A battery state of charge prediction system and method based on multi-model fusion

Abstract

This invention relates to a battery state-of-charge (SOC) prediction system and method based on multi-model fusion, belonging to the field of battery management technology. The system includes a data acquisition module, a voltage model prediction module, a capacity accumulation calculation module, a charging detection module, an adaptive calibration module, a rationality check module, and a data output module. A fourth-order polynomial regression model is used to predict the SOC based on voltage and temperature for initialization and charging calibration. A capacity accumulation method based on dynamic effective capacity is used for real-time prediction. Through a charging-triggered adaptive calibration mechanism, the SOC is recalibrated using the voltage model, and the effective capacity is updated, eliminating accumulated errors and tracking battery aging. This invention ensures prediction accuracy while having low memory footprint and low computational load, making it suitable for 8-bit / 16-bit microcontrollers. It also features automatic calibration and adaptability to the entire battery lifecycle.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and specifically to a battery state of charge prediction system and method based on multi-model fusion. Background Technology

[0002] Battery state of charge (SOC) is a core parameter of the battery management system (BMS), directly impacting battery lifespan, safety, and user experience. Accurate SOC prediction is crucial for applications such as electric vehicle range estimation, energy storage system energy dispatch, and portable device battery level display. Currently, various SOC prediction technologies exist, each with its own limitations.

[0003] 1. Open-circuit voltage method: This method measures the open-circuit voltage of the battery in a static state and obtains the SOC value by looking up a table using a pre-calibrated OCV-SOC curve. This method is simple in principle, has high accuracy (static error <3%) after the battery has been sufficiently rested, and does not rely on a current sensor. However, its fatal flaw is that it requires the battery to be rested for a long time (usually 2-4 hours) to reach electrochemical equilibrium, making real-time prediction during battery use impossible. Furthermore, the OCV-SOC curve is significantly affected by temperature and battery aging, requiring frequent recalibration. In the middle SOC region (30%~70%), the voltage change is gradual, and the prediction accuracy decreases significantly.

[0004] 2. Ampere-hour integration method: This method calculates the capacity change by integrating the charging and discharging current over time, and then extrapolates the State of Charge (SOC) change. It offers real-time calculation, fast response, and simple computation, making it easy to implement in embedded systems. However, it suffers from serious technical drawbacks: it relies on the accuracy of the initial SOC, and initial errors accumulate and propagate; the actual battery capacity changes with temperature and aging, and using a fixed nominal capacity leads to systematic accumulated errors; current sensor measurement errors accumulate over time, reaching 10%–20% after prolonged use; the algorithm lacks a self-calibration mechanism, and errors cannot be automatically corrected; the charging efficiency (coulombic efficiency) is not 100%, but the algorithm typically assumes it to be 100%, further introducing bias.

[0005] 3. Kalman Filtering: This method models the battery as a dynamic system, describing the State of Charge (SOC) evolution process through state equations and observation equations. It utilizes filtering algorithms to fuse multi-source observation data such as voltage and current, recursively estimating the SOC state. This method fuses multi-source data with high accuracy (error <5%) and possesses some self-calibration capability. However, it has high computational complexity, requiring matrix operations (updating the state covariance matrix), large memory footprint (typically requiring storage of the covariance matrix, >100 bytes), reliance on an accurate battery model, and difficulty in model parameter identification, making it unsuitable for low-cost embedded systems (such as 8-bit / 16-bit microcontrollers). Furthermore, the need for online identification of model parameters as they change with temperature and aging further increases the computational burden.

