A buck-boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation

Abstract

The present application relates to the technical field of lithium battery charging under photovoltaic power supply, in particular to a Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation, wherein the charging stage is divided into trickle, constant current and constant voltage charging stages according to the voltage of the lithium battery; in the constant current stage, the ambient temperature, solar irradiance and maximum power point voltage data are collected, the data are preprocessed, the maximum power point voltage prediction machine learning model is trained, the compressed model is obtained by distilling the model and is deployed on the microcontroller; the real-time environmental data are input into the model to predict the maximum power point voltage as the reference voltage, the photovoltaic panel output voltage is sampled, and the innovative PWM control system is used to control the input voltage of the Buck-Boost circuit, that is, the duty cycle of the PWM pulse is controlled by the microcontroller, the input impedance of the Buck-Boost circuit is adjusted, and the photovoltaic panel output voltage, i.e. the switching power supply input voltage, is adjusted to the maximum power point voltage according to the environmental conditions; the lithium battery is dynamically switched in different charging stages according to the battery state and input power. The present application integrates machine learning into the prediction of the maximum power point voltage and is suitable for the intelligent charging scene of the lithium battery under photovoltaic power supply.

Description

Technical Field

[0001] This invention relates to the field of lithium battery charging technology under photovoltaic power supply, specifically a Buck-Boost intelligent charging method for lithium batteries based on ML-MPPT and dynamic reference voltage regulation. Background Technology

[0002] MPPT is a key technology in photovoltaic power generation systems, and its role is to ensure that solar cell modules can output electrical energy with optimal efficiency.

[0003] Common algorithms for MPPT (Maximum Power Point Tracking) technology include: Perturbation and Observation (P&O) algorithm, which periodically adjusts the output voltage and observes power changes to gradually approach the maximum power point; Incremental Conductivity (IC) method, which calculates incremental conductance to determine whether the photovoltaic cell has reached its maximum power point; Constant Voltage Tracking (CVT) method, which controls the output of the photovoltaic cell by preset voltage and is suitable for environments with uniform illumination; Model Predictive Control (MPTC), which establishes a mathematical model of the photovoltaic system and predicts changes in light intensity over a future period; and Intelligent Algorithms, such as Artificial Neural Networks (ANN), Genetic Algorithms (GA), and Fuzzy Logic Control (FLC), to optimize the power output of the photovoltaic system.

[0004] For example, patent CN119356475A proposes a strategy for real-time adjustment of the disturbance step size based on electrical information. By dynamically adjusting the duty cycle disturbance amplitude through the power change sensitivity coefficient, it solves the power oscillation problem of the traditional disturbance observation method under fluctuating illumination, enabling fast and accurate maximum power point tracking of fluctuating photovoltaic arrays. The paper "Application of BP Neural Network in MPPT Control of Photovoltaic Power Generation" constructs a nonlinear mapping model between temperature-irradiance input and maximum power point voltage output to achieve high-precision voltage tracking in complex environments. Patent CN119270990A uses a time delay averaging method to handle system time delay and establishes a quantitative relationship of controller parameters by combining ordinary differential equation stability analysis. This solves the problems of low tracking accuracy, slow tracking speed, and inability to cope with complex working environments, perturbations of battery array parameters, and time delays in the system caused by traditional maximum power point tracking control methods. Patent CN119496381A proposes a dynamic reference power adjustment mechanism based on DC-DC circuits. By generating a disturbance voltage through real-time comparison of output power and reference power, it can easily and quickly track the photovoltaic maximum power point and realize hardware DC-DC MPPT tracking.

[0005] In summary, the goal of photovoltaic MPPT technology is to ensure that solar cell modules always operate at their maximum power point under different lighting conditions, thereby maximizing the power generation efficiency of the photovoltaic system.

[0006] The Perturbation and Observation (P&O) algorithm is simple to implement, but it suffers from oscillations and insufficient performance near the maximum power point. In multi-peak conditions, it may be misled to a local maximum power point and fail to find the global maximum power point. The Incremental Conductance (IC) method is sensitive to parameters and may get stuck in a local maximum power point in multi-peak conditions. The Constant Voltage Tracking (CVT) method ignores the influence of temperature changes. Fuzzy control is difficult to design. Although it is simple to operate, designing a good fuzzy control system requires rich experience and skills, placing high demands on the designer. Its performance is affected by parameters: certain parameters in fuzzy control, such as membership functions and fuzzy rules, have a significant impact on system performance. Improper parameter selection may lead to a decrease in system performance. Traditional machine learning algorithms cause delays in MPPT due to computation.

[0007] The output power and voltage curves of photovoltaic modules under different illumination conditions are as follows: Figure 5 As shown, by Figure 6 It can be seen that, under the same temperature, the stronger the illuminance, the greater the output power of a photovoltaic module, and the maximum power point increases with the increase of illuminance; the charging process of a lithium-ion battery can be divided into four stages: trickle charging (low-voltage pre-charging), constant current charging, constant voltage charging, and charging termination, as follows... Figure 7 As shown:

[0008] Phase 1: Trickle Charging – Trickle charging is used to pre-charge (recovery charging) fully discharged battery cells. Trickle charging is used when the battery voltage is below approximately 3V. The trickle charging current is one-tenth of the constant current charging current, i.e., 0.1c (for example, if the constant charging current is 1A, then the trickle charging current is 100mA).

[0009] Over-discharging a battery can negatively impact its performance. The primary reason lithium batteries require restorative charging is to protect battery life, improve performance, and ensure proper device operation. Improper handling during power restoration, such as excessive charging current or unstable voltage, can damage the battery's internal structure, further reducing performance and even shortening its lifespan.

[0010] Phase 2: Constant Current Charging – When the battery voltage rises above the trickle charging threshold, the charging current is increased for constant current charging. The constant current charging current is between 0.2C and 1.0C. The battery voltage gradually increases during the constant current charging process; typically, this voltage is set to 3.0-4.2V for a single cell.

[0011] Phase 3: Constant Voltage Charging – When the battery voltage rises to 4.2V, constant current charging ends, and constant voltage charging begins. The current decreases gradually from its maximum value as charging continues, depending on the cell's saturation level. Charging is considered complete when the current decreases to 0.01C. (C is a method of expressing current relative to the battery's nominal capacity; for example, if the battery has a capacity of 1000mAh, 1C means a charging current of 1000mA.)

