Energy storage battery health feature extraction and state evaluation method based on transfer learning

Abstract

The invention discloses an energy storage battery health feature extraction and state evaluation method based on transfer learning. The method comprises the following steps: S1, constructing a source domain health feature library and pre-training a model; according to the method, dependence on complete cyclic data is broken through, high precision and robustness are still achieved under the conditions of data sparsity and working condition difference, and the method is suitable for intelligent operation and maintenance and predictive maintenance of an energy storage power station. And meanwhile, common incomplete and partial charge and discharge data fragments under actual working conditions can be directly utilized for feature extraction and state evaluation, dependence on complete charge and discharge cycles is avoided, and the application scene of the data driving method is greatly widened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage battery management, in particular to a lithium iron phosphate energy storage battery health feature extraction and state evaluation method based on transfer learning, which is suitable for accurately evaluating the battery health state and predicting the remaining useful life through incomplete or partial charge-discharge data of the battery under actual operating conditions. BACKGROUND

[0002] With the wide application of renewable energy and the promotion of smart grid construction, the role of energy storage systems in power systems is becoming increasingly important. Among them, lithium iron phosphate batteries have become one of the mainstream choices for large-scale energy storage systems due to their high safety, long cycle life, moderate cost and other advantages. In the operation and management of energy storage batteries, accurately evaluating the health state (State of Health, SoH) and predicting the remaining useful life (Remaining Useful Life, RUL) of the battery is of great significance to the safe operation and economy of the system.

[0003] Traditional battery health state evaluation methods mainly rely on complete charge-discharge cycle data to evaluate the health state by analyzing the battery capacity attenuation curve or internal resistance change. However, in actual energy storage system operation, the battery is often in irregular partial charge-discharge state, and there is little opportunity for complete charge-discharge cycle, which makes it difficult to apply traditional methods. In addition, the working conditions, load characteristics and environmental factors of batteries in different energy storage systems are different, resulting in significant differences in battery degradation patterns.

[0004] In recent years, data-driven deep learning methods have shown great potential in the field of battery health state evaluation. However, such methods usually require a large amount of historical operation data for model training, while newly commissioned energy storage systems often lack sufficient historical data. In addition, there are distribution differences between laboratory test data and actual operation data, and direct application of models trained in laboratory environment to actual systems may result in performance degradation.

[0005] Therefore, there is an urgent need for a method that can utilize limited and incomplete actual operation data, combine existing battery test data, and achieve accurate battery health state evaluation through knowledge transfer. SUMMARY

[0006] The purpose of the present application is to provide a lithium iron phosphate energy storage battery health feature extraction and state evaluation method based on transfer learning to address the above deficiencies. The existing technology has deficiencies in lithium iron phosphate energy storage battery health state evaluation, and the specific technical problems to be solved include: The problem of incomplete actual operating condition data: In actual energy storage systems, batteries are mostly in a floating charge, shallow charge and discharge or irregular partial charge and discharge state, making it difficult to obtain the complete and standardized charge and discharge cycle data necessary for health status assessment, causing traditional methods that rely on complete cycle data to fail.

[0007] The problem of sparsity in historical data for new systems: Newly deployed energy storage systems lack long-term operational history data, which cannot meet the large number of labeled samples required for traditional data-driven models (such as deep learning models) to be trained from scratch, resulting in difficulty in building models or low accuracy.

[0008] Data distribution discrepancy issue: Battery aging data collected in a laboratory environment (source domain data) and actual operating data of the energy storage system (target domain data) differ significantly in terms of operating temperature, load characteristics, and charge / discharge rate, leading to inconsistent data distribution (i.e., "domain drift"). Directly applying models trained on laboratory data to real-world systems will result in significant evaluation errors due to insufficient generalization ability.

[0009] Poor static adaptability of the model: Battery aging is a dynamic and non-linear process, and its degradation mode may change over time. A static model trained once is difficult to adapt to the changes in battery characteristics throughout its entire lifespan, which may lead to a decrease in evaluation accuracy over time.

