Power battery SOC estimation system and method

Abstract

The invention provides a power battery SOC estimation system and method, and relates to the technical field of new energy automobiles. The system comprises a data input and preprocessing module, a deep learning SOC preliminary prediction module, a PNGV model module and a filtering module, and the data input and preprocessing module is used for preprocessing original battery data and outputting a preprocessed data stream; the deep learning SOC preliminary prediction module is used for outputting a deep learning SOC preliminary prediction value at the current moment; the PNGV model module is used for outputting a PNGV model SOC preliminary prediction value at the current moment; and the filtering module is used for performing fusion filtering calculation and outputting a fused SOC estimation result. According to the system, the SOC estimation precision can be remarkably improved, the robustness and adaptability of SOC estimation are enhanced, the response capability to complex working conditions is improved, the dependence on model parameters is relatively reduced, and a more reliable basis is provided for battery full-life-cycle management.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicles, and in particular to a power battery SOC estimation system and method. Background Art

[0002] The state of charge (SOC) of a power battery is one of the most critical parameters in a battery management system (BMS). It directly impacts an electric vehicle's range display, energy management strategy, and safe battery operation. Currently, commonly used SOC estimation methods include: 1. The ampere-hour integration method (coulomb counting): This method estimates SOC by accumulating the amount of charge flowing through the battery. This method is simple and easy to use, but its accuracy gradually decreases due to factors such as inaccurate initial SOC, accumulated current measurement errors, and changes in battery capacity with aging and temperature. 2. The open-circuit voltage method (OCV-SOC): This method estimates SOC using the relatively stable relationship between the open-circuit voltage and the SOC after the battery has been at rest for a sufficient period of time. This method is highly accurate, but requires the battery to be at rest for an extended period of time and is not suitable for dynamic operating conditions. 3. Model-based estimation methods: These methods establish a battery equivalent circuit model (such as the Thevenin model, PNGV model, or second-order RC model) and combine it with a Kalman filter (KF) and its variants (such as the extended Kalman filter (EKF) and the unscented Kalman filter (UKF)) to estimate SOC. These methods achieve good results when the model is accurate and parameter identification is precise. However, battery model accuracy is affected by multiple factors, including temperature, aging, and charge / discharge rate. Parameter identification is complex, and model mismatch is particularly prominent under complex dynamic operating conditions. 4. Data-driven estimation methods: With the recent development of artificial intelligence (AI) technology, SOC estimation methods based on machine learning algorithms, such as neural networks and support vector machines, have received widespread attention. These methods can learn the complex nonlinear behavior of batteries from large amounts of historical data without requiring a precise physical model. However, purely data-driven methods can suffer from insufficient generalization, require high training data coverage, and exhibit poor interpretability under certain extreme operating conditions.

[0003] In existing technologies, a single method often struggles to achieve both accuracy and robustness under all operating conditions. Therefore, integrating the strengths of different methods to develop a solution capable of achieving highly accurate and robust SOC estimation across a wide temperature range, throughout the entire lifecycle, and under complex dynamic conditions remains a hot topic and a key challenge in current BMS research. Summary of the Invention

[0004] The purpose of the present invention is to provide a power battery SOC estimation system and method to solve the SOC estimation problem in view of the above-mentioned deficiencies in the prior art.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a power battery SOC estimation system, which includes: a data input and preprocessing module, a deep learning SOC preliminary prediction module, a PNGV model module, and a filtering module. The data input and preprocessing module is used to receive raw battery data collected from the internal or external sensor network of the battery pack, preprocess the raw battery data, and output a preprocessed data stream. The preprocessed data stream includes: feature sequence data containing historical information and current information, as well as the real-time driving current value at the current moment; The deep learning SOC preliminary prediction module is used to receive the feature sequence data from the data input and preprocessing module, and output the deep learning SOC preliminary prediction value at the current moment. The deep learning SOC preliminary prediction module adopts a hybrid deep learning model based on dilated convolution and bidirectional gated recurrent unit; The PNGV model module is used to receive the current real-time driving current value from the data input and preprocessing module and the SOC state value fed back by the filtering module at the previous moment, and output the preliminary predicted SOC value of the PNGV model at the current moment; The filtering module is used to receive the deep learning SOC preliminary prediction value from the deep learning SOC preliminary prediction module, the PNGV model SOC preliminary prediction value from the PNGV model module, and the battery terminal voltage and battery current actually measured at the current moment from the data input and preprocessing module, and perform fusion filtering calculations to output the fused SOC estimation result. The filtering module is also used to feed back the fused SOC estimation result to the PNGV model module for correcting the internal SOC state of the PNGV model module.

[0006] Optionally, the hybrid deep learning model includes an expanded convolution layer, a BiGRU layer, and a fully connected output layer. The expanded convolution layer is used to extract multi-scale local features from the input feature sequence data. The BiGRU layer is used to process the feature sequence extracted by the expanded convolution layer to capture the temporal dynamics and contextual information in the sequence. The fully connected output layer is used to map the output of the BiGRU layer to an SOC prediction value. The hybrid deep learning model uses mean square error or mean absolute error as a loss function to measure the difference between the model-predicted SOC and the actual SOC.