[0006] 4. Neural Network Method: This method trains deep neural networks (such as LSTM and GRU) to learn the nonlinear mapping relationship between input features such as voltage, current, and temperature and the State of Charge (SOC). This method offers high accuracy (error <2%), can learn complex nonlinear relationships, and is highly adaptable. However, its model size is large (typically >10MB), making it unsuitable for resource-constrained embedded devices; inference computation is high, requiring floating-point units, resulting in high power consumption and unsuitability for battery-powered devices; it requires a large amount of labeled data for training, leading to high data acquisition costs; and the model has poor interpretability, making it difficult to analyze the reasons for prediction failures. Summary of the Invention

[0007] To address the challenges of existing battery state-of-charge (SOC) prediction technologies in resource-constrained embedded systems, which struggle to balance accuracy, real-time performance, self-calibration capabilities, and adaptability to battery aging and temperature variations, this paper presents a lightweight, high-precision SOC prediction system and method with adaptive calibration capabilities.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A battery state of charge prediction system based on multi-model fusion includes: a data acquisition module for acquiring battery voltage, current, temperature and capacity data at a preset sampling period; The voltage model prediction module is used to predict the state of charge based on the current voltage and current temperature using a fourth-order polynomial regression model. The capacity accumulation calculation module is used to calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the dynamically updated effective capacity. The charging detection module is used to determine whether the battery is in a charging state based on the comparison result between the capacity increment and the preset charging threshold, and to trigger adaptive calibration when the battery is in a charging state. An adaptive calibration module is used to call the voltage model prediction module to obtain the calibration state of charge when charging, infer the new effective capacity based on the current capacity, and update the effective capacity through a filter. The rationality check module is used to impose boundary constraints on the predicted state of charge and effective capacity; The data output module is used to output the predicted state of charge and effective capacity.

[0009] Furthermore, it also includes a multi-strategy initialization module, which selects the initialization method based on available data when the system starts up: if initial capacity data exists, the initial state of charge is calculated based on the ratio of the initial capacity to the nominal capacity, and the initial effective capacity is deduced in reverse. If no initial capacity data exists, the voltage model prediction module is invoked to estimate the initial state of charge based on the initial voltage and temperature, and the nominal capacity is set as the initial effective capacity.

[0010] Furthermore, the prediction equation of the fourth-order polynomial regression model is: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting.

[0011] Furthermore, the adaptive calibration module updates the effective capacity according to the current capacity in the following manner. and Back-inferring new effective capacity , , The charge state is predicted by a fourth-order polynomial regression model; and the effective capacity is updated through a first-order low-pass filter: , where α is the filter coefficient, and α is set between 0.1 and 0.3.

[0012] A battery state-of-charge prediction method based on multi-model fusion includes the following steps: S1: System initialization, acquiring initial voltage, temperature and capacity data, and estimating initial state of charge and initial effective capacity; S2: Collect current voltage, current, temperature and capacity data at a preset sampling period; S3: Calculate the capacity increment. If the capacity increment is greater than the preset charging threshold, it is determined to be in a charging state and step S4 is executed; otherwise, step S5 is executed. S4: Charging calibration path: Call the fourth-order polynomial regression model to predict the calibration state of charge based on the current voltage and temperature, back-calculate the new effective capacity based on the current capacity, update the effective capacity through the filter, and set the current state of charge as the calibration state of charge. S5: Discharge accumulation path: Calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the current effective capacity, and add it to the state of charge at the previous moment. S6: Perform a reasonable check on the state of charge and effective capacity; S7: Output the state of charge and effective capacity, update the capacity of the previous moment, and return to step S2.

[0013] Furthermore, in step S1, if initial capacity data exists, the initial state of charge is calculated using the capacity ratio method; if initial capacity data does not exist, the initial state of charge is calculated using a fourth-order polynomial regression model.

[0014] Furthermore, the data acquisition in step S2 includes outlier removal preprocessing: if the difference between the current voltage and the voltage at the previous moment is greater than a preset voltage threshold, the current sampling point is discarded and the voltage data at the previous moment is used instead.

[0015] Furthermore, the prediction equation of the fourth-order polynomial regression model is: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting.

