A power distribution method and system for an energy storage device

Abstract

The application provides a power distribution method and system of an energy storage device, the method comprising: identifying a current operating condition based on power grid frequency data, electricity price signal data and system power demand change rate; dynamically adjusting weight coefficients corresponding to each optimization target according to the identified operating condition to obtain an operating condition adaptive weight; inputting health state data, temperature data and cycle count data of each energy storage unit into a preset nonlinear health mapping network to obtain health state estimated values and high-dimensional hidden state vectors of each energy storage unit; inputting index type features, image type features, the operating condition adaptive weight and the corrected health state estimated values into a multi-model integration unit to obtain power distribution weights of each energy storage unit; and distributing power to each energy storage unit according to the power distribution weights. Through operating condition adaptive weight adjustment and dynamic switching of optimization targets, the power distribution strategy can adapt to different operating scenarios such as power grid frequency fluctuation and electricity price change.

Description

Technical Field

[0001] This invention relates to the field of energy management technology for electrochemical energy storage systems, and specifically to a power distribution method and system for energy storage devices. Background Technology

[0002] With the continuous expansion of renewable energy grid connection and the increasing demand for power system flexibility, electrochemical energy storage technology has become a key infrastructure supporting the safe and stable operation of the power grid. Currently, the power allocation of energy storage systems generally adopts a proportional allocation method based on state of charge or rated capacity, while some advanced systems introduce simple health state estimation as a correction factor.

[0003] However, existing feature extraction methods struggle to effectively depict the full picture of energy storage unit operation characteristics, lacking a systematic representation of power fluctuation patterns, operational dynamics, and external factors, resulting in insufficient information for power allocation decisions. Fixed optimization strategies cannot adapt to the switching needs of multiple scenarios such as peak shaving, demand response, and frequency support. They lag in response under conditions such as grid frequency exceeding limits or drastic electricity price fluctuations, failing to dynamically adjust according to real-time operating conditions such as grid frequency, electricity price signals, and power demand changes, making it difficult to adapt to complex and ever-changing grid operation scenarios. Traditional health status assessment models make strong linear assumptions, making it difficult to capture the nonlinear effects of multiple factors such as temperature and power ratio coupling, and they do not consider the temporal evolution trend of health status, easily causing assessment results to deviate from the actual aging state, leading to power allocation imbalance and accelerating the overall performance degradation of the energy storage system. Therefore, this invention studies and designs a power allocation method and system for energy storage devices. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the above-mentioned defects, thereby providing a power distribution method and system for energy storage devices.

[0005] To address the above problems, the present invention provides a power distribution method for an energy storage device, comprising the following steps:

[0006] Acquire operational data, grid frequency data, and electricity price signal data from each energy storage unit in the energy storage device;

[0007] Based on the operational data, operational morphology features, power fluctuation features, and operational correlation features are extracted to construct index-type features; and typical operational curves are converted into graphical features.

[0008] Based on the power grid frequency data, electricity price signal data, and system power demand change rate, the current operating condition is identified;

[0009] Based on the identified operating conditions, the weight coefficients corresponding to each optimization objective are dynamically adjusted to obtain the operating condition adaptive weights.

[0010] The health status data, temperature data, and cycle count data of each energy storage unit are input into a preset nonlinear health mapping network to obtain the health status estimate and high-dimensional hidden state vector of each energy storage unit.

[0011] The health status estimate and the high-dimensional hidden state vector are input into the second long short-term memory network to obtain the health status evolution rate of each energy storage unit. A composite loss function is constructed based on allocation error loss, evolution consistency loss and monotonicity constraint loss to correct the mapping result.

[0012] The index-type features, image-based features, operating condition adaptive weights, and corrected health status estimates are input into the multi-model integration unit to obtain the power allocation weights of each energy storage unit.

[0013] Power is allocated to each energy storage unit according to the power allocation weight.

[0014] Preferably, the extraction of operational morphological features includes:

[0015] Based on the typical operating curve of the energy storage unit, calculate the daily average power, peak-to-average power ratio, and depth of charge / discharge ratio.

[0016] Divide the day into several time periods and calculate the average power of each time period in the typical operating curve;

[0017] Based on historical operating data, calculate the maximum and minimum average power for each time period, and calculate the first difference ratio between the maximum average power for each time period and the average power for the corresponding time period of the typical operating curve, as well as the second difference ratio between the minimum average power for each time period and the average power for the corresponding time period of the typical operating curve.

[0018] The daily average power, peak-to-average power ratio, charge-to-discharge depth ratio, average power for each time period, first difference ratio, and second difference ratio constitute the operating characteristics.

[0019] Preferably, the extraction of power fluctuation features includes:

[0020] For each time period, calculate the standard deviation of instantaneous power within that time period, and average it over all operating cycles to obtain the intraday standard deviation of fluctuation for that time period;

[0021] For each time period, calculate the power standard deviation at the same moment in different operating cycles, and average it over all moments in the time period to obtain the cross-cycle stability index for that time period.

[0022] The power fluctuation characteristics are composed of the intraday standard deviation of fluctuations and the cross-cycle stability index for each time period.

[0023] Preferably, the extraction of operational relevance features includes:

[0024] Calculate the correlation coefficient between the output power of the energy storage unit and the grid frequency to obtain the first correlation coefficient;

[0025] Calculate the correlation coefficient between the output power of the energy storage unit and the electricity price signal to obtain the second correlation coefficient;

[0026] Separate the weekday and holiday operation data, construct typical weekday operation curves and typical holiday operation curves respectively, calculate the correlation coefficient between the two, and obtain the third correlation coefficient;

[0027] The first correlation coefficient, the second correlation coefficient, and the third correlation coefficient are combined to form the operational correlation feature.