[0010] This invention is achieved through the following scheme: A method for extracting health features and assessing the state of energy storage batteries based on transfer learning includes the following steps: Step S1, Source Domain Health Feature Library Construction and Model Pre-training: Obtain long-term full-cycle test data of at least one source battery; extract health feature sequences from the data, wherein the health features include at least the feature curves and their statistics obtained through incremental capacity analysis or differential voltage analysis; and train at least one pre-trained deep learning model for learning the evolution law of the features based on the health feature sequences. Step S2, target domain health feature extraction: obtain incomplete or partial charge and discharge data fragments of the target energy storage battery under actual operating conditions; process the data fragments to extract the fragmented health features of the target battery; Step S3, knowledge transfer and model adjustment based on domain adaptation: using a domain adaptation algorithm to align the data distribution of source domain health features and target domain fragmented health features; using some or all parameters in the pre-trained deep learning model as initial parameters, and adjusting the model using the fragmented health features of the target domain to obtain at least one health status assessment model adapted to the target battery. Step S4, State Assessment and Prediction: Input the fragmented health characteristics of the target battery into the adjusted health state assessment model, and output the current battery health state assessment result and / or remaining service life prediction result.

[0011] The health characteristics in step S1 specifically include the peak position, peak height, and area under the peak obtained by analyzing the incremental capacity curve, and / or the valley position obtained by analyzing the differential voltage curve. Valley depth Distance between valleys .

[0012] In step S1, the pre-trained deep learning model is a hybrid model containing convolutional neural network layers and long short-term memory network layers; wherein, the convolutional neural network layer is used to automatically extract local and shape-related spatial features from the incremental capacity / differential voltage curve, and the long short-term memory network layer is used to learn the time series pattern of these features evolving with the number of iterations.

[0013] The core of the fragmented health feature extraction in step S2 is the calculation of the incremental capacity curve, which is mathematically defined as the derivative of voltage with respect to capacity. In actual calculations, it is approximated by difference.

[0014] in, and These represent the capacity and voltage at the k-th sampling point, respectively. and These represent the capacity and voltage at the (k+1)th sampling point, respectively.

[0015] Step S2 further includes the following steps: Step S21, Data preprocessing: Noise is filtered out from the collected voltage, current, and time data segments, and smoothing is performed using a filtering algorithm; Step S22, Fragmented IC curve generation: Based on the preprocessed data, calculate IC curve segments within the available voltage segments.

[0016] The domain adaptation algorithm in step S3 includes minimizing the distribution of source domain health features. and the distribution of fragmented health features in the target domain The loss function is used to align the feature space by maximizing the average difference between them. Represented as:

[0017] in, and These are health feature samples extracted from the source domain and the target domain, respectively. To map features to the reproducing kernel Hilbert space kernel function, and These represent the number of samples in the source domain and the target domain, respectively.

[0018] The model adjustment in step S3 aims to minimize the prediction loss function in the target domain. Its mathematical expression is:

[0019] in, The true labels for the target domain samples. For the entire deep learning model, For the input fragmented health features, For convolutional neural network layer parameters fixed from a pre-trained model, For the parameters of the Long Short-Term Memory network layer to be optimized, This represents the number of samples in the target domain.

[0020] This method also includes an incremental learning step. When new data fragments of the target battery are acquired, the model is updated using an elastic weight merging algorithm, and its total loss function... for:

[0021] in, It's the task loss on new data. It is the regularization coefficient. It is the model number One parameter, This is the optimal value for this parameter in previous tasks. These are the diagonal elements of the Fisher information matrix, used to measure parameters. Importance to previous tasks.

