Method, system and equipment for predicting battery attenuation curve based on machine learning

Abstract

The invention belongs to the field of battery attenuation prediction, and particularly relates to a method, a system and equipment for predicting a battery attenuation curve based on machine learning. Extracting the recovery rate, the standing time proportion and the health state characteristics of the battery from the historical power curve, and constructing a time sequence characteristic matrix; a local mode is extracted through one-dimensional convolution, global time sequence dependence is captured through a multi-head self-attention mechanism, and a global feature sequence fused with context information is generated. And a bidirectional gating mechanism is used to further extract the time sequence dynamic state, and a high-dimensional feature vector is obtained. And finally, respectively outputting a short-term attenuation prediction sequence and a long-term attenuation curve through a full-connection network, and realizing multi-time-scale accurate prediction. The method is especially suitable for battery health management of the energy storage converter under complex working conditions, and provides data support for optimization control and maintenance strategies.

Description

Technical Field

[0001] This invention belongs to the field of battery degradation prediction, specifically relating to methods, systems, and devices for predicting battery degradation curves based on machine learning. Background Technology

[0002] With the widespread application of energy storage systems in power systems, new energy grid connection, electric vehicles and other fields, the health status management and life prediction of batteries in energy storage converters have become particularly important.

[0003] Traditional battery degradation prediction methods rely heavily on empirical models or simple statistical regressions, which are difficult to accurately capture the dynamic degradation behavior of batteries, especially lacking the ability to effectively model the nonlinear response of batteries under conditions such as rapid charging and discharging and static recovery.

[0004] In existing technologies, electrochemical models involve large computational loads, and parameters are difficult to identify online, making them unsuitable for real-time prediction. They also struggle to accurately describe the degradation behavior of batteries under dynamic operating conditions. Furthermore, they fail to extract sufficient time-series features and cannot effectively integrate the dynamic behavior characteristics of battery resilience and rest adequacy, resulting in limited accuracy in long-term degradation prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and device for predicting battery degradation curves based on machine learning.

[0006] The method for predicting battery degradation curves based on machine learning includes the following steps: S1. Collect the output power curves of the battery over several consecutive time windows in the past, extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage. S2. The recovery rate, battery health status and rest time ratio extracted from different time windows are spliced ​​and normalized according to the time windows to obtain the feature matrix. S3. Perform two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. Perform multi-head self-attention processing on the convolutional feature sequence. For each target time window in the sequence, calculate its correlation weight with all time windows in the sequence, and perform weighted fusion of the features of all windows based on the weight to obtain a global feature sequence. S4. According to the forward time sequence, obtain the hidden state of the previous time window of the last window in the global feature sequence, calculate the reset gate and update gate of the global feature sequence, multiply the reset gate with the hidden state of the previous time window element by element to obtain the filtering history information, concatenate the filtering history information with the global feature sequence, obtain the positive candidate hidden state through nonlinear transformation, and obtain the positive state by weighted fusion of the hidden state of the previous time window and the positive candidate hidden state through the update gate. In reverse chronological order, obtain the hidden state of the next time window after the first window in the global feature sequence, execute the process of obtaining the positive state, and obtain the reverse candidate hidden state. By concatenating the forward and reverse states, a high-dimensional feature vector is obtained. S5. Perform fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. Perform linear mapping on the long-term prediction matrix to obtain the battery decay curve and complete the battery decay prediction.

[0007] S5 specifically refers to: S5.1 performs a first-layer fully connected processing on the high-dimensional feature vector to obtain short-term decay prediction sequences for multiple time windows; S5.2 performs a second fully connected layer on the high-dimensional feature vector to obtain a long-term decay prediction matrix for multiple time windows, and performs linear mapping to obtain the battery decay curve. S5.3 According to the time window, substitute the short-term degradation prediction sequence into the degradation curve and calculate the mean square error of the two in the same time window. If the mean square error is greater than the variance threshold, it is determined whether the degradation curve meets the physical rationality constraint. If it does, the difference between the detection data of the next time window and the values ​​of the short-term degradation prediction sequence and the corresponding degradation curve is calculated respectively. If both differences are less than the preset error range, the weighted sum is used as the final degradation curve and short-term degradation prediction sequence. If there is one or more differences that are not less than the preset error range, the detection data of the next time window is substituted, and S1 to S4 are executed again to obtain the battery degradation curve and complete the battery degradation prediction.