[0028] Preferably, the step of converting typical operating curves into graphical features includes:

[0029] An improved recursive graph is constructed based on the typical operating curve, serving as the red channel of the three-channel color image;

[0030] The typical running curve is scaled to a preset range, and an angle difference sine matrix is ​​generated by the Gram differential angle field algorithm as the green channel.

[0031] The typical running curve is discretized into multiple quantile states. The first-order transition probabilities between states are statistically analyzed, and a Markov transition field matrix is ​​constructed as the blue channel.

[0032] Based on the red, green, and blue channels, a three-channel color image feature is generated.

[0033] Preferably, identifying the current operating condition includes:

[0034] The input feature vector is constructed based on grid frequency fluctuations, electricity price signals, and the rate of change in system power demand.

[0035] A fuzzy inference system is used to process the input feature vector and output the recognition result of the current operating condition.

[0036] The operating conditions include at least one of peak shaving and valley filling mode, demand response mode, and emergency backup power mode.

[0037] Preferably, the nonlinear health mapping network includes:

[0038] The feature segmenter is used to assign independent weight vectors and bias vectors to each input feature in the health status data, temperature data, and cycle count data of each energy storage unit, and to project the data of different physical attributes onto a unified high-dimensional semantic space to obtain feature embedding vectors.

[0039] The deep learning module, which includes a multi-head attention mechanism and a feedforward neural network, is used to perform global feature interaction processing on the feature embedding vector to obtain an updated feature vector.

[0040] A first long short-term memory network is used to perform time-series prediction processing on the updated feature vector to obtain the health status estimate and high-dimensional hidden state vector of each energy storage unit.

[0041] Preferably, constructing the composite loss function includes:

[0042] Construct an allocation error loss to quantify the deviation between the power allocation value and the system demand;

[0043] An evolutionary consistency loss is constructed to constrain the predicted evolution rate of healthy states to remain consistent with the actual evolution rate.

[0044] A monotonic constraint loss is constructed to penalize power allocation decisions that violate the physical law of monotonically decreasing health status.

[0045] The weighted sum of the allocation error loss, evolution consistency loss, and monotonicity constraint loss yields the composite loss function.

[0046] Preferably, the multi-model integration unit includes:

[0047] Multiple base learners, including at least a linear mapping model based on health status, a nonlinear mapping model based on health status, a support vector machine model based on power fluctuation features, and a random forest model based on operational morphology features;

[0048] The meta-learner, employing a decision tree model, is used to integrate and fuse the outputs of each base learner.

[0049] In this process, each base learner uses a cross-training method to obtain the out-of-fold prediction results, and the meta-learner is trained using the out-of-fold prediction results of each base learner as input.

[0050] The present invention also provides a power distribution system for an energy storage device, comprising the power distribution method for the energy storage device described in any of the preceding claims, including:

[0051] The data acquisition module is used to acquire the operating data of each energy storage unit in the energy storage device, grid frequency data, and electricity price signal data;

[0052] The feature construction module is used to extract operational morphology features, power fluctuation features, and operational correlation features based on the operational data, construct index-type features, and convert typical operational curves into graphical features.

[0053] The operating condition identification module is used to identify the current operating condition based on the power grid frequency data, electricity price signal data, and system power demand change rate.

[0054] The dynamic weight adjustment module is used to dynamically adjust the weight coefficients corresponding to each optimization objective based on the identified operating conditions, so as to obtain the operating condition adaptive weights.

[0055] The nonlinear health mapping module, which includes a feature segmenter, a deep learning module and a first long short-term memory network, is used to map the health status data of each energy storage unit into a health status estimate and a high-dimensional hidden state vector.

[0056] The evolution rate constraint module includes a second long short-term memory network to predict the evolution rate of the health state and corrects the mapping results based on a composite loss function.

[0057] The multi-model ensemble module contains multiple base learners and meta-learners, used to determine the power allocation weights of each energy storage unit;

[0058] The power allocation execution module is used to allocate power to each energy storage unit according to the power allocation weight.

[0059] The power distribution method and system for energy storage devices provided by this invention have the following beneficial effects:

[0060] 1. This invention acquires comprehensive information about the operation of an energy storage system through multi-source data acquisition, and then performs feature extraction and operating condition identification in parallel. The feature extraction branch transforms the raw operating data into index-type features and image-based features, while the operating condition identification branch determines the current operating mode and generates adaptive weights. Simultaneously, a nonlinear health mapping network performs in-depth estimation of the health status of the energy storage units, and an evolution rate constraint module corrects the health assessment results and provides aging trend information. A multi-model integration unit integrates all the above information and outputs the power allocation weights for each unit. Finally, the execution module completes the power allocation. The entire process forms a closed-loop power allocation mechanism that is data-driven, operating condition adaptive, and health-aware.

[0061] 2. This invention also achieves a comprehensive characterization of the energy storage unit's operating characteristics through multi-dimensional feature extraction, including operating mode features, power fluctuation features, operating correlation features, and image features; and achieves dynamic switching of optimization targets through adaptive weight adjustment of operating conditions, enabling the power allocation strategy to adapt to different operating scenarios such as grid frequency fluctuations and electricity price changes.

