A fusion SOC estimation method based on cloud initial value calibration

Abstract

This invention relates to the field of electric vehicle battery state of charge estimation technology, and discloses a fusion SOC estimation method based on cloud-based initial value calibration. The method includes calculating the current battery capacity and battery health status, extracting initial point data for driving segments to form a high-quality time-series dataset; extracting driving data from the time-series dataset to analyze operating condition characteristics and perform operating condition clustering, calculating operating condition influencing factors corresponding to different operating conditions; obtaining the baseline true value of SOC under a static stable state using the battery OCVSOC curve, expanding the true value sample using the ampere-hour integral method, and dividing the dataset; performing feature construction, screening, and standardization preprocessing based on battery data to obtain core input features; selecting the best-performing model for semi-supervised training and adaptive weighted fusion optimization to construct a high-precision initial SOC estimation model; and achieving accurate SOC estimation for all driving segments through collaborative work between cloud-based initial SOC calibration and on-vehicle aging condition adaptive ampere-hour integral calculation.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle battery state of charge estimation technology, specifically a fusion SOC estimation method based on cloud-based initial value calibration. Background Technology

[0002] Against the backdrop of the rapid development of the new energy vehicle industry, lithium-ion batteries have become the mainstream energy storage solution for electric vehicles. State of charge (SOC) estimation, as a core function of battery management systems (BMS), is crucial for battery safety, energy management, and lifespan extension. However, SOC estimation for lithium iron phosphate batteries faces two major challenges: firstly, the nonlinear relationship between battery parameters and SOC is easily affected by factors such as aging and temperature; secondly, data loss due to sensor failures and communication interruptions can severely reduce the accuracy and reliability of traditional estimation algorithms.

[0003] The invention patent with publication number CN121454338A proposes a SOC detection method based on electrochemical impedance spectroscopy. This method constructs a fractional-order equivalent circuit model by acquiring impedance spectral data across the entire state of charge range, thereby improving the accuracy of describing the electrochemical behavior of the battery. However, this method relies on complete impedance spectral measurement data. If the impedance spectral data is missing due to data acquisition interruption or sensor failure in practical applications, the model will not function properly. Moreover, this method does not consider the impact of battery aging on the impedance spectral-SOC relationship. As the number of battery cycles increases, the impedance spectral characteristics drift, which leads to a continuous decrease in estimation accuracy, making it difficult to meet the usage requirements throughout the entire battery lifecycle.

[0004] The invention patent with publication number CN121385656A is based on a second-order RC equivalent circuit model and uses a dual high-order capacitive Kalman filter to jointly estimate the State of Charge (SOC) and State of Harshness (SOH), and introduces a time-scale transformation factor to optimize the dynamic response. Although this method considers the corrective effect of SOH on the SOC estimation, the parameters of the equivalent circuit model are sensitive to changes in operating conditions. Under severe operating conditions such as rapid acceleration and deceleration, the dynamic characteristics of the battery will change abruptly, making it difficult to match the model parameters in real time, and easily leading to the accumulation of estimation errors. At the same time, the Kalman filter algorithm has extremely high requirements for the continuity of input data. If there are jumps or missing current and voltage data, the stability of the filter will be destroyed, and the robustness of the estimation results will be greatly reduced.

[0005] The model-based methods described above essentially rely on accurate physical models and complete input data. However, in real-world applications, with their complex scenarios involving dynamic changes in operating conditions, battery aging, and missing data, it is difficult to simultaneously achieve both accuracy and robustness. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a fusion SOC estimation method based on cloud-based initial value calibration. This method enables precise SOC management across all scenarios through collaborative work between the cloud and the vehicle, ensuring the accuracy of SOC measurement and calculation. This provides key technical support for the safe operation of battery systems, energy optimization management, and extended battery life.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a fusion SOC estimation method based on cloud-based initial value calibration, comprising the following steps: Step 1: Obtain raw battery data, complete data preprocessing and calculate the current battery capacity and battery health status, extract the initial point data of driving segments, and form a high-quality time series dataset; Step 2: Extract driving data from the time series dataset to conduct multimodal dynamic working condition analysis, extract working condition features and complete working condition clustering, and calculate the working condition influencing factors corresponding to different working conditions. Step 3: Utilizing battery-powered OCV The SOC curve is used to obtain the SOC baseline true value under static stable conditions. The true value sample is expanded to cover the entire driving segment by the ampere-hour integration method. The training set and the validation set are divided according to the time sequence principle. Step 4: Based on the preprocessed battery data, perform feature construction, filtering, and standardization preprocessing to obtain core input features suitable for model training; Step 5: Train and compare the performance of multiple regression models, select the model with the best performance for semi-supervised training and adaptive weighted fusion optimization, and build a high-precision SOC initial value estimation model. Step 6: Initial SOC calibration via cloud and vehicle-side aging The adaptive ampere-hour integral calculation works in conjunction with the operating conditions to achieve accurate SOC estimation for all driving segments in all scenarios.

[0008] Compared with existing technologies, this invention provides a fusion SOC estimation method based on cloud-based initial value calibration, which has the following beneficial effects: 1. Improved SOC estimation accuracy and robustness: This invention utilizes cloud-based initial value calibration and powerful cloud computing to construct a high-precision SOC initial value estimation model. By combining semi-supervised learning to fully utilize unlabeled data, and employing SOC interval adaptive dynamic weighted fusion and nearest neighbor correction strategies, the accuracy of SOC initial value estimation is effectively improved. On the vehicle side, the ampere-hour integration method is corrected by combining battery health status and operating condition influencing factors, solving the problems of large cumulative error and sensitivity to operating conditions and aging in the traditional ampere-hour integration method. This significantly improves the accuracy and robustness of SOC estimation under different operating conditions and different battery aging levels. Especially in scenarios with missing data, high estimation accuracy can still be maintained through edge-cloud collaboration and model optimization.