[0012] Phase 4: Charging Termination – There are two typical charging termination methods: using the minimum charging current as a criterion or using a timer (or a combination of both). The minimum current method monitors the charging current during the constant voltage charging phase and terminates charging when the charging current decreases to the range of 0.02C to 0.07C. The second method starts timing at the beginning of the constant voltage charging phase and terminates the charging process after two hours of continuous charging.

[0013] Current photovoltaic MPPT (Multi-Level Testing) focuses on ensuring that solar cell modules can output power at the best efficiency under current environmental conditions. It does not consider the relationship between the voltage under the photovoltaic panel MPPT, the DC-DC circuit, and the charging curve of the lithium battery. This will limit the maximum power point of the photovoltaic module, the maximum power point (MPP) voltage of the photovoltaic panel. With lithium battery charging voltage There is a natural mismatch. Summary of the Invention

[0014] The purpose of this invention is to provide a Buck-Boost intelligent charging method for lithium batteries based on ML-MPPT and dynamic reference voltage regulation, in order to solve the problem mentioned in the background art that the existing MPPT technology of photovoltaic panels does not take into account the relationship between the MPPT voltage of the photovoltaic panel, the DC-DC circuit, and the charging curve of the lithium battery.

[0015] To achieve the above objectives, the present invention provides the following technical solution:

[0016] A smart charging method for Buck-Boost lithium batteries based on ML-MPPT and dynamic reference voltage regulation includes the following steps:

[0017] S10: Detects the lithium battery voltage and enters the corresponding charging stage according to the voltage range. When the voltage of a single lithium battery is below 3V, it enters the trickle charging stage; when the voltage of a single lithium battery is between 3.0-4.2V, it enters the constant current charging stage; when the voltage of a single lithium battery is above 4.2V, it enters the constant voltage charging stage.

[0018] S20: During the constant current charging stage, the photovoltaic panel MPPT charging mode is activated, and temperature, illuminance and corresponding maximum power point data are collected under all working conditions. After data preprocessing, a machine learning model for maximum power point voltage prediction is trained. The model is then distilled to obtain a compressed model and deployed on the microcontroller.

[0019] S30: Input the real-time collected temperature and illuminance data into the compression model to obtain the maximum power point voltage and use it as the reference voltage. Control the duty cycle of the PWM pulse through the microcontroller and adjust the input impedance of the Buck-Boost circuit so that the photovoltaic panel voltage reaches the maximum power point voltage.

[0020] S40: During each charging stage, based on the battery status and input power, it dynamically switches between the MPPT mode stage and other different charging stages to ensure charging safety and efficiency.

[0021] The data preprocessing in step S20 includes outlier handling, missing value imputation, and normalization. The normalization operation maps the data to the range [0,1]. The normalization formula is:

[0022] ,

[0023] Where x' is the normalized data, x is the data to be normalized, and x(min) and x(max) represent the minimum and maximum values ​​in the data to be normalized, respectively.

[0024] Preferably, the machine learning model in step S20 includes BP neural network, RBF neural network, SVM, random forest, decision tree or GRU, and the preprocessed dataset is divided into 70% training set, 15% validation set and 15% test set during model training.

[0025] Preferably, the BP neural network adopts a 3-layer architecture, including an input layer, a hidden layer and an output layer. The input layer receives ambient temperature, light radiation illuminance and the output voltage value corresponding to the maximum power point, and the output is the output voltage value corresponding to the maximum power point.

[0026] The model is evaluated using metrics such as Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Mean Squared Error (MSE), and R² score. The model evaluation formula is as follows:

[0027] ,

[0028] ,

[0029] ,

[0030] ,

[0031] Where n represents the number of samples. This represents the true value of the i-th sample. The predicted value for the sample; This is the average value of all samples of the output voltage corresponding to the maximum power point.

[0032] Preferably, the model distillation process in step S20 is as follows: using the trained and converged BP neural network as the teacher model, constructing a lightweight RBF network as the student model, and extracting the ambient temperature from the teacher model. Radiant Illuminance Output voltage value corresponding to the maximum power point Construct a distillation training dataset, which involves running a converged backpropagation neural network in the training environment and collecting ambient temperature data. Radiant Illuminance Record the output voltage value corresponding to the maximum power point of the BP neural network. The data is divided into training and validation sets. The distillation training process involves supervised learning to make the output of the RBF network approximate the actions of the BP neural network, using MSE as the loss function. The formula for the MSE loss function is:

[0033] ,

[0034] Mini-batch SGD was used with the Adam optimizer and a low learning rate to train the RBF student model.

[0035] Preferably, in step S30, the output voltage of the photovoltaic panel is collected in real time through voltage sampling and A / D conversion circuit, and used as the front feedback voltage input to the microcontroller to adjust the duty cycle of the PWM pulse.

[0036] Preferably, a microcontroller with a neural network processing unit (NPU) and high operating frequency, such as the STM32N6, is selected to perform model inference, data acquisition, and A / D conversion in real time, converting the reference voltage predicted by the machine learning model into a digital signal. Stored in the CCRx register as the target duty cycle value, the input voltage of the switching power supply. The output voltage of the photovoltaic panel serves as the feedback voltage. This voltage is read by the ADC and used to dynamically adjust the duty cycle. Closed-loop control is achieved by updating the CCRx register via software.

[0037] Preferably, in step S30, the maximum power point voltage output by the compression model is inversely normalized. The inverse normalization formula is as follows: , where x is the original data, x' is the normalized data, and x(min) and x(max) are the minimum and maximum values ​​of the data, respectively.

[0038] Preferably, the connection between MPPT and trickle charging in step S40 is as follows: when the battery voltage is detected to be below 3V, the MPPT algorithm is paused, the system enters trickle charging mode, and charges with a current of 0.05C. When the battery voltage rises back to above 3V, it switches to MPPT charging mode.

[0039] Preferably, the dynamic balancing method between MPPT and constant current charging in step S40 includes: when there is sufficient light, the MPPT algorithm is enabled to track the maximum power point, and the output current is kept constant by adjusting the PWM duty cycle of Buck-Boost; when there is insufficient light, the boost characteristic of Buck-Boost is used to charge the lithium battery with a weak charging current, or the adapter or battery pack is used as a supplementary power source to maintain constant current charging.

[0040] Preferably, in the constant voltage charging stage of step S40, the MPPT algorithm stops running, and the Buck-Boost circuit reduces the charging current by adjusting the duty cycle. When the charging current decreases to the range of 0.02C to 0.07C, charging is terminated.