[0062] 3. This invention also achieves decoupling of functions such as data acquisition, feature construction, operating condition identification, health mapping, and power allocation through modular architecture design, thereby reducing system complexity and improving maintainability and scalability. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0064] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0065] like Figure 1 As shown, the present invention provides a power distribution method and system for an energy storage device, which includes the following steps:

[0066] The system acquires operational data, grid frequency data, and electricity price signal data from each energy storage unit within the energy storage device. The operational data includes the voltage, current, power, state of charge, temperature, and historical charge / discharge records of each energy storage unit. The sampling frequency is adjustable from 1 second to 1 minute and is collected by the battery management unit and transmitted to the central controller via the communication bus. The grid frequency data is acquired through a synchronous phasor measurement device with a sampling accuracy of 0.001 Hz. The electricity price signal data comes from the electricity market trading system and includes real-time electricity prices, time-of-use electricity prices, and demand response price signals.

[0067] Based on the operational data, operational morphology features, power fluctuation features, and operational correlation features are extracted to construct index-type features; and typical operational curves are converted into graphical features; operational morphology features reflect the macroscopic operational mode of the energy storage unit, power fluctuation features depict the statistical characteristics of power changes, and operational correlation features reveal the degree of correlation between the energy storage unit and external power grid and market signals; graphical features convert one-dimensional time series into two-dimensional images, preserving the local correlation and global pattern of time series data.

[0068] Based on the power grid frequency data, electricity price signal data, and system power demand change rate, the current operating condition is identified; the system power demand change rate is obtained by calculating the difference between the total system power demand at the current moment and the previous moment and dividing it by the time interval; the operating conditions include peak shaving and valley filling mode, demand response mode, and emergency backup power mode.

[0069] Based on the identified operating conditions, the weight coefficients corresponding to each optimization objective are dynamically adjusted to obtain the operating condition adaptive weights. The optimization objectives include economy, response speed, cycle life, etc., and different operating conditions correspond to different weight configurations.

[0070] The health status data, temperature data, and cycle count data of each energy storage unit are input into a pre-defined nonlinear health mapping network to obtain the health status estimate and high-dimensional hidden state vector of each energy storage unit. The nonlinear health mapping network includes a feature segmenter, a deep learning (Transformer) module, and a first long short-term memory network. The feature segmenter converts multi-source heterogeneous data into a unified-dimensional embedding vector. The deep learning module captures long-distance dependencies between features through a multi-head self-attention mechanism. The deep learning module contains a 4- to 8-layer encoder structure, with 8 to 16 attention heads configured in each layer. The dimension of the feedforward network hidden layer is 4 times the dimension of the embedding vector. The first long short-term memory network is used to model the temporal evolution of the health status. The first long short-term memory network contains 2 to 3 layers of LSTM units, with 64 to 256 hidden units in each layer. It is used to model the temporal evolution of the health status and outputs the health status estimate and the high-dimensional hidden state vector. The dimension of the high-dimensional hidden state vector is set to 32 to 128 dimensions.

[0071] The estimated health status and the high-dimensional hidden state vector are input into the second long short-term memory network to obtain the health status evolution rate of each energy storage unit. A composite loss function is constructed based on allocation error loss, evolution consistency loss, and monotonicity constraint loss to correct the mapping result. The allocation error loss quantifies the deviation between the power allocation value and the system demand, the evolution consistency loss constrains the predicted health status evolution rate to be consistent with the actual evolution rate, and the monotonicity constraint loss ensures that the health status decreases monotonically with the number of cycles. The structure of the second long short-term memory network is similar to that of the first long short-term memory network, but the input is the concatenation result of the estimated health status and the high-dimensional hidden state vector, and the output is the health status evolution rate, that is, the change in health status per unit time.

[0072] The index-type features, image features, operating condition adaptive weights, and corrected health status estimates are input into the multi-model integration unit to obtain the power allocation weights of each energy storage unit. The multi-model integration unit adopts a stacking integration strategy. The base learners include linear regression models, support vector machines, random forests, and gradient boosting trees. The meta-learners adopt logistic regression or lightweight neural networks. Each base learner obtains out-of-bounds predictions through cross-validation. The meta-learner learns the optimal combined weights based on these predictions.

[0073] Power is allocated to each energy storage unit according to the power allocation weight; the power allocation command is sent to the converter of each unit through the controller local area network bus or Ethernet communication, and the converter adjusts the charging and discharging power according to the received power set value to realize closed-loop control.

[0074] Specifically, comprehensive information on the operation of the energy storage system is acquired through multi-source data acquisition, followed by parallel feature extraction and operating condition identification. The feature extraction branch transforms the raw operating data into index-type features and image-based features, while the operating condition identification branch determines the current operating mode and generates adaptive weights. Simultaneously, a nonlinear health mapping network performs in-depth estimation of the health status of the energy storage units, and the evolution rate constraint module corrects the health assessment results and provides aging trend information. The multi-model integration unit integrates all the above information and outputs the power allocation weights for each unit, which are finally completed by the execution module. The entire process forms a closed-loop power allocation mechanism that is data-driven, operating condition adaptive, and health-aware.

[0075] In some implementations, the extraction of operational characteristics includes: calculating the daily average power, peak-to-average power ratio, and charge / discharge depth ratio based on the typical operating curve of the energy storage unit; dividing the day into several time periods and calculating the average power of each time period in the typical operating curve; calculating the maximum and minimum values ​​of the average power for each time period based on historical operating data, and calculating the first difference ratio between the maximum average power for each time period and the average power for the corresponding time period of the typical operating curve, and the second difference ratio between the minimum average power for each time period and the average power for the corresponding time period of the typical operating curve; and combining the daily average power, peak-to-average power ratio, charge / discharge depth ratio, average power for each time period, first difference ratio, and second difference ratio to form operational characteristics.

