating the effectiveness of the algorithm.

[0324] The final charging curve optimization algorithm not only considers the current aging state of the battery, but also its aging trend and operating environment, achieving truly personalized charging optimization. Implemented on an embedded system, the algorithm's computation time is controlled within one second. It can quickly generate optimal charging parameters before each charging session, providing a "tailor-made" charging solution for the battery, maximizing battery service life, and reducing the operating costs of industrial forklifts.

[0325] Step S5.4: Perform parameter mapping and dynamic adjustment processing on the charging curve optimization algorithm to achieve real-time adaptive control of charging parameters and obtain the adjusted charging parameters.

[0326] The theoretically optimal parameters generated by the charging curve optimization algorithm need to be transformed into control parameters during the actual charging process and dynamically adjusted according to the real-time status to cope with various changes and disturbances. This step designs a parameter mapping and dynamic adjustment mechanism to achieve real-time adaptive control of charging parameters, ensuring that the optimization effect is fully realized during the actual charging process. First, a parameter mapping relationship is established to convert the abstract parameters output by the optimization algorithm into control parameters that the charger can directly execute. The mapping process takes into account the characteristics and limitations of the charging equipment, including current regulation accuracy, voltage control range, power limitations, and control delay.

[0327] The parameter mapping employs a combination of piecewise linear mapping and lookup tables to ensure accuracy and computational efficiency. For example, for multi-stage charging current, the theoretically optimal current value is mapped to the discrete current levels supported by the charger; for voltage parameters, compensation adjustments are made considering charging cable voltage drop and measurement errors; and for time parameters, quantization is performed considering the control execution cycle. The mapping function is experimentally calibrated to ensure consistency between theoretical and actual execution parameters, preserving the optimization effect to the maximum extent.

[0328] The core of real-time adaptive control is a dynamic parameter adjustment mechanism, which adjusts preset charging parameters online based on real-time status and environmental changes during the charging process. Dynamic adjustment is based on closed-loop feedback control principles and mainly considers the following real-time information: battery temperature changes (automatically reducing charging current or pausing charging when the battery temperature exceeds a preset threshold or the temperature rise rate is too rapid to prevent overheating damage); voltage response characteristics (monitoring the battery's voltage response to charging current, identifying abnormal polarization behavior, and adjusting the charging strategy in a timely manner); current acceptance capability (observing the battery's current drop curve during the constant voltage phase, evaluating charging acceptance capability, and adjusting charging termination conditions); and ambient temperature changes (dynamically adjusting the temperature compensation coefficients of charging voltage and current based on ambient temperature fluctuations).

[0329] Dynamic adjustment employs a combination of rule-based and model-predictive methods. Rule-based adjustment uses a pre-defined condition-action mapping table, such as "if the battery temperature rise rate > 1℃ / min, then reduce the charging current by 20%." These rules are simple, intuitive, and responsive, suitable for handling clearly defined abnormal situations. Model-predictive adjustment, on the other hand, is based on a battery state prediction model that predicts the future state evolution under current charging parameters in real time. When the prediction results indicate a potential adverse situation, the charging parameters are adjusted in advance to achieve preventative control.

[0330] To improve the robustness of adaptive control, a multi-layered safety mechanism was designed. The first layer is parameter boundary checking to ensure that all adjusted parameters are within safe limits. The second layer is trend monitoring to detect the changing trends of key parameters (such as temperature and voltage) and provide early warnings of potential risks. The third layer is anomaly detection, which identifies abnormal patterns in the charging process based on statistical methods and model predictions. The fourth layer is emergency protection, which immediately interrupts charging and issues an alarm when a serious anomaly is detected to prevent safety accidents.

[0331] Adaptive control also considers the specific needs of different charging stages. In the initial charging stage, the focus is on accurately assessing the battery's initial state, including open-circuit voltage, initial temperature, and initial internal resistance, to provide a benchmark for subsequent charging. In the constant current stage, attention is paid to the battery's voltage rise characteristics and temperature changes, and the current value is adjusted in a timely manner. In the constant voltage stage, the charging current decline curve is monitored to optimize the charging termination conditions. In the final charging stage, the charging completion rate and battery state changes are evaluated to provide a reference for the next charging.