[0022] Step S1: Construction of the Source Domain Health Feature Library and Model Pre-training; This step aims to establish a general battery aging knowledge base; specifically, it includes the following steps: Step S11, Source Domain Data Acquisition: Acquire full life cycle aging test data of at least one lithium iron phosphate battery (as the source battery) under standard laboratory conditions. This data should include multiple complete charge-discharge cycle records from brand new to end of life (e.g., capacity decay to 80% of nominal capacity), with each record point containing precise voltage, current, capacity, and timestamp; Step S12, Source Domain Health Feature Extraction: Incremental capacity analysis (ICA) is performed on the charge or discharge data for each cycle. The incremental capacity (IC) curve is obtained by differential calculation of the voltage-capacity curve. The formula for calculating the IC curve is as follows: From the IC curve of each cycle, key statistics characterizing the internal electrochemical state of the battery are extracted as health indicators (HI), such as the valley positions of one or more major peaks. Valley depth Distance between valleys These characteristics change systematically with battery aging; Step S13: Arrange the health feature vectors of all cycles of the source battery in chronological order to form one or more time series, constituting a source domain health feature library. Based on this feature library, construct and pre-train a deep learning model to learn the mapping relationship from health features to battery state of health (SoH) or remaining useful life (RUL). Preferably, the model adopts a hybrid structure of convolutional neural network (CNN) and long short-term memory network (LSTM). CNN layer: Used to automatically extract deep, abstract spatial correlation features within the input health feature vector, enhancing the robustness of the features.

[0023] LSTM layer: Receives features extracted by the CNN layer and learns the evolution and long-term dependencies of these features over time, thereby capturing the dynamic process of battery aging. After pre-training, a pre-trained model containing general aging knowledge is obtained.

[0024] Step S2 Target Domain Health Feature Extraction; This step aims to extract effective information from actual operating target batteries with incomplete data, and specifically includes the following steps; Step S21, Target Domain Data Acquisition: Acquire voltage, current, and time data segments from the target energy storage battery system to be evaluated under actual operating conditions. These segments may only cover a very small voltage range and vary in duration and depth; Step S22, Fragmented Feature Extraction: The collected data fragments are cleaned and filtered, retaining valid charging or discharging fragments. For each valid fragment, even if it is incomplete, its local IC curve is calculated. From these "fragmented IC curves," health features (such as features of local peaks) that are the same as those defined in the source domain are extracted to form a fragmented health feature vector for the target domain.

[0025] Step S3, knowledge transfer and model adjustment based on domain adaptation, is the core of this invention. It aims to adapt general knowledge to a specific target battery and includes the following steps: Step S31, Feature Space Alignment: To address the data distribution differences between the source and target domains, a domain adaptation algorithm is introduced. Maximum Mean Discrepancy (MMD) is preferably used as the metric. During model training, an MMD loss term is added to the loss function to minimize the distribution distance between source and target domain features in the high-dimensional feature space mapped by the CNN layers. This forces the model to learn a "domain-invariant" feature representation, thereby eliminating the impact of differences in operating conditions. Step S32, Model Transfer and Tuning: The parameters of the pre-trained model are used as the initial weights of the new model. Then, the model is fine-tuned using a small (or even very small) number of fragmented health features collected from the target domain and their corresponding true SoH labels (if available, obtained through offline calibration or several initial full charge-discharge cycles). The overall loss function for fine-tuning is... It consists of two parts: in, It is the SoH prediction loss (such as mean square error) in the target domain, used to make the model fit the characteristics of the target battery; This is the domain adaptation loss; and the weighting coefficients that balance the two. By optimizing this overall loss, the model inherits general aging knowledge while also achieving personalized adaptation to the target battery characteristics. Step S4, State Assessment and Prediction, is the application and continuous optimization phase of the method. Specifically, it can be: online assessment: deploying the fine-tuned model into the battery management system (BMS); during battery operation, collecting the latest data fragments in real time, extracting fragmented health features, and inputting them into the model to obtain a high-precision assessment value of the current SoH and / or a predicted value of RUL.

[0026] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. Strong applicability and breakthrough in data integrity limitations: This invention can directly utilize incomplete or partial charge and discharge data fragments commonly found in actual working conditions for feature extraction and state assessment, eliminating the dependence on complete charge and discharge cycles and greatly expanding the application scenarios of data-driven methods.