[0009] 2. Adapting to vehicle computing power constraints: Considering that the deployment of AI models is limited by the computing power of vehicle chips, this invention deploys the computationally complex SOC initial value estimation model in the cloud. The vehicle only needs to perform lightweight ampere-hour integral recursive calculation and operating condition matching, which effectively reduces the computing power requirements of vehicle hardware and avoids the problem of model deployment failure or low operating efficiency due to insufficient computing power of vehicle chips, thus balancing estimation accuracy and engineering practicality.

[0010] 3. Achieve precise SOC management across all scenarios: Through multimodal dynamic operating condition analysis, typical operating conditions are classified and operating condition influencing factors are calculated. Combined with the battery SOH value, aging-operating condition adaptive SOC calculation is realized. At the same time, the edge-cloud collaborative working mode covers the full-scenario SOC management needs of electric vehicles from start-up to real-time recursion during driving, and then to continuous optimization of cloud models. This ensures the accuracy and reliability of SOC estimation throughout the entire life cycle, providing strong support for battery safe operation, energy optimization management and extended battery life.

[0011] 4. Sufficient Data Utilization and Strong Scalability: This invention improves data quality by achieving battery type identification, vehicle model classification, and anomaly data handling through data preprocessing; the SOC ground truth calibration method expands the number of ground truth samples, providing sufficient data for model training; feature engineering selects core features, improving model efficiency; and semi-supervised learning fully utilizes unlabeled data, reducing reliance on large amounts of labeled data. Furthermore, the model construction process compares various regression models, allowing for flexible selection of the optimal model based on actual data conditions and accuracy requirements, demonstrating strong scalability and adaptability to different types of lithium-ion batteries and vehicle models. Attached Figure Description

[0012] Figure 1 This is a flowchart of the technical route of the present invention; Figure 2 This is a flowchart of the factors affecting operating conditions; Figure 3 This is a comparison chart of specific working condition factors in the dataset; Figure 4 The dataset contains plots of actual and predicted values ​​of lithium iron phosphate batteries, histograms of residual distribution, and box plots of residuals. Figure 5 It includes a graph showing the true and predicted values ​​of ternary lithium batteries in the dataset, a histogram of residual distribution, and a box plot of residuals. Figure 6 This is a diagram of an edge-cloud collaborative system; Figure 7 This is a diagram illustrating the steps of the method of the present invention. Detailed Implementation

[0013] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It is worth noting that this application also relates to prior art. Since prior art is well known to those skilled in the art, it will not be described in detail in this application.