[0076] Specifically, the daily average power is the arithmetic mean of all sampling points on the typical operating curve over 24 hours; the peak-to-average power ratio is the ratio of the absolute value of the maximum power on the typical operating curve to the absolute value of the daily average power, reflecting the severity of power fluctuations; the depth of charge / discharge ratio is the proportion of times the depth of discharge exceeds 50% of the rated capacity on the typical operating curve, or the ratio of the average depth of discharge to the rated capacity, reflecting the utilization intensity of the energy storage unit; the time period division can be set to 4 to 12 time periods according to the grid load characteristics. The first difference ratio and the second difference ratio reflect the degree of deviation between the current typical operating mode and historical extreme cases, used to assess the boundary of operational risk; among them, the operational morphology characteristics are constructed through a hierarchical statistical method; at the macro level, the daily average power, peak-to-average power ratio, and depth of charge / discharge ratio describe the overall operational intensity; at the meso level, the time-period average power characterizes the intraday operating mode; at the micro level, the first difference ratio and the second difference ratio quantify the deviation from historical extremes; the combination of multi-scale features comprehensively reflects the operational morphology of the energy storage unit, providing a basis for subsequent power allocation decisions.

[0077] In some implementations, the extraction of power fluctuation characteristics includes: for each time period, calculating the standard deviation of instantaneous power within that time period and averaging it over all operating cycles to obtain the intraday fluctuation standard deviation for that time period; for each time period, calculating the power standard deviation at the same moment in different operating cycles and averaging it over all moments within that time period to obtain the cross-cycle stability index for that time period; and combining the intraday fluctuation standard deviation and the cross-cycle stability index for each time period to form power fluctuation characteristics.

[0078] Specifically, the intraday fluctuation standard deviation reflects the intensity of the instantaneous power surge experienced by the energy storage unit during that period. The standard deviation of these power data is calculated to obtain the degree of fluctuation within a single operating cycle. Then, for the same period of the most recent N operating cycles, such as the most recent 30 days, the average of the standard deviations is calculated as the intraday fluctuation standard deviation for that period. This indicator reflects the intensity of the instantaneous power surge experienced by the energy storage unit during that period. The cross-cycle stability index reflects the repeatability and predictability of the power pattern across different operating cycles. The smaller the index value, the more stable the power pattern and the easier it is to plan power allocation. The time period division is consistent with the aforementioned time period division to ensure spatiotemporal alignment between features. The dimension of the power fluctuation feature is twice the number of time periods, with each time period corresponding to two feature values: the intraday fluctuation standard deviation and the cross-cycle stability index.

[0079] In some implementations, the extraction of operational correlation features includes: calculating the correlation coefficient between the output power of the energy storage unit and the grid frequency to obtain a first correlation coefficient; calculating the correlation coefficient between the output power of the energy storage unit and the electricity price signal to obtain a second correlation coefficient; separating weekday and holiday operation data, constructing typical weekday operation curves and typical holiday operation curves respectively, calculating the correlation coefficient between the two to obtain a third correlation coefficient; and combining the first correlation coefficient, the second correlation coefficient, and the third correlation coefficient to form operational correlation features.

[0080] Specifically, the first correlation coefficient reflects the degree to which the energy storage unit participates in grid frequency regulation. A larger absolute value indicates a more significant frequency response characteristic. The first correlation coefficient is calculated using the Pearson correlation coefficient method, with the formula being the product of the covariance of the energy storage unit's output power and the grid frequency divided by the standard deviation of both. The calculation time window selects the operating data from the most recent 7 to 30 days, and the sampling interval is consistent with the power data sampling interval. The second correlation coefficient reflects the degree to which the energy storage unit responds to electricity price signals. A positive coefficient indicates a tendency to discharge during high electricity prices and charge during low electricity prices. The second correlation coefficient also uses the Pearson correlation coefficient. The third correlation coefficient reflects... The similarity of operating modes between weekdays and holidays is assessed. A coefficient closer to 1 indicates a more stable mode, while a coefficient closer to 0 or a negative number indicates a greater difference in modes. The third correlation coefficient is calculated by first marking historical operating data as weekday data or holiday data based on calendar information, and then averaging the two types of data by time to obtain typical operating curves for weekdays and holidays. Both curves are vectors with a length of 24 hours multiplied by the number of sampling points per hour. The Pearson correlation coefficient between these two vectors is calculated. Correlation analysis is established from three dimensions: grid interaction, market response, and scenario differences. The combination of these three factors comprehensively characterizes the operational characteristics of the energy storage unit.

[0081] In some implementations, converting the typical operating curve into image features includes: constructing an improved recursive graph based on the typical operating curve, serving as the red channel of a three-channel color image; scaling the typical operating curve to a preset interval and generating an angle difference sine matrix using the Gram differential angle field algorithm, serving as the green channel; discretizing the typical operating curve into multiple quantile states, statistically analyzing the first-order transition probabilities between states, and constructing a Markov transition field matrix, serving as the blue channel; and combining the red, green, and blue channels to generate three-channel color image features.