[0041] Preferably, the Buck-Boost circuit includes a switching transistor, an inductor, a diode, a capacitor, and a PWM driver. When the switching transistor is closed, the inductor is directly connected to both ends of the photovoltaic panel, and the current gradually increases. The output terminal relies on its own discharge to power the lithium battery. When the switching transistor is turned off, the inductor powers the output capacitor and the lithium battery through the diode, and the inductor volt-second conservation is satisfied when the system is working stably.

[0042] When the switch Q1 is closed, the inductor L1 is directly connected to the two ends of the photovoltaic panel at the input terminal. At this time, the inductor current gradually increases, and during the conduction transient... Enhanced, at the output end, capacitor C1 provides energy to the lithium battery by discharging itself;

[0043] When the switching transistor Q1 is turned off, since the current of the inductor L1 cannot change abruptly, the inductor supplies power to the output capacitor C1 and the lithium battery through the diode D1.

[0044] After the system stabilizes, the inductor voltage-second constant is maintained. When the switching transistor Q1 is turned on, the inductor voltage equals the input voltage. When the switching transistor Q1 is turned off, the inductor voltage is equal to the output voltage. Let T be the period. For conduction time, D is the off time, and D is the duty cycle. );

[0045] From the conservation of volt-seconds in inductance, we have:

[0046] ,

[0047] ,

[0048] Therefore, we can conclude that:

[0049] ,

[0050] ,

[0051] When the duty cycle is less than 0.5, the output voltage is stepped down; when the duty cycle is greater than 0.5, the output voltage is stepped up.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation provided by this invention combines machine learning maximum power point tracking (ML-MPPT) technology with dynamic reference voltage regulation and relies on Buck-Boost circuits to achieve intelligent charging of lithium batteries, which has the following beneficial effects:

[0054] I. Improved charging efficiency: The constant current charging stage adopts the ML-MPPT mode. By training a machine learning model and obtaining a compressed model through distillation, it can quickly and accurately predict the maximum power point voltage based on real-time temperature and illuminance data. This voltage is used as a reference to adjust the input impedance of the Buck-Boost circuit, ensuring that the photovoltaic panel always works at the maximum power point, significantly improving the utilization rate of photovoltaic energy and thus enhancing the overall charging efficiency.

[0055] II. Achieving Intelligent Dynamic Switching: Based on the range of lithium battery voltage, the charging process is divided into trickle charging, constant current charging, and constant voltage charging stages. According to the battery status and input power, the MPPT constant current charging stage can be dynamically switched with other charging stages. This not only enables efficient utilization of photovoltaic energy when there is sufficient sunlight, but also ensures that the charging process matches the battery charging curve characteristics in different charging stages, thus balancing charging efficiency and safety.

[0056] 3. Optimize model deployment and operation: Distill the trained machine learning model to obtain a compressed model, which greatly reduces the model parameters and the occupation of hardware resources such as microcontrollers. This ensures that the model can be deployed stably and efficiently on embedded devices while maintaining high prediction accuracy to meet the requirements of real-time charging control.

[0057] IV. Ensuring Charging Safety: Targeted control measures are taken at each charging stage. For example, in the trickle charging stage, a small current is used to avoid battery damage. In the constant voltage stage, the MPPT is turned off and the circuit parameters are dynamically adjusted to keep the voltage stable within a safe range. Furthermore, the PWM duty cycle is adjusted in real time through voltage sampling feedback to effectively prevent overcharging, overvoltage, and other problems, thereby extending battery life.

[0058] V. Enhanced adaptability and flexibility: Employing various machine learning models (such as BP neural networks, RBF neural networks, etc.) can adapt to the data analysis needs under different operating conditions; the Buck-Boost circuit can realize the step-up and step-down functions through duty cycle adjustment, adapting to the matching requirements of photovoltaic panel output voltage and lithium battery charging voltage, thus improving the applicability of the method in complex environments and under different equipment configurations. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0060] Figure 1 This is a diagram of the lithium battery charging technology architecture that integrates machine learning MPPT and PWM in this invention.

[0061] Figure 2 This is a diagram of the BP neural network architecture of the present invention;

[0062] Figure 3 This is a diagram of the BP neural network distillation into RBF network architecture of the present invention;

[0063] Figure 4 This is a flowchart illustrating the charging process using the BUCK-BOOST buck-boost topology in this invention.

[0064] Figure 5 Figure 1 shows the IV curves of the photovoltaic module under different temperatures;

[0065] Figure 6 PV curves of photovoltaic modules under different temperature conditions;

[0066] Figure 7 This is a charging curve diagram for a lithium battery. Detailed Implementation

[0067] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0068] The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation of the present invention consists of a BUCK-BOOST circuit, an MPPT control module, and trickle charging, constant current charging, and constant voltage charging modules for the lithium battery.

[0069] The lithium battery charging technology solution proposed in this invention, which integrates machine learning MPPT and PWM, is as follows: Figure 1 As shown, the technical solution includes a BUCK-BOOST circuit composed of a switching transistor Q1, an inductor L1, a diode D1, a capacitor C1, and a PWM generator; the MPPT control circuit consists of a machine learning model for predicting the MPPT voltage, a photovoltaic panel output voltage (V) sampling and A / D conversion circuit, a temperature (°C) sampling and A / D conversion circuit, and a solar irradiance (W / m²) control circuit. 2 The system consists of sampling and A / D conversion circuits and an MCU; the efficient charging and safety protection of the lithium battery is composed of trickle, constant current and constant voltage control modules.

[0070] When the PWM controller detects that the voltage of a single lithium battery cell is below approximately 3V, trickle charging is used; when the voltage of a single lithium battery cell is between 3.0V and 4.2V, constant current charging is used; and when the voltage of a single lithium battery cell is above 4.2V, constant voltage charging is used.

[0071] During the constant current charging phase of the lithium battery, the MPPT charging mode of the photovoltaic panel is activated. Temperature, illuminance, and corresponding maximum power point data are collected across all operating conditions. Abnormal data processing and missing value imputation are performed, and machine learning models are trained. Finally, machine learning models such as SVM, BP neural network, and RBF are obtained. These models are then distilled to obtain a compressed model suitable for running on microcontrollers (MCUs) or other resource-constrained hardware platforms. Temperature and illuminance data are input into the distilled model to obtain the maximum power point voltage. and the maximum power point voltage As a reference voltage, the voltage of the photovoltaic panel is used as the feedback voltage and input to the MCU-based control circuit. This controls the duty cycle of the PWM pulses and adjusts the input impedance of the BUCK-BOOST circuit, causing the photovoltaic panel voltage to quickly reach its maximum power point voltage. .