[0007] Optionally, the filtering module is a fusion extended Kalman filter module.

[0008] In a second aspect, the present invention further provides a method for estimating the SOC of a power battery, which is used in the system according to the first aspect, and comprises the following steps: The power battery SOC estimation system is started and initialized. When the vehicle's battery management system is powered on, the weight parameters of the pre-trained deep learning model are loaded, the parameters of the PNGV model are loaded, the state variables of the filter module are initialized, and the process noise covariance matrix and measurement noise covariance matrix of the filter module are set; Data acquisition and preprocessing: The power battery SOC estimation system enters a real-time estimation loop, collects the battery's raw voltage, current, and temperature data in real time through sensors, calls the data input and preprocessing module to preprocess the collected raw data, and outputs preprocessed feature sequence data containing historical and current information as well as the current real-time driving current; The deep learning SOC preliminary prediction module performs preliminary SOC prediction, inputs the preprocessed feature sequence data into the loaded deep learning model, and the deep learning model performs forward propagation calculation and outputs the deep learning SOC preliminary prediction value at the current moment; The PNGV model module performs a preliminary SOC prediction, using the current real-time current and the SOC state fused by the filter module at the previous moment. According to the PNGV model, the preliminary SOC prediction value of the PNGV model at the current moment is calculated; The filtering module performs fusion filtering. The filtering module is called and first performs the state prediction step: based on the fused SOC state and current current at the previous moment, the state equation of the PNGV model is used to predict the current moment's prior SOC state and prior covariance. Then, the measurement update step is performed: the deep learning SOC preliminary prediction value, the PNGV model SOC preliminary prediction value, and the current actual measured battery terminal voltage are used as the composite measurement value of the filtering module; the Kalman gain is calculated, and the prior SOC state and covariance are updated based on the residual between the composite measurement value and the prediction value to obtain the current moment's a posteriori, fused SOC estimate. Output the SOC estimation result, and use the fused SOC estimation value output by the filtering module as the final SOC estimation result at the current moment; Optionally, after outputting the SOC estimation result, the method further includes the following steps: Check the battery management system operating status. If the battery management system is still working properly, proceed to the next round of SOC estimation. If the battery management system is turned off, the estimation process is stopped and the current final SOC estimation result is saved in the non-volatile memory for use at the next startup.

[0009] The beneficial effects of the present invention include: The power battery SOC estimation system provided by the present invention includes: a data input and preprocessing module, a deep learning SOC preliminary prediction module, a PNGV model module, and a filtering module. The data input and preprocessing module is used to receive raw battery data collected from the internal or external sensor network of the battery pack, preprocess the raw battery data, and output a preprocessed data stream. The preprocessed data stream includes: feature sequence data containing historical information and current moment information, and the real-time driving current value at the current moment; the deep learning SOC preliminary prediction module is used to receive feature sequence data from the data input and preprocessing module, and output a deep learning SOC preliminary prediction value at the current moment. The deep learning SOC preliminary prediction module adopts a hybrid deep convolution based on dilated convolution and bidirectional gated recurrent unit. The PNGV model module receives the current real-time driving current value from the data input and preprocessing module and the SOC state value fed back by the filtering module at the previous moment, and outputs the current PNGV model SOC preliminary prediction value. The filtering module receives the deep learning SOC preliminary prediction value from the deep learning SOC preliminary prediction module, the PNGV model SOC preliminary prediction value from the PNGV model module, and the current actual measured battery terminal voltage and battery current from the data input and preprocessing module, performs fusion filtering calculations, and outputs the fused SOC estimation result. The filtering module is also used to feed the fused SOC estimation result back to the PNGV model module for correcting the internal SOC state of the PNGV model module. This system can significantly improve the accuracy of SOC estimation, enhance the robustness and adaptability of SOC estimation, improve the responsiveness to complex operating conditions, and relatively reduce the dependence on model parameters, providing a more reliable foundation for battery life cycle management. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 The figure shows the overall architecture of the power battery SOC estimation system provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a hybrid deep learning model provided by an embodiment of the present invention is shown; Figure 3 A schematic diagram of the structure of a PNGV model provided by an embodiment of the present invention is shown; Figure 4 The figure shows a flow chart of a method for estimating the SOC of a power battery provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0013] Existing SOC estimation technologies suffer from the following major drawbacks: 1. Accumulated error in the ampere-hour integration method: Factors such as current sensor accuracy limitations, inaccurate initial integration values, and variations in coulombic efficiency lead to cumulative SOC estimation errors after prolonged operation, reducing reliability. 2. Application limitations of the open-circuit voltage method: This requires the battery to remain stationary for extended periods, making it inadequate for real-time SOC estimation during dynamic driving conditions. 3. Model mismatch and parameter sensitivity in traditional model methods: The parameters of the battery equivalent circuit model (such as internal resistance and capacitance) are highly sensitive to temperature, SOC, and state of aging (SOH), making parameter identification difficult and online adaptive adjustment complex. Under complex operating conditions or after battery aging, model mismatch can significantly reduce SOC estimation accuracy. 4. Limitations in interpretability and generalization of single data-driven methods: While methods like deep learning can fit complex nonlinear relationships, their "black box" nature can sometimes make the estimation results difficult to interpret. Furthermore, their generalization capabilities can be limited in edge cases not covered by the training data, posing robustness challenges. 5. Sensitivity to sensor noise and data quality: Some methods are sensitive to the noise of voltage and current sensors. Improper preprocessing may lead to large fluctuations in the estimation results.

