Cross-scale joint estimation method and system for dynamic power scheduling of large energy storage system

Abstract

The invention discloses a cross-scale joint estimation method for dynamic power scheduling of a large-scale energy storage system, and the method comprises the steps: obtaining operation data in the energy storage system, and constructing a corresponding first data set; constructing a first regression model used for predicting SOC and a second regression model used for predicting SOH based on an extreme gradient boosting tree of Bayesian optimization; forming a second data set by the working condition data and the corresponding predicted SOC sequence and predicted SOH sequence; constructing a bidirectional interaction neural network, wherein the bidirectional interaction neural network comprises a main feature coding module and a coupling type interaction modeling module; and deploying the prediction model at an operation control end of the energy storage system, and predicting the current SOH and SOC according to the collected working condition data. The invention also provides a cross-scale joint estimation system. According to the method provided by the invention, high-precision combined prediction of the battery state can be realized, and the safety, precision and robustness of power scheduling of the energy storage system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery state estimation technology, and in particular to a cross-scale joint estimation method and system for dynamic power scheduling of large-scale energy storage systems. Background Technology

[0002] With the continuous increase in the proportion of renewable energy connected to the grid, large-scale battery energy storage systems have become an important support for power system peak shaving and frequency regulation, emergency backup, and renewable energy consumption. However, in actual operation, energy storage systems face complex challenges such as frequent charge-discharge switching, high-rate load changes, and long-term aging and degradation, which place higher demands on the safety, accuracy, and dynamic response capabilities of power dispatch.

[0003] The power dispatching process of energy storage systems relies heavily on accurate perception of battery status. State of Charge (SOC) reflects the current remaining available energy, while State of Health (SOH) describes the degree of long-term performance degradation, such as capacity decay and increased internal resistance. Estimating both together is crucial for avoiding overcharging and over-discharging, improving power utilization efficiency, and achieving lifetime-aware dispatching.

[0004] However, current mainstream scheduling methods generally have the following shortcomings: Based solely on SOC scheduling, ignoring the impact of SOH, the power capability cannot be correctly assessed after battery aging, which can easily lead to overload or mismatch. SOC and SOH estimations are usually performed separately, without taking into account their coupling relationship and time scale differences, making it difficult to achieve unified modeling and high-precision prediction. In systems with drastic changes in operating conditions and significant differences between clusters, traditional static models are difficult to effectively reflect the state inconsistencies across cells and modules, and scheduling strategies lack flexibility and robustness.

[0005] Patent CN 119355530 A discloses a biLSTM-based joint estimation method and system for SOC-SOH. This method first acquires the time-series data of voltage, temperature, current, SOC, and SOH within the current cycle of the energy storage device. The voltage and temperature time-series data are then input into a temperature-voltage prediction sub-network, outputting the voltage and temperature prediction curves for the next cycle. The voltage, temperature, and current time-series data are then input into the SOC prediction sub-network, outputting the SOH prediction curve. Finally, the voltage, temperature, current, and SOH prediction curves are input into the SOH prediction sub-network, outputting the SOC prediction curve. However, this method only specifies the order of SOH and SOC predictions and does not perform collaborative optimization, failing to reflect the coupling influence between SOH and SOC, resulting in limited prediction accuracy.

[0006] Patent CN 120121995 A discloses a method and system for jointly estimating the State of Charge (SOC) and State of Hypothesis (SOH) of all cells in an energy storage system. This method first measures the terminal voltage of each battery, selecting the cells with the highest and lowest terminal voltages as characteristic cells. Then, it establishes circuit models for these characteristic cells. Based on these circuit models, it jointly estimates the SOC and SOH of the characteristic cells. Finally, based on the circuit models of each cell, it performs a fusion estimation of the SOC and SOH of each cell. The SOC and SOH estimates of each cell obtained from the joint estimation are used to correct the fusion estimation results, which are then used as the final joint estimation result for the SOC and SOH of all cells in the energy storage system. However, because the terminal voltage and circuit model parameters change significantly with battery aging, affecting the selection of characteristic cells and the accuracy of the cell circuit models, this method has limited estimation accuracy during long-term operation of the energy storage system. Summary of the Invention

[0007] The purpose of this invention is to provide a cross-scale joint estimation method and system for dynamic power dispatching of large-scale energy storage systems. This method can achieve high-precision joint prediction of battery state and enhance the security, accuracy and robustness of power dispatching of energy storage systems.