[0008] The physical rationality constraints include: the capacity decay curve should be monotonically non-increasing, the decay rate should not exceed the rate change threshold, and the predicted health status of the battery should be greater than the end of its lifespan.

[0009] Extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve segment. The specific operation is as follows: If there exists a point in time where the output power is greater than the output power at the time step before this point and greater than the output power at the time step after this point, then this point is considered a local maximum point, and from this point onwards, the output power continues to decrease. If there exists a point in time where the output power is less than the output power at the time step before this point and less than the output power at the time step after this point, then this point is a local minimum point, and from this point onwards, the output power continues to rise. At the same time, the decline is greater than the minimum decline threshold, the recovery is greater than the minimum recovery threshold, and the entire process of decline followed by recovery is within the preset time range. For the output power curve of segment k, which first decreases and then recovers, calculate its recovery rate: , in, The output power at the local maximum point. This represents the time point corresponding to a local maximum. The output power at the local minimum point. This represents the time point corresponding to the local minimum.

[0010] The percentage of time the battery is in a static state out of the total time is defined as the static time percentage. The specific steps are as follows: , The total time of the time window. The duration of the static state to satisfy the static state judgment condition.

[0011] Conditions for determining a static state: The absolute value of the battery's average current is lower than the current threshold, the battery's average temperature change rate is lower than the temperature threshold, and the battery's open-circuit voltage change rate is lower than the change rate threshold. The above conditions must be maintained for at least the threshold duration.

[0012] The specific calculation method for the positive state is as follows: , in, To update the door, This refers to the hidden state of the window that preceded the last window. This is a positive candidate hidden state.

[0013] The calculation method for resetting and updating doors is as follows: Reset Door: , in, It is the Sigmoid activation function. To reset the weight matrix of the gate, To reset the gate bias vector, The hidden state of the window preceding the last window is the global feature sequence; Update Gate: , in, To update the gate weight matrix, To update the gate bias vector.

[0014] The system for predicting battery degradation curves based on machine learning is used to implement the above-mentioned method for predicting battery degradation curves based on machine learning. It includes: a data acquisition module, which collects the output power curves of the battery in several consecutive time windows in the past, extracts all output power curves that first decrease and then recover, and calculates the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage. The splicing module extracts the recovery rate, battery health status, and rest time ratio from different time windows, and performs feature splicing and normalization according to the time windows to obtain a feature matrix. The feature extraction module performs two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. The convolutional feature sequence is then processed by multi-head self-attention. For each target time window in the sequence, the correlation weight between it and all time windows in the sequence is calculated. Based on this weight, the features of all windows are weighted and fused to obtain a global feature sequence. The high-dimensional feature acquisition module, in ascending time order, obtains the hidden state of the previous time window of the last window in the global feature sequence, calculates the reset gate and update gate of the global feature sequence, multiplies the reset gate element-wise with the hidden state of the previous time window to obtain filtering history information, concatenates the filtering history information with the global feature sequence, and obtains the positive candidate hidden state through nonlinear transformation. The update gate then performs a weighted fusion of the hidden state of the previous time window and the positive candidate hidden state to obtain the positive state. In reverse time order, it obtains the hidden state of the next time window of the first window in the global feature sequence, performs the process of obtaining the positive state, and obtains the reverse candidate hidden state. Finally, it concatenates the positive state and the reverse state to obtain the high-dimensional feature vector. The prediction module performs fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. It then performs linear mapping on the long-term prediction matrix to obtain the battery decay curve, thus completing the battery decay prediction.