[0027] 2. Efficiently utilize knowledge to solve the problem of data sparsity: Through transfer learning, this method will transfer general aging knowledge learned from a large amount of laboratory data to the target battery. Only a small amount of target domain data is needed to complete the model adaptation, which effectively solves the problem of insufficient historical data for new energy storage systems and achieves rapid and accurate evaluation under "cold start".

[0028] 3. High accuracy and robustness: By introducing the domain adaptation algorithm (MMD), this method effectively reduces the distribution difference between laboratory data and actual operating data, enabling the model to learn the essential features that are invariant to the domain, thereby significantly improving the evaluation accuracy and robustness of the model under complex and ever-changing actual working conditions.

[0029] 4. Adaptive capabilities ensure long-term effectiveness: By combining an online learning strategy (EWC), the model can continuously optimize and adjust itself during use, adapting to the characteristic changes of the target battery at different stages of its life cycle, thus ensuring the long-term accuracy of the evaluation results throughout the battery's entire life cycle.

[0030] 5. High degree of automation and easy to implement in engineering: The entire method, from feature extraction to model evaluation, is data-driven, with a clear process and can achieve a high degree of automation. It is easy to integrate into existing energy storage battery management systems (BMS), providing strong technical support for intelligent operation and maintenance and predictive maintenance of energy storage systems. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the battery health status assessment method based on transfer learning proposed in this invention. Figure 2 This is a schematic diagram illustrating the extraction of incremental capacity (IC) curves and related health characteristics from battery charge and discharge data; Figure 3 This is a schematic diagram of the CNN-LSTM hybrid neural network model structure used in this embodiment of the invention; Figure 4 This is a schematic diagram showing the alignment of the feature distributions of the source domain and the target domain before and after the domain adaptation process. Detailed Implementation

[0032] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0033] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0034] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a predetermined orientation, or be constructed and operated in a predetermined orientation. Therefore, they should not be construed as limitations on this invention.

[0035] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0036] Example 1 A method for extracting health features and assessing the state of energy storage batteries based on transfer learning; (Refer to...) Figure 1 The overall process of this method includes four main stages: source domain health feature library construction and model pre-training, target domain health feature extraction, domain adaptation-based knowledge transfer and model adjustment, and state evaluation and prediction. Step S1: Construction of source domain health feature library and model pre-training; Step S11, Source Domain Data Acquisition: First, acquire complete cycle aging test data of one or more sets of the same type of lithium iron phosphate batteries (source batteries) under standard laboratory conditions (such as constant temperature and standard charge / discharge rate). This data includes detailed voltage, current, and capacity records for each charge / discharge cycle.

[0037] Step S12, Source Domain Health Feature Extraction: For each cycle, extract the voltage-capacity curve of its charging or discharging process. Based on this curve, calculate its incremental capacity (IC) curve. Figure 2 As shown, the IC curve typically contains several distinct peaks, the shape (position, height, area) of which are closely related to the electrochemical reactions and aging state inside the battery. Therefore, by extracting the voltage position, peak height, and area under the peak for each major peak, these three factors, or combinations thereof, can be used to construct a health feature vector for a single cycle.

[0038] Step S13, Source Domain Feature Sequence Construction and Model Pre-training: Arrange the health feature vectors of the source battery in chronological order for all cycles from brand new to end of life to form the source domain health feature sequence.

[0039] Build a deep learning model to learn the evolutionary patterns of battery aging. For example... Figure 3 As shown, this embodiment employs a CNN-LSTM hybrid model. The CNN layer automatically extracts deeper, more robust local spatial features from the input health feature vector; the LSTM layer learns the evolution and long-term dependencies of these features over time. The model is trained using the source domain health feature sequence and the corresponding real SoH / RUL labels, resulting in a pre-trained model.

[0040] Step S2, target domain health feature extraction; Step S21, Target Domain Data Acquisition: In an actual energy storage power station, collect voltage, current, and time data of the target lithium iron phosphate battery during daily operation. These data are typically irregular and incomplete charge and discharge segments.