[0008] To achieve the first objective of this invention, the following technical solution is provided: a cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems, comprising the following steps: The operating data of the energy storage system is acquired, including operating condition data, historical SOC sequence and historical SOH sequence. The operating condition data is rearranged in charge-discharge cycle order and constructed with the historical SOC sequence and historical SOH sequence to form a corresponding first dataset. Based on Bayesian-optimized extreme gradient boosting trees, a first regression model for predicting SOC and a second regression model for predicting SOH are constructed. The operating condition data are input into the first regression model and the second regression model respectively to output the predicted SOC sequence and the predicted SOH sequence. The operating condition data and the corresponding predicted SOC sequence and predicted SOH sequence are combined to form the second dataset. Construct a bidirectional interactive neural network, including a main feature encoding module and a coupled interactive modeling module; The main feature encoding module is used to extract multimodal features from the input running data to generate a joint embedding representation; The coupled interactive modeling module includes a SOC-dominant path unit, a SOH-dominant path unit, and a data augmentation unit. The SOC-dominant path unit takes historical SOC sequences and charge / discharge behavior as input to predict SOH results. The SOH-dominant path unit takes historical SOH sequences and operating condition data and predicts SOC results by back-calculating SOH back to SOC. The data augmentation unit is used to take the output of one dominant path unit as the self-attention input of another dominant path unit to adjust the conditions of the other dominant path unit. The bidirectional interactive neural network was initially trained using the second dataset, and then trained a second time using the first dataset to obtain a predictive model for predicting SOH and SOC. The predictive model is deployed at the operation control terminal of the energy storage system, and the current SOH and SOC are predicted based on the collected operating data.

[0009] This invention uses Bayesian optimization and XGBoost regression algorithms to fuse the short-term operating state (SOC) and long-term health state (SOH) characteristics of the battery for high-precision joint prediction of battery status. It also introduces uncertainty quantification and online closed-loop correction mechanisms to improve the model's adaptability to dynamic operating conditions and battery degradation behavior, thereby enhancing the safety, accuracy, and robustness of power scheduling in energy storage systems.

[0010] Specifically, the operating data includes voltage, current, temperature, charge / discharge quantity, and number of cycles.

[0011] Specifically, the bidirectional interactive neural network is constructed based on a cooperative Transformer structure, which simulates the dynamic feedback relationship between the two through residual coupling blocks, and integrates the nonlinear mapping between charge-capacity and impedance-health in the battery degradation evolution process.

[0012] Specifically, during the second training process, a joint loss function was used to fine-tune the initial model.

[0013] Specifically, the expression for the joint loss function is as follows: ; in, For the total loss, and Weighting coefficients set by humans. Weak labels for the outputs of the first regression model for predicting SOC and the second regression model for predicting SOH. , Losses caused by guidance The loss is guided by the true SOC and SOH labels. The loss of physical coupling consistency is constructed based on the known thermodynamic and electrochemical relationships of the battery.

[0014] Specifically, the expression for the SOC dominant path unit is as follows:

[0015] in, : This represents the charge and discharge charge for that cycle; This represents the depth of discharge during that cycle; k 1, k 2, , These are empirical parameters related to battery type, temperature, and operating conditions.

[0016] Specifically, the prediction model also includes uncertainty estimation based on the Monte Carlo Dropout method before outputting the prediction results.

[0017] Specifically, the parameters of the prediction model deployed in the energy storage system are iteratively updated through a lightweight error backpropagation mechanism and a gradient constraint update strategy.

