Lithium battery capacity state evaluation method based on multi-time scale double neural network
By combining LSTM and Transformer neural networks, a dual neural network architecture with multiple time scales is constructed, which solves the problem of multi-time scale differences in lithium battery capacity state estimation, achieves high-precision and high-robust capacity state assessment, and reduces errors and complexity.
Patent Information
- Application Number
- CN202511659839.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-02-06
AI Technical Summary
Existing lithium battery capacity state estimation methods are difficult to effectively handle differences across multiple time scales, leading to increased algorithm complexity and training difficulty. Furthermore, traditional methods are computationally cumbersome and time-consuming.
A dual neural network architecture based on multiple time scales is adopted, combining LSTM neural network and Transformer neural network. By acquiring the voltage and current data of lithium battery in real time, the charge features and capacity features of multiple time scales are extracted, and capacity state estimation is performed using multi-head attention mechanism and position encoder.
It significantly reduced capacity estimation error, achieving high-precision and robust online evaluation. The capacity estimation error was controlled within 0.0378 Ah, improving the accuracy and robustness of lithium battery capacity status evaluation.
Smart Images
Figure CN121476966A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of lithium battery state prediction, and particularly relates to a lithium battery capacity state evaluation method based on a multi-time scale double neural network. BACKGROUND
[0002] As the core of the global clean energy system, the importance of lithium batteries is increasingly prominent. Under the promotion of energy transformation and sustainable development goals, its application has expanded from electric vehicles to energy storage systems, and its performance and safety are related to the stability and efficiency of the entire energy system. The capacity state of lithium batteries is a key parameter of the battery management system, which directly affects the battery life, charging and discharging strategy and safety. How to quickly and accurately estimate the capacity state is a problem to be solved.
[0003] In order to solve this challenge, it is particularly important to build a capacity state estimation algorithm that takes into account the characteristics of lithium batteries. Traditional capacity state estimation methods rely on empirical models or simple statistical methods, which are difficult to cope with complex battery aging processes. Intelligent algorithms based on neural networks can achieve high-precision capacity state estimation through a large amount of data training. However, how to design a neural network model that can cope with capacity decay and quickly estimate is still a technical problem to be solved.
[0004] For the estimation of lithium ion battery capacity state, the commonly used capacity characteristic extraction method is the incremental capacity method, but this method requires a complete charging process, which takes a long time and is complicated to calculate. In addition, there are also research methods based on equivalent impedance and data-driven, such as recursive least squares method, LSTM and DNN, etc. Although data-driven algorithms have achieved some success in lithium battery capacity estimation, there are still some problems to be solved. The existing neural network algorithm does not consider the physical characteristics of long-term capacity decay and short-term relative stability, which leads to the increase of algorithm complexity and training difficulty. SUMMARY
[0005] In order to solve the technical problems existing in the prior art, the application provides a lithium battery capacity state evaluation method based on a multi-time scale double neural network, which solves the technical problem of processing multi-time scale differences in the current lithium ion battery capacity state evaluation.
[0006] To achieve the above purpose, the application provides a lithium battery capacity state evaluation method based on a multi-time scale double neural network, which includes: measuring and obtaining working condition voltage and current data of the lithium battery; inputting the working condition voltage and current data into an LSTM neural network, and calculating the state of charge of the lithium battery through the LSTM neural network; According to the state of charge and the voltage and current data, multi-time scale charge characteristics and multi-time scale capacity characteristics are extracted based on multi-time scale characteristics; The multi-time scale charge characteristics and the multi-time scale capacity characteristics are input into a Transformer neural network, and a capacity state of the lithium battery is calculated through the Transformer neural network.
[0007] Preferably, the process of measuring and acquiring the working condition voltage and current data of the lithium battery comprises: The voltage and current data of the lithium battery during the charging and discharging process are acquired in real time through a data acquisition card with a sampling frequency of 10 Hz.
[0008] Preferably, the process of acquiring the state of charge of the lithium battery comprises: The sampled voltage and current at time k are input into a neural network with 5 layers and 128 LSTM units per layer, and the state of charge at time k is obtained through a calculation function of the neural network.