[0016] Furthermore, the model coefficients are obtained through the following steps: Step (1): Collect battery experimental data under different discharge rates and temperature conditions, including voltage, temperature and true SOC; Step (2): Construct a training dataset with voltage V and temperature T as features and real SOC as labels; Step (3): Fit the fourth-order polynomial regression model using the least squares method to obtain the coefficients c0 to c5; Step (4): Store the coefficients in non-volatile memory for real-time prediction.

[0017] Furthermore, the formula for calculating the change in state of charge in step S5 is as follows: , For the current capacity, The capacity at the previous time step. For dynamically updated effective capacity; in step S4, a fourth-order polynomial regression model is called to predict. To deduce the new effective capacity The effective capacity is updated using a first-order low-pass filter: ,renew , where α is the filter coefficient, and α is set between 0.1 and 0.3.

[0018] The beneficial effects of this invention are: 1. Significantly Improved Prediction Accuracy: A fourth-order polynomial regression model replaces the simple OCV lookup table, and a temperature compensation term is introduced to eliminate the impact of temperature on prediction accuracy. Simultaneously, adaptive effective capacity replaces the fixed nominal capacity, eliminating systematic errors caused by capacity decay. Compared to the traditional ampere-hour integration method (error 10%~20%), this invention reduces prediction error to MAE < 2% and RMSE < 3%, improving accuracy by approximately 85%. Technical Basis: The fourth-order polynomial model accurately describes the nonlinear characteristics of the lithium battery OCV-SOC curve; the adaptive effective capacity tracks actual capacity changes, eliminating systematic biases.

[0019] 2. Extremely low resource consumption, suitable for embedded deployment: Because it uses polynomial operations instead of matrix operations (Kalman filtering) and neural network inference, the algorithm only requires basic arithmetic operations and does not require floating-point units. Compared to Kalman filtering (RAM > 100 bytes), this invention uses approximately 13 bytes of RAM (SOC_current: 4 bytes). 4 bytes The model size is 4 bytes (1 byte for the initialized flag), reducing the size by approximately 87%. Compared to neural networks (model > 10MB), the model coefficients of this invention require only 24 bytes (6 floats), reducing the size by approximately 99.9998%. Polynomial operations only require addition, subtraction, multiplication, and division, which can be executed by an 8-bit microcontroller (such as ATmega328P), with a computation time of <1ms@16MHz.

[0020] 3. Features automatic calibration capabilities and prevents error accumulation: Employing a charge-triggered adaptive calibration mechanism, the system automatically recalibrates the State of Charge (SOC) using the voltage model during each charge, eliminating accumulated errors. Simultaneously, a first-order low-pass filter updates the effective capacity and tracks battery aging. Compared to the ampere-hour integration method (which lacks self-calibration and allows errors to accumulate over time), this invention automatically returns the error to zero after each charge, maintaining accuracy over long-term use. The correlation between battery voltage and SOC during charging is more stable (voltage changes are more pronounced towards the end of charging), resulting in higher voltage model calibration accuracy. The low-pass filter smooths out the noise impact of single calibrations.

[0021] 4. Adaptive to battery aging and temperature changes: Employs dynamic effective capacity. Replace fixed nominal capacity The actual capacity change is tracked by a low-pass filter; the voltage model includes a temperature compensation term, adapting to a wide temperature range. Compared to fixed-parameter methods (which cannot adapt to aging), this invention maintains prediction accuracy throughout the battery's entire lifespan; compared to Kalman filtering, which requires online parameter identification, the adaptive mechanism of this invention has extremely low computational cost (only one multiplication and one addition). The first-order low-pass filter is the simplest and most effective parameter tracking method, with a computational cost of O(1), making it suitable for embedded implementation. Attached Figure Description

[0022] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the SOC prediction algorithm; Figure 3 This is a flowchart of the adaptive effective capacity update mechanism; Figure 4 It is a multi-strategy initialization mechanism decision tree; Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0023] Embodiments of the present invention: like Figures 1 to 4 As shown, the battery state of charge prediction system based on multi-model fusion includes: The data acquisition module is used to acquire battery voltage, current, temperature and capacity data at a preset sampling period.