[0014] Please see Figures 1-7 A fusion SOC estimation method based on cloud-based initial value calibration includes the following steps: Step 1: Acquire raw battery data, complete data preprocessing, calculate the current battery capacity and battery health status, extract initial point data for driving segments, and form a high-quality time-series dataset. S1.1 Battery type identification and vehicle model classification; Battery type identification: First, extract the individual cell voltage data from the collected raw battery data, and determine the battery type by the individual cell voltage range: if the individual cell voltage is between 3.2-3.65V, it is identified as a lithium iron phosphate battery; if the individual cell voltage is between 3.7-4.2V, it is identified as a ternary lithium battery. Vehicle Classification: Based on the different battery types, the dataset is initially divided into vehicles equipped with lithium iron phosphate batteries and vehicles equipped with ternary lithium batteries. If the total voltage data of a certain vehicle is not within the reasonable total voltage range that the battery pack of that vehicle should have, the vehicle will be further subdivided according to different total voltage platforms. Based on the defined vehicle models, the datasets corresponding to each vehicle model are sorted in chronological order to form a time-series dataset. S1.2 Charging segment identification and abnormal data processing; Charging segment identification: Identify charging segments from the time-series dataset. The definition of a charging segment is based on the following criteria: the current data is continuously negative and the SOC shows an increasing trend. Anomaly data processing: Plot a scatter plot of the identified charging segment data with current as the vertical axis and time as the horizontal axis to visually identify outliers in the data. S1.3, Current battery capacity Battery health status calculate; ; in, This refers to the current battery capacity. for The charging current at any given moment; This is the start time of the charging segment; This refers to the end time of the charging segment. This represents the change in SOC corresponding to this charging segment; ; in, Battery health status; This is the initial design capacity; S1.4, Initial point data processing for driving segments; Initial point data extraction and processing were performed on the driving segments excluding non-charging segments in the time series dataset, including the current, voltage, temperature, and SOC at the initial moment. Step 2: Extract driving data from the time-series dataset to conduct multimodal dynamic operating condition analysis, extract operating condition features and complete operating condition clustering, and calculate the operating condition influencing factors corresponding to different operating conditions, including: S2.1 Time window division; For driving segments that require SOC estimation, the time series is first rounded according to fixed time intervals, and then the time windows are divided according to the rounded time intervals to cut the driving segment into several time window segments. S2.2 Selection of operating condition characteristics; Referring to the driving characteristic parameters specified in the Chinese national standard for driving conditions (GB / T38146.1-2019), and considering the influence of temperature on the performance of automotive batteries, the following 14 multimodal driving condition characteristics were selected; S2.3, Working condition clustering; The K-means++ clustering algorithm was used to perform clustering analysis on the segmented driving time window data according to the above 14 working condition features, and the working conditions were divided into 4 categories: urban congestion, urban expressway, suburban road, and highway. S2.4 Calculate the working condition influence factors corresponding to different working conditions; Using the mileage achievable per unit SOC as the core indicator for evaluating the economic efficiency of operating conditions, the operating condition with the highest mileage per unit SOC is defined as the most economical operating condition among the four operating condition clusters obtained from operating condition clustering. For the other three operating conditions, the ratio of the mileage per unit SOC to that of the most economical operating condition is calculated, which is the operating condition impact factor corresponding to that operating condition. The calculation formula is as follows: ; in, Operating condition influencing factors; This refers to the mileage driven per unit SOC under this operating condition. The driving range per unit SOC under the most economical operating conditions; Step 3: Utilizing battery-powered OCV The SOC curve obtains the baseline true value of SOC under static steady-state conditions. The true value sample is expanded using the ampere-hour integration method to cover the entire driving segment. The training set and validation set are divided according to time sequence principles, including: S3.1 Selection of SOC truth value based on OCV curve; Obtain the OCV-SOC curves for lithium iron phosphate battery models and ternary lithium battery models respectively; For lithium iron phosphate batteries, the SOC value in the high SOC range of the OCV-SOC curve is selected as the true value. For ternary lithium batteries, the SOC value of the OCV-SOC curve in most of the SOC range is selected as the true value. In the time series dataset, segments where the battery is in an open circuit state are selected, and the corresponding SOC value is found on the corresponding OCV-SOC curve based on the OCV value at this time, which is taken as the true SOC value of the segment. S3.2, SOC truth value extension based on ampere-hour integral; The SOC truth value is extended using the ampere-hour integration method, and the calculation formula is as follows: ; in, for The truth value of SOC at time step; Initial time The truth value of SOC; This refers to the battery's rated capacity. for The discharge current at any given moment; S3.3 Training and validation set partitioning; The dataset with the calibrated SOC ground truth value is labeled as the vehicle battery dataset. It is divided into training set and validation set according to a certain ratio. The division process should follow the principle of time sequence, that is, ensure that the time of the training set is earlier and the time of the validation set is later. Step 4: Based on the preprocessed battery data, perform feature construction, filtering, and standardization preprocessing to obtain core input features suitable for model training; Step four includes: S4.1 Original feature extraction and derived feature construction; For the vehicle battery dataset, the units of the original current, voltage, and temperature are first converted, and the results of outlier cleanup are combined to obtain clean original feature data. Based on the business logic of battery operation and the influencing factors of SOC changes, new core features are derived. S4.2, Four-layer feature selection strategy; Through a four-layer feature screening strategy—preliminary screening using a business whitelist, importance screening using the LightGBM model, weight threshold filtering, and manual removal of low-contribution features—new core features are screened, redundant and low-contribution features are eliminated, and core features are retained. S4.3 Feature standardization processing; The StandardScaler method was used to standardize the core features after screening, and then applied to the validation and test sets. S4.4 Model parameter tuning and regularization; The leave-one-out cross-validation method selects one sample as the validation set each time, and uses all other samples as the training set. This process is repeated until all samples have been used as the validation set once. Set the parameter tuning range as follows: learning rate 0.01-0.2, number of trees 50-2000, maximum tree depth 2-15, minimum number of samples per leaf node 5-100, feature sampling ratio 0.5-1.0, data sampling ratio 0.5-1.0, L1 regularization coefficient 0-10, L2 regularization coefficient 0-10. Optimization objective: To minimize the root mean square error of the validation set, the optimal combination of model parameters is found through grid search or random search methods. Regularization enhancement: During parameter tuning, increase the L1 and L2 regularization coefficients to limit model complexity and avoid overfitting the training data. At the same time, adopt an early stopping strategy and set an early stopping round number. If the RMSE of the validation set does not improve in a set number of consecutive iterations, stop model training and select the current best model. S4.5 Result Output and Dataset Saving; After the model training is completed, the RMSE index of the optimal model, the weights of the core features, and the visualization charts are output. At the same time, the selected feature dataset after feature selection and standardization is saved. Step 5: Train and compare the performance of various regression models, select the best-performing model for semi-supervised training and adaptive weighted fusion optimization, and construct a high-precision SOC initial value estimation model, including: S5.1, Multi-model training and performance comparison; Seven commonly used regression models were constructed and trained, including: Bayesian Ridge Regression, Gaussian Process Regression, Kernel Ridge Regression, LightGBM Regression, XGBoost Regression, Random Forest Regression, and Multilayer Perceptron Regression. The above 7 types of models are trained using the training set calibrated in step 3. The performance of each model is evaluated on the validation set. The evaluation metrics include root mean square error, mean absolute error, and coefficient of determination. The sum of the evaluation metrics is used as the evaluation result. The two models with the highest evaluation results are extracted as the two optimal models and named Model A and Model B, respectively. S5.2, Semi-supervised Co-training; In the semi-supervised co-training phase, the original training set is first divided into 70% labeled and 30% unlabeled portions at a ratio of 0.3, and the SOC ground truth values ​​of the unlabeled samples are removed to construct an unlabeled dataset. Then, the selected A and B models are initially trained using the labeled dataset. The two models are then used to predict the unlabeled data. A quantile screening method is used to select high-confidence samples whose predicted values ​​are within the 30%-70% quantile range and generate pseudo-labels. The pseudo-label samples generated by model A are added to the training set of model B, and the pseudo-label samples generated by model B are added to the training set of model A. The model parameters are then retrained and updated. The process of pseudo-label generation, training set expansion, and model update is repeated. The maximum number of iterations is set to 10. If the RMSE of the validation set does not decrease for three consecutive iterations, the model converges early. Finally, the two basic models A and B, which have been optimized by semi-supervised training, are obtained. S5.3, SOC interval adaptive dynamic weighted fusion; Based on the true distribution of SOC values ​​of labeled samples in the training set, the SOC range is divided into multiple target SOC intervals and other intervals. Then, the average absolute error of prediction of model A and model B in each SOC interval is calculated, and the exclusive dynamic weights of model A and model B are assigned to each SOC interval. Finally, the prediction results of model A and model B are weighted and fused to generate the initial SOC estimate. S5.4 Dynamic weighted optimization correction; First, in the training set, samples belonging to the same SOC interval as the sample to be predicted are selected. The Euclidean distance is used to calculate the feature distance between the sample to be predicted and these samples in the same interval. The five closest samples are selected, and the weight of each nearest neighbor sample is calculated based on the reciprocal of the feature distance. Then, based on the true SOC value of the nearest neighbor sample and the corresponding weight, the weighted SOC value is calculated to correct the SOC estimate, and the final initial SOC estimate result is obtained. Step 6: Initial SOC calibration via cloud and vehicle-side aging The adaptive ampere-hour integral calculation works in conjunction with the operating condition to achieve accurate SOC estimation for all driving segments across all scenarios, including: S6.1 Design of vehicle-side data upload and communication mechanism; At the start of a driving segment, the vehicle-side BMS system collects initial feature data, removes outliers exceeding reasonable limits, preprocesses the data, and encapsulates it into a JSON data packet tagged with the initial feature data. The communication module automatically selects either the 4G / 5G primary link or the V2X backup link to transmit data based on the RSSI signal strength, establishes a connection using the TCP protocol, and sets up timeout retransmission and local caching mechanisms. After receiving the data, the cloud gateway parses it through VIN code authentication and CRC32 integrity verification, and returns confirmation information via the UDP protocol. The overall system adopts a dual-mode communication architecture of "4G / 5G as the primary and V2X as the secondary," and simultaneously uses SM4 end-to-end encryption, SM2 key negotiation, vehicle-cloud two-way identity authentication, and CRC32 verification. S6.2, Cloud-based SOC initial value calculation and distribution process; The cloud-based data processing module inputs the parsed vehicle-side data into the pre-built SOC initial value estimation model, calls the training parameters to calculate the SOC initial value accurate to 0.01%, and then verifies the reasonableness of the result. If the initial value is within the range of 0%-100%, it is considered valid. If it is outside the range, it triggers a maximum of 2 recalculations. If it is still invalid, the SOC initial value of historical similar working conditions with an error within ±1% is used as a backup. Then, the SOC initial value and auxiliary information are encapsulated into a high-priority JSON data packet with SOC initial value data tags. It is sent down using an adaptive link consistent with the upload from the vehicle. After receiving it, the vehicle verifies the checksum and VIN code and parses the initial value. If the confidence is ≥95%, it is directly used for SOC recursive calculation. If the confidence is <95%, a local warning is triggered and the initial value is still used. At the same time, a receipt confirmation is returned to the cloud, completing the sending process to provide a reliable initial SOC baseline. S6.3, Vehicle-side Operating Condition Analysis and Value matching; While receiving the initial SOC value from the cloud, the vehicle-side BMS system performs real-time analysis of the operating conditions of the current driving segment. Based on the 14 operating condition features determined in step two (multimodal dynamic operating condition analysis), it collects operating condition data of the current driving segment in real time, calculates the feature vector of the current operating condition, compares the current operating condition feature vector with the central feature vectors of the four previously clustered operating conditions to calculate the Euclidean distance, finds the nearest operating condition category, determines the operating condition type of the current driving segment, and accurately matches the corresponding operating condition influencing factors. S6.4, Vehicle-side aging - Adaptive ampere-hour integral calculation under operating conditions; The vehicle-side system uses the initial SOC value provided by the cloud, the current SOH value of the battery, and the matched operating condition influencing factors. The change in SOC during driving was calculated using the ampere-hour integral method that is adaptive to aging and operating conditions.