[0072] 1. BUCK-BOOST module

[0073] The main factors affecting photovoltaic power generation include solar irradiance and ambient temperature. Solar irradiance is not continuous, such as being blocked by buildings or clouds, resulting in insufficient sunlight. There are also differences in irradiance in the morning, noon and evening. To adapt to complex operating conditions, the buck-boost circuit is used to provide sufficient charging power for lithium batteries.

[0074] When the switch Q1 is closed, the inductor L1 is directly connected to the two ends of the photovoltaic panel at the input terminal, and the inductor current gradually increases. During the conduction transient... Enhancement, at the output end, It relies on its own discharge to provide energy for the lithium battery.

[0075] When the switching transistor Q1 is turned off, since the current in inductor L1 cannot change abruptly, the current flows from the inductor to the output capacitor through diode D1. Powered by lithium batteries.

[0076] Once the system is operating stably, the inductance volt-seconds are conserved. When the circuit is turned on, the inductor voltage is equal to the input voltage. ; When turned off, the inductor voltage is equal to the output voltage. Let T be the period. For conduction time, D is the off time, and D is the duty cycle. ).

[0077] From the conservation of volt-seconds in inductance, we have:

[0078] ,

[0079] ,

[0080] Therefore, we can conclude that:

[0081] ,

[0082] ,

[0083] When the duty cycle is less than 0.5, the output voltage is stepped down; when the duty cycle is greater than 0.5, the output voltage is stepped up.

[0084] 2. MPPT control module

[0085] The MPPT control module consists of a voltage sampling and A / D module, an MPPT machine learning model, and an MCU module. The voltage sampling and A / D module samples the output voltage of the photovoltaic panel and performs digital-to-analog conversion. The solar irradiance and ambient temperature are input into the MPPT machine learning model to predict the MPPT voltage of the photovoltaic panel. The MPPT voltage is used as the reference voltage of the PWM generator. The duty cycle is adjusted in real time by the PWM, and the output voltage of the photovoltaic panel quickly tracks the MPPT voltage, thus realizing the tracking of the maximum power point of the photovoltaic panel.

[0086] 2.1 Voltage Sampling and A / D Module

[0087] The voltage sampling and A / D conversion module of the MCU is used to collect the output voltage of the photovoltaic panel in real time and perform digital-to-analog conversion.

[0088] 2.2 MPPT Machine Learning Model

[0089] 2.2.1 Data Acquisition

[0090] Data acquisition includes ambient temperature, solar irradiance, and photovoltaic panel output voltage. Irradiance can be sampled using a mature integrated module such as the BH1750FVI.

[0091] 2.2.2 Data Preprocessing

[0092] Missing values ​​were imputed in the collected photovoltaic panel data, and outlier data was processed. The dataset was normalized to scale features and labels at different scales to the range of [0,1] to eliminate the influence of units of measurement.

[0093] ,

[0094] Where x is the data to be normalized. and These represent the maximum and minimum values ​​in the data to be normalized, respectively. Finally, the data is mapped to the range between 0 and 1 and used as feature data to input into the model for training.

[0095] The preprocessed dataset is divided into training, validation, and test sets for use during training. A common ratio is 70% training set, 15% validation set, and 15% test set.

[0096] 2.2.3 Model Building and Evaluation

[0097] MPPT machine learning models can be built using machine learning algorithms such as BP, RBF, SVM, RF (Random Forest), DT (Decision Tree), and GRU. BP, RBF, SVM, RF, DT, and GRU are preferred. Here, we will use a BP neural network as an example. The BP neural network uses a three-layer architecture, including an input layer, hidden layers, and an output layer. The input layer receives ambient temperature, light irradiance, and the output voltage value corresponding to the maximum power point. The output is the output voltage value corresponding to the maximum power point. The model is evaluated using metrics such as Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Mean Squared Error (MSE), and R² score. The model evaluation formula is as follows:

[0098] ,

[0099] ,

[0100] ,

[0101] ,

[0102] Where n represents the number of samples. This represents the true value of the i-th sample. The predicted value for the sample; This represents the average of all samples for the output voltage value corresponding to the maximum power point. The trained BP, RBF, SVM, RF (Random Forest), DT (Decision Tree), and GRU models require significant computational resources and storage space. These models need to be compressed and optimized to fit embedded devices with limited memory, computing power, and energy consumption, reducing model size and computational requirements without impacting performance.

[0103] 2.2.4 Model Distillation

[0104] Model distillation is a method of model compression. After the trained BP network converges, the BP network is distilled into an RBF network.

[0105] After the BP neural network is trained and converged, it serves as the teacher model. A lightweight RBF network is then constructed as the student model. The ambient temperature is extracted from the teacher model. Radiant Illuminance Output voltage value corresponding to the maximum power point Construct a distillation training dataset, which involves running a converged backpropagation neural network in the training environment and collecting ambient temperature data. Radiant Illuminance Record the output voltage value corresponding to the maximum power point of the BP neural network. The data is divided into training and validation sets; the distillation training process involves supervised learning to make the RBF network output approximate the actions of the BP neural network, using MSE as the loss function. We used Mini-batch SGD with the Adam optimizer and a low learning rate to train the RBF student model.

[0106] Through the distillation process described above, the RBF network can inherit the control strategy of the BP neural network while significantly reducing computational complexity, thus meeting the real-time control requirements of the power converter. This method, while retaining the performance advantages of the BP neural network, solves the resource bottleneck problem of the BP neural network in embedded deployments.

[0107] 2.2.5 Model Deployment

[0108] To minimize model size and computational complexity, TinyML technology is employed to run machine learning models on microcontrollers (MCUs) or other resource-constrained hardware platforms, extending AI technology to Internet of Things (IoT) devices and enabling them to possess intelligent sensing and processing capabilities, thereby achieving wider applications.

[0109] 2.2.6 Model Application

[0110] Real-time measured ambient temperature Radiant Illuminance Input the RBF model and perform Real-time prediction and inverse normalization operation.

[0111]

[0112] Where x is the original data, x′ is the normalized data, and x(min) and x(max) are the minimum and maximum values ​​of the data, respectively.