[0015] The device for predicting battery degradation curves based on machine learning includes a processor and a memory, wherein the processor implements the method for predicting battery degradation curves based on machine learning when executing a computer program stored in the memory.

[0016] The advantages of this application compared to the prior art are as follows: By extracting the initial decline followed by a rebound pattern and its rebound rate from the battery output power curve, and combining this with the ratio of battery health status to rest time, a dynamic feature system capable of characterizing battery resilience and rest adequacy was constructed. These features reflect the battery's degradation behavior at the physical mechanism level, providing more interpretable input for the model.

[0017] The feature extraction mechanism, which combines one-dimensional convolution with multi-head self-attention, can capture local temporal patterns and model the dependencies between global time windows, effectively improving the model's ability to represent the dynamic process of battery degradation.

[0018] Combining reset gates and update gates enhances the ability to filter historical information and infer future states, making it suitable for predicting tasks such as battery degradation, which have time-dependent and lag effects.

[0019] By using a fully connected network to output short-term decay sequences and long-term decay curves, and combining them with physical rationality constraints for post-processing, the real-time nature of short-term predictions is ensured, as well as the rationality and stability of long-term decay trends. Detailed Implementation

[0020] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments.

[0021] This invention relates to a method for predicting battery degradation curves based on machine learning, comprising the following steps: S1. Collect the output power curves of the battery over several consecutive time windows in the past, extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage.

[0022] This application, considering the characteristics of energy storage converters, selects the battery data to be collected. The battery operating characteristics of energy storage converters include frequent and random power throughput, responding to grid frequency regulation, smoothing renewable energy fluctuations, and peak shaving and valley filling. Battery power commands need to frequently and rapidly switch between charging and discharging, with random amplitudes. This operation accelerates battery aging, and the battery's recovery rate directly reflects its internal chemical state recovery after being subjected to shocks. This is an operating condition not present in electric vehicle batteries and consumer electronics batteries.

[0023] Furthermore, the batteries in energy storage converters may be in a state of prolonged quiescence or low power when grid demand is low. During this quiescence period, chemical repair processes such as Li+ concentration rebalancing occur inside the battery. The proportion of quiescence time directly quantifies the battery's chances of undergoing repair.

[0024] Specifically, extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve. The specific operation is as follows: if there is an output power at a certain time point that is greater than the output power at the time step before this time point and greater than the output power at the time step after this time point, then this point is a local maximum point. From this point onwards, the output power continues to decrease. If there exists a point in time where the output power is less than the output power at the time step before this point and less than the output power at the time step after this point, then this point is a local minimum point, and from this point onwards, the output power continues to rise. At the same time, the decline is greater than the minimum decline threshold, the recovery is greater than the minimum recovery threshold, and the entire process of decline followed by recovery lasts within the preset time range.

[0025] For the output power curve of segment k, which first decreases and then recovers, calculate its recovery rate: , in, The output power at the local maximum point. This represents the time point corresponding to a local maximum. The output power at the local minimum point. This represents the time point corresponding to the local minimum.

[0026] Recovery rate characterizes battery resilience; a high recovery rate indicates good kinetics of the battery's internal chemical system, enabling rapid repair after stress. In predictive models, this should be associated with a gentler long-term degradation slope and a later degradation inflection point. A low recovery rate implies a high proportion of irreversible damage or a decline in material activity. This should be associated with a steeper degradation trend.

[0027] The proportion of time the battery spends in a static state to the total time is called the static time proportion. The static state is determined by the following conditions: the absolute value of the battery's average current is lower than the current threshold, the battery's average temperature change rate is lower than the temperature threshold, and the battery's open-circuit voltage change rate is lower than the change rate threshold. The above conditions must be maintained for at least the static time threshold.