[0024] The voltage model prediction module is used to predict the state of charge based on the current voltage and current temperature using a fourth-order polynomial regression model (fourth-order polynomial voltage-temperature model). The capacity accumulation calculation module is used to calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the dynamically updated effective capacity. The charging detection module is used to determine whether the battery is in a charging state based on the comparison result between the capacity increment and the preset charging threshold, and to trigger adaptive calibration when the battery is in a charging state. An adaptive calibration module is used to call the voltage model prediction module to obtain the calibration state of charge when charging, infer the new effective capacity based on the current capacity, and update the effective capacity through a filter. The rationality check module is used to impose boundary constraints on the predicted state of charge and effective capacity; The data output module is used to output the predicted state of charge and effective capacity.

[0025] The multi-strategy initialization module is used to select the initialization method based on available data when the system starts up: if initial capacity data exists, the initial state of charge is calculated based on the ratio of initial capacity to nominal capacity, and the initial effective capacity is deduced in reverse; if initial capacity data does not exist, the voltage model prediction module is called to estimate the initial state of charge based on the initial voltage and temperature, and the nominal capacity is set as the initial effective capacity.

[0026] The prediction equation for the fourth-order polynomial regression model is as follows: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting. This model performs two functions: initial SOC estimation and charging calibration, addressing the lack of initial values ​​and self-calibration capabilities in the ampere-hour integration method. The fourth-order polynomial was chosen because first- and second-order polynomials cannot accurately describe the nonlinear characteristics of the lithium battery OCV-SOC curve (especially the plateau region of SOC between 20% and 80%), while fifth-order and higher-order polynomials are prone to overfitting and increase computational load. The temperature term c5·T is used to compensate for the influence of temperature on the battery's electrochemical characteristics, ensuring the model maintains accuracy over a wide temperature range. The model coefficients c0 to c5 were obtained through offline training and stored in Flash (6 float coefficients, totaling 24 bytes). Prediction requires only 6 multiplications and 5 additions, resulting in minimal computational load, making it suitable for execution by an 8-bit microcontroller.

[0027] The capacity accumulation calculation module addresses the issue of accumulated errors caused by using a fixed nominal capacity in the traditional ampere-hour integration method. The calculation formula is as follows:

[0028] Input is the current capacity value (mAh) and capacity value at the previous moment (mAh), and current effective capacity (mAh), the output is the change in SOC ΔSOC (%). Unlike the traditional ampere-hour integration method that uses a fixed nominal capacity, this invention uses dynamically updated effective capacity. Replaces fixed nominal capacity.

[0029] Effective capacity The initial value is set to the nominal capacity. It is updated via a first-order low-pass filter during each charge calibration. When battery aging causes a decrease in actual capacity... It will automatically track the actual capacity to avoid accumulated errors.

[0030] Implementation details: Capacity accumulation calculation requires only one subtraction, one division, and one multiplication, resulting in minimal computational complexity. Effective capacity. Stored in RAM (1 float variable, 4 bytes).

[0031] The adaptive calibration module addresses the issue of increased prediction errors due to battery capacity changes with aging and temperature, enabling automatic tracking and updating of effective capacity. Implementation mechanism: Step (1): Charging detection; calculate capacity increment. If ΔQ > CHARGE_THRESHOLD (default 100mAh), it is determined to be in charging state, triggering the calibration process. The selection of the charging threshold of 100mAh is based on the fact that normal discharge noise is usually <50mAh, while the capacity increment during charging is usually >200mAh, and 100mAh can effectively distinguish between the two states.

[0032] Step (2): Voltage model calibration SOC; call the fourth-order polynomial regression model to re-estimate based on the current voltage and temperature. .

[0033] Step (3): Back-calculate the effective capacity; based on the current capacity and calibration To deduce the new effective capacity: .