[0018] To achieve the second objective of this invention, the following technical solution is provided: a cross-scale joint estimation system for performing the steps of the above-described cross-scale SOC and SOH joint estimation method for dynamic power scheduling of large-scale energy storage systems.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the complex operating conditions of frequent power regulation, high-speed response, and battery degradation in large-scale energy storage systems, a cross-scale joint estimation model integrating SOC and SOH was constructed. By introducing Bayesian optimization parameter tuning, a SOC–SOH cascade structure, an uncertainty quantification mechanism, and an online closed-loop correction strategy, high-precision, probabilistic, and adaptive prediction of battery state is achieved, providing more reliable, flexible, and health-friendly state support for power scheduling strategies. This significantly improves prediction reliability and system scheduling safety margins, demonstrating good engineering practicality and deployment adaptability. Attached Figure Description

[0020] Figure 1 This is a flowchart of a cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems provided in this embodiment; Figure 2 A schematic diagram illustrating model training provided in this embodiment; Figure 3 The flowchart of the hyperparameter Bayesian optimization of the extreme gradient boosting tree provided in this embodiment is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown in this embodiment, a cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems is provided, including the following steps: Feature extraction and preprocessing. Battery operation data is collected from large-scale energy storage systems to construct an input feature set for joint prediction of state of charge (SOC) and state of health (SOH). The features include: Voltage (V): Records the change in terminal voltage of the battery during charging and discharging, and is a direct representation of the state of charge. Current (I): Captures load changes and energy flow direction, serving as an important input for dynamic behavior; Temperature (T): Considering the battery's thermosensitive characteristics, it reflects the impact of high-temperature aging and low-temperature performance degradation; Charge / discharge quantity (Ah): The amount of charge accumulated in each charge / discharge cycle, which helps in modeling short-term energy change processes; Cycle count (N): Records the number of complete charge-discharge cycles the battery has completed, used to measure the degree of degradation; State of charge (SOC): An approximate estimate that can be directly obtained for some systems, and can also be used as an auxiliary input; Battery State of Health (SOH): A historical health indicator that can be used to build an initial SOH prediction baseline.

[0023] The seven categories of features are organized into a time series based on charge / discharge cycles, constructing a dataset with a temporal structure. To improve the consistency and completeness of the feature input, the data needs to undergo the following preprocessing before entering the model: Time series reconstruction uses a sliding window approach to reconstruct features from multiple consecutive cycles into a fixed-length time series input, preserving dynamic trends and long-term correlation information.

[0024] Normalization: All features are uniformly normalized to the interval [0,1] or [-1,1] to improve the convergence stability of the model.

[0025] Missing and outlier values ​​are handled by using methods such as interpolation, annealing, or local averaging to fill in some missing data and removing voltage / current abrupt changes.

[0026] The consistency enhancement mechanism aligns the features of multiple cells and modules uniformly according to the index, and can integrate extended dimensions such as cell number and module ID to improve the generalization capability across battery systems.

[0027] Through the above feature extraction and preprocessing steps, stable, complete, and usable multidimensional time series samples for joint prediction model input can be obtained, providing a high-quality data foundation for subsequent parameter optimization and SOC / SOH modeling.

[0028] Model Construction and Parameter Optimization. A dual-output regression model based on extreme gradient boosting trees (XGBoost) is constructed to jointly predict the State of Charge (SOC) and State of Harm (SOH) in energy storage battery systems. This model incorporates physical mechanism coupling and a data-driven fusion strategy, and further utilizes Bayesian optimization to automatically search for the optimal hyperparameter combination, thereby improving prediction accuracy and generalization ability. The specific steps are as follows: like Figure 2 The diagram illustrates the coupled modeling and bidirectional prediction of SOC-SOH. Extreme gradient boosting (XGBoost) is used to construct a pair of high-confidence SOC and SOH regression models, outputting preliminary prediction results for SOC and SOH. and , as an auxiliary label for neural networks.

[0029] The XGBoost model uses multi-source operating condition variables from historical observation sequences as input, including: voltage V Current I ,temperature T , Charge quantity Δ Q Number of loops N In addition to some SOC derivative information ΔSOC, the following two prediction sub-models were trained respectively:

[0030]

[0031] in, This represents the SOC estimate output by the XGBoost model. This represents the SOH estimate output by the XGBoost model.