[0028] The specific steps for calculating the settling time ratio are as follows: , The total time of the time window. The duration of the static state to satisfy the static state judgment condition.

[0029] A high resting percentage indicates that the battery has sufficient time for internal balancing and self-repair. This should help slow down the degradation rate during forecasting; a low resting percentage indicates that the battery is under fatigue from continuous operation or frequent scheduling, and determines whether accelerated degradation will occur.

[0030] S2. Extract the recovery rate, battery health status, and rest time ratio from different time windows, and perform feature splicing and normalization according to the time windows to obtain the feature matrix.

[0031] The different data extracted from S1 are fused and normalized, and the data from each time window are combined into a comprehensive feature vector. The feature vectors of the past T consecutive time windows are time-aligned and stacked in chronological order to construct a feature matrix of dimension [T, D], where D is the total dimension of the feature vectors. This matrix completely preserves the temporal evolution information of the battery's recent operating state.

[0032] S3. Perform two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. Perform multi-head self-attention processing on the convolutional feature sequence. For each target time window in the sequence, calculate its correlation weight with all time windows in the sequence, and perform weighted fusion of the features of all windows based on the weight to obtain a global feature sequence.

[0033] Within each time window and its local neighborhood, a first-layer one-dimensional convolution is performed to fuse the recovery rate, battery health status, and rest time ratio on the time scale, generating a series of primary local features that characterize the joint state of each time window and its neighboring windows.

[0034] The second convolutional layer, building upon the previous one, further expands the receptive field, enabling the learning of more complex and abstract combinational patterns.

[0035] Multi-head self-attention processing actively focuses on which windows in all data are most important for predicting current or future decay and performs information fusion.

[0036] Multi-head attention focuses on the similarity of resting time ratios, the similarity of recovery rates for each curve segment, and the continuity of the long-term battery health decline trend. This captures complex global dependencies from multiple semantic subspaces.

[0037] S4. According to the forward time sequence, obtain the hidden state of the previous time window of the last window in the global feature sequence, calculate the reset gate and update gate of the global feature sequence, multiply the reset gate with the hidden state of the previous time window element by element to obtain the filtering history information, concatenate the filtering history information with the global feature sequence, obtain the positive candidate hidden state through nonlinear transformation, and obtain the positive state by weighted fusion of the hidden state of the previous time window and the positive candidate hidden state through the update gate. In reverse chronological order, obtain the hidden state of the next time window after the first window in the global feature sequence, execute the process of obtaining the positive state, and obtain the reverse candidate hidden state. By concatenating the forward and reverse states, a high-dimensional feature vector is obtained.

[0038] The reset gate and update gate of the global feature sequence are calculated as follows: Reset Door: , in, It is the Sigmoid activation function. To reset the weight matrix of the gate, To reset the gate bias vector, The hidden state of the window preceding the last window is the global feature sequence; Update Gate: , in, To update the gate weight matrix, To update the gate bias vector.

[0039] The specific calculation method for the positive state is as follows: , in, To update the door, This refers to the hidden state of the window that preceded the last window. This is a positive candidate hidden state.

[0040] S5. Perform fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. Perform linear mapping on the long-term prediction matrix to obtain the battery decay curve and complete the battery decay prediction.

[0041] Specifically: The high-dimensional feature vector H ∈ S4 output {T×D} , where T is the number of time windows and D is the feature dimension.

[0042] S5.1 performs a first-layer fully connected processing on the high-dimensional feature vector to obtain short-term decay prediction sequences for multiple time windows.

[0043] Treating H as a sequence, the feature vector h for each time window t... t The sequence is processed independently, mapped to a high-dimensional space, and activated. A 128-dimensional space can be used to obtain a short-term decay prediction sequence S of length M. t = [s {t,1} , s {t,2} , ..., s {t,M} ] represents the short-term health status prediction value for the next M windows, starting from window t.

[0044] Short-term health forecasts only involve short-term health forecasts for the next few days and are used to determine whether the batteries in an energy storage converter are nearing the end of their lifespan.