[0082] Specifically, the improved recursive graph introduces an exponential weighting mechanism to assign higher weights to nearest points in distance calculation, highlighting the repetitive patterns and periodic structures of the running state. The specific steps are as follows: normalize the typical running curve of length L to the interval between 0 and 1; construct an L×L recursive matrix, with matrix elements R. ijThe similarity between the states at time i and time j is calculated using a Gaussian kernel function, with the similarity being exp(-squared distance divided by twice the squared bandwidth), where the bandwidth is automatically determined using the Silverman rule. A thresholding process is applied to the recursion matrix, retaining the top 10% to 20% of similarity elements and setting the rest to zero, resulting in a sparse recursive graph. Morphological dilation is then performed on the recursive graph to enhance connected regions, which serve as the red channel, highlighting the repetitive patterns and periodic structures of the running states. The Gram difference angle field maps the time series to a polar coordinate system using trigonometric functions, calculating the sine value of the angle difference to preserve temporal dependence and amplitude variation information. Specifically, the time series X is converted to polar coordinates, with the radius r being the time index divided by the sequence length, the angle φ being arccos(X), and X being the normalized amplitude. The Gram difference angle field matrix G is constructed, with matrix elements G... ij sin(φ) i –φ j The range of matrix G is linearly mapped from -1 to 1 to 0 to 255, serving as the green channel. The green channel preserves the temporal dependence and amplitude variation information of the time series. The Markov transition field reflects the transition law and dynamic evolution characteristics of the operating state through quantile state division and state transition probability statistics. The specific steps are: calculating the quantile boundaries of the time series, discretizing the series into Q state labels; constructing a Q×Q transition frequency matrix, with elements M ab The frequency matrix represents the number of transitions from state a to state b. The frequency matrix is ​​normalized to a probability matrix with a row sum of 1. The Q×Q matrix is ​​expanded to an L×L matrix using bilinear interpolation, serving as the blue channel. The blue channel reflects the transition patterns and dynamic evolution characteristics of the running states. Regarding the three-channel combination, the red, green, and blue channels are superimposed in RGB format to generate an L×L color image. The RGB values ​​of each pixel in the image correspond to three different temporal feature representations, and the convolutional neural network can learn to capture the feature correlations between different channels.

[0083] In some implementations, identifying the current operating condition includes: constructing an input feature vector based on grid frequency fluctuations, electricity price signals, and the rate of change of system power demand; processing the input feature vector using a fuzzy inference system and outputting the identification result of the current operating condition; wherein the operating condition includes at least one of peak shaving and valley filling mode, demand response mode, and emergency backup power mode.

[0084] Specifically, the power grid frequency fluctuation characteristics include the absolute value of frequency deviation and the rate of frequency change; the electricity price signal characteristics include the relative value of electricity price level and the rate of electricity price change; the system power demand change rate is obtained by calculating the difference between the total power demand at the current moment and the previous moment and dividing it by the time interval. These features are then normalized and concatenated into a 3D input feature vector. The fuzzy inference system adopts a Mamdani-type fuzzy inference structure, comprising three modules: fuzzification, rule-based inference, and defuzzification. The fuzzification module defines the fuzzy sets of the input variables: the absolute value of frequency deviation is divided into three fuzzy sets: small, medium, and large, with membership functions using triangular or Gaussian types; the rate of frequency change is divided into negative large and negative small. The system uses five fuzzy sets: zero, positive small, and positive large; the relative value of electricity price level is divided into three fuzzy sets: low, medium, and high; the rate of change of electricity price is divided into three fuzzy sets: decreasing, stable, and increasing; the rate of change of power demand is divided into three fuzzy sets: decreasing, stable, and increasing; the rule base contains several IF-THEN rules, for example: if the frequency deviation is large and the frequency change rate is positive, the operating condition is emergency backup mode; if the electricity price level is high and the rate of change of electricity price is increasing, the operating condition is demand response mode; if the frequency deviation is small and the electricity price is stable, the operating condition is peak shaving and valley filling mode. Rule reasoning uses the minimum-maximum synthesis method, defuzzification uses the centroid method, and the output is the membership degree or clear classification result of each operating condition.

[0085] In some embodiments, the nonlinear health mapping network includes: a feature segmenter, used to assign independent weight vectors and bias vectors to each input feature in the health status data, temperature data, and cycle count data of each energy storage unit, projecting data with different physical attributes onto a unified high-dimensional semantic space to obtain feature embedding vectors; a deep learning module, including a multi-head attention mechanism and a feedforward neural network, used to perform global feature interaction processing on the feature embedding vectors to obtain updated feature vectors; and a first long short-term memory network, used to perform temporal prediction processing on the updated feature vectors to obtain health status estimates and high-dimensional hidden state vectors for each energy storage unit.

[0086] Specifically, the feature segmenter projects features with different physical dimensions into a unified dimensional space by assigning independent learnable weights and biases to each feature. The transformation formula is: the embedding vector is equal to the weight matrix multiplied by the input feature value plus the bias vector, and then processed by layer normalization and ReLU activation function.

[0087] The Transformer module employs an encoder structure, comprising 4 to 8 layers. Each layer includes a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention mechanism calculates the correlation weights between features, achieving global feature interaction. The first Long Short-Term Memory (LSTM) network uses a 2- to 3-layer stacked structure, with each layer containing 64 to 256 LSTM units. Information flow is controlled through input gates, forget gates, output gates, and cell states, effectively modeling long-term temporal dependencies. For network training, mean squared error is used as the basic loss function, and the Adam optimizer is employed. The initial learning rate is set to 0.001, and cosine annealing is used for adjustment. Training data includes historical running data and laboratory accelerated aging test data. Transfer learning is used to improve generalization ability. A feature segmenter eliminates the dimensional differences of multi-source heterogeneous data, achieving a unified spatial projection. Then, the global attention mechanism of the Transformer module captures complex nonlinear relationships between features, such as the coupled influence of temperature and charge / discharge rate on health status. Finally, the LSTM network models the temporal evolution of health status, outputting the current health status estimate and deep implicit information. The three-layer structure has a clear division of labor, realizing a complete mapping process from data fusion and feature interaction to time series modeling.