[0009] Preferably, the LSTM unit controls information flow through an input gate, a forgetting gate and an output gate, and adopts a sigmoid function and a tanh function as activation functions; wherein the sigmoid function converts input features to a nonlinear interval of 0 to 1, and the tanh function converts input features to a nonlinear interval of -1 to 1.
[0010] Preferably, the process of acquiring the multi-time scale charge characteristics comprises: According to the difference between the state of charge at time k and the state of charge at time k-n, the multi-time scale charge characteristics at time k are obtained, wherein n is a cumulative time scale constant.
[0011] Preferably, the process of acquiring the multi-time scale capacity characteristics comprises: According to the current accumulation amount of multiple sampling periods before time k, the multi-time scale capacity characteristics at time k are obtained, wherein the sampling period is m.
[0012] Preferably, the process of acquiring the capacity state of the lithium battery comprises: The multi-time scale charge characteristics and the multi-time scale capacity characteristics are input into a Transformer neural network containing a multi-head attention mechanism, and the capacity state at time k is calculated through a function of the neural network.
[0013] Preferably, the multi-head attention mechanism is realized by the following way: After linear transformation of a query matrix Q, a key matrix K and a value matrix V through a weight matrix, a scaled dot-product attention is calculated, linear transformation is performed on the outputs of multiple attention heads, and a multi-head attention output is obtained.
[0014] Preferably, the Transformer neural network adopts a position encoder to process sequence order information. The position encoder encodes positions using sine and cosine functions of different frequencies.
[0015] Compared with the prior art, the present application has the following advantages and technical effects: The present application adopts a dual-network collaborative architecture composed of an LSTM neural network and a Transformer neural network to decouple real-time state of charge estimation and long-term capacity state prediction, fully utilizes the advantages of LSTM in processing time-series voltage and current data and the ability of Transformer in capturing long-distance dependencies, significantly reduces the complexity and training difficulty of a single model, controls the capacity estimation error within 0.0378 Ah, and realizes high-precision and high-robustness online evaluation of the capacity state of a lithium battery.
[0016] The present application uses a data acquisition card with a sampling frequency of 10 Hz to acquire real-time lithium battery charging and discharging condition data, ensures the timeliness and accuracy of the input data, and provides a high-quality data basis for subsequent neural network models, effectively reducing the evaluation error caused by data delay or distortion.
[0017] The present application processes voltage and current inputs by constructing a deep network structure with 5 layers and 128 LSTM units in each layer, balances the model expression ability and computational efficiency, fully excavates the nonlinear mapping relationship between voltage, current and state of charge, and improves the accuracy of state of charge estimation.
[0018] The present application introduces three gating mechanisms of input gate, forget gate and output gate in the LSTM unit, and uses sigmoid function and tanh function as activation functions, realizes selective memory and forgetting of time-series information, enhances the learning ability of the network to key features, and effectively fits the dynamic characteristics of the battery.
[0019] The present application extracts the difference between the state of charge at time k and the state of charge at time k-n as a multi-time scale capacity feature, accumulates the small changes in a short time into a significant change in a long time scale, eliminates the influence of instantaneous fluctuations, and makes the feature better reflect the inherent trend of capacity degradation.
[0020] The present application calculates the current accumulation amount of multiple sampling periods before time k as a multi-time scale capacity feature, smooths the sampling error and transient interference as the time scale increases, improves the signal-to-noise ratio of the capacity feature, and enhances the corresponding relationship between the feature and the capacity state.
[0021] This invention effectively captures key information in long sequences by inputting multi-timescale charge features and capacity features into a Transformer neural network and using its self-attention mechanism to dynamically allocate weights for features at different time steps, thereby improving the robustness and accuracy of capacity state prediction.
[0022] This invention employs a multi-head attention mechanism in the Transformer to project input features into multiple subspaces for parallel computation, enabling the model to focus on different feature dimensions from different representation subspaces, thereby enhancing the diversity of feature extraction and the ability to recognize complex patterns.