[0014] The present invention aims to propose a power battery SOC estimation system and method based on the fusion of a deep learning model and the PNGV (Partnership for a New Generation of Vehicles) equivalent circuit model, so as to significantly improve the accuracy, robustness and environmental adaptability of SOC estimation.

[0015] Figure 1 FIG. 1 shows a schematic diagram of the overall architecture of a power battery SOC estimation system provided by an embodiment of the present invention. Figure 1 As shown, the power battery SOC estimation system provided by the present invention includes: a data input and preprocessing module 100, a deep learning SOC preliminary prediction module 200_DL, a PNGV model module 300, and a filtering module 400.

[0016] The data input and preprocessing module 100 is used to receive raw battery data 10 collected from a sensor network inside or outside the battery pack, preprocess the raw battery data 10, and output a preprocessed data stream. The preprocessed data stream includes: feature sequence data containing historical information and current time information, as well as the real-time drive current value at the current moment. Raw battery data 10 is typically provided in the form of a time series. Key data includes, but is not limited to: a sequence of terminal voltages of battery cells or modules, a sequence of charge and discharge currents flowing through the battery (precisely distinguishing between charging and discharging directions, e.g., charging is positive and discharging is negative), and a sequence of temperatures at key temperature measurement points within the battery pack (e.g., the temperature of battery cell surfaces, module centers, buses, etc.).

[0017] The data input and preprocessing module 100 is capable of communicating directly with the underlying hardware or sensor network of the battery management system (BMS). This may involve parsing specific communication protocols (such as CAN bus messages, SPI, and I2C) to accurately extract raw measurement data. The acquisition frequency depends on the BMS design and application requirements. For example, current signals may be acquired at a frequency of 10-100 Hz to capture dynamic changes, while voltage and temperature signals may be acquired at a frequency of 1-10 Hz.

[0018] The original sensor signal is often accompanied by various noises, such as electromagnetic interference, sensor noise itself, etc. The data input and preprocessing module 100 first performs preliminary signal conditioning and denoising on the collected original signal. Commonly used digital filtering algorithms may include but are not limited to: mean filtering or sliding average filtering: suitable for smoothing high-frequency random noise, by calculating the average value of the data in a fixed-size window as the filtering result of the current point. The selection of window size requires a trade-off between the denoising effect and the signal lag caused by smoothing. Median filtering: particularly effective for removing pulse noise or glitch data (spikes) accidentally generated by the sensor. It replaces the value of the current point with the median of the data in the window. Gaussian filtering: a linear smoothing filter that is better at suppressing noise that follows a normal distribution. Its weight is determined according to the Gaussian function. The goal of this stage is to improve the signal-to-noise ratio and reliability of the data required for subsequent processing steps.

[0019] Because the sampling times of different sensors or different measurement units within the BMS may have slight time deviations, or different signals may have different sampling rates, the data input and preprocessing module 100 includes a critical data alignment and synchronization sub-function. This is usually based on precise timestamp information. Through interpolation (such as linear interpolation or spline interpolation) or resampling techniques, all relevant input data (voltage, current, temperature, etc.) are unified to the same time base and sampling frequency, forming a strictly time-aligned, multi-dimensional synchronized data set, which is crucial for subsequent timing model analysis.

[0020] In actual operation, sensors may generate obvious abnormal data points or outliers due to faults, interference or extreme operating conditions. If these outliers are not processed, they will seriously affect the training accuracy of the subsequent model and the accuracy of SOC estimation. In this step, the present invention adopts a strategy of directly eliminating obvious outliers in the identified raw data (for example, voltage values ​​that exceed the physical allowable range of the battery, current mutations far exceeding the normal rate, etc.). After elimination, the missing state of the time point can be selectively retained or filled with simple adjacent values ​​to avoid these abnormal data from causing adverse propagation effects on subsequent model training and SOC estimation. The identification process can be based on a preset physical threshold, a statistical method (such as the 3-sigma principle) or a simple jump detection.