[0113] Denormalization involves transforming the normalized data back to the original scale. The specific steps depend on the normalization method used during preprocessing. Ensure that the normalization parameters are saved during training and used for denormalization during prediction to obtain predictions at the original scale.

[0114] 2.2.7 Generating PWM signals

[0115] Real-time monitoring of ambient temperature Radiant Illuminance The MPPT voltage is predicted by inputting it into the RBF model. ,Will As a reference voltage, the output voltage of the photovoltaic panel is tracked in real time. The voltage at the maximum power point is obtained.

[0116] 3. Trickle, constant current, and constant voltage charging modules for lithium batteries

[0117] This invention employs a BUCK-BOOST buck-boost topology for charging, which automatically adjusts the operating mode when the battery voltage is lower, higher, or equal to the input voltage, ensuring the battery is always in an optimal charging state. During the charging process, the chip supports multiple stages, including trickle charging (TC), constant current charging (CC), constant voltage charging (CV), and charging termination.

[0118] 3.1 Connection between MPPT and trickle charging

[0119] When the MCU detects that the battery voltage is below a safe threshold (e.g., a single lithium battery cell voltage < 3.0V), the MPPT algorithm pauses to avoid high-voltage input impacting the battery, forcing the charging system into trickle charging mode. This mode uses a small current (typically 0.05C) to restore battery activity and prevent battery damage caused by deep discharge. Once the battery voltage rises above the trickle threshold, the system switches to MPPT charging mode, adjusting the input voltage / current through the BUCK-BOOST circuit to maximize the solar panel's output power.

[0120] 3.2 Dynamic Balance between MPPT and Constant Current Charging

[0121] During the constant current phase, the upper limit of the MPPT target power is... The photovoltaic output power needs to be brought close to this value by dynamically adjusting the PWM duty cycle.

[0122] 3.2.1 When there is sufficient light

[0123] When the battery voltage returns to the normal range and the input source is a photovoltaic module, the MPPT algorithm is enabled to track the maximum power point. By adjusting the PWM duty cycle of the BUCK-BOOST circuit, the energy conversion efficiency is optimized. The BUCK-BOOST circuit converts the input power into a constant current output (such as 1C), which quickly boosts the battery voltage.

[0124] When the charging current exceeds the battery's allowable threshold (such as 1C) in MPPT mode, it automatically switches to constant current charging to protect battery life.

[0125] 3.2.2 When there is insufficient light

[0126] If the input power cannot maintain the constant current requirement, the buck-boost circuit utilizes its boost characteristic to charge the lithium battery with a weak charging current, or uses an adapter or battery pack as a supplementary power source to maintain constant current charging. In dual-input source scenarios (such as solar power + AC mains power), the buck-boost circuit prioritizes solar MPPT power supply, with AC mains power as a backup, maximizing energy utilization efficiency.

[0127] 3.3 Power Adaptation During Constant Voltage Charging

[0128] When the battery voltage approaches its saturation value (e.g., 4.2V for a lithium battery), the system switches to constant voltage mode, and the MPPT algorithm stops running. The BUCK-BOOST circuit reduces the charging current by adjusting the duty cycle to prevent overvoltage risks.

[0129] 3.4 Charging terminated

[0130] Monitor the charging current during the constant voltage charging phase and terminate charging when the charging current decreases to the range of 0.02C to 0.07C.

[0131] If the battery temperature or voltage is abnormal, immediately interrupt MPPT and terminate charging.

[0132] Through the above mechanism, the BUCK-BOOST charging circuit achieves seamless switching between trickle charging, constant current charging, and constant voltage charging stages, which not only makes full use of renewable energy but also ensures the safety and efficiency of the battery charging process.

[0133] Example 1: Intelligent charging solution based on BP neural network and RBF distillation model

[0134] Experimental conditions:

[0135] Photovoltaic panel parameters: maximum power 200W, open circuit voltage 36V, short circuit current 6.8A;

[0136] Lithium battery: a single 3.7V / 10Ah lithium iron phosphate battery, connected in series to form a 12V battery pack;

[0137] Environmental conditions: Temperature range -10~50℃, irradiance 200~1000W / m² (simulating cloudy to sunny weather).

[0138] Technical details:

[0139] Data acquisition: 5000 sets of temperature, irradiance and corresponding MPP voltage data were collected. After outlier removal (3σ criterion) and missing value filling by linear interpolation, the data were divided into training set, validation set and test set according to 70%:15%:15%.

[0140] Model training: A 3-layer BP neural network (2 neurons in the input layer, 10 neurons in the hidden layer, and 1 neuron in the output layer) was used, with MSE as the loss function, and the Adam optimizer (learning rate 0.001) was used to train until convergence (MAPE < 2%).

[0141] Model distillation: Using a BP network as the teacher model, a 2-layer RBF network (2 neurons in the input layer and 5 neurons in the hidden layer) was constructed as the student model. The model was trained for 500 rounds (learning rate 0.0005), and the model parameters were reduced by 60% after compression.

[0142] Charging control: During the constant current stage (3.0~4.2V), ML-MPPT is enabled to collect temperature (DS18B20 sensor) and irradiance (BH1750FVI module) in real time, and the reference voltage is output through the RBF model, with the PWM duty cycle adjustment step size being 0.1%.

[0143] Experimental results:

[0144] When illumination fluctuates (e.g., from a sudden change from 500 to 800 W / m²), the MPP tracking response time is <0.5s and the power fluctuation amplitude is <3%;

[0145] The overall charging efficiency (PV panel output power / battery absorbed power) reaches 92.3%, which is 8.5% higher than the traditional P&O algorithm.

[0146] Example 2: Dual-input source (solar power + AC power) collaborative charging solution

[0147] Experimental conditions:

[0148] New hardware addition: a 220V to 12V adapter (100W) as a backup power source, with input switching via a relay;

[0149] Lithium battery: a single 4.2V / 20Ah ternary lithium battery, connected in series to form a 24V battery pack;

[0150] Environmental conditions: Simulate low light in the early morning and evening (irradiance < 300 W / m²) and strong light at noon (irradiance > 800 W / m²).

[0151] Technical details:

[0152] Mode switching logic:

[0153] When the output power of the photovoltaic MPPT is greater than or equal to the constant current requirement (1C=20A), only solar charging is used;

[0154] When the power is <0.7C (14A), switch to mains power supplement (photovoltaic + mains power combined power supply).