[0332] To adapt to the specific needs of different battery types, a battery type adaptive mechanism was designed. For lead-acid batteries, special attention is paid to gas generation and temperature distribution. The charging current is adjusted in real time to control the gas generation rate, and the temperature differences in various parts of the battery are monitored to prevent local overheating. For lithium-ion batteries, the voltage plateau characteristics and current acceptance capability at the end of the charging process are monitored to identify potential lithium deposition risks. If necessary, charging is terminated early or the charging voltage is reduced.

[0333] The execution of real-time adaptive control adopts a hierarchical architecture, consistent with the hierarchical execution strategy in step S4.5.4. The fast control layer performs basic parameter adjustments, such as current fine-tuning and protection response, at a high frequency (10-100ms); the mid-level control layer performs strategy adjustments, such as stage transitions and parameter re-optimization, at a medium frequency (1-10s); and the high-level control layer performs global evaluations, such as charging progress analysis and effect prediction, at a low frequency (30-60s). This hierarchical execution ensures both the real-time nature and comprehensiveness of the control.

[0334] Through parameter mapping and dynamic adjustment, a set of charging parameters that change in real time during the charging process is obtained, including the current charging current value, voltage limit value, stage transition conditions, and termination criteria. These parameters directly control the charger's output, achieving precise control of the charging process. Practical application tests show that, compared to static parameter control, real-time adaptive control can better cope with changes in battery state and environmental disturbances, resulting in a smoother charging process, more precise temperature rise control, a 5-10% increase in charging efficiency, and a further 3-8% extension of battery life, while maintaining high charging completion rate and safety reliability.

[0335] Step S5.5: Based on the adjusted charging parameters, analyze and process the comparative experiment and life model to verify the effect on extending battery life and obtain the battery state assessment and adaptive adjustment of charging parameters.

[0336] To verify the actual effectiveness of battery state assessment and adaptive adjustment of charging parameters, a verification scheme combining comparative experiments and lifetime model analysis was designed, and a multi-level evaluation system was adopted to ensure the reliability of the conclusions.

[0337] Comparative experiments selected battery samples from the same batch, with the same specifications and aging levels, and divided them into an experimental group (adaptive charging scheme) and a control group (traditional constant current-constant voltage scheme). Three types of tests were conducted: Short-term performance tests showed that for moderately aged lead-acid batteries (SOH approximately 75%), the adaptive charging scheme improved charging efficiency by 7.2%, reduced maximum temperature fluctuation by 5.3℃, increased charging time by 12.5%, and improved completion rate by 4.8%; for moderately aged lithium batteries (SOH approximately 80%), the efficiency improved by 5.5%, the temperature fluctuation by 3.8℃, the charging time increased by 8.7%, and the completion rate improved by 3.2%. Accelerated aging tests, after 300 cycles (equivalent to 1-2 years of aging), for lead-acid batteries with an initial SOH of 85%, the SOH in the control group decreased to 62%, and in the experimental group it decreased to 71%, resulting in a relative lifespan extension of 28.7%; for lithium batteries with an initial SOH of 90%, the SOH in the control group decreased to 76%, and in the experimental group it decreased to 82%, resulting in a relative lifespan extension of 42.9%. In a real-world application test conducted over a 6-month period in an industrial forklift environment, the experimental group showed a 18-25% lower battery capacity decay rate, a 15-22% lower internal resistance growth rate, a 12-18% increase in operating time, and a 20-30% reduction in maintenance requirements.

[0338] A lifespan model analysis established an aging mechanism model, correlating charging conditions with aging rates. Monte Carlo simulations showed that adaptive charging can extend the lifespan of lead-acid batteries by 25-40% and lithium batteries by 20-35%, consistent with experimental data. In summary, this solution can significantly extend battery life and reduce costs, making it suitable for industrial forklift applications. Its modular design offers good adaptability and upgrade potential.

[0339] like Figure 3 As shown, the present invention also provides a power control device for an industrial forklift charger, comprising:

[0340] LLC resonant converter power conversion module 601 is used to acquire industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including a primary half-bridge circuit, a resonant network and a secondary rectifier circuit.