[0034] Step (4): First-order low-pass filter update; To avoid the influence of noise introduced by a single calibration, a first-order low-pass filter is used to smoothly update the effective capacity: The filtering coefficient α = 0.2 (the value of α ranges from 0.1 to 0.3). The smaller the value of α, the stronger the filtering effect, but the slower the tracking speed; α = 0.2 strikes a balance between tracking speed and filtering effect.

[0035] Step (5): Reasonableness check; if the updated If the value exceeds the reasonable range [C_min, C_max] (default [30000mAh, 60000mAh]), the original value will be kept unchanged to prevent abnormal data from contaminating the effective capacity estimation.

[0036] The multi-strategy initialization module addresses the issue of unknown initial SOC upon system power-up, providing a reliable initial SOC estimate. Inputs are the initially acquired voltage, temperature, and capacity data; outputs are the initial SOC and initial effective capacity.

[0037] A priority strategy is adopted to select the optimal initialization method based on available data: Priority 1 (Capacity data available): If the BMS provides the current capacity (mAh), then the initial SOC is estimated by the capacity ratio: Simultaneously, the initial effective capacity is calculated backwards: This method offers the highest accuracy because the capacity data directly reflects the remaining battery power.

[0038] Priority 2 (No Capacity Data): If the BMS does not provide capacity data, a fourth-order polynomial regression model is called to estimate the initial SOC; this method is less accurate and is suitable for simplified BMSs without capacity sensors.

[0039] The above describes the basic structure of the system. The hardware components include a Battery Management Unit (BMU), a voltage acquisition module, a current acquisition module, and a temperature acquisition module. The voltage acquisition module uses a high-precision ADC (such as a 16-bit Σ-Δ ADC) with a sampling accuracy ≤5mV and a sampling frequency of 1Hz, used to acquire the voltage of individual battery cells or the average voltage of the battery pack. The current acquisition module uses a Hall effect current sensor or a shunt with a measurement accuracy ≤0.5% and a sampling frequency of 1Hz, used to acquire the charging and discharging current (positive values ​​indicate charging, negative values ​​indicate discharging). The temperature acquisition module uses an NTC thermistor or a digital temperature sensor (such as DS18B20) with a measurement accuracy ≤0.5℃ and a sampling frequency of 0.1Hz, used to acquire the battery temperature. The Battery Management Unit uses an 8-bit / 16-bit / 32-bit microcontroller (such as STM32F103, Arduino, ESP32), with ≥512 bytes of RAM, ≥32KB of Flash, and an operating frequency ≥16MHz.

[0040] The collected data is transmitted to a host computer or cloud server via communication protocols such as CAN bus, RS485, Bluetooth, and WiFi.

[0041] The overall operation process is as follows: This invention relates to a battery state of charge prediction method based on multi-model fusion, comprising the following steps: S1: System initialization, acquiring initial voltage, temperature, and capacity data, estimating initial state of charge and initial effective capacity; after initialization, the system enters the real-time prediction loop. If initial capacity data exists, the initial state of charge is calculated using the capacity ratio method; if initial capacity data does not exist, the initial state of charge is calculated using the voltage model prediction module.

[0042] S2: Collect current voltage, current, temperature and capacity data at a preset sampling period.

[0043] The sampling period can be set to 1 second to read the voltage. (mV), current (mA), temperature (°C), Capacity (mAh); The data acquisition in step S2 includes outlier removal preprocessing: if the difference between the current voltage and the voltage at the previous moment is greater than a preset voltage threshold, such as... If the current sampling point is not found, the voltage data from the previous moment will be discarded and replaced with the current sampling point.

[0044] S3: Calculate the capacity increment. If the capacity increment is greater than the preset charging threshold, determine that the system is in a charging state and proceed to step S4; otherwise, proceed to step S5. Capacity Increment , For the current capacity, The capacity at the previous moment.

[0045] S4: Charging calibration path: Call the fourth-order polynomial regression model to predict the calibration state of charge based on the current voltage and temperature, back-calculate the new effective capacity based on the current capacity, update the effective capacity through a filter, and set the current state of charge as the calibration state of charge.