[0032] To improve model performance, a Bayesian optimization algorithm is introduced to efficiently search for the hyperparameters of the XGBoost model. The optimization process includes: (1) Define the hyperparameter set of the XGBoost model as follows: θ ={ θ 1, θ 2,..., θ n}, including but not limited to: Learning rate (0.01~0.3): controls the step size of model updates; Maximum tree depth (3~10): limits model complexity; Subsampling rate (0.6~1.0): controls sample utilization; L1 and L2 regularization coefficients (0~10): suppress overfitting; Construct a mapping function from hyperparameters to model performance as the objective function. f ( θ ),Right now

[0033] in, Indicates the sample data ( x , y The mathematical expectation of ) represents the average performance of the performance metric on the training / validation dataset; The performance metric function is commonly represented by the mean squared error (MSE) under cross-validation or the negative log-likelihood (NLL). y For SOC or SOH, x The input features of the model include voltage. V Current I ,temperature T , Charge quantity Δ Q Number of loops N wait.

[0034] (2) Construct a Gaussian process (GP) model to approximate the hyperparameter combination. θ With the objective function value f ( x The process models the distribution of the objective function value using evaluated samples, providing joint predictive power for mean and variance.

[0035]

[0036] in, m ( θ ) indicates that the Gaussian process is applied to θ The predicted mean is an estimate of the objective function; This is a kernel function that characterizes the correlation between different combinations of hyperparameters.

[0037] (3) The expected improvement (EI) is used as the acquisition function to guide the search, and the hyperparameter combination with the largest potential improvement space is selected for evaluation to achieve efficient convergence. The definition is as follows:

[0038] in, f best This represents the currently known optimal objective function value. The function uses the posterior distribution of the generalized system (GP) to calculate the "expected improvement" brought by each candidate hyperparameter combination, selecting the one with the largest EI. * As a point for the next round of evaluation.

[0039] If EI is less than the set threshold or the maximum number of iterations is reached, the search terminates and the optimal hyperparameter combination is output. * ,exist * We then retrained two XGBoost models, SOC and SOH, to serve as auxiliary label sources for subsequent neural network prediction models, thereby achieving the embedding and guidance of Bayesian prior knowledge.

[0040] S2-2, Physically Coupled Guided Interactive Neural Network Modeling. , The features are concatenated to the main model input and then fed into the constructed collaborative Transformer structure. This structure introduces a bidirectional influence modeling mechanism of "SOC→SOH" and "SOH→SOC", simulating the dynamic feedback relationship between the two through residual coupling blocks, and integrating the nonlinear mapping between charge-capacity and impedance-health in the battery degradation evolution process.

[0041] Specifically, it includes the following parts: (1) Main feature encoding module: This module extracts the joint embedding representation of multimodal features such as voltage, current, and temperature with prior labels (XGBoost predictions) as the basis for subsequent interactive modeling.

[0042] (2) Coupled Interactive Modeling Module: This module is used to model the relationship between the influence of SOC on SOH and the influence of SOH on SOC, and includes the following parts: a. Dominant SOC pathway: SOC → SOH.

[0043] By constructing SOC, Δ Q The Transformer branch, taking historical SOC sequences and charge / discharge behavior as input, captures the impact of the current state of charge (SOC) on the battery degradation rate. This branch extracts time-related features and constructs the following physically-guided nonlinear degradation relationship:

[0044] in, : This represents the charge and discharge charge for that cycle; This represents the depth of discharge during that cycle; k 1, k 2, , These are empirical parameters related to battery type, temperature, and operating conditions. This differential equation characterizes the battery degradation rate under different SOC charge / discharge states. The Transformer network learns the complex dynamic response relationship through its multi-head attention mechanism, thereby outputting accurate SOH evolution prediction results.