[0015] Example To verify the effectiveness of the method of this invention, actual operating data from two electric vehicles with different battery types were selected as experimental samples for verification. The experimental dataset recorded the actual operating data of the two vehicles over 10 days, covering vehicle operating parameters. The data sampling frequency was 10 seconds, which could accurately capture the dynamic changes of the battery and the characteristics of vehicle driving conditions. One vehicle was equipped with a lithium iron phosphate battery (hereinafter referred to as "Vehicle A"), and the other vehicle was equipped with a ternary lithium battery (hereinafter referred to as "Vehicle B").

[0016] 1. Data preprocessing implementation The experimental data were preprocessed according to the procedure in step one. The key processing results are as follows: Battery type identification and vehicle model classification: Individual battery voltage data for vehicle A was extracted. Its voltage range was stable at 3.2-3.62V, consistent with the voltage characteristics of lithium iron phosphate batteries, thus vehicle A was identified as model A. Individual battery voltage data for vehicle B was extracted. Its voltage range was 3.72-4.18V, consistent with the voltage characteristics of ternary lithium batteries, thus vehicle B was identified as model B. Further verification of total voltage: Vehicle A's battery pack consists of 96 individual batteries connected in series. The reasonable range for the total voltage is 3.2×96-3.65×96=307.2-350.4V. The actual total voltage is within the range of 308.5-349.2V, with no abnormalities. Vehicle B's battery pack consists of 72 individual batteries connected in series. The reasonable range for the total voltage is 3.7×72-4.2×72=266.4-302.4V. The actual total voltage is within the range of 267.8-301.5V, with no abnormalities, and no further vehicle model classification is needed. Finally, the data from the two vehicles were sorted by time to form a time-series dataset.

[0017] Charging Segment Identification and Abnormal Data Processing: Using the condition of "continuously negative current + increasing SOC", 28 charging segments were identified from the time-series data of vehicle A, and 32 charging segments were identified from the time-series data of vehicle B. Visualization using current scatter plots revealed 3 abnormal segments in vehicle A (positive current lasting more than 30 seconds) and 2 abnormal segments in vehicle B (frequent current oscillations), all of which were cut and deleted. Gaussian filtering was applied to the remaining valid charging segments (the standard deviation of the filter was set to 0.8 for vehicle A and 1.0 for vehicle B; experiments verified that this parameter minimized the deviation between the filtered current and the actual charging current). Through triple verification (length threshold of 30 seconds, effective data volume ratio threshold of 80%, numerical range: current ≤ 3C, voltage ≤ single-cell upper limit), 32 valid charging segments were ultimately retained for vehicle A, and 16 valid charging segments were retained for vehicle B.

[0018] SOH Calculation: The capacity of 32 effective charging segments for vehicle A is calculated. The capacities of these effective charging segments over the past two months are 98.2Ah, 97.5Ah, 97.2Ah, ..., 98.5Ah (32 data points in total). The median of 97.66Ah is taken as the current capacity of vehicle A. The initial design capacity of vehicle A's battery is 100Ah, therefore, the SOH of vehicle A = (97.66 / 100) × 100% = 97.66%. The capacity of 16 effective charging segments for vehicle B is calculated. The capacities of these effective charging segments over the past two months are 83.1Ah, 82.5Ah, 81.9Ah, ..., 82.3Ah (16 data points in total). The median of 82.4Ah is taken as the current capacity of vehicle B. The initial design capacity of vehicle B's battery is 85Ah, therefore, the SOH of vehicle B = (82.4 / 85) × 100% = 96.94%.