[0021] After the aforementioned series of processing, the data input and preprocessing module 100 outputs a structured, high-quality preprocessed data stream. This preprocessed data stream can be specifically divided into: 1) Feature sequence data 101a, containing historical and current information: This is typically a multidimensional time series consisting of normalized voltage, current, temperature, and possible derived features. This data stream primarily serves as input to the subsequent deep learning SOC preliminary prediction module 200_DL and may also provide current observations (such as terminal voltage V_m,k and temperature T_k) to the filtering module 400. 2) Current real-time driving current value (Ik) 101b: This is the processed and verified battery charge and discharge current value at the current sampling moment. It primarily serves as input to the PNGV model module 300's dynamic model and the state prediction phase of the filtering module 400.

[0022] The deep learning SOC preliminary prediction module 200_DL is used to receive the feature sequence data 101a from the data input and preprocessing module 100, and output the deep learning SOC preliminary prediction value 201 at the current moment. The deep learning SOC preliminary prediction module 200_DL adopts a hybrid deep learning model based on dilated convolution and bidirectional gated recurrent unit. The feature sequence data 101a usually contains the battery voltage, current, temperature and possible other derived features within a window of a period of time in the past. The deep learning SOC preliminary prediction module 200_DL adopts a hybrid deep learning model based on dilated convolution and bidirectional gated recurrent unit (BiGRU), which is denoted as dilated convolution BiGRU model. The model structure is as follows: Figure 2As shown, the hybrid deep learning model includes an expanded convolution layer (specifically including: expanded convolution layer 200b-1, expanded convolution layer 200b-2, expanded convolution layer 200b-3), a BiGRU layer 200c and a fully connected output layer 200d. The expanded convolution layer is used to extract multi-scale local features from the input feature sequence data. The BiGRU layer 200c is used to process the feature sequence extracted by the expanded convolution layer to capture the temporal dynamics and contextual information in the sequence. The fully connected output layer 200d is used to map the output of the BiGRU layer 200c to the SOC prediction value. The hybrid deep learning model uses mean square error or mean absolute error as the loss function to measure the difference between the model-predicted SOC and the actual SOC.

[0023] Specifically, the dilated convolution layer extracts multi-scale local features from the input time series data. By setting different dilation rates (e.g., 1, 2, or 4), the convolution kernel can capture a wider receptive field without increasing the number of parameters or performing pooling operations. This enables the model to effectively learn long-range dependencies in the input sequence while preserving the resolution of the time series. Multiple layers of dilated convolution are stacked, with the dilation rate increasing exponentially (e.g., 1, 2, or 4). Each dilated convolution layer is typically followed by batch normalization and an activation function (e.g., ReLU or LeakyReLU) to improve the model's stability and nonlinear expressiveness. The BiGRU layer 200c further processes the feature sequence extracted by the dilated convolution layer, capturing the temporal dynamics and contextual information within the sequence. The GRU is an improved recurrent neural network (RNN) that effectively mitigates the vanishing / exploding gradient problem through a gating mechanism. The bidirectional GRU (BiGRU) processes the sequence simultaneously in both the forward and backward directions, allowing the current output to leverage both past and future contextual information (within the input sequence), thereby providing a more comprehensive understanding of battery dynamic characteristics. The fully connected output layer 200d is used to map the output of the BiGRU layer (usually the hidden state of the last time step or the pooled result of the hidden state of all time steps) to the SOC prediction value. The output layer uses one or more fully connected layers (Dense Layer) and uses the Sigmoid activation function to normalize the output to the range of [0, 1], representing the SOC value of the battery ( Figure 1 Preliminary predictions of deep learning in SOC_DL 201, Figure 2 SOC prediction value in 203).

[0024] This deep learning model requires offline training to learn its parameters. Training data utilizes a large, diverse battery operating data set covering various driving conditions (such as UDDS, NEDC, and WLTP), ambient temperatures, battery aging, and charge / discharge rates. Mean squared error (MSE) or mean absolute error (MAE) is typically used as a loss function to measure the difference between the model's predicted SOC and the true SOC (determined through high-precision laboratory testing or reliable calibration methods). Adaptive learning rate optimization algorithms such as Adam and RMSprop are used to iteratively optimize model parameters. Training strategies include setting an appropriate batch size, epochs, learning rate and its decay strategy, and possible regularization methods (such as dropout and L2 regularization) to prevent overfitting. The model automatically learns the nonlinear mapping between battery SOC and multi-dimensional inputs (such as voltage, current, and temperature) from complex historical data. It has low reliance on the accuracy of the battery physical model and is good at capturing long-term dependencies and battery dynamic characteristics. By combining dilated convolution and BiGRU, the model can efficiently process multi-scale features and temporal dynamics, improving the accuracy and robustness of SOC prediction. The model output is the current deep learning SOC preliminary prediction value SOC_DL ( Figure 1 Preliminary predictions of deep learning in SOC_DL 201, Figure 2 SOC prediction value in 203).

[0025] The PNGV model module 300 is used to receive the current real-time driving current value (Ik) 101b from the data input and preprocessing module 100 and the SOC state value fed back by the filtering module 400 (or maintained by itself) at the previous moment, and output the preliminary predicted SOC value of the PNGV model at the current moment.