[0341] The magnetic component optimization design module 602 is used to construct a multi-objective optimization stochastic surrogate model for magnetic component parameters based on the operating frequency range and power density requirements of the LLC resonant converter through a probabilistic learning algorithm and a stochastic surrogate model. The module then optimizes the core size, air gap length, and winding structure of the stochastic surrogate model to obtain a high-frequency, low-loss magnetic component.

[0342] The random neural control module 603 is used to integrate the high-frequency low-loss magnetic components into the LLC resonant converter, and to perform random neural network structure design and control barrier function construction based on the physical characteristics and operating data of the LLC resonant converter, thereby realizing the random neural control barrier function algorithm, performing dynamic power control on the charging process, and generating a dynamic power control strategy.

[0343] The data-driven predictive control module 604 is used to process the real-time data of voltage, current and temperature during the charging process based on the dynamic power control strategy, perform noise-resistant data filtering algorithm and predictive model construction on the real-time data, realize the data-driven hybrid predictive control strategy, and generate charging parameter optimization instructions.

[0344] The battery adaptive charging module 605 is used to perform charging curve optimization algorithms and parameter mapping processing based on the charging parameter optimization instructions, combined with battery charging and discharging characteristic data and battery aging state assessment models, so as to realize battery state assessment and adaptive adjustment of charging parameters.

[0345] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A power control method for an industrial forklift charger, characterized in that, Includes the following steps: Acquire industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including a primary half-bridge circuit, a resonant network and a secondary rectifier circuit. Based on the operating frequency range and power density requirements of the LLC resonant converter, a stochastic surrogate model for multi-objective optimization of magnetic component parameters is formed by constructing a probabilistic learning algorithm and a stochastic surrogate model. The stochastic surrogate model for multi-objective optimization of magnetic component parameters is then optimized in terms of core size, air gap length and winding structure to obtain a high-frequency, low-loss magnetic component. The high-frequency, low-loss magnetic components are integrated into the LLC resonant converter. Based on the physical characteristics and operating data of the LLC resonant converter, a stochastic neural network structure is designed and a control barrier function is constructed to realize the stochastic neural control barrier function algorithm, which performs dynamic power control during the charging process and generates a dynamic power control strategy. Based on the real-time data of voltage, current and temperature during the charging process under the execution of the dynamic power control strategy, the real-time data is processed by noise-resistant data filtering algorithm and prediction model construction to realize a data-driven hybrid predictive control strategy and generate charging parameter optimization instructions. Based on the charging parameter optimization instructions, and combined with battery charging and discharging characteristic data and battery aging state assessment model, a charging curve optimization algorithm and parameter mapping processing are performed to achieve battery state assessment and adaptive adjustment of charging parameters.

2. The method according to claim 1, characterized in that, The process involves acquiring industrial forklift charging demand data and LLC resonant topology characteristic parameters, performing resonant parameter optimization calculations and frequency modulation control design on the industrial forklift charging demand data and the LLC resonant topology characteristic parameters, and constructing an LLC resonant converter including a primary half-bridge circuit, a resonant network, and a secondary rectifier circuit, including: Circuit topology analysis and component selection were performed on the industrial forklift charging demand data to design an LLC converter topology. Based on the LLC converter topology and the LLC resonant topology characteristic parameters, the parameters of resonant inductance, resonant capacitance and magnetizing inductance are determined by resonant frequency calculation and quality factor optimization. By selecting control strategies and designing circuit parameters based on the AC input characteristics, a PFC preamplifier circuit with a power factor greater than 0.99 is achieved, resulting in a stable DC input for the PFC preamplifier circuit. Based on the parameters of the resonant inductor, resonant capacitor, and magnetizing inductor, and the stable DC input of the PFC pre-amplifier circuit, the operating frequency range is determined and the modulation strategy is designed to achieve zero-voltage switching and zero-current switching operation across the entire load range, thus obtaining the LLC resonant converter.

3. The method according to claim 1, characterized in that, Based on the operating frequency range and power density requirements of the LLC resonant converter, a stochastic surrogate model for multi-objective optimization of magnetic component parameters is constructed using a probabilistic learning algorithm and a stochastic surrogate model, including: Based on the operating frequency range and power density requirements of the LLC resonant converter, the material performance of PC40 ferrite material was compared and the loss mechanism was analyzed to determine the basic material characteristics of the magnetic core. Based on the aforementioned core material property data, a stochastic surrogate model capable of predicting core performance parameters is constructed through Gaussian process regression algorithm and kernel function parameter optimization. The random surrogate model capable of predicting magnetic core performance parameters is subjected to cross-validation evaluation and generalization ability testing to obtain a random surrogate model for multi-objective optimization of the magnetic component parameters.