[0019] Initial point data processing for driving segments: Initial point data for 81 driving segments were extracted from the time series data of vehicle A, and initial point data for 110 driving segments were extracted from the time series data of vehicle B. All initial point data were complete, providing a complete input for subsequent cloud-based SOC initial value estimation.

[0020] 2. Implementation of Multimodal Dynamic Operating Condition Analysis Following the procedure in step two, a multimodal dynamic condition analysis was performed on the driving segments of the two vehicles. The key results are as follows: Time window segmentation: Considering that the cycle of vehicle driving conditions usually lasts within ten minutes, the time rounding interval is set to 10 minutes, and all 208 driving segments are divided into 622 time window segments (each segment is 10 minutes).

[0021] Clustering based on operating conditions: The K-means++ algorithm was used to cluster all time window segments. Based on the elbow method, the number of clusters was determined to be 4.

[0022] Calculation of operating condition impact factors: Calculate the unit SOC mileage for each operating condition. Among all four operating conditions, condition 4 has the highest unit SOC mileage, reaching 3.32 km / %, and is defined as the most economical operating condition. The unit SOC mileages for conditions 1, 2, and 3 are 2.6 km / %, 2.82 km / %, and 3.27 km / %, respectively. Therefore, the corresponding operating condition impact factors are α0 = 2.6 / 3.32 ≈ 78.2%, α1 = 2.82 / 3.32 ≈ 84.9%, and α2 = 3.27 / 3.32 ≈ 98.4%. Specific comparisons of operating condition factors are provided below. Figure 3 As shown, this provides a basis for subsequent operating condition corrections in vehicle-side SOC calculations.

[0023] 3. Implementation of SOC truth value calibration Following the procedure in step three, the SOC true value was calibrated for the data from both vehicles. The key results are as follows: Selection of true SOC values ​​based on OCV curves: Based on the OCV-SOC curves of lithium iron phosphate batteries, 18 true SOC values ​​for vehicle A were selected from the dataset; based on the OCV-SOC curves of ternary lithium batteries, 22 true SOC values ​​for vehicle B were selected.

[0024] SOC truth value extension based on ampere-hour integration: For the 18 segments with truth values ​​of vehicle A, 5 consecutive driving segments are selected after each segment (stopping when another truth value or charging segment is encountered), and the SOC truth value of these segments is calculated by ampere-hour integration. Vehicle A has a total of 81 SOC truth value samples; For the 22 segments with truth values ​​of vehicle B, 5 consecutive driving segments are selected after each segment (stopping when another truth value or charging segment is encountered), vehicle B has a total of 110 SOC truth value samples.

[0025] Training and validation set partitioning: The 81 ground truth SOC samples of car A were partitioned into a training set (65 samples) and a validation set (16 samples) in a ratio of 8:2; the 110 ground truth SOC samples of car B were partitioned into a training set (77 samples) and a validation set (33 samples) in a ratio of 7:3.

[0026] 4. Feature Engineering Implementation Following the process in step four, feature engineering was performed, and the feature selection results for the two vehicle data are as follows: Feature construction: 24 core features were constructed for samples of car A and car B respectively.

[0027] Four-layer feature filtering: Feature selection for Vehicle A: Initial screening using a business whitelist retained 21 features (removing one feature with extremely low correlation to SOC changes: single-unit voltage variation coefficient); LightGBM model importance screening selected the top 14 features; weight threshold filtering retained 14 features; manual review revealed that the single-unit voltage variation coefficient contributed little to the samples for Vehicle A, and some samples with a mean of 0 were set to NaN and removed. Finally, 13 core features were determined for Vehicle A, as follows: Conditional starting package voltage, segment starting package voltage, conditional starting current, conditional starting single-cell maximum voltage, conditional starting single-cell minimum voltage, segment starting single-cell maximum voltage, conditional starting single-cell maximum temperature, conditional starting single-cell minimum temperature, mean of single-cell temperature list, standard deviation of single-cell temperature list, mean of single-cell parameter list, standard deviation of single-cell parameter list, mean of single-cell parameter list; Vehicle B Feature Selection: After initial screening using the business whitelist, 19 features were retained; the top 13 features were selected based on importance using the LightGBM model; after weighted threshold filtering, 12 features were retained; manual review revealed that the correlation coefficient between the initial package voltage and the initial package voltage of the segment was 0.91, and the package voltage difference was difficult to obtain, so it was removed. Finally, 11 core features were determined for Vehicle B, as follows: Conditional starting current, segment starting current, conditional starting maximum voltage of a single cell, segment starting maximum voltage of a single cell, segment starting minimum voltage of a single cell, single cell voltage variation coefficient, mean of the single cell temperature list, standard deviation of the single cell temperature list, standard deviation of the single cell parameter list for conditional starting, range of the single cell parameter list for conditional starting, range of the single cell parameter list for segment starting. Feature standardization and model tuning: Z-score standardization was performed on the core features of the two vehicles; the LightGBM model was tuned by leave-one-out cross-validation. The optimal parameters for vehicle A were: learning rate 0.05, number of trees 500, maximum tree depth 6, L1 regularization coefficient 0.5, and L2 regularization coefficient 1.0; the optimal parameters for vehicle B were: learning rate 0.03, number of trees 800, maximum tree depth 5, L1 regularization coefficient 0.3, and L2 regularization coefficient 0.8. After tuning, the RMSE of the models on the validation set were 6.1813% (vehicle A) and 1.6867% (vehicle B), respectively, laying the foundation for subsequent multi-model training.

[0028] 5. Model Building and Implementation Following the procedure in step five, the initial SOC estimation model was constructed, and the implementation results for the two vehicle data are as follows: Multi-model training and performance comparison: Seven regression models were trained on the training sets of both vehicles, and their performance was evaluated on their respective validation sets (metrics: RMSE, MAE, R²). Table 1 shows the performance comparison results of vehicle A. XGBoost regression and random forest regression performed best. Therefore, XGBoost regression was selected as model A for vehicle A, and random forest regression was selected as model B for vehicle A. Table 2 shows the performance comparison results of vehicle B. XGBoost regression and LightGBM regression performed best. Similarly, XGBoost regression was selected as model A for vehicle A, and random forest regression was selected as model B for vehicle A.