[0023] This invention introduces a position encoder based on sine and cosine functions into the Transformer, injecting positional information into the feature sequence, enabling the model to recognize the temporal order relationship of features at multiple time scales, and ensuring the integrity of sequence processing. Attached Figure Description
[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the estimation results of the state of charge in an embodiment of the present invention; Figure 3 This is a schematic diagram of the capacity state estimation results of the lithium battery training set used in an embodiment of the present invention; Figure 4 This is a schematic diagram of the capacity state estimation results of the lithium-ion battery test set used in an embodiment of the present invention; Detailed Implementation
[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0026] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0027] like Figure 1 As shown, this embodiment provides a lithium battery capacity state assessment method based on a multi-timescale dual neural network, including: Measure and acquire the operating voltage and current data of the lithium battery; The operating voltage and current data are input into the LSTM neural network, and the state of charge of the lithium battery is calculated by the LSTM neural network. Based on the state of charge and voltage and current data, features covering the time length are extracted based on the multi-timescale characteristics of lithium batteries, including multi-timescale charge features and multi-timescale capacity features. The multi-timescale charge characteristics and multi-timescale capacity characteristics are input into the Transformer neural network, and the capacity state of the lithium battery is calculated by the Transformer neural network.
[0028] This embodiment addresses the challenge of lithium battery state-of-charge (SOC) estimation by integrating two deep learning techniques and leveraging the strengths of each network architecture. First, the charging and discharging current and voltage of the lithium battery are acquired using a test platform. Second, the LSTM algorithm is employed to estimate the SOC of the lithium battery. Third, based on the multi-timescale characteristics of lithium batteries, features covering the time duration are extracted. Finally, a Transformer neural network is used to estimate the SOC of the lithium battery. The proposed dual-neural network collaborative algorithm can control the capacity estimation error within 0.0378 AH.
[0029] Furthermore, the process of measuring and acquiring the operating voltage and current data of the lithium battery includes: The voltage and current data of the lithium battery during charging and discharging are acquired in real time using a data acquisition card with a sampling frequency of 10Hz.
[0030] Furthermore, this embodiment first acquires the charging and discharging current and voltage of the lithium battery based on a test platform. The test platform uses a data acquisition card with a sampling frequency of 10Hz, enabling real-time and accurate acquisition of current and voltage data during the charging and discharging process. Secondly, an LSTM (Long Short-Term Memory) algorithm is used to estimate the state of charge (SOC) of the lithium battery. After acquiring the charging and discharging current and voltage data, these are input to a trained LSTM model. Thirdly, based on the multi-timescale characteristics of the lithium battery, charge and capacity features covering different time periods are extracted. Finally, a Transformer neural network is used to estimate the SOC of the lithium battery. The Transformer neural network receives the multi-timescale charge and capacity features, and through learning and analysis of these features, estimates the SOC of the lithium battery.
[0031] Furthermore, such as Figure 2 As shown, the process of obtaining the state of charge of a lithium battery includes: The sampled voltage and sampled current at time k are input into a neural network with 5 layers and 128 LSTM units per layer. The state of charge at time k is obtained through the calculation function of the neural network.
[0032] Furthermore, this embodiment uses an LSTM neural network to estimate the state of charge. Specifically, it employs a 5-layer neural network with 128 LSTMs per layer to estimate the state of charge. The activation functions of the LSTM are set as sigmoid(σ) and tanh, respectively. ; Wherein, σ and tanh functions transform the charged input characteristics to the nonlinear intervals of 0 to 1 and -1 to 1, respectively, z represents the input variable, and e is the natural constant; The formula for estimating the state of charge using an LSTM neural network is shown below; In the formula: It is the state of charge at time k. It is the computation function of LSTM. Let be the sampled voltage at time k. Let be the sampled current at time k, where k is the sampling time.
[0033] Furthermore, the LSTM unit controls the flow of information through input gates, forget gates, and output gates, and uses the sigmoid function and the tanh function as activation functions; the sigmoid function transforms the input features to a non-linear range of 0 to 1, and the tanh function transforms the input features to a non-linear range of -1 to 1.