[0041] Step S22, Fragmented Feature Extraction: The acquired data fragments are preprocessed to filter out valid charging or discharging fragments. For each fragment, even if it only covers a small voltage range, its IC curve is calculated. From these fragmented IC curves, features corresponding to the source domain features, such as the features of locally occurring peaks, are extracted as much as possible. These features constitute the fragmented health features of the target domain.

[0042] Step S3: Knowledge transfer and model adjustment based on domain adaptation; Step S31, Feature Space Alignment: Due to differences between laboratory conditions and actual operating conditions, the distributions of source domain features and target domain features may differ. For example... Figure 4 As shown on the left, the distributions of the two are offset. To address this issue, an MMD loss term is introduced into the loss function during model training. By minimizing the MMD loss, the feature distribution of the target domain can be pulled towards the feature distribution of the source domain at the output of the feature extraction network (such as a CNN layer), thus achieving... Figure 4 The alignment effect shown on the right.

[0043] Step S32, Model Transfer and Adjustment: The pre-trained model parameters obtained in Step S1 are used as initial values. Then, the model is adjusted using a small number of fragmented healthy features of the target domain and their corresponding (if any) true SoH labels. The loss function in the adjustment process consists of two parts: one part is the prediction loss on the target domain, used to adapt the model to the data in the target domain; the other part is the MMD loss, used to ensure domain alignment.

[0044] (Preferred Solution) Online Update and Elastic Weight Combination (EWC): To enable the model to continuously adapt to changes in the target domain, an online learning approach can be employed. When new data fragments from the target domain are collected, the EWC strategy is used to update the model. EWC evaluates the importance of each parameter to the old task (source domain task) using the Fisher information matrix, and penalizes important parameters when updating them, thereby preventing the model from forgetting key knowledge learned from the source domain when learning new knowledge (i.e., the "catastrophic forgetting" problem).

[0045] Step S4, State Assessment and Prediction: When it is necessary to assess the current health state of the target battery, the latest data fragment is collected, fragmented health features are extracted, and input into the adjusted final assessment model. The model will output an accurate assessment value for the current SoH or a predicted value for RUL.

[0046] Through the above steps, this invention successfully transfers the universal aging knowledge learned from laboratory big data sets to real-world application scenarios with sparse data and complex operating conditions, solving the pain point that traditional methods are difficult to apply and achieving high-precision and robust assessment of the health status of energy storage batteries.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for energy storage battery health feature extraction and state assessment based on transfer learning, characterized in that, The method comprises the following steps: Step S1, source domain health feature library construction and model pre-training: obtaining long-term full cycle test data of at least one source battery; extracting a health feature sequence from the data, the health feature at least including a feature curve and a statistical quantity thereof obtained by incremental capacity analysis or differential voltage analysis; Based on the health feature sequence, at least one pre-trained deep learning model for learning the evolution law of the feature is trained; Step S2, target domain health feature extraction: obtaining an incomplete or partial charge and discharge data segment of a target energy storage battery under actual operation conditions; processing the data segment to extract a fragmented health feature of the target battery; Step S3, domain adaptation-based knowledge transfer and model adjustment: using a domain adaptation algorithm to align the data distribution of the source domain health feature and the target domain fragmented health feature; using the fragmented health feature of the target domain to adjust the pre-trained deep learning model with part or all of the parameters as initial parameters, to obtain at least one health state evaluation model adapted to the target battery; Step S4, state evaluation and prediction: inputting the fragmented health feature of the target battery into the adjusted health state evaluation model to output the health state evaluation result and / or the remaining service life prediction result of the current battery.

2. The method of claim 1, wherein the method is based on transfer learning. The health features in the step S1 specifically include a peak position, a peak height, and a peak area obtained by analyzing the incremental capacity curve, and / or a valley position, a valley depth, and a valley distance obtained by analyzing the differential voltage curve. .​​ 3. The method of claim 1, wherein the method is based on transfer learning. In step S1, the pre-trained deep learning model is a hybrid model comprising a convolutional neural network layer and a long short-term memory network layer; wherein the convolutional neural network layer is used to automatically extract local and shape-related spatial features from the incremental capacity / differential voltage curve, and the long short-term memory network layer is used to learn the time sequence law of the evolution of these features with the cycle number.