[0045] S5.2 performs a second fully connected layer on the high-dimensional feature vector to obtain a long-term decay prediction matrix for multiple time windows, and performs linear mapping to obtain the battery decay curve.

[0046] Treating H as a sequence, the feature vector h for each time window t... t Perform independent processing, map to a high-dimensional space and activate; 256 dimensions can be taken, resulting in L∈ {T×N×K} For each time window t, an N×K matrix L is output. t .

[0047] Each L t Each row corresponds to a future long-term window n, containing K parameters, used to define a local decay function segment. Through linear mapping, the parameters of each window are mapped to the predicted battery health status at the end of the window. Connecting all N points forms a long-term decay curve starting from the current time t.

[0048] The long-term degradation curve here includes battery health degradation values ​​for the next few months, which can help estimate how long the battery will last before it reaches the end of its life, allowing for proactive measures.

[0049] S5.3 According to the time window, substitute the short-term degradation prediction sequence into the degradation curve and calculate the mean square error of the two in the same time window. If the mean square error is greater than the variance threshold, it is determined whether the degradation curve meets the physical rationality constraint. If it does, the difference between the detection data of the next time window and the values ​​of the short-term degradation prediction sequence and the corresponding degradation curve is calculated respectively. If both differences are less than the preset error range, the weighted sum is used as the final degradation curve and short-term degradation prediction sequence. If there is one or more differences that are not less than the preset error range, the detection data of the next time window is substituted, and S1 to S4 are executed again to obtain the battery degradation curve and complete the battery degradation prediction.

[0050] For each time window t, take the short-term series S. t The first predicted value, i.e., the value in window t+1, is compared with the value of the decay curve in window t+1, and the root mean square error is calculated. If the root mean square error is greater than the variance threshold of 0.01, it indicates that short-term fluctuations conflict with long-term trends. Further judgment is needed on whether the decay curve meets the physical rationality constraints. The capacity decay curve should be monotonically non-increasing, the decay rate should not be greater than the rate change threshold, and the predicted health state of the battery should be greater than the end of its lifespan.

[0051] Since the decay curve is composed of multiple long-term predictions, the possibility of deviation is much greater than that of short-term prediction sequences. Therefore, the decay curve is first validated here. If the validation passes, it indicates that the decay curve is reasonable, and the noise may be causing the mean squared error to exceed the variance threshold; if it fails, a re-prediction is triggered.

[0052] After the test is passed, the difference between the detection data of the next time window and the values ​​of the short-term decay prediction sequence and the corresponding decay curve is calculated. If both differences are less than the preset error range, the weighted sum is used as the final decay curve and short-term decay prediction sequence. If there is one or more differences that are not less than the preset error range, the detection data of the next time window is substituted into the data, and S1 to S4 are executed again to obtain the battery decay curve and complete the battery decay prediction.

[0053] The system for predicting battery degradation curves based on machine learning is used to implement the above-mentioned method for predicting battery degradation curves based on machine learning. It includes: a data acquisition module, which collects the output power curves of the battery in several consecutive time windows in the past, extracts all output power curves that first decrease and then recover, and calculates the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage. The splicing module extracts the recovery rate, battery health status, and rest time ratio from different time windows, and performs feature splicing and normalization according to the time windows to obtain a feature matrix. The feature extraction module performs two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. The convolutional feature sequence is then processed by multi-head self-attention. For each target time window in the sequence, the correlation weight between it and all time windows in the sequence is calculated. Based on this weight, the features of all windows are weighted and fused to obtain a global feature sequence. The high-dimensional feature acquisition module, in ascending time order, obtains the hidden state of the previous time window of the last window in the global feature sequence, calculates the reset gate and update gate of the global feature sequence, multiplies the reset gate element-wise with the hidden state of the previous time window to obtain filtering history information, concatenates the filtering history information with the global feature sequence, and obtains the positive candidate hidden state through nonlinear transformation. The update gate then performs a weighted fusion of the hidden state of the previous time window and the positive candidate hidden state to obtain the positive state. In reverse time order, it obtains the hidden state of the next time window of the first window in the global feature sequence, performs the process of obtaining the positive state, and obtains the reverse candidate hidden state. Finally, it concatenates the positive state and the reverse state to obtain the high-dimensional feature vector. The prediction module performs fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. It then performs linear mapping on the long-term prediction matrix to obtain the battery decay curve, thus completing the battery decay prediction.