[0026] The PNGV model module 300 uses the PNGV (Partnership for a New Generation of Vehicles) equivalent circuit model. This is a commonly used battery model that can better describe the dynamic voltage response characteristics of the battery. Figure 3As shown, a typical PNGV model typically consists of the following components: an open-circuit voltage source Uoc, whose voltage is a function of SOC and temperature T, namely, Uoc(SOC, T); a series resistor R0, which represents the battery's ohmic internal resistance and can also be a function of SOC and temperature, namely, R0(SOC, T); and an RC network, which uses a first-order or second-order RC network to simulate electrochemical polarization and concentration polarization effects. The second-order model includes a first RC network (R1, C1) with a voltage of U_RC1 and a second RC network (R2, C2) with a voltage of U_RC2. The parameters R1, C1, R2, and C2 can vary with SOC and temperature. The battery terminal voltage U0 can be expressed as: U0 = Uoc(SOC, T) - I*R0(SOC, T) - U_RC1 - U_RC2.

[0027] The update of SOC is mainly based on the ampere-hour integration method: SOC_k=SOC_{k-1}-(η*I_{k-1}*Δt) / C_n, where η is the coulomb efficiency, C_n is the rated or current available capacity of the battery, and Δt is the sampling time interval.

[0028] The accuracy of the PNGV model is highly dependent on the accuracy of its parameters (Uoc-SOC curve, R0, R1, C1, R2, C2, etc.). These parameters are typically identified through offline experiments (such as hybrid pulse power characterization (HPPC) testing, electrochemical impedance spectroscopy (EIS), and open-circuit voltage testing) at different SOC, temperature, and aging conditions. These parameters are then stored as lookup tables (LUTs) or fitted as functions. In advanced applications, these parameters can be updated in real time or periodically using recursive least squares (RLS) to adapt to changes in battery state.

[0029] The PNGV model can model battery behavior based on clear physical and chemical meanings, is relatively computationally intensive, and is easy to implement. When the model parameters are accurate, it can provide relatively reliable SOC tracking.

[0030] like Figure 1 As shown, the output of the PNGV model module 300 is the current PNGV model SOC preliminary prediction value SOC_PNGV 301. This output is the SOC obtained based on the ampere-hour integration and the model voltage response.

[0031] The filtering module 400 is configured to receive the deep learning preliminary SOC prediction value SOC_DL(Z_dl,k) 201 from the deep learning preliminary SOC prediction module 200_DL, the PNGV model preliminary SOC prediction value SOC_PNGV(Z_pngv,k) 301 from the PNGV model module 300, and the currently measured battery terminal voltage (V_m,k) 101a and battery current (Ik) 101b from the data input and preprocessing module 100. The filtering module 400 is also configured to receive the fused SOC_{k-1}_fused from the previous moment, perform fusion filtering calculations, and output a fused SOC estimation result. The filtering module 400 is further configured to feed the fused SOC estimation result back to the PNGV model module 300 for correcting the internal SOC state of the PNGV model module 300. For example, the filtering module 400 may be a fused extended Kalman filter module.

[0032] The core algorithm of filtering module 400 is the Extended Kalman Filter (EKF). The EKF is an optimization estimation algorithm for nonlinear systems. It achieves an optimal estimate of the system state by linearizing the nonlinear state equation and observation equation. In this invention, the EKF is used to fuse SOC information from the deep learning model and the PNGV model, combined with actual voltage and current measurements, to obtain a more accurate SOC estimate.

[0033] The construction of the EKF state and measurement equations (one possible implementation) is as follows: State variables (x_k): The core state variable is the battery's true SOC. To more accurately exploit the dynamic characteristics of the PNGV model, the state variables can be expanded to x_k = [SOC_k, U_RC1_k, U_RC2_k]^T (taking PNGV as a second-order RC model as an example), where U_RC1_k and U_RC2_k are the voltages across the two RC networks, respectively.

[0034] State equation (x_k=f(x_{k-1}, u_k)+w_k): SOC_k=SOC_{k-1}-(η*I_{k-1}*Δt) / C_n+w_soc_k U_RC1_k=U_RC1_{k-1}*exp(-Δt / (R1*C1))+R1*(1-exp(-Δt / (R1*C1)))*I_{k-1}+w_u1_k U_RC2_k=U_RC2_{k-1}*exp(-Δt / (R2*C2))+R2*(1-exp(-Δt / (R2*C2)))*I_{k-1}+w_u2_k where u_k is the current input current I_k (or I_{k-1}, depending on the discretization method), w_k is the process noise vector, and its covariance matrix is ​​Q. The parameters of the PNGV model (R1, C1, R2, C2, η, Cn) are used here and may be adjusted according to temperature and aging state.