[0088] In some implementations, constructing the composite loss function includes: constructing an allocation error loss to quantify the deviation between the power allocation value and the system demand; constructing an evolution consistency loss to constrain the predicted health state evolution rate to be consistent with the actual evolution rate; constructing a monotonicity constraint loss to penalize power allocation decisions that violate the physical law of monotonically decreasing health states; and weighted summing the allocation error loss, evolution consistency loss, and monotonicity constraint loss to obtain the composite loss function.

[0089] Specifically, the allocation error loss takes the form of absolute error or relative error to ensure that the total power allocation meets the system requirements. The calculation method is: allocation error loss equals the system required power minus the absolute value of the sum of the allocated power of each unit, then divided by the system required power. Evolutionary consistency loss is defined as the mean square error between the predicted evolution rate and the actual evolution rate, constraining the accuracy of predicting the health state evolution trend. First, the health state evolution rate is predicted based on the second long short-term memory network, i.e., the change in health state per unit time. Simultaneously, the actual evolution rate is calculated using actual observation data, i.e., the difference between the current health state estimate and the previous estimate divided by the time interval. Evolutionary consistency loss is defined as the mean square error between the predicted evolution rate and the actual evolution rate. Error loss can be mitigated by using Huber loss to reduce the impact of outliers; monotonic constraint loss is constructed based on the physical law that the health state monotonically decreases with the number of iterations. It is implemented as follows: for any two times t1 and t2, if the number of iterations at time t2 is greater than that at time t1, then the health state at time t2 should be less than or equal to that at time t1. The monotonic constraint loss is defined as the sum of the penalty terms for all sample pairs that violate this constraint. The penalty function uses the ReLU function, that is, the loss value is max(0, health state t2 minus health state t1); the composite loss function is a weighted sum of the above three losses. The weight coefficients can be set fixedly or dynamically adjusted, such as focusing on allocating error loss in the early stage of training and increasing the weight of physical constraint loss in the later stage.

[0090] In some embodiments, the multi-model ensemble unit includes: multiple base learners, including at least a linear mapping model based on health status, a nonlinear mapping model based on health status, a support vector machine model based on power fluctuation features, and a random forest model based on operational morphology features; and a meta-learner, employing a decision tree model, used to integrate and fuse the outputs of each base learner; wherein each base learner obtains the extrapolation prediction results through cross-training, and the meta-learner is trained using the extrapolation prediction results of each base learner as input.

[0091] Specifically, the linear mapping model based on health status employs a multiple linear regression structure, with the health status estimate as input and the power allocation ratio as output. The nonlinear mapping model based on health status uses a feedforward neural network structure, containing 2 to 3 hidden layers with 32 to 128 neurons per layer. The activation function is ReLU or Sigmoid, and the output layer uses a Softmax function to ensure the sum of the allocation weights is 1, capturing the nonlinear relationship between health status and power allocation. The support vector machine model based on power fluctuation features uses a support vector regression structure, with power fluctuation features as input and a kernel function used to establish... The mapping relationship between power fluctuation characteristics and weight allocation; the random forest model based on operational morphology features adopts an ensemble decision tree structure, with operational morphology features as input, capturing the complex nonlinear relationship between operational morphology and power allocation; cross-training uses K-fold cross-validation, where K is 5 to 10, dividing the training data into K subsets, using K-1 subsets to train the base learner each time, and making predictions on the remaining 1 subset to obtain the out-of-fold prediction results; the meta-learner adopts a decision tree model, such as CART decision tree, or a logistic regression model, and uses the out-of-fold prediction results of each base learner as input to learn the optimal combination strategy.

[0092] This application achieves a comprehensive characterization of the energy storage unit's operating characteristics through multi-dimensional feature extraction, including operating mode features, power fluctuation features, operating correlation features, and image features; and realizes dynamic switching of optimization targets through adaptive weight adjustment of operating conditions, enabling the power allocation strategy to adapt to different operating scenarios such as grid frequency fluctuations and electricity price changes.

[0093] This application achieves accurate estimation of battery health status through a nonlinear health mapping network, including a feature segmenter, a deep learning module, and a long short-term memory network. It captures the nonlinear characteristics and temporal dependence of battery aging, outperforming traditional linear models. By using evolution rate constraints and a composite loss function, it ensures that the health status estimation results conform to physical laws and avoids erroneous decisions caused by abnormal estimations.

[0094] like Figure 2 As shown, the present invention also provides a power distribution system for an energy storage device, including the power distribution method for the energy storage device described in any of the preceding claims, comprising:

[0095] The data acquisition module is used to acquire operating data, grid frequency data, and electricity price signal data of each energy storage unit in the energy storage device. The hardware components include a battery management unit, a synchronization phasor measurement device, and a communication interface unit. The battery management unit collects operating data of the energy storage units through voltage and current sensors, with a sampling frequency adjustable from 1Hz to 1kHz, and communicates with the central controller through the controller area network bus. The synchronization phasor measurement device uses the Global Positioning System for time synchronization to ensure that the time synchronization accuracy of frequency measurement is better than 1 microsecond. The communication interface unit supports multiple communication protocols such as Modbus and IEC61850 to realize data interaction with the grid dispatching system and the power trading system.

[0096] The feature construction module is used to extract operational morphology features, power fluctuation features, and operational correlation features based on the operational data, construct index-type features, and convert typical operational curves into graphical features. An edge computing architecture is adopted, with feature extraction completed in the local controller to reduce data transmission latency. The calculation of operational morphology features, power fluctuation features, and operational correlation features is implemented in a digital signal processor or embedded processor, and the generation of graphical features is accelerated by a graphics processor. The feature data is compressed and then uploaded to the central processing unit.