[0046] The prediction equation of the fourth-order polynomial regression model is: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting.

[0047] The model coefficients are obtained through the following steps: Step (1): Collect battery experimental data under different discharge rates and temperature conditions, including voltage, temperature and true SOC; Step (2): Construct a training dataset with voltage V and temperature T as features and real SOC as labels; Step (3): Fit the fourth-order polynomial regression model using the least squares method to obtain the coefficients c0 to c5; Step (4): Store the coefficients in non-volatile memory for real-time prediction.

[0048] Calling a fourth-order polynomial regression model for prediction To deduce the new effective capacity The effective capacity is updated using a first-order low-pass filter: ,renew , where α is the filter coefficient, and α is set to 0.2.

[0049] S5: Discharge accumulation path: Calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the current effective capacity, and add it to the state of charge at the previous moment.

[0050] Enter the current capacity Previous timeframe capacity Current effective capacity ;calculate , cumulative Output the updated version .

[0051] S6: Perform a reasonableness check on the state of charge and effective capacity.

[0052] For prediction Process it, Limiting to the range [0, 100], Limited to Within the range.

[0053] S7: Output the state of charge and effective capacity, update the capacity of the previous moment, return to step S2, and enter the next sampling period.

[0054] Application Scenario 1: Electric Vehicle BMS 1. Hardware Configuration Terminal Equipment: Battery Management Unit (BMU): STM32F103C8T6 microcontroller, ARM Cortex-M3 core, 72MHz clock speed, 20KB RAM, 64KB Flash. Voltage Acquisition Module: 16-bit Σ-Δ ADC (ADS1115), sampling accuracy ±5mV, sampling frequency 1Hz, I2C interface. Current Acquisition Module: Hall current sensor (ACS758), range ±200A, accuracy 0.5%. Temperature Acquisition Module: NTC thermistor (B value 3950K), measurement accuracy ±0.5℃. Battery Pack: 14-cell 20Ah lithium iron phosphate battery pack, nominal capacity 51200mAh, nominal voltage 51.8V.

[0055] Data transmission: CAN bus (500kbps baud rate).

[0056] 2. Software parameters Voltage model prediction module coefficients: c0=-1247.3856, c1=1089.2341, c2=-356.7892, c3=52.1234, c4=-2.8765, c5=0.1523; Nominal capacity: = 51200.0 mAh; Charging detection threshold: CHARGE_THRESHOLD = 100.0 mAh; Capacity filter coefficient: α = 0.2; Effective capacity range: [30000.0, 60000.0] mAh.

[0057] 3. Steps Step (1): System initialization, vehicle power-on, read V0=52.3V, T0=25℃, Q0=45000mAh; Calculate: Conversely: ; Step (2): Real-time prediction (discharge), 10th second: V 10 =51.8V, T 10 =26℃, Q 10 =44500mAh; It was determined to be a discharge. , ; Step (3): Charging calibration, 1000th second: V1000=54.2V, T1000=28℃, Q1000=48000mAh Trigger charging calibration (Calculated by the voltage model prediction module) , .

[0058] Effect: Accuracy: MAE < 2%, RMSE < 3%; Real-time performance: SOC updated once per second; Resource consumption: RAM only 13 bytes, Flash only 24 bytes; Adaptability: Maintains stable accuracy throughout the entire battery life.

[0059] Application Scenario 2: Energy Storage System 1. Hardware Configuration Terminal Equipment: Battery Management Unit: ESP32-WROOM-32, dual-core Xtensa LX6, 240MHz clock speed, 520KB RAM. Voltage Acquisition Module: 12-bit SAR ADC (built into ESP32), sampling accuracy ±10mV. Current Acquisition Module: Shunt (75mV / 100A), accuracy 0.5%. Temperature Acquisition Module: DS18B20 digital temperature sensor, accuracy ±0.5℃. Battery Pack: 16-cell 100Ah lithium iron phosphate battery pack, nominal capacity 100,000mAh.