[0035] Measurement equation (z_k=h(x_k,u_k)+v_k): The measurement vector z_k can contain information from multiple sources: z_k=[V_m,k; SOC_DL,k; SOC_PNGV,k]^T The corresponding measurement function h(x_k,u_k) is: h1(x_k,u_k)=Uoc(SOC_k)-I_k*R0-U_RC1_k-U_RC2_k (terminal voltage predicted by PNGV model) h2(x_k,u_k)=SOC_k (for fusion SOC_DL) h3(x_k,u_k)=SOC_k (used for fusion of SOC_PNGV) Where v_k is the measurement noise vector, and its covariance matrix is ​​R. The diagonal elements of the R matrix can be set based on the confidence level in the measured voltage, SOC_DL, and SOC_PNGV, and can even be adjusted dynamically. For example, when operating conditions change dramatically, the noise variance of the corresponding SOC_DL item can be appropriately increased.

[0036] EKF execution steps: including state prediction, covariance prediction, calculation of Kalman gain, state update and covariance update and other standard steps.

[0037] The main technical features of filter module 400 are its ability to effectively integrate multiple information sources, effectively smooth noise, and enable online SOC estimation and correction. By rationally designing the state equation, observation equation, and noise covariances Q and R, it is possible to achieve an optimal combination of the deep learning model output and the PNGV model output.

[0038] The output of the filter module 400 is the fused high-precision SOC estimation result (SOC_fused,k) 401. At the same time, the fused SOC can be fed back to the PNGV model module 300 (e.g. Figure 1 The SOC is corrected by the internal SOC state and even used for online parameter calibration.

[0039] The data flow and connection relationship between modules are as follows: Raw battery data 10, serving as the system's initial input, is fed into the data input and preprocessing module 100. After processing, the module 100 outputs two data streams: preprocessed feature sequence data 101a (including historical and current voltage, current, temperature, etc.) is provided to the deep learning SOC preliminary prediction module 200_DL as its primary sequence input and to the filtering module 400 as the current measured voltage V_m,k required in its observation equation. The current real-time drive current (Ik) 101b (which may also include the current temperature Tk) is provided to the PNGV model module 300 as its model drive input and to the filtering module 400 as the current input required in its state prediction equation. The deep learning SOC preliminary prediction module 200_DL calculates the preliminary SOC_DL 201 based on the input 101a and outputs it to the filtering module 400. The PNGV model module 300 calculates a preliminary SOC_PNGV 301 based on the current real-time drive current (Ik) 101b (and possible feedback SOC) and outputs it to the filtering module 400. The filtering module 400 receives SOC_DL from 201, SOC_PNGV from 301, Vm,k from 101a, and Ik from 101b, performs fusion filtering calculations, and ultimately outputs a fused high-precision SOC estimation result 401, which is the high-precision SOC 500 output by the system.

[0040] Figure 4 FIG. 1 shows a flow chart of a method for estimating the SOC of a power battery provided by an embodiment of the present invention, as shown in FIG. Figure 4 As shown, the power battery SOC estimation method provided by the present invention is used in the system according to the above aspects, and the method includes the following steps: S401: Power battery SOC estimation system startup and initialization. When the vehicle's battery management system (BMS) is powered on or the SOC estimation function is enabled, the system is initialized: the weight parameters of the pre-trained deep learning model (dilated convolution BiGRU model) are loaded; the parameters of the PNGV model are loaded, including the open circuit voltage-SOC curve (Uoc-SOC lookup table), ohmic internal resistance R0, polarization internal resistance R1 / R2, polarization capacitance C1 / C2, and other parameters (these parameters may be functions of temperature and aging state and stored in the form of a lookup table or function); the state variables of the EKF filter module are initialized. For example, the initial SOC can be read from the last memorized value in EEPROM or roughly estimated using the open circuit voltage method after quiescence; the state covariance matrix P is initialized; and the EKF process noise covariance matrix Q and measurement noise covariance matrix R are set.

[0041] S402: Data acquisition and preprocessing. The power battery SOC estimation system enters a real-time estimation cycle, and collects the original voltage, current, and temperature data of the battery in real time through sensors (i.e. Figure 1 The raw battery data 10 in the data input and preprocessing module 100 is called to perform a series of preprocessing operations such as denoising, time alignment, abnormal data removal, and normalization on the collected raw data; and the preprocessed feature sequence data 101a containing historical and current time information and the real-time driving current 101b at the current time are output.

[0042] S403a: The deep learning SOC preliminary prediction module 200_DL performs a preliminary SOC prediction. The preprocessed feature sequence data 101a (typically containing data within a fixed-length time window) is input into the loaded deep learning model (i.e., the dilated convolutional BiGRU model). The deep learning model performs forward propagation calculations (inference) and outputs the current deep learning preliminary SOC prediction value SOC_DL 201.

[0043] S403b: The PNGV model module 300 performs a preliminary SOC prediction. Using the current real-time current (Ik) 101b (and possibly the current temperature Tk for parameter adjustment) and the SOC state fused by the EKF filter module at the previous moment (or the SOC state maintained by the PNGV module itself), the PNGV model is used to calculate the current PNGV model's preliminary SOC prediction value, SOC_PNGV 301, based on the PNGV model, specifically the mathematical equations of the PNGV model (including the ampere-hour integral and voltage response components). Step S403b can be performed in parallel or serially with step S403a.