[0155] Safety control: During the trickle charge phase (<3.0V), the voltage is charged at 0.05C (1A). During the constant voltage phase (>4.2V), the MPPT is turned off, and the PWM duty cycle is dynamically adjusted until the output voltage is stable at 4.2V±0.02V.

[0156] Experimental results:

[0157] In low-light conditions (200W / m²), charging time is reduced by 40% compared to a pure solar power solution;

[0158] No overcharging throughout the entire process (voltage fluctuation <0.05V), and the battery cycle life test (1000 cycles) shows a capacity retention rate of ≥85%.

[0159] Example 3: Extreme Environment Adaptability Optimization Scheme

[0160] Experimental conditions:

[0161] Environmental simulation: high temperature (60℃), low temperature (-20℃) and rapid shading (irradiance from 1000 to 200W / m² within 10s).

[0162] Lithium battery: Single 3.2V / 5Ah lithium titanate battery (wide temperature range).

[0163] Technical details:

[0164] Model optimization: 2000 extreme temperature (-20~60℃) samples were added to the training data. A GRU network (5s time window) was used to predict the trend of irradiance change, and the reference voltage was adjusted 50ms in advance.

[0165] Circuit compatibility: The Buck-Boost inductor uses a ferrite core with a temperature resistance of 125℃, and the switching transistor uses a SiCMOSFET (with a voltage resistance of 600V) to reduce high-temperature losses.

[0166] Experimental results:

[0167] Under extreme temperatures, the MPPT tracking accuracy (measured MPP voltage / theoretical value) is ≥98%, while traditional IC algorithms only achieve 90%.

[0168] During rapid blocking, the power drop is less than 5%, the recovery time is less than 0.3s, and there is no obvious oscillation.

[0169] Comparative Example 1: Traditional Perturbation and Observation (P&O) Charging Scheme

[0170] Experimental conditions: Same as in Example 1, but the MPPT algorithm was replaced with a fixed step size (0.5V) perturbation and observation method.

[0171] Technical defects:

[0172] When the illumination fluctuates (e.g., from 800W / m² to 500W / m²), the power oscillation amplitude reaches 15%, and the tracking time is >2s;

[0173] During the constant current phase, the battery voltage demand could not be dynamically matched, resulting in a charging efficiency of only 83.8%, which is 8.5% lower than that of Example 1.

[0174] Multi-peak illumination (partial shading) can easily lead to localized MPP, resulting in power loss of up to 12%.

[0175] Comparative Example 2: Model-free distillation ML-MPPT scheme

[0176] Experimental conditions: Same as in Example 1, but with direct deployment of an undistilled BP neural network (parameter size 1.2MB).

[0177] Technical defects:

[0178] The microcontroller (STM32N6, 64KB RAM) frequently overflows during operation, with a response delay >1s;

[0179] The model inference power consumption reaches 80mA, which is 300% higher than the distilled RBF model (20mA), and does not meet the low power consumption requirements;

[0180] During the constant current phase, the voltage regulation overshoot reaches 5%, posing a risk of battery overvoltage.

[0181] A comparison table of key parameters and results for the three embodiments and two sets of comparative examples above:

[0182]

[0183] Based on a comparison of key data from the above embodiments and comparative examples, the specific analysis is as follows:

[0184] I. Impact of Core Algorithms on Performance

[0185] Comparison of ML-MPPT and traditional algorithms

[0186] Examples 1-3 all employ the MPPT scheme combining machine learning (ML) and model distillation techniques, while Comparative Example 1 uses the traditional perturbation and observation (P&O) method. Data shows:

[0187] Tracking speed: The MPP tracking response time of the example (0.3-0.5s) is much faster than that of Comparative Example 1 (>2s). This is because the ML model can directly predict the maximum power point (MPP) by training with historical data, without having to go through repeated perturbations and trial and error like the P&O algorithm, thus reducing adjustment delay.

[0188] Power stability: The power fluctuation amplitude of the embodiment (<3%-5%) is significantly lower than that of Comparative Example 1 (15%), especially in scenarios with sudden changes in illumination (such as rapid occlusion in Example 3). The ML model adjusts the reference voltage in advance through time-series prediction (such as GRU network), avoiding the oscillation problem of the P&O algorithm.

[0189] Efficiency advantage: The charging efficiency of the example (90.8%-92.3%) is 7%-8.5% higher than that of Comparative Example 1 (83.8%). The core reason is that the ML model can accurately lock the global MPP under complex lighting (such as multi-peak shadows), while P&O is prone to getting trapped in local optima (power loss of up to 12%).

[0190] The necessity of model distillation

[0191] Examples 1-3 all employed model distillation (e.g., BP neural network distillation using a lightweight RBF network), while Comparative Example 2 directly used the undistilled BP model. The results show that:

[0192] Resource compatibility: After distillation, the model parameters are reduced by 60% (Example 1), which can be adapted to the limited memory (64KB) of microcontrollers (such as STM32N6); while the undistilled model (1.2MB) in Comparative Example 2 exceeds the memory limit, resulting in response delay (>1s) and runtime overflow, and cannot work stably.

[0193] Safety: The voltage fluctuations in Examples 1-3 were all <0.05V, with no risk of overcharging; Comparative Example 2, due to model delay and adjustment lag, had a voltage overshoot of 5%, posing a serious risk of overcharging.

[0194] II. Scene Adaptability Analysis

[0195] Normal environment (Example 1)

[0196] In typical scenarios with temperatures ranging from -10 to 50°C and irradiance from 200 to 1000 W / m², Example 1 exhibits the best overall performance.

[0197] The efficiency (92.3%) was the highest among the three sets of examples, thanks to the accurate fit of the BP+RBF model to the normal working conditions;

[0198] The power fluctuation is less than 3%, indicating that the model's adjustment accuracy is higher under stable lighting conditions than that of the dual-input source (Example 2, <5%) and the extreme environment scheme (Example 3, <5%).

[0199] Low light and dual-source synergy (Example 2)

[0200] For low-light scenarios with irradiance <300W / m², Example 2 solves the problem of low single-source charging efficiency through dual-input synergy of "photovoltaic + mains power":

[0201] Charging time under low light is reduced by 40% compared to the pure solar energy solution, while the efficiency remains at 91.5% (only 0.8% lower than Example 1).

[0202] After 1000 cycles, the capacity retention rate is ≥85%, proving that the dual-source switching logic (photovoltaic priority, mains power supplement) does not affect the battery life and is better than the overcharge risk caused by voltage fluctuation (±0.1V) in Comparative Example 1.