[0060] Data transmission: WiFi (802.11n).

[0061] 2. Software parameters Voltage model prediction module coefficients: c0=-1350.2341, c1=1120.5678, c2=-368.9012, c3=54.3456, c4=-3.0123, c5=0.1678; Nominal capacity: = 100000.0 mAh; Charging detection threshold: CHARGE_THRESHOLD = 200.0 mAh; Capacity filtering coefficient: α = 0.15; Effective capacity range: [70000.0, 120000.0] mAh.

[0062] 3. Steps Step (1): System initialization, power-on of the energy storage system, read V0=52.8V, T0=20℃, Q0=80000mAh. , .

[0063] Step (2): Photovoltaic charging, 600th second: V 600 =53.5V, T 600 =22℃, Q 600 =82000mAh, Trigger charging calibration - , .

[0064] Step (3): Nighttime discharge, 3600 seconds: V 3600 =51.5V, T 3600 =18℃, Q 3600 =70000mAh It was determined to be a discharge. , .

[0065] Effect: Long-term accuracy: MAE < 2.5% within a 3-year usage period; Temperature adaptability: Accuracy variation < 1% within the range of -10℃ to 45℃; Cloud monitoring: SOC data is uploaded in real time via WiFi.

[0066] Application Scenario 3: Portable Devices 1. Hardware Configuration Terminal Devices: Battery Management Unit: Arduino Nano (ATmega328P), 8-bit AVR core, 16MHz clock speed, 2KB RAM. Voltage Acquisition Module: 10-bit ADC (built into ATmega328P), sampling accuracy ±20mV. Current Acquisition Module: INA219 current sensor, range ±3.2A, accuracy 0.5%. Temperature Acquisition Module: LM35 temperature sensor, accuracy ±0.5℃. Battery Pack: 3 sets of 2600mAh 18650 lithium batteries, nominal capacity 2600mAh.

[0067] Data transmission: Bluetooth (HC-05 module).

[0068] 2. Software parameters Voltage model prediction module coefficients: c0=-890.1234, c1=780.5678, c2=-245.3456, c3=35.6789, c4=-1.9012, c5=0.0987; Nominal capacity: = 2600.0 mAh; Charging detection threshold: CHARGE_THRESHOLD = 10.0 mAh; Capacity filter coefficient: α = 0.25; Effective capacity range: [1800.0, 3000.0] mAh.

[0069] 3. Steps Step (1): System initialization, device power-on, read V0=11.8V, T0=25℃, no capacity data, use voltage model prediction module: , .

[0070] Step (2): Usage process (discharge), 300 seconds: V 300 =11.5V, T 300 =26℃, no capacity data, calling voltage model prediction module: SOC 300 = 58.76%.

[0071] Step (3): During the charging process, the charging current is detected. The state is marked as charging, and the voltage model prediction module is called to update the SOC each time a sample is taken.

[0072] Effect: Ultra-low resource consumption: only 13 bytes of RAM and 24 bytes of Flash; Fast response: operation time <1ms@16MHz; Simplified deployment: no capacity sensor required.

[0073] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0074] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery state of charge prediction system based on multi-model fusion, characterized in that, include: The data acquisition module is used to acquire battery voltage, current, temperature and capacity data at a preset sampling period; The voltage model prediction module is used to predict the state of charge based on the current voltage and current temperature using a fourth-order polynomial regression model. The capacity accumulation calculation module is used to calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the dynamically updated effective capacity. The charging detection module is used to determine whether the battery is in a charging state based on the comparison result between the capacity increment and the preset charging threshold, and to trigger adaptive calibration when the battery is in a charging state. An adaptive calibration module is used to call the voltage model prediction module to obtain the calibration state of charge when charging, infer the new effective capacity based on the current capacity, and update the effective capacity through a filter. The rationality check module is used to impose boundary constraints on the predicted state of charge and effective capacity; The data output module is used to output the predicted state of charge and effective capacity.