[0044] S404: Filtering Module Fusion Filtering. Filtering module 400 is called, and the EKF first performs a state prediction step: Based on the fused SOC state (and other states of the PNGV model, such as polarization voltage) and the current current (Ik) 101b, the PNGV model's state equation is used to predict the current moment's prior SOC state and prior covariance. Then, a measurement update step is performed: the deep learning SOC preliminary prediction value SOC_DL 201, the PNGV model SOC preliminary prediction value SOC_PNGV 301, and the current actual measured battery terminal voltage Vm,k (from 101a) are used as the composite measurement value of the EKF filtering module. The Kalman gain is calculated, and the prior SOC state and covariance are updated based on the residual between the composite measurement value and the prediction value, resulting in the current moment's a posteriori, fused SOC estimate (SOC_fused,k).

[0045] S405: Output the SOC estimation result. The fused SOC estimate (SOC_fused,k) 401 output by the filter module 400 is used as the final high-precision SOC estimation result 500 at the current moment. This result can be used by other functional modules of the BMS, such as energy control strategy, range calculation, and fault diagnosis. Simultaneously, this fused SOC value serves as the "previous SOC state" required for the EKF calculation at the next sampling moment (k+1).

[0046] Optionally, after outputting the SOC estimation result, the method further includes the following steps: S406: Determine whether to continue estimating. Check the operating status of the battery management system. If the battery management system BMS is still operating normally and the vehicle is not powered off, return to step S402 and continue with the next round of SOC estimation.

[0047] S407: End estimation. If the battery management system is shut down or the SOC estimation task is completed, the estimation process is stopped and the current final SOC estimation result is saved in the non-volatile memory for use at the next startup.

[0048] Through the above steps, the present invention can combine the deep learning model's powerful fitting ability for battery nonlinear characteristics and the PNGV model's ability to describe battery electrochemical behavior, and effectively fuse them through EKF, thereby achieving high-precision and high-robustness SOC estimation.

[0049] The power battery SOC estimation system and method based on deep learning and PNGV model fusion proposed in the present invention can achieve at least one or more of the following beneficial effects through the aforementioned technical solutions: 1. Significantly Improved SOC Estimation Accuracy: The deep learning model (dilated convolutional BiGRU) can learn the complex nonlinear dynamic characteristics of batteries from large amounts of data. Its initial SOC estimation capability is superior to traditional algorithms that rely on simplified models, especially under conditions with drastic current and temperature fluctuations. The PNGV model provides physically meaningful constraints for SOC estimation, providing a stable reference, especially in areas where data is relatively stable or where the DL model may overfit. The EKF optimally fuses the preliminary SOC estimation results of the two heterogeneous models with the actual measured voltage, effectively combining the fitting capabilities of the data-driven approach with the mechanistic advantages of the model approach. This results in a final SOC estimation result that is more accurate than a single approach alone. The EKF's filtering properties also help smooth noise and reduce jumps in the estimation results.

[0050] 2. Enhance the robustness and adaptability of SOC estimation: The deep learning model is trained on data containing various operating conditions, temperatures and aging states, so that it has a certain degree of adaptability to these changing factors. The parameters of the PNGV model can be designed to dynamically adjust with temperature and aging state (for example, through a lookup table or compensation function), which enhances the model's adaptability to battery changes. The EKF fusion framework can dynamically adjust the weights of different model outputs according to the credibility of their outputs (by adjusting the measurement noise covariance R). For example, it can give the deep learning model a greater weight in operating conditions where it performs better, and vice versa. This adaptive fusion strategy improves the overall robustness of the system to different operating conditions and battery state changes. Feedback correction of the SOC state inside the PNGV model is performed through EKF (such as Figure 1 The optional feedback in the ampere-hour integration method can effectively suppress the error accumulation problem inherent in the ampere-hour integration method and improve the reliability of long-term operation.

[0051] 3. Improved responsiveness to complex operating conditions: Dilated convolutions capture long-range dependencies in the input sequence, while BiGRUs understand the bidirectional context of the sequence. This enables deep learning models to better understand and predict battery SOC changes under complex, dynamic charge and discharge conditions. The PNGV model describes the battery's rapid dynamic responses (such as ohmic voltage drop and polarization effects), providing a foundation for SOC estimation under dynamic conditions. The combination of these two, along with the rapid iterative updates of the EKF, enables this method to quickly and accurately respond to drastic load changes.

[0052] 4. Relatively reduced dependence on model parameters: Although the PNGV model still requires parameters, the introduction of deep learning models allows the system to relax the accuracy requirements of these parameters. Deep learning can compensate for the deviations of the PNGV model to a certain extent. The EKF fusion process itself also has a certain degree of error correction capability.