[0097] The operating condition identification module is used to identify the current operating condition based on the power grid frequency data, electricity price signal data, and system power demand change rate; the rule base of the fuzzy inference system supports online updates and can adjust rule parameters according to power grid operation experience or scheduling strategies; the operating condition identification module outputs the operating condition identification result to the dynamic weight adjustment module to trigger the weight coefficient update.

[0098] The dynamic weight adjustment module dynamically adjusts the weight coefficients corresponding to each optimization objective based on the identified operating conditions, obtaining adaptive weights for each condition. Weight mapping is implemented using a lookup table or a lightweight neural network. Optimized weight combinations for different operating conditions are pre-stored, and the identification results serve as the lookup address, outputting the corresponding weight vector in real time. The weight update cycle is synchronized with the operating condition identification cycle, ranging from 100 milliseconds to 1 second.

[0099] The nonlinear health mapping module, comprising a feature segmenter, a Transformer module, and a first long short-term memory network, is used to map the health status data of each energy storage unit into a health status estimate and a high-dimensional latent state vector. It employs an industrial control computer or server, with the feature segmenter, Transformer module, and first long short-term memory network deployed in software, supporting online updates of model parameters. The nonlinear health mapping module takes real-time monitoring data from each energy storage unit as input and outputs a health status estimate and a high-dimensional latent state vector, with an update cycle of 1 to 15 minutes.

[0100] The evolution rate constraint module includes a second long short-term memory network, which is used to predict the evolution rate of health status and correct the mapping results based on a composite loss function. The second long short-term memory network shares some parameters with the first long short-term memory network or is trained independently. The input is a sequence of health status estimates and the output is the predicted evolution rate. The calculation of the composite loss function is performed during the training phase, and only forward inference is performed during the online operation phase.

[0101] The multi-model integration module contains multiple base learners and meta learners to determine the power allocation weights of each energy storage unit. The model files of the base learners and meta learners are stored in non-volatile memory and support hot updates. The module receives feature vectors output by the feature construction module and the nonlinear health mapping module, and outputs the power allocation weights of each energy storage unit. The calculation cycle is 1 second to 1 minute.

[0102] The power allocation execution module is used to allocate power to each energy storage unit according to the power allocation weight; it includes a power allocation instruction generation unit and a converter control interface. The instruction generation unit converts the allocation weight into a power setpoint, which is then sent after being processed by amplitude limiting and slope limiting. The converter control interface supports pulse width modulation signal output or communication protocol transmission, and the control period is 10 milliseconds to 100 milliseconds.

[0103] Specifically, the data acquisition module and the feature construction module are connected via a fieldbus. The feature construction module, the operating condition identification module, the nonlinear health mapping module, and the multi-model integration module are connected via a high-speed data bus. The dynamic weight adjustment module is connected to the multi-model integration module. The evolution rate constraint module and the nonlinear health mapping module form a feedback loop. The multi-model integration module and the power allocation execution module are connected via a control bus. Through modular architecture design, the functions of data acquisition, feature construction, operating condition identification, health mapping, and power allocation are decoupled, reducing system complexity and improving maintainability and scalability.

[0104] The above description is merely 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. The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the protection scope of the present invention.

Claims

1. A power distribution method for an energy storage device, characterized in that, Includes the following steps: Acquire operational data, grid frequency data, and electricity price signal data from each energy storage unit in the energy storage device; Based on the operational data, operational morphology features, power fluctuation features, and operational correlation features are extracted to construct index-type features; And convert typical operating curves into graphical features; Based on the power grid frequency data, electricity price signal data, and system power demand change rate, the current operating condition is identified; Based on the identified operating conditions, the weight coefficients corresponding to each optimization objective are dynamically adjusted to obtain the operating condition adaptive weights. The health status data, temperature data, and cycle count data of each energy storage unit are input into a preset nonlinear health mapping network to obtain the health status estimate and high-dimensional hidden state vector of each energy storage unit. The health status estimate and the high-dimensional hidden state vector are input into the second long short-term memory network to obtain the health status evolution rate of each energy storage unit. A composite loss function is constructed based on allocation error loss, evolution consistency loss and monotonicity constraint loss to correct the mapping result. The index-type features, image-based features, operating condition adaptive weights, and corrected health status estimates are input into the multi-model integration unit to obtain the power allocation weights of each energy storage unit. Power is allocated to each energy storage unit according to the power allocation weight; The extracted operational morphology features include: Based on the typical operating curve of the energy storage unit, calculate the daily average power, peak-to-average power ratio, and depth of charge / discharge ratio. Divide the day into several time periods and calculate the average power of each time period in the typical operating curve; Based on historical operating data, calculate the maximum and minimum average power for each time period, and calculate the first difference ratio between the maximum average power for each time period and the average power for the corresponding time period of the typical operating curve, as well as the second difference ratio between the minimum average power for each time period and the average power for the corresponding time period of the typical operating curve. The daily average power, peak-to-average power ratio, charge-discharge depth ratio, average power for each time period, first difference ratio, and second difference ratio constitute the operating morphology characteristics. The extracted power fluctuation features include: For each time period, calculate the standard deviation of instantaneous power within that time period, and average it over all operating cycles to obtain the intraday standard deviation of fluctuation for that time period; For each time period, calculate the power standard deviation at the same moment in different operating cycles, and average it over all moments in the time period to obtain the cross-cycle stability index for that time period. The power fluctuation characteristics are composed of the intraday standard deviation of fluctuations and the cross-cycle stability index for each time period. The extracted operational relevance features include: Calculate the correlation coefficient between the output power of the energy storage unit and the grid frequency to obtain the first correlation coefficient; Calculate the correlation coefficient between the output power of the energy storage unit and the electricity price signal to obtain the second correlation coefficient; Separate the weekday and holiday operation data, construct typical weekday operation curves and typical holiday operation curves respectively, calculate the correlation coefficient between the two, and obtain the third correlation coefficient; The first correlation coefficient, the second correlation coefficient, and the third correlation coefficient are combined to form the operational correlation feature; The nonlinear health mapping network includes: The feature segmenter is used to assign independent weight vectors and bias vectors to each input feature in the health status data, temperature data, and cycle count data of each energy storage unit, and to project the data of different physical attributes onto a unified high-dimensional semantic space to obtain feature embedding vectors. The deep learning module, which includes a multi-head attention mechanism and a feedforward neural network, is used to perform global feature interaction processing on the feature embedding vector to obtain an updated feature vector. A first long short-term memory network is used to perform time-series prediction processing on the updated feature vector to obtain the health status estimate and high-dimensional hidden state vector of each energy storage unit.