4. The method according to claim 1, characterized in that, The stochastic surrogate model for multi-objective optimization of the magnetic element parameters is used to optimize the core size, air gap length, and winding structure to obtain a high-frequency, low-loss magnetic element, including: Based on the stochastic surrogate model and design constraints of the multi-objective optimization of the magnetic element parameters, the Pareto optimal solution search process is performed on the multi-objective Bayesian optimization algorithm to obtain the optimal parameter combination of core size, air gap length and winding structure. Based on the core size and air gap length in the optimal parameter combination, and combined with the conductor structure design and winding process optimization of the high-frequency current distribution characteristics, a Litz wire winding structure with low AC resistance and low eddy current loss is realized, and the winding design parameters are obtained. Based on the core geometry determined by the optimal parameter combination and the winding design parameters, loss analysis and calculation are performed on the core loss and copper loss to obtain the loss calculation results. Based on the loss calculation results, the thermal resistance network is analyzed and the cooling structure is designed to ensure that the temperature rise of the magnetic component under full load conditions is controlled within a safe range, thus obtaining the high-frequency low-loss magnetic component.

5. The method according to claim 1, characterized in that, The process of integrating the high-frequency, low-loss magnetic components into the LLC resonant converter, and designing a stochastic neural network structure and constructing a control barrier function based on the physical characteristics and operating data of the LLC resonant converter to realize a stochastic neural control barrier function algorithm includes: Based on the physical characteristics and operating data of the LLC resonant converter, a system dynamic model containing parameter uncertainties is established through system identification methods and model construction. The power control requirements of the system dynamic model are processed by Bayesian neural network structure design and variational inference method to construct a stochastic neural network model that can handle uncertainty. Based on the system safety constraints and operational boundary conditions determined by the system dynamic model, the barrier function is designed and the safety domain is defined to construct a control barrier function that ensures that the system state is always within the safety region. The random neural network model and the control barrier function are subjected to control law design and parameter tuning to realize the random neural control barrier function algorithm.

6. The method according to claim 1, characterized in that, The real-time data of voltage, current, and temperature during the charging process based on the dynamic power control strategy are processed using an anti-noise data filtering algorithm and a prediction model to realize a data-driven hybrid predictive control strategy and generate charging parameter optimization instructions, including: Based on a network of voltage, current, and temperature sensors, the signal is conditioned and the data is preprocessed to obtain real-time data of the charging process. Statistical analysis and feature extraction are performed on the real-time data to establish a charging system noise model that includes electromagnetic interference noise, power fluctuation noise, and sensor noise. Based on the charging system noise model, effective filtering of charging data is achieved through unscented Kalman filter design and adaptive adjustment of the covariance matrix. The filtered historical data is processed by a bidirectional LSTM network structure and Monte Carlo Dropout technique to build a prediction model that can predict the evolution of charging state and quantify prediction uncertainty. Based on the aforementioned prediction model, nonlinear model predictive control and stochastic model predictive control are combined to achieve a hybrid predictive control strategy that combines deterministic and stochastic control. Based on the control decision of the hybrid predictive control strategy, the charging voltage setpoint, charging current setpoint, and switching frequency are optimized and calculated to generate the charging parameter optimization command.

7. The method according to claim 6, characterized in that, The process involves designing a bidirectional LSTM network structure and processing the filtered historical data using Monte Carlo Dropout technology to construct a predictive model capable of predicting charging state evolution and quantifying prediction uncertainty, including: Based on the battery charge and discharge characteristic data, a second-order RC equivalent circuit physical model describing the battery dynamic characteristics is established through equivalent circuit parameter identification and model fitting. Based on filtered historical data, bidirectional LSTM layer design and fully connected layer configuration are performed on the normalized multivariate time series to establish a neural network structure that captures forward and backward temporal dependencies, and generate a data-driven network architecture. Based on the battery parameters of the second-order RC equivalent circuit physical model and the data-driven network architecture, the physical constraint embedding and neural network parameter optimization are fused to form a hybrid prediction model structure that combines physical constraints and data-driven approaches. Based on the hybrid prediction model structure, the Monte Carlo Dropout technique is subjected to multiple forward propagation and probability distribution calculations to obtain the prediction model that can predict the evolution of charging state and quantify prediction uncertainty.