[0054] The device for predicting battery degradation curves based on machine learning includes a processor and a memory, wherein the processor implements the method for predicting battery degradation curves based on machine learning when executing a computer program stored in the memory.

Claims

1. A method for predicting battery degradation curves based on machine learning, characterized in that, Includes the following steps: S1. Collect the output power curves of the battery over several consecutive time windows in the past, extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage. S2. The recovery rate, battery health status and rest time ratio extracted from different time windows are spliced ​​and normalized according to the time windows to obtain the feature matrix. S3. Perform two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. Perform multi-head self-attention processing on the convolutional feature sequence. For each target time window in the sequence, calculate its correlation weight with all time windows in the sequence, and perform weighted fusion of the features of all windows based on the weight to obtain a global feature sequence. S4. According to the forward time sequence, obtain the hidden state of the previous time window of the last window in the global feature sequence, calculate the reset gate and update gate of the global feature sequence, multiply the reset gate with the hidden state of the previous time window element by element to obtain the filtering history information, concatenate the filtering history information with the global feature sequence, obtain the positive candidate hidden state through nonlinear transformation, and obtain the positive state by weighted fusion of the hidden state of the previous time window and the positive candidate hidden state through the update gate. In reverse chronological order, obtain the hidden state of the next time window after the first window in the global feature sequence, execute the process of obtaining the positive state, and obtain the reverse candidate hidden state. By concatenating the forward and reverse states, a high-dimensional feature vector is obtained. S5. Perform fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. Perform linear mapping on the long-term prediction matrix to obtain the battery decay curve and complete the battery decay prediction.

2. The method for predicting battery degradation curves based on machine learning according to claim 1, characterized in that, S5 specifically refers to: S5.1 performs a first-layer fully connected processing on the high-dimensional feature vector to obtain short-term decay prediction sequences for multiple time windows; S5.2 performs a second fully connected layer on the high-dimensional feature vector to obtain a long-term decay prediction matrix for multiple time windows, and performs linear mapping to obtain the battery decay curve. S5.3 According to the time window, substitute the short-term degradation prediction sequence into the degradation curve and calculate the mean square error of the two in the same time window. If the mean square error is greater than the variance threshold, it is determined whether the degradation curve meets the physical rationality constraint. If it does, the difference between the detection data of the next time window and the values ​​of the short-term degradation prediction sequence and the corresponding degradation curve is calculated respectively. If both differences are less than the preset error range, the weighted sum is used as the final degradation curve and short-term degradation prediction sequence. If there is one or more differences that are not less than the preset error range, the detection data of the next time window is substituted, and S1 to S4 are executed again to obtain the battery degradation curve and complete the battery degradation prediction.

3. The method for predicting battery degradation curves based on machine learning according to claim 2, characterized in that, The physical rationality constraints include: the capacity decay curve should be monotonically non-increasing, the decay rate should not exceed the rate change threshold, and the predicted health status of the battery should be greater than the end of its lifespan.