[0053] 5. Provide a more reliable foundation for full battery lifecycle management: More accurate SOC estimation helps the BMS implement more refined battery management strategies, such as optimized charge and discharge control, precise range prediction, and effective balancing management, thereby helping to extend the battery life and ensure safe operation.

[0054] In summary, the present invention effectively overcomes the limitations of existing single SOC estimation methods through an innovative model fusion architecture, and achieves high-precision and high-robustness estimation of power battery SOC under a variety of complex conditions. It has important theoretical research value and broad practical application prospects.

[0055] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable ordinary technicians in this field to understand the content of the present invention and implement it. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A power battery SOC estimation system, characterized in that: The system includes: data input and preprocessing module, deep learning SOC preliminary prediction module, PNGV model module, filtering module, The data input and preprocessing module is used to receive raw battery data collected from the internal or external sensor network of the battery pack, preprocess the raw battery data, and output a preprocessed data stream. The preprocessed data stream includes: feature sequence data containing historical information and current information, as well as the real-time driving current value at the current moment; The deep learning SOC preliminary prediction module is used to receive the feature sequence data from the data input and preprocessing module, and output the deep learning SOC preliminary prediction value at the current moment. The deep learning SOC preliminary prediction module adopts a hybrid deep learning model based on dilated convolution and bidirectional gated recurrent unit; The PNGV model module is used to receive the current real-time driving current value from the data input and preprocessing module and the SOC state value fed back by the filtering module at the previous moment, and output the preliminary predicted SOC value of the PNGV model at the current moment; The filtering module is used to receive the deep learning SOC preliminary prediction value from the deep learning SOC preliminary prediction module, the PNGV model SOC preliminary prediction value from the PNGV model module, and the battery terminal voltage and battery current actually measured at the current moment from the data input and preprocessing module, and perform fusion filtering calculations to output the fused SOC estimation result. The filtering module is also used to feed back the fused SOC estimation result to the PNGV model module for correcting the internal SOC state of the PNGV model module.

2. The power battery SOC estimation system according to claim 1, characterized in that: The hybrid deep learning model includes an expanded convolution layer, a BiGRU layer, and a fully connected output layer. The expanded convolution layer is used to extract multi-scale local features from the input feature sequence data. The BiGRU layer is used to process the feature sequence extracted by the expanded convolution layer to capture the temporal dynamics and contextual information in the sequence. The fully connected output layer is used to map the output of the BiGRU layer to the SOC prediction value. The hybrid deep learning model uses mean square error or mean absolute error as the loss function to measure the difference between the model-predicted SOC and the actual SOC.

3. The power battery SOC estimation system according to claim 1, characterized in that: The filtering module is a fusion extended Kalman filter module.

4. A power battery SOC estimation method, characterized in that: The method is used in a system according to any one of claims 1 to 3, and the method comprises the following steps: The power battery SOC estimation system is started and initialized. When the vehicle's battery management system is powered on, the weight parameters of the pre-trained deep learning model are loaded, the parameters of the PNGV model are loaded, the state variables of the filter module are initialized, and the process noise covariance matrix and measurement noise covariance matrix of the filter module are set; Data acquisition and preprocessing: The power battery SOC estimation system enters a real-time estimation loop, collects the battery's raw voltage, current, and temperature data in real time through sensors, calls the data input and preprocessing module to preprocess the collected raw data, and outputs preprocessed feature sequence data containing historical and current information as well as the current real-time driving current; The deep learning SOC preliminary prediction module performs preliminary SOC prediction, inputs the preprocessed feature sequence data into the loaded deep learning model, and the deep learning model performs forward propagation calculation and outputs the deep learning SOC preliminary prediction value at the current moment; The PNGV model module performs a preliminary SOC prediction, using the current real-time current and the SOC state fused by the filter module at the previous moment. According to the PNGV model, the preliminary SOC prediction value of the PNGV model at the current moment is calculated; The filtering module performs fusion filtering. The filtering module is called and first performs the state prediction step: based on the fused SOC state and current current at the previous moment, the state equation of the PNGV model is used to predict the current moment's prior SOC state and prior covariance. Then, the measurement update step is performed: the deep learning SOC preliminary prediction value, the PNGV model SOC preliminary prediction value, and the current actual measured battery terminal voltage are used as the composite measurement value of the filtering module; the Kalman gain is calculated, and the prior SOC state and covariance are updated based on the residual between the composite measurement value and the prediction value to obtain the current moment's a posteriori, fused SOC estimate. Output the SOC estimation result, and use the fused SOC estimation value output by the filtering module as the final SOC estimation result at the current moment.

5. The power battery SOC estimation method according to claim 4, characterized in that: After outputting the SOC estimation result, the method further includes the following steps: Check the battery management system operating status. If the battery management system is still working properly, proceed to the next round of SOC estimation. If the battery management system is turned off, the estimation process is stopped and the current final SOC estimation result is saved in the non-volatile memory for use at the next startup.

Images