8. The method according to claim 7, characterized in that, The process of combining nonlinear model predictive control and stochastic model predictive control based on the prediction model to achieve a hybrid predictive control strategy that combines deterministic and stochastic control includes: Based on a nonlinear model predictive control framework, optimization and constraint processing are performed on the prediction time domain to realize a deterministic controller for handling nominal operating conditions; Based on the stochastic model predictive control framework, the probability constraints and expected values ​​are optimized to realize a stochastic controller that can handle uncertainties and disturbances. The sigmoid function mapping and dynamic weight calculation are applied to the measurement of prediction uncertainty to achieve adaptive fusion of deterministic and stochastic controllers. Based on the system's real-time requirements and the complexity of the control algorithm, the control task is divided into layers and the execution cycle is allocated, forming a layered execution strategy that includes a fast control layer, a predictive control layer, and an adaptive layer. Based on the adaptive fusion and the hierarchical execution strategy, the control weights are dynamically adjusted and the control sequence generation process is optimized to obtain the hybrid predictive control strategy.

9. The method according to claim 1, characterized in that, The process of optimizing charging parameters based on the charging parameter optimization instructions, combined with battery charge / discharge characteristic data and a battery aging state assessment model, involves a charging curve optimization algorithm and parameter mapping to achieve battery state assessment and adaptive adjustment of charging parameters. This includes: Characteristic analysis and database construction processing were performed on test data of batteries of different types and aging levels to establish a battery characteristic database including lead-acid batteries and lithium batteries. Based on the battery characteristic database, features are extracted and model construction is performed to design a battery aging state assessment model that can evaluate the battery health status. Based on the evaluation results of the battery aging state assessment model and the charging parameter optimization instructions, multi-objective optimization and constraints are processed to design a charging curve optimization algorithm that can extend battery life. The charging curve optimization algorithm is subjected to parameter mapping and dynamic adjustment to achieve real-time adaptive control of charging parameters and obtain the adjusted charging parameters. Based on the adjusted charging parameters, comparative experiments and lifespan models are analyzed and processed to verify the system's effect on extending battery lifespan, and to obtain the battery state assessment and adaptive adjustment of charging parameters.

10. A power control device for an industrial forklift charger, characterized in that, include: LLC resonant converter power conversion module is used to acquire industrial forklift charging demand data and LLC resonant topology characteristic parameters, perform resonant parameter optimization calculation and frequency modulation control design on the industrial forklift charging demand data and LLC resonant topology characteristic parameters, and construct an LLC resonant converter including primary half-bridge circuit, resonant network and secondary rectifier circuit. The magnetic component optimization design module is used to construct a multi-objective optimization stochastic surrogate model for magnetic component parameters based on the operating frequency range and power density requirements of the LLC resonant converter through probabilistic learning algorithms and stochastic surrogate models. The module then optimizes the core size, air gap length, and winding structure of the stochastic surrogate model to obtain a high-frequency, low-loss magnetic component. A stochastic neural control module is used to integrate the high-frequency, low-loss magnetic components into the LLC resonant converter. Based on the physical characteristics and operating data of the LLC resonant converter, a stochastic neural network structure is designed and a control barrier function is constructed to realize the stochastic neural control barrier function algorithm, which performs dynamic power control during the charging process and generates a dynamic power control strategy. The data-driven predictive control module is used to process the real-time data of voltage, current and temperature during the charging process based on the dynamic power control strategy, perform noise-resistant data filtering algorithm and predictive model construction on the real-time data, realize the data-driven hybrid predictive control strategy, and generate charging parameter optimization instructions. The battery adaptive charging module is used to perform charging curve optimization algorithms and parameter mapping processing based on the charging parameter optimization instructions, combined with battery charging and discharging characteristic data and battery aging state assessment models, so as to realize battery state assessment and adaptive adjustment of charging parameters.

Images