4. The method of claim 1, wherein the method is based on transfer learning. The core of the fragmented health feature extraction in step S2 is to calculate the incremental capacity curve, which is mathematically defined as the derivative of voltage with respect to capacity, and in actual calculation, it is approximated by difference: wherein, and V(k) and V(k+1) are the voltage at the kth and (k+1)th sampling points, respectively. and C(k) and C(k+1) are the capacitance at the kth and (k+1)th sampling points, respectively.

5. The method of claim 4, wherein the method is based on transfer learning. The step S2 further comprises the following steps: Step S21, data preprocessing: performing noise filtering on the collected voltage, current and time data segments, and performing smoothing processing by using a filtering algorithm; Step S22, fragmented IC curve generation: based on the preprocessed data, calculating an IC curve segment within the voltage segment. 6.The method of claim 1, wherein: The domain adaptation algorithm in the step S3 includes aligning the feature space by minimizing the maximum mean discrepancy between the source domain health feature distribution and the target domain fragmented health feature distribution with a loss function represented as: wherein, and are healthy feature samples extracted from the source domain and the target domain, respectively, is a kernel function that maps the features to a reproducing kernel Hilbert space , and are the number of samples of the source domain and the target domain, respectively.

7. The method of claim 1, wherein the method is based on transfer learning. The model adjustment in the step S3 aims to minimize a prediction loss function on the target domain The mathematical expression is: wherein, is a true label of a target domain sample, is a whole deep learning model, is an input fragmented health feature, is a fixed convolutional neural network layer parameter from a pre-training model, is a long short-term memory network layer parameter to be optimized, is a target domain sample quantity. 8.The method of claim 1, wherein: An incremental learning step is also included, when a data segment of the target battery is obtained, the model is updated using an elastic weight combination algorithm, and the total loss function of the model is : wherein, is the task loss on new data, is a regularization coefficient, is the model’s th parameter, is the optimal value of this parameter in the previous task, is the diagonal element of the Fisher information matrix, measuring the importance of the parameter for the previous task. 9.The method of claim 1, wherein: Step S1 specifically comprises the following steps: Step S11, source domain data acquisition: obtaining full life cycle cycle aging test data of at least one source battery under laboratory standard conditions; the data contains multiple complete charge and discharge cycle records from a brand new state to the end of life, and each record point contains voltage, current, capacity and time stamp; Step S12, source domain health feature extraction: performing incremental capacity analysis on the charge or discharge data of each cycle; by differentiating the voltage-capacity curve, the incremental capacity is obtained; the calculation formula of the incremental capacity is: from the incremental capacity curve of each cycle, the key statistical quantity representing the internal electrochemical state of the battery is extracted as the health feature; Step S13, source domain feature sequence construction and model pre-training: arranging the health feature vectors of all cycle periods of the source battery in time sequence to form one or more time sequences, to constitute a source domain health feature library; based on the source domain health feature library, constructing and pre-training at least one deep learning model, which is used to learn the mapping relationship from the health feature to the battery health state or the remaining useful life.

10. The method of claim 1, wherein the method is based on transfer learning. Step S3 specifically includes the following steps: Step S31, feature space alignment: introducing a domain adaptation algorithm; using maximum mean discrepancy as a measurement criterion; in the model training process, by adding a maximum mean discrepancy loss term in the loss function, to minimize the distribution distance of the source domain features and the target domain features in the high-dimensional feature space mapped by the convolutional neural network layer; Step S32, model migration and adjustment: taking the parameters of the pre-trained model as the initial weights of the new model; using the fragmented health features collected in the target domain and their corresponding real SoH labels to adjust the model; the overall loss function of the adjustment Comprises two parts: Wherein, is the SoH prediction loss on the target domain, used to make the model fit the characteristics of the target battery; is the domain adaptation loss; is the weight coefficient of balancing the two; by optimizing the overall loss, the model not only inherits the general aging knowledge, but also completes the individualized adaptation to the characteristics of the target battery.

Images