[0045] b. Dominant pathway of SOH: SOH → SOC This approach models the feedback effect of SOH changes on SOC measurement accuracy. It is based on the following charge balance relationship:

[0046] in, I Represents a current signal. C nominal This represents the nominal battery capacity. When the actual battery's State of Health (SOH) decreases, it can lead to deviations in the State of Charge (SOC) measurement. Therefore, this branch takes historical SOH sequences, current sequences, temperature, and other state variables as inputs and uses a Transformer to construct a backtracking mechanism from SOH to SOC, achieving high-precision SOC estimation under dynamic capacity degradation.

[0047] c. Residual Coupling Mechanism and Dynamic Interaction Unit The two branches mentioned above are not only trained in parallel, but also achieve cross-branch feature fusion through a residual coupling block. This coupling block takes the attention output of the current branch as input and conditionsally adjusts the other branch, enabling the network to capture the impact of SOC changes on the SOH evolution rate and the feedback of SOH decay on the SOC calculation bias.

[0048] S2-3, Prior Fusion and Two-Stage Training Mechanism. The first stage uses XGBoost output. and As a weak label, it guides the neural network to quickly fit the main changing trends of SOC and SOH.

[0049] The second stage, after the network has acquired basic predictive capabilities, switches to real SOC and SOH measurement labels for the main training phase. Fine-tuning is performed using a joint loss function, the form of which is as follows:

[0050] in, For the total loss, and Weighting coefficients set manually; Weak labels for the first stage of XGBoost output , Losses caused by guidance For the second phase of real SOC and SOH labels , Losses caused by guidance. The loss of physical coupling consistency is constructed based on the known thermodynamic and electrochemical relationships of the battery. These are shown below.

[0051]

[0052] in, , This represents the final prediction result of the model; MSE stands for Mean Squared Error. t 0 indicates the start time of the charging / discharging process; t Indicates the current time point.

[0053] The model ultimately outputs the main prediction result. , The system simultaneously estimates the confidence interval and uncertainty index to support risk awareness and robust scheduling in scenarios such as peak shaving and frequency regulation. Uncertainty estimation is primarily based on the Monte Carlo Dropout method. The specific steps are as follows: (1) Enable the Dropout layer in the neural network during both the training and inference phases, so that the network randomly discards some neuron connections during each forward inference.

[0054] (2) Repeat the process for the same input feature sequence. T After one forward propagation, a set of output prediction sequences is obtained:

[0055] (3) Use the sample mean of the predicted sequence as the final predicted value:

[0056] (4) Using the sample variance of the predicted sequence as a measure of prediction uncertainty, upper and lower confidence intervals are further generated:

[0057] (5) Combining the aforementioned uncertainty indicators, a prediction interval is constructed with “main prediction value ± multiple standard deviations” as the boundary, so as to realize safety margin assessment and reliability indication for risk control scenarios of energy storage systems.

[0058] S3. Online Feedback Adaptive Correction and Model Correction. The trained SOC-SOH coupled estimation network is deployed at the energy storage system's operation and control terminal. Based on the state acquisition results after actual power scheduling behavior, the error between the predicted value and the sampled value is compared in real time. Through a lightweight error backpropagation mechanism and gradient constraint update strategy, the key parameters of the network are iteratively updated, forming a closed-loop self-evolving structure of "prediction → scheduling → feedback → correction", achieving continuous adaptation and highly robust response to battery aging trends and unsteady dynamic conditions.

[0059] S3-1 Error Feedback Mechanism: After each scheduling cycle, the predicted SOC and SOH values ​​of the model within that cycle are recorded along with the corresponding values ​​actually collected by the system, forming an error comparison sample. .

[0060] S3-2. Online Correction Mechanism: A gradient descent-based model weight update strategy is adopted to incrementally adjust the model parameters. As shown below:

[0061] in, η L is the learning rate, and L is the loss function (such as mean squared error). θ This represents the trainable parameters.

[0062] S3-3. Closed-loop scheduling structure: The above correction steps are automatically executed after each round of prediction tasks, so that the model parameters can be dynamically adjusted over time, forming a closed-loop process of prediction-scheduling-feedback-correction, which significantly enhances the model's adaptability to non-stationary factors such as battery performance degradation and changes in the operating environment.