[0029]

[0030] Table 1

[0031] Table 2 Semi-supervised collaborative training: The training sets of the two vehicles are divided into unlabeled data at a ratio of 30%, and 10 rounds of collaborative training are carried out (STOP_ROUNDS=3).

[0032] SOC Interval Adaptive Dynamic Weighted Fusion and Correction: The SOC range of vehicle A is divided into four intervals: 0%-20% (low SOC), 20%-40% (low-medium SOC), 40%-80% (high-medium SOC, critical interval), and 80%-100% (high SOC). The MAE of models A and B in each interval is calculated. The SOC range of vehicle B is also divided into four intervals: 0%-10% (low SOC), 10%-40% (low-medium SOC), 40%-80% (high-medium SOC), and 80%-100% (high SOC).

[0033] The graphs showing the true and predicted values ​​of car A and car B, the histogram of residual distribution, and the box plot of residuals are shown below. Figure 4 , Figure 5 As shown in the figure, the predicted values ​​closely match the actual values, the residuals exhibit a normal distribution, and are concentrated between -2% and 2%, demonstrating excellent model accuracy and stability. The ternary lithium group shows high accuracy, with an RMSE as low as 0.825% and a MAE of 0.621%; after iterative optimization, the lithium iron phosphate group achieves an RMSE as low as 1.648% and a MAE of 0.752%.

[0034] 6. Implementation of SOC estimation combining vehicle and cloud platforms Following the process in step six of Part One, vehicle-cloud collaborative SOC estimation was performed. A driving segment of vehicle A (10 minutes long, belonging to Class 1 working condition, corresponding to α0=78.2%) and a driving segment of vehicle B (10 minutes long, belonging to Class 2, corresponding to α1=84.9%) were selected for instance verification. The results show that the estimation error of the method of the present invention is less than 0.5%.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fusion SOC estimation method based on cloud-based initial value calibration, characterized in that, Includes the following steps: Step 1: Obtain raw battery data, complete data preprocessing and calculate the current battery capacity and battery health status, extract the initial point data of driving segments, and form a high-quality time series dataset; Step 2: Extract driving data from the time series dataset to conduct multimodal dynamic working condition analysis, extract working condition features and complete working condition clustering, and calculate the working condition influencing factors corresponding to different working conditions. Step 3: Utilizing battery-powered OCV The SOC curve is used to obtain the SOC baseline true value under static stable conditions. The true value sample is expanded to cover the entire driving segment by the ampere-hour integration method. The training set and the validation set are divided according to the time sequence principle. Step 4: Based on the preprocessed battery data, perform feature construction, filtering, and standardization preprocessing to obtain core input features suitable for model training; Step 5: Train and compare the performance of multiple regression models, select the model with the best performance for semi-supervised training and adaptive weighted fusion optimization, and build a high-precision SOC initial value estimation model. Step 6: Initial SOC calibration via cloud and vehicle-side aging The adaptive ampere-hour integral calculation works in conjunction with the operating conditions to achieve accurate SOC estimation for all driving segments in all scenarios.

2. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 1, characterized in that, Step one includes: S1.1 Battery type identification and vehicle model classification; Battery type identification: First, extract the individual cell voltage data from the collected raw battery data, and determine the battery type by the individual cell voltage range: if the individual cell voltage is between 3.2-3.65V, it is identified as a lithium iron phosphate battery; if the individual cell voltage is between 3.7-4.2V, it is identified as a ternary lithium battery. Vehicle Classification: Based on the different battery types, the dataset is initially divided into vehicles equipped with lithium iron phosphate batteries and vehicles equipped with ternary lithium batteries. If the total voltage data of a certain vehicle is not within the reasonable total voltage range that the battery pack of that vehicle should have, the vehicle will be further subdivided according to different total voltage platforms. Based on the defined vehicle models, the datasets corresponding to each vehicle model are sorted in chronological order to form a time-series dataset. S1.2 Charging segment identification and abnormal data processing; Charging segment identification: Identify charging segments from the time-series dataset. The definition of a charging segment is based on the following criteria: the current data is continuously negative and the SOC shows an increasing trend. Anomaly Data Processing: Identified charging segment data is plotted as a scatter plot with current on the ordinate and time on the abscissa. The scatter plot visually identifies outliers in the data. Common anomalies include: positive or negative currents lasting longer than expected; current oscillation frequencies between positive and negative values ​​exceeding the expected normal frequency; and peaks in the current data that do not conform to physical laws. For data exhibiting common anomalies, the data is cut and deleted. The data sequence is then reorganized and recalculated starting from the cut time points. Gaussian filtering is applied to the initially processed valid charging current data to eliminate random noise. Simultaneously, the charging segment data undergoes triple verification of length, valid data volume, and numerical value to filter out invalid data. S1.3, Current battery capacity Battery health status calculate; ； in, This refers to the current battery capacity. for The charging current at any given moment; This is the start time of the charging segment; This refers to the end time of the charging segment. This represents the change in SOC corresponding to this charging segment; ； in, Battery health status; This is the initial design capacity; S1.4, Initial point data processing for driving segments; Initial point data is extracted and organized from the driving segments in the time series dataset, excluding non-charging segments, including the current, voltage, temperature, and SOC at the initial moment.

3. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 2, characterized in that, Step two includes: S2.1 Time window division; For driving segments that require SOC estimation, the time series is first rounded according to fixed time intervals, and then the time windows are divided according to the rounded time intervals to cut the driving segment into several time window segments. S2.2 Selection of operating condition characteristics; Referring to the driving characteristic parameters specified in the Chinese national standard for driving conditions (GB / T38146.1-2019), and considering the impact of temperature on automotive battery performance, the following 14 multimodal driving condition characteristics were selected: total duration of segment, segment mileage, average speed, maximum speed of segment, average acceleration of segment, maximum acceleration of segment, minimum acceleration of segment, proportion of idling time of segment, proportion of constant speed driving time of segment, proportion of acceleration driving time of segment, proportion of deceleration driving time of segment, mean of lowest single cell temperature of segment, mean of highest single cell temperature of segment, and standard deviation of single cell temperature of segment. S2.3, Working condition clustering; The K-means++ clustering algorithm was used to perform clustering analysis on the segmented driving time window data according to the above 14 working condition features, and the working conditions were divided into 4 categories: urban congestion, urban expressway, suburban road, and highway. S2.4 Calculate the working condition influence factors corresponding to different working conditions; Using the mileage achievable per unit SOC as the core indicator for evaluating the economic efficiency of operating conditions, the operating condition with the highest mileage per unit SOC is defined as the most economical operating condition among the four operating condition clusters obtained from operating condition clustering. For the other three operating conditions, the ratio of the mileage per unit SOC to that of the most economical operating condition is calculated, which is the operating condition impact factor corresponding to that operating condition. The calculation formula is as follows: ； in, Operating condition influencing factors; This refers to the mileage driven per unit SOC under this operating condition. The unit SOC driving range is the most economical operating condition.

4. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 3, characterized in that, Step three includes: S3.1 Selection of SOC truth value based on OCV curve; Obtain the OCV-SOC curves for lithium iron phosphate battery models and ternary lithium battery models respectively; For lithium iron phosphate batteries, the SOC value in the high SOC range of the OCV-SOC curve is selected as the true value. For ternary lithium batteries, the SOC value of the OCV-SOC curve in most of the SOC range is selected as the true value. In the time series dataset, segments where the battery is in an open circuit state are selected, and the corresponding SOC value is found on the corresponding OCV-SOC curve based on the OCV value at this time, which is taken as the true SOC value of the segment. S3.2, SOC truth value extension based on ampere-hour integral; The SOC truth value is extended using the ampere-hour integration method, and the calculation formula is as follows: ； in, for The truth value of SOC at time step; Initial time The truth value of SOC; This refers to the battery's rated capacity. for The discharge current at any given moment; S3.3 Training and validation set partitioning; The dataset with the calibrated SOC ground truth value is labeled as the vehicle battery dataset. It is then divided into a training set and a validation set according to a certain ratio. The division process must follow the principle of temporal order, that is, ensure that the time of the training set is earlier and the time of the validation set is later.

5. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 4, characterized in that, Step four includes: S4.1 Original feature extraction and derived feature construction; For the vehicle battery dataset, the units of the original current, voltage, and temperature were first converted, and combined with the outlier cleanup results to obtain clean raw feature data. Based on the business logic of battery operation and the influencing factors of SOC changes, new core features were derived, specifically including: Conditional starting total voltage, segment starting total voltage, conditional starting current, segment starting current, conditional starting maximum voltage of individual unit, conditional starting minimum voltage of individual unit, segment starting maximum voltage of individual unit, segment starting minimum voltage of individual unit, conditional starting maximum temperature of individual unit, conditional starting minimum temperature of individual unit, segment starting maximum temperature of individual unit, segment starting maximum temperature of individual unit, segment starting minimum temperature of individual unit, total voltage difference, individual unit voltage variation coefficient, mean of individual unit temperature list, standard deviation of individual unit temperature list, range of individual unit temperature list, mean of conditional starting individual unit parameter list, standard deviation of conditional starting individual unit parameter list, range of conditional starting individual unit parameter list, mean of segment starting individual unit parameter list, standard deviation of segment starting individual unit parameter list, range of segment starting individual unit parameter list; S4.2, Four-layer feature selection strategy; Through a four-layer feature screening strategy—preliminary screening using a business whitelist, importance screening using the LightGBM model, weight threshold filtering, and manual removal of low-contribution features—new core features are screened, redundant and low-contribution features are eliminated, and core features are retained. S4.3 Feature standardization processing; The StandardScaler method was used to standardize the core features after screening, and then applied to the validation and test sets. S4.4 Model parameter tuning and regularization; The leave-one-out cross-validation method selects one sample as the validation set each time, and uses all other samples as the training set. This process is repeated until all samples have been used as the validation set once. Set the parameter tuning range as follows: learning rate 0.01-0.2, number of trees 50-2000, maximum tree depth 2-15, minimum number of samples per leaf node 5-100, feature sampling ratio 0.5-1.0, data sampling ratio 0.5-1.0, L1 regularization coefficient 0-10, L2 regularization coefficient 0-10. Optimization objective: To minimize the root mean square error of the validation set, the optimal combination of model parameters is found through grid search or random search methods. Regularization enhancement: During parameter tuning, increase the L1 and L2 regularization coefficients to limit model complexity and avoid overfitting the training data. At the same time, adopt an early stopping strategy and set an early stopping round number. If the RMSE of the validation set does not improve in a set number of consecutive iterations, stop model training and select the current best model. S4.5 Result Output and Dataset Saving; After the model training is completed, the RMSE index of the optimal model, the weights of the core features, and the visualization charts are output. At the same time, the selected feature dataset after feature selection and standardization is saved.

6. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 5, characterized in that, Step five includes: S5.1, Multi-model training and performance comparison; Seven commonly used regression models were constructed and trained, including: Bayesian Ridge Regression, Gaussian Process Regression, Kernel Ridge Regression, LightGBM Regression, XGBoost Regression, Random Forest Regression, and Multilayer Perceptron Regression. The above 7 types of models are trained using the training set calibrated in step 3. The performance of each model is evaluated on the validation set. The evaluation metrics include root mean square error, mean absolute error, and coefficient of determination. The sum of the evaluation metrics is used as the evaluation result. The two models with the highest evaluation results are extracted as the two optimal models and named Model A and Model B, respectively. S5.2, Semi-supervised Co-training; In the semi-supervised co-training phase, the original training set is first divided into 70% labeled and 30% unlabeled portions at a ratio of 0.3, and the SOC ground truth values ​​of the unlabeled samples are removed to construct an unlabeled dataset. Then, the selected A and B models are initially trained using the labeled dataset. The two models are then used to predict the unlabeled data. A quantile screening method is used to select high-confidence samples whose predicted values ​​are within the 30%-70% quantile range and generate pseudo-labels. The pseudo-label samples generated by model A are added to the training set of model B, and the pseudo-label samples generated by model B are added to the training set of model A. The model parameters are then retrained and updated. The process of pseudo-label generation, training set expansion, and model update is repeated. The maximum number of iterations is set to 10. If the RMSE of the validation set does not decrease for three consecutive iterations, the model converges early. Finally, the two basic models A and B, which have been optimized by semi-supervised training, are obtained. S5.3, SOC interval adaptive dynamic weighted fusion; Based on the true distribution of SOC values ​​of labeled samples in the training set, the SOC range is divided into multiple target SOC intervals and other intervals. Then, the average absolute error of prediction of model A and model B in each SOC interval is calculated, and the exclusive dynamic weights of model A and model B are assigned to each SOC interval. Finally, the prediction results of model A and model B are weighted and fused to generate the initial SOC estimate. S5.4 Dynamic weighted optimization correction; First, in the training set, samples belonging to the same SOC interval as the sample to be predicted are selected. The feature distance between the sample to be predicted and these samples in the same interval is calculated using Euclidean distance. The five closest samples are selected, and the weight of each nearest neighbor sample is calculated based on the reciprocal of the feature distance. Then, based on the true SOC value of the nearest neighbor sample and the corresponding weight, the weighted SOC value is calculated to correct the SOC estimate, and the final initial SOC estimate is obtained.

7. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 6, characterized in that, In S5.3, the formula for calculating the mean absolute error of prediction is: ； in, To predict the mean absolute error, This represents the number of samples within the SOC interval; For the first The true SOC value of each sample; For the model to the first SOC prediction value for each sample; The formula for calculating the dedicated dynamic weight is: ； ； in, The weights of model A in this interval. The weights of model B in this interval. For model A in this interval , For model B in this interval , This is a minimum value, used to avoid... A division-by-zero error occurs when the value is 0; for non-target SOC intervals, a balanced weight of 0.5:0.5 is used by default. The formula for calculating the initial estimate of SOC is: ； in, This is the initial estimate of the fused SOC; These are the predicted values ​​from model A; These are the predicted values ​​from Model B.

8. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 6, characterized in that, In S5.4, the formula for calculating the weight of each nearest neighbor sample is as follows: ； in, For the first The weights of the nearest neighbor samples; For the sample to be predicted and the first The feature distance of the nearest neighbor samples; The power of the distance; This is the weight normalization factor; The final initial SOC estimate is calculated using the following formula: ； ； in, The weighted SOC value of the nearest neighbor samples. Let SOC be the ground truth value of the k-th nearest neighbor sample. This is the final initial value estimate for SOC.

9. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 6, characterized in that, Step six includes: S6.1 Design of vehicle-side data upload and communication mechanism; At the start of a driving segment, the vehicle-side BMS system collects initial feature data, removes outliers exceeding reasonable limits, preprocesses the data, and encapsulates it into a JSON data packet tagged with the initial feature data. The communication module automatically selects either the 4G / 5G primary link or the V2X backup link to transmit data based on the RSSI signal strength, establishes a connection using the TCP protocol, and sets up timeout retransmission and local caching mechanisms. After receiving the data, the cloud gateway parses it through VIN code authentication and CRC32 integrity verification, and returns confirmation information via the UDP protocol. The overall system adopts a dual-mode communication architecture of "4G / 5G as the primary and V2X as the secondary," and simultaneously uses SM4 end-to-end encryption, SM2 key negotiation, vehicle-cloud two-way identity authentication, and CRC32 verification. S6.2, Cloud-based SOC initial value calculation and distribution process; The cloud-based data processing module inputs the parsed vehicle-side data into the pre-built SOC initial value estimation model, calls the training parameters to calculate the SOC initial value accurate to 0.01%, and then verifies the reasonableness of the result. If the initial value is within the range of 0%-100%, it is considered valid. If it is outside the range, it triggers a maximum of 2 recalculations. If it is still invalid, the SOC initial value of historical similar working conditions with an error within ±1% is used as a backup. Then, the SOC initial value and auxiliary information are encapsulated into a high-priority JSON data packet with SOC initial value data tags. It is sent down using an adaptive link consistent with the upload from the vehicle. After receiving it, the vehicle verifies the checksum and VIN code and parses the initial value. If the confidence is ≥95%, it is directly used for SOC recursive calculation. If the confidence is <95%, a local warning is triggered and the initial value is still used. At the same time, a receipt confirmation is returned to the cloud, completing the sending process to provide a reliable initial SOC baseline. S6.3, Vehicle-side Operating Condition Analysis and Value matching; While receiving the initial SOC value from the cloud, the vehicle-side BMS system performs real-time analysis of the operating conditions of the current driving segment. Based on the 14 operating condition features determined in step two (multimodal dynamic operating condition analysis), it collects operating condition data for the current driving segment in real time, calculates the feature vector of the current operating condition, compares the current operating condition feature vector with the central feature vectors of the four previously clustered operating conditions to calculate the Euclidean distance, finds the nearest operating condition category, determines the operating condition type of the current driving segment, and accurately matches the corresponding operating condition influencing factors. ; S6.4, Vehicle-side aging - Adaptive ampere-hour integral calculation under operating conditions; The vehicle-side system uses the initial SOC value provided by the cloud, the current SOH value of the battery, and the matched operating condition influencing factors. The change in SOC during driving was calculated using the ampere-hour integral method that is adaptive to aging and operating conditions.

10. The fusion SOC estimation method based on cloud-based initial value calibration according to claim 9, characterized in that, In S6.3, the formula for calculating the change in SOC is: ； in, The change in SOC; From the start time of the driving segment up to the current moment Total discharge capacity; The total energy of the battery; The current health status of the battery. Operating condition influencing factors; From arrive The change in SOC at time t.

Images