2. The power distribution method for the energy storage device according to claim 1, characterized in that, The process of converting typical operating curves into graphical features includes: An improved recursive graph is constructed based on the typical operating curve, serving as the red channel of the three-channel color image; The typical running curve is scaled to a preset range, and an angle difference sine matrix is ​​generated by the Gram differential angle field algorithm as the green channel. The typical running curve is discretized into multiple quantile states. The first-order transition probabilities between states are statistically analyzed, and a Markov transition field matrix is ​​constructed as the blue channel. Based on the red, green, and blue channels, a three-channel color image feature is generated.

3. The power distribution method for the energy storage device according to claim 1, characterized in that, The identification of the current operating condition includes: The input feature vector is constructed based on grid frequency fluctuations, electricity price signals, and the rate of change in system power demand. A fuzzy inference system is used to process the input feature vector and output the recognition result of the current operating condition. The operating conditions include at least one of peak shaving and valley filling mode, demand response mode, and emergency backup power mode.

4. The power distribution method for an energy storage device according to claim 1, characterized in that, The construction of the composite loss function includes: Construct an allocation error loss to quantify the deviation between the power allocation value and the system demand; An evolutionary consistency loss is constructed to constrain the predicted evolution rate of healthy states to remain consistent with the actual evolution rate. A monotonic constraint loss is constructed to penalize power allocation decisions that violate the physical law of monotonically decreasing health status. The weighted sum of the allocation error loss, evolution consistency loss, and monotonicity constraint loss yields the composite loss function.

5. The power distribution method for an energy storage device according to claim 1, characterized in that, The multi-model integration unit includes: Multiple base learners, including at least a linear mapping model based on health status, a nonlinear mapping model based on health status, a support vector machine model based on power fluctuation features, and a random forest model based on operational morphology features; The meta-learner, employing a decision tree model, is used to integrate and fuse the outputs of each base learner. In this process, each base learner uses a cross-training method to obtain the out-of-fold prediction results, and the meta-learner is trained using the out-of-fold prediction results of each base learner as input.

6. A power distribution system for an energy storage device, characterized in that: The power distribution method for the energy storage device according to any one of claims 1-5 includes: The data acquisition module is used to acquire the operating data of each energy storage unit in the energy storage device, grid frequency data, and electricity price signal data; The feature construction module is used to extract operational morphology features, power fluctuation features, and operational correlation features based on the operational data, construct index-type features, and convert typical operational curves into graphical features. The operating condition identification module is used to identify the current operating condition based on the power grid frequency data, electricity price signal data, and system power demand change rate. The dynamic weight adjustment module is used to dynamically adjust the weight coefficients corresponding to each optimization objective based on the identified operating conditions, so as to obtain the operating condition adaptive weights. The nonlinear health mapping module, which includes a feature segmenter, a deep learning module and a first long short-term memory network, is used to map the health status data of each energy storage unit into a health status estimate and a high-dimensional hidden state vector. The evolution rate constraint module includes a second long short-term memory network to predict the evolution rate of the health state and corrects the mapping results based on a composite loss function. The multi-model ensemble module contains multiple base learners and meta-learners, used to determine the power allocation weights of each energy storage unit; A power allocation execution module is used to allocate power to each energy storage unit according to the power allocation weight; The operational characteristics include: the daily average power, peak-to-average power ratio, and charge / discharge depth ratio calculated based on the typical operating curve of the energy storage unit; the average power of each time period calculated in the typical operating curve after dividing the day into several time periods; and the first difference ratio between the maximum value of the average power of each time period and the average power of the corresponding time period of the typical operating curve, and the second difference ratio between the minimum value of the average power of each time period and the average power of the corresponding time period of the typical operating curve, calculated based on historical operating data. The power fluctuation characteristics include: the intraday fluctuation standard deviation of the period obtained by calculating the standard deviation of the instantaneous power in each period and averaging it over all operating cycles; and the cross-cycle stability index of the period obtained by calculating the standard deviation of the power at the same moment in different operating cycles and averaging it over all moments in the period. The operational correlation characteristics include: the first correlation coefficient between the energy storage unit output power and the grid frequency, the second correlation coefficient between the energy storage unit output power and the electricity price signal, and the third correlation coefficient between the typical operating curve on weekdays and the typical operating curve on holidays.

Images