2. The battery state of charge prediction system based on multi-model fusion according to claim 1, characterized in that, It also includes a multi-strategy initialization module, which selects the initialization method based on available data when the system starts up: if initial capacity data exists, the initial state of charge is calculated based on the ratio of the initial capacity to the nominal capacity, and the initial effective capacity is deduced in reverse. If no initial capacity data exists, the voltage model prediction module is invoked to estimate the initial state of charge based on the initial voltage and temperature, and the nominal capacity is set as the initial effective capacity.

3. The battery state of charge prediction system based on multi-model fusion according to claim 1, characterized in that: The prediction equation of the fourth-order polynomial regression model is: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting.

4. The battery state of charge prediction system based on multi-model fusion according to claim 1, characterized in that: The adaptive calibration module updates the effective capacity based on the current capacity in the following manner. and Back-inferring new effective capacity , , The charge state is predicted by a fourth-order polynomial regression model; and the effective capacity is updated through a first-order low-pass filter: , where α is the filter coefficient, and α is set between 0.1 and 0.

3.

5. A battery state-of-charge prediction method based on multi-model fusion, characterized in that, Includes the following steps: S1: System initialization, acquiring initial voltage, temperature and capacity data, and estimating initial state of charge and initial effective capacity; S2: Collect current voltage, current, temperature and capacity data at a preset sampling period; S3: Calculate the capacity increment. If the capacity increment is greater than the preset charging threshold, it is determined to be in a charging state and step S4 is executed; otherwise, step S5 is executed. S4: Charging calibration path: Call the fourth-order polynomial regression model to predict the calibration state of charge based on the current voltage and temperature, back-calculate the new effective capacity based on the current capacity, update the effective capacity through the filter, and set the current state of charge as the calibration state of charge. S5: Discharge accumulation path: Calculate the change in state of charge based on the difference between the current capacity and the capacity at the previous moment, as well as the current effective capacity, and add it to the state of charge at the previous moment. S6: Perform a reasonable check on the state of charge and effective capacity; S7: Output the state of charge and effective capacity, update the capacity of the previous moment, and return to step S2.

6. The battery state of charge prediction method based on multi-model fusion according to claim 5, characterized in that, In step S1, if initial capacity data exists, the initial state of charge is calculated using the capacity ratio method; if initial capacity data does not exist, the initial state of charge is calculated using a fourth-order polynomial regression model.

7. The battery state of charge prediction method based on multi-model fusion according to claim 5, characterized in that, The data acquisition in step S2 includes outlier removal preprocessing: if the difference between the current voltage and the voltage at the previous moment is greater than a preset voltage threshold, the current sampling point is discarded and the voltage data at the previous moment is used instead.

8. The battery state of charge prediction method based on multi-model fusion according to claim 5, characterized in that: The prediction equation of the fourth-order polynomial regression model is: Where V is voltage and T is temperature. to These are the model coefficients obtained through least-squares polynomial fitting.

9. The battery state of charge prediction method based on multi-model fusion according to claim 8, characterized in that, The model coefficients are obtained through the following steps: Step (1): Collect battery experimental data under different discharge rates and temperature conditions, including voltage, temperature and true SOC; Step (2): Construct a training dataset with voltage V and temperature T as features and real SOC as labels; Step (3): Fit the fourth-order polynomial regression model using the least squares method to obtain the coefficients c0 to c5; Step (4): Store the coefficients in non-volatile memory for real-time prediction.

10. The battery state of charge prediction method based on multi-model fusion according to claim 5, characterized in that: The formula for calculating the change in state of charge in step S5, , For the current capacity, The capacity at the previous time step. For dynamically updated effective capacity; in step S4, a fourth-order polynomial regression model is called to predict. To deduce the new effective capacity The effective capacity is updated using a first-order low-pass filter: ,renew , where α is the filter coefficient, and α is set between 0.1 and 0.3.

Images