[0203] Extreme Environment (Example 3)

[0204] In scenarios with temperatures ranging from -20°C to 60°C and rapid occlusion, the GRU network in Example 3 (including training with extreme samples) exhibits strong adaptability:

[0205] The MPP tracking response time is <0.3s, which is the fastest among all solutions, because the timing prediction capability of GRU can adjust the reference voltage 50ms in advance;

[0206] The tracking accuracy was ≥98% under extreme temperatures, which is much higher than the 90% of Comparative Example 1, indicating that the model optimization for extreme data (sample expansion of 2000 groups) effectively offset the impact of temperature on the characteristics of photovoltaic panels.

[0207] III. Model Resource Consumption and Hardware Adaptability

[0208] The key role of distillation technology

[0209] The models in Examples 1-2, after distillation, have 60% fewer parameters and can be directly deployed on a microcontroller (such as STM32N6). However, Comparative Example 2, due to lack of distillation (parameters 1.2MB), exceeds the hardware memory (64KB), resulting in runtime overflow and latency (>1s). This indicates that:

[0210] Model distillation is a core technology for balancing prediction accuracy and hardware resources. Examples 1-2 maintain high performance even after parameter compression, verifying the practicality of the "teacher-student" model architecture.

[0211] Example 3 is designed to adapt to extreme environments, with the parameter scale increased by 20%, but it is still kept within the range that the microcontroller can handle, demonstrating the flexibility of the algorithm design.

[0212] Limitations of model-free solutions

[0213] Comparative Example 1 (P&O algorithm) requires no model resources but relies on hardware logic to implement perturbation adjustment, resulting in:

[0214] Under multi-peak illumination (partial shading), the global MPP cannot be identified, resulting in a power loss of up to 12%.

[0215] Voltage fluctuations of ±0.1V pose a risk of overcharging, while the embodiment uses dynamic reference voltage regulation based on an ML model to control fluctuations to <0.05V, resulting in superior safety.

[0216] IV. Battery Protection and Lifespan Performance

[0217] Voltage stability

[0218] The voltage fluctuations in Examples 1-3 were all <0.05V, and no overcharge was recorded, while Comparative Example 1 showed fluctuations of ±0.1V and Comparative Example 2 showed an overshoot of 5%. This indicates that:

[0219] The ML model can precisely control the PWM duty cycle of the Buck-Boost circuit by outputting a reference voltage in real time, thus avoiding the adjustment lag of traditional algorithms.

[0220] Dynamic switching logic (such as disabling MPPT during the constant voltage stage in Example 2) further ensures voltage stability at the end of charging.

[0221] Cycle life

[0222] Example 2 demonstrates a capacity retention rate of ≥85% after 1000 cycles, which is superior to the industry average (approximately 80%), proving that:

[0223] The fine division of charging stages (from trickle charging 0.05C to constant current 1C to constant voltage 4.2V±0.02V) effectively reduces battery loss;

[0224] The smooth transition (no current surge) during dual-source switching reduces damage to the electrode structure, while the fixed step size adjustment of Comparative Example 1 is prone to current fluctuations and accelerates decay.

[0225] V. Overall Conclusion

[0226] Technical advantages: The ML-MPPT of this invention, combined with dynamic reference voltage regulation, is superior to traditional P&O algorithms and undistilled model schemes in terms of tracking speed (<0.5s), efficiency (90.8%-92.3%), and safety (fluctuation <0.05V).

[0227] Scene adaptability: Through algorithm optimization (such as GRU timing prediction) and hardware collaboration (dual input sources), it can meet the diverse needs of normal, low light and extreme environments;

[0228] Engineering value: Model distillation technology solves the resource limitations of embedded deployment, making the method feasible for practical application and providing an efficient and reliable solution for intelligent charging of lithium batteries.

[0229] The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation provided by this invention combines machine learning maximum power point tracking (ML-MPPT) technology with dynamic reference voltage regulation and relies on Buck-Boost circuits to achieve intelligent charging of lithium batteries, which has the following beneficial effects:

[0230] I. Improved charging efficiency: The ML-MPPT mode is adopted in the constant current charging stage. By training the machine learning model and obtaining the compressed model through distillation, it can quickly and accurately predict the maximum power point voltage based on real-time temperature and illuminance data. This voltage is used as a reference to adjust the input impedance of the Buck-Boost circuit, so that the photovoltaic panel always works near the maximum power point, which significantly improves the utilization rate of photovoltaic energy and thus improves the overall charging efficiency.

[0231] 2. Achieve intelligent dynamic switching: Divide the charging stages into trickle, constant current, and constant voltage based on the range of lithium battery voltage, and achieve dynamic switching between MPPT mode and each charging stage according to battery status and input power. This can not only make efficient use of photovoltaic energy when there is sufficient sunlight, but also ensure that the charging process matches the battery characteristics in different charging stages, taking into account both charging efficiency and safety.

[0232] 3. Optimize model deployment and operation: Distill the trained machine learning model to obtain a compressed model, which greatly reduces the model parameters and the occupation of hardware resources such as microcontrollers. This ensures that the model can be deployed stably and efficiently on embedded devices while maintaining high prediction accuracy to meet the requirements of real-time charging control.

[0233] IV. Ensuring Charging Safety: Targeted control measures are taken at each charging stage. For example, in the trickle charging stage, a small current is used to avoid battery damage. In the constant voltage stage, the MPPT is turned off and the circuit parameters are dynamically adjusted to keep the voltage stable within a safe range. Furthermore, the PWM duty cycle is adjusted in real time through voltage sampling feedback to effectively prevent overcharging, overvoltage, and other problems, thereby extending battery life.

[0234] V. Enhanced adaptability and flexibility: Employing various machine learning models (such as BP neural networks, RBF neural networks, etc.) can adapt to the data analysis needs under different operating conditions; the Buck-Boost circuit can realize the step-up and step-down functions through duty cycle adjustment, adapting to the matching requirements of photovoltaic panel output voltage and lithium battery charging voltage, thus improving the applicability of the method in complex environments and under different equipment configurations.