[0063] This embodiment also provides a cross-scale joint estimation system for executing the steps of the cross-scale SOC and SOH joint estimation method for dynamic power scheduling of large-scale energy storage systems provided in the above embodiments. The system includes: a data acquisition and preprocessing module, a priori modeling module, a coupled interactive modeling module, a two-stage training and physical consistency constraint module, an uncertainty estimation module, an online feedback correction module, and an application interface and scheduling integration module.

[0064] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0065] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 Devices that specify the functions in one or more boxes. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] Compared with existing technologies, this invention has significant advantages in prediction accuracy, uncertainty quantification, model adaptability and deployment efficiency, and is particularly suitable for battery state perception and scheduling optimization in energy storage systems.

Claims

1. A cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems, characterized in that, Includes the following steps: The operating data of the energy storage system is acquired, including operating condition data, historical SOC sequence and historical SOH sequence. The operating condition data is rearranged in charge-discharge cycle order and constructed with the historical SOC sequence and historical SOH sequence to form a corresponding first dataset. Based on Bayesian-optimized extreme gradient boosting trees, a first regression model for predicting SOC and a second regression model for predicting SOH are constructed. The operating condition data are input into the first regression model and the second regression model respectively to output the predicted SOC sequence and the predicted SOH sequence. The operating condition data and the corresponding predicted SOC sequence and predicted SOH sequence are combined to form the second dataset. Construct a bidirectional interactive neural network, including a main feature encoding module and a coupled interactive modeling module; The main feature encoding module is used to extract multimodal features from the input running data to generate a joint embedding representation; The coupled interactive modeling module includes a SOC-dominant path unit, a SOH-dominant path unit, and a data augmentation unit. The SOC-dominant path unit takes historical SOC sequences and charge / discharge behavior as input to predict SOH results. The SOH-dominant path unit takes historical SOH sequences and operating condition data and predicts SOC results by back-calculating SOH back to SOC. The data augmentation unit is used to take the output of one dominant path unit as the self-attention input of another dominant path unit to adjust the conditions of the other dominant path unit. The bidirectional interactive neural network was initially trained using the second dataset, and then trained a second time using the first dataset to obtain a predictive model for predicting SOH and SOC. The predictive model is deployed at the operation control terminal of the energy storage system, and the current SOH and SOC are predicted based on the collected operating data.

2. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The operating data includes voltage, current, temperature, charge / discharge quantity, and number of cycles.

3. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The bidirectional interactive neural network is constructed based on a cooperative Transformer structure. It simulates the dynamic feedback relationship between the two through residual coupling blocks and integrates the nonlinear mapping between charge-capacity and impedance-health in the battery degradation evolution process.

4. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, During the second training process, a joint loss function was used to fine-tune the initial model.

5. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 4, characterized in that, The expression for the joint loss function is as follows: ;in, For the total loss, and Weighting coefficients set by humans. Weak labels for the outputs of the first regression model for predicting SOC and the second regression model for predicting SOH. , Losses caused by guidance The loss is guided by the true SOC and SOH labels. The loss of physical coupling consistency is constructed based on the known thermodynamic and electrochemical relationships of the battery.

6. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The expression for the SOC dominant path unit is as follows: ;in, : This represents the charge and discharge charge for that cycle; This represents the depth of discharge during that cycle; k 1, k 2, and These are preset empirical parameters.

7. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The expression for the SOH dominant path unit is as follows: ； in, I Represents a current signal. C nominal This indicates the nominal battery capacity.

8. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The prediction model also includes uncertainty estimation based on the Monte Carlo Dropout method before outputting the prediction results.

9. The cross-scale joint estimation method for dynamic power dispatching of large-scale energy storage systems according to claim 1, characterized in that, The parameters of the prediction model deployed in the energy storage system are iteratively updated using a lightweight error backpropagation mechanism and a gradient constraint update strategy.

10. A cross-scale joint estimation system, characterized in that, The steps are for performing the cross-scale SOC and SOH joint estimation method for dynamic power dispatch of large-scale energy storage systems as described in any one of claims 1-9.

Images