4. The method for predicting battery degradation curves based on machine learning according to claim 1, characterized in that, Extract all output power curves that first decrease and then recover, and calculate the recovery rate of each curve segment. The specific operation is as follows: If there exists a point in time where the output power is greater than the output power at the time step before this point and greater than the output power at the time step after this point, then this point is considered a local maximum point, and from this point onwards, the output power continues to decrease. If there exists a point in time where the output power is less than the output power at the time step before this point and less than the output power at the time step after this point, then this point is a local minimum point, and from this point onwards, the output power continues to rise. At the same time, the decline is greater than the minimum decline threshold, the recovery is greater than the minimum recovery threshold, and the entire process of decline followed by recovery is within the preset time range. For the output power curve of segment k, which first decreases and then recovers, calculate its recovery rate: ， in, The output power at the local maximum point. This represents the time point corresponding to a local maximum. The output power at the local minimum point. This represents the time point corresponding to the local minimum.

5. The method for predicting battery degradation curves based on machine learning according to claim 1, characterized in that, The percentage of time the battery is in a static state out of the total time is defined as the static time percentage. The specific steps are as follows: , The total time of the time window. The duration of the static state to satisfy the static state judgment condition.

6. The method for predicting battery degradation curves based on machine learning according to claim 5, characterized in that, Conditions for determining a static state: The absolute value of the battery's average current is lower than the current threshold, the battery's average temperature change rate is lower than the temperature threshold, and the battery's open-circuit voltage change rate is lower than the change rate threshold. The above conditions must be maintained for at least the threshold duration.

7. The method for predicting battery degradation curves based on machine learning according to claim 1, characterized in that, The specific calculation method for the positive state in S4 is as follows: ， in, To update the door, This refers to the hidden state of the window that preceded the last window. This is a positive candidate hidden state.

8. The method for predicting battery degradation curves based on machine learning according to claim 1, characterized in that, The calculation method for reset and update gates in S4 is as follows: Reset Door: , in, It is the Sigmoid activation function. To reset the weight matrix of the gate, To reset the gate bias vector, The hidden state of the window preceding the last window is the global feature sequence; Update Gate: , in, To update the gate weight matrix, To update the gate bias vector.

9. A system for predicting battery degradation curves based on machine learning, used to implement the method for predicting battery degradation curves based on machine learning as described in any one of claims 1-8, characterized in that, include: The data acquisition module collects the battery's output power curves over several consecutive time windows, extracts all output power curves that first decrease and then recover, and calculates the recovery rate of each curve segment. Obtain the battery health status over several past time windows, and use the percentage of time the battery was idle as the idle time percentage. The splicing module extracts the recovery rate, battery health status, and rest time ratio from different time windows, and performs feature splicing and normalization according to the time windows to obtain a feature matrix. The feature extraction module performs two-layer one-dimensional convolution on the feature matrix to extract local features and obtain a convolutional feature sequence. The convolutional feature sequence is then processed by multi-head self-attention. For each target time window in the sequence, the correlation weight between it and all time windows in the sequence is calculated. Based on this weight, the features of all windows are weighted and fused to obtain a global feature sequence. The high-dimensional feature acquisition module, in ascending time order, obtains the hidden state of the previous time window of the last window in the global feature sequence, calculates the reset gate and update gate of the global feature sequence, multiplies the reset gate element-wise with the hidden state of the previous time window to obtain filtering history information, concatenates the filtering history information with the global feature sequence, and obtains the positive candidate hidden state through nonlinear transformation. The update gate then performs a weighted fusion of the hidden state of the previous time window and the positive candidate hidden state to obtain the positive state. In reverse time order, it obtains the hidden state of the next time window of the first window in the global feature sequence, performs the process of obtaining the positive state, and obtains the reverse candidate hidden state. Finally, it concatenates the positive state and the reverse state to obtain the high-dimensional feature vector. The prediction module performs fully connected processing on the high-dimensional feature vectors to obtain short-term decay prediction sequences and long-term prediction matrices for multiple time windows. It then performs linear mapping on the long-term prediction matrix to obtain the battery decay curve, thus completing the battery decay prediction.

10. A device for predicting battery degradation curves based on machine learning, characterized in that, The system includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the method for predicting battery degradation curves based on machine learning as described in any one of claims 1-8.

Images