[0235] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart charging method for Buck-Boost lithium batteries based on ML-MPPT and dynamic reference voltage regulation, characterized in that: Includes the following steps: S10: Detects the lithium battery voltage and enters the corresponding charging stage according to the voltage range. When the voltage of a single lithium battery is below 3V, it enters the trickle charging stage and charges at a current of 0.05C. When the voltage of a single lithium battery is between 3.0-4.2V, it enters the constant current charging stage. When the voltage of a single lithium battery is above 4.2V, it enters the constant voltage charging stage. S20: During the constant current charging stage, the photovoltaic panel MPPT charging mode is activated, and temperature, illuminance and corresponding maximum power point data under all working conditions are collected. After data preprocessing, a machine learning model is trained, and the model is distilled to obtain a compressed model and deployed on the microcontroller. S30: Input the real-time collected temperature and illuminance data into the compression model to obtain the maximum power point voltage and use it as the reference voltage. Collect the output voltage of the photovoltaic panel in real time and use it as the front feedback voltage to input to the microcontroller. Use the PWM control system to control the input voltage of the Buck-Boost circuit. That is, control the duty cycle of the PWM pulse through the microcontroller to adjust the input impedance of the Buck-Boost circuit so that the output voltage of the photovoltaic panel, i.e. the input voltage of the switching power supply, is adjusted to the maximum power point voltage according to the environmental conditions. S40: In each charging stage, based on the battery status and input power, the MPPT mode stage is dynamically switched with other charging stages to ensure charging safety and efficiency. The MPPT mode stage is a constant current stage. The data preprocessing in step S20 includes outlier handling, missing value imputation, and normalization. The normalization operation maps the data to the range [0,1]. The normalization formula is: ， Where x' is the normalized data, x is the data to be normalized, and x(min) and x(max) represent the minimum and maximum values ​​in the data to be normalized, respectively. The model distillation process in step S20 is as follows: A convergent BP neural network is used as the teacher model, and a lightweight RBF network is constructed as the student model. Ambient temperature and light irradiance are extracted from the teacher model. Output voltage value corresponding to the maximum power point Construct a distillation training dataset, which involves running a converged backpropagation neural network in the training environment and collecting data on ambient temperature and light irradiance. Record the output voltage value corresponding to the maximum power point of the BP neural network. The data is divided into training and validation sets. The distillation training process involves supervised learning to make the output of the RBF network approximate the actions of the BP neural network, using MSE as the loss function. The formula for the MSE loss function is: ， Mini-batch SGD was used with the Adam optimizer and a low learning rate to train the RBF student model. In step S30, the output voltage of the photovoltaic panel, i.e. the input voltage of the switching power supply, is collected in real time through voltage sampling and A / D conversion circuit. It is used as the front feedback voltage input to the microcontroller and compared with the reference voltage predicted by the machine learning model, thereby adjusting the duty cycle of the PWM pulse. In step S30, the maximum power point voltage output by the compression model is inversely normalized. The inverse normalization formula is as follows: , where x is the original data, x' is the normalized data, and x(min) and x(max) are the minimum and maximum values ​​of the data, respectively; The Buck-Boost circuit includes a switch Q1, an inductor L1, a diode D1, a capacitor C1, and a PWM generator. When the switch Q1 is closed, the inductor L1 is directly connected to both ends of the photovoltaic panel, and the current gradually increases. The output terminal supplies power to the lithium battery by discharging itself. When the switch Q1 is turned off, the inductor L1 supplies power to the capacitor C1 and the lithium battery through the diode D1, and the inductor volt-second conservation is satisfied when the system is working stably. When the switch Q1 is closed, the inductor L1 is directly connected to the two ends of the photovoltaic panel at the input terminal. At this time, the inductor current gradually increases, and during the conduction transient... Enhanced, at the output end, capacitor C1 provides energy to the lithium battery by discharging itself; When the switch Q1 is turned off, since the current of the inductor L1 cannot change abruptly, the inductor L1 supplies power to the capacitor C1 and the lithium battery through the diode D1. After the system stabilizes, the inductor voltage-second constant is maintained. When the switching transistor Q1 is turned on, the inductor voltage equals the input voltage. When the switching transistor Q1 is turned off, the inductor voltage is equal to the output voltage. Let T be the period. For conduction time, D is the off time, and D is the duty cycle. ; From the conservation of volt-seconds in inductance, we have: ， ， Therefore, we can conclude that: ， ， When the duty cycle is less than 0.5, the output voltage is stepped down; when the duty cycle is greater than 0.5, the output voltage is stepped up.

2. The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation according to claim 1, characterized in that: In step S20, the machine learning model includes BP neural network, RBF neural network, SVM, random forest, decision tree or GRU. During model training, the preprocessed dataset is divided into a 70% training set, a 15% validation set and a 15% test set.

3. The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation according to claim 2, characterized in that: The BP neural network adopts a three-layer architecture, including an input layer, a hidden layer, and an output layer. The input layer receives ambient temperature and light radiation illuminance, and the output is the output voltage value corresponding to the maximum power point. The mean absolute percentage error (MAPE), mean absolute error (MAE), mean square error (MSE), and R are used to measure the mean absolute percentage error (MAPE), mean absolute error (MAE), mean square error (MSE), and R. 2 The scoring metrics are used to evaluate the model, and the scoring metric models mentioned above are V. mpp_mape _V mpp_mae V mpp_mse , The model evaluation formula is as follows: ， ， ， ， Where n represents the number of samples. This represents the true value of the i-th sample. The predicted value for the sample; This is the average value of all samples of the output voltage corresponding to the maximum power point.

4. The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation according to claim 1, characterized in that: The connection between MPPT mode and trickle charging in step S40 is as follows: when the battery voltage is detected to be below 3V, the MPPT algorithm is paused, the system enters trickle charging mode, and charges with a current of 0.05C. When the battery voltage rises back to above 3V, it switches back to MPPT mode.

5. The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation according to claim 1, characterized in that: The dynamic balancing method between MPPT mode and constant current charging in step S40 includes: when there is sufficient light, the MPPT algorithm is enabled to track the maximum power point, and the output current is kept constant by adjusting the PWM duty cycle of Buck-Boost; when there is insufficient light, the boost characteristic of Buck-Boost is used to charge the lithium battery with a weak charging current, or the adapter or battery pack is used as a supplementary power source to maintain constant current charging.

6. The Buck-Boost lithium battery intelligent charging method based on ML-MPPT and dynamic reference voltage regulation according to claim 1, characterized in that: In step S40, during the constant voltage charging stage, the MPPT mode stops operating, and the Buck-Boost circuit reduces the charging current by adjusting the duty cycle. When the charging current decreases to the range of 0.02C to 0.07C, charging is terminated.

Images