[0034] Furthermore, the LSTM involved in this embodiment introduces three gates with unique functions: the input gate, the forget gate, and the output gate, to better fit the predicted state with the input features. The calculation formula is shown below: In the formula, W i Let be the weight vector of the input gate. h t-1 This represents the output value of the LSTM unit at the previous moment. x t The input features at the current time, b i For the bias term of the input gate, W f Let be the weight vector of the forget gate. b f For the bias term of the input gate, W o This is the weight vector of the output gate. b o This is the bias term for the output gate. tThe calculation time is shown below. The calculation of the cell input information at the current time is as follows: In the formula, The information input into the cell at the current moment is... W c This is the weight vector of the current input information. b c This is the bias term for the current input information. The cell's current input information is closely related to the cell's output value at the previous time step and the input features at the current time step, and contains new information brought about by the current input. The update formula for the cell's memory unit at the current time step is as follows: In the formula, C t As the cell's memory unit at the current moment, C t -1 represents the cellular memory unit from the previous moment. h t This is the output value of the LSTM unit at the current moment.
[0035] The current state of charge has been calculated as follows: In the formula: It is the state of charge at time k. It is a computation function for 5 layers, each with 128 LSTMs. Let be the sampled voltage at time k. Let be the sampled current at time k, where k is the sampling time.
[0036] Furthermore, the process of obtaining multi-timescale charge characteristics includes: Based on the difference between the state of charge at time k and the state of charge at time kn, the multi-timescale charge characteristics at time k are obtained, where n is the cumulative timescale constant.
[0037] Furthermore, the multi-timescale charge feature extraction process in this embodiment includes: the state of charge does not change significantly within a short period. Using short-term changes in the state of charge as input features to the neural network would increase the difficulty of data processing. Conversely, the cumulative changes in the state of charge across multiple timescales will result in a larger change, thus better reflecting the capacity status. The multi-timescale charge feature extraction formula is as follows: In the formula: It is the multi-timescale charge characteristic at time k. It is the state of charge at time k. It represents the state of charge at time kn, where n is the cumulative time scale constant.
[0038] Furthermore, the process of obtaining multi-timescale capacity features includes: Based on the cumulative current over multiple sampling periods prior to time k, the multi-timescale capacity characteristics at time k are obtained, where the sampling period is m.
[0039] Furthermore, the multi-timescale capacity feature extraction process in this embodiment includes: calculating the charge and discharge amounts over a short period of time, which is insufficient to reflect capacity changes. Within a short timeframe, sampling errors and interference can lead to significant errors in the calculated value of the charge change. As the timescale increases, this error gradually decreases, thus better reflecting the correspondence with the capacity state. The multi-timescale capacity feature extraction formula is as follows: In the formula: It is a multi-timescale capacity feature at time k. m The data sampling period.
[0040] Furthermore, such as Figure 3 and Figure 4 As shown, the process of obtaining the state of capacity of a lithium battery includes: The multi-timescale charge features and multi-timescale capacity features are input into a Transformer neural network containing a multi-head attention mechanism, and the capacity state at time k is obtained by calculating the function of the neural network.
[0041] Furthermore, the multi-head attention mechanism is implemented in the following ways: After linearly transforming the query matrix Q, key matrix K, and value matrix V through the weight matrix, the scaled dot product attention is calculated. The outputs of multiple attention heads are then concatenated and linearly transformed to obtain the multi-head attention output.
[0042] Furthermore, the calculation formula for the multi-head attention mechanism in the Transformer neural network of this embodiment is as follows: Where Q is the query matrix, K is the key matrix, and V is the value matrix. It is a scaling factor used to mitigate gradient vanishing, ensuring that the input to the Softmax function remains within a reasonable range.
[0043] Multi-head attention is an extension of the self-attention mechanism. It projects the input data onto multiple subspaces, applies the self-attention mechanism to each subspace, and finally combines the outputs of each subspace using the following linear transformation.
[0044] in, , , This corresponds to the weight matrix Q, K, and V. After completing the multi-head attention mechanism calculation, a feedforward neural network is used to further process the self-attention output: Based on this, the position encoder uses sine and cosine functions of different frequencies to process the sequence information, as shown below: Where PE stands for position encoder, and i is the dimension index in the position encoding. This represents the number of hidden units in the model. sin is the sine function, and cos is the cosine function.
[0045] This leads to capacity state prediction based on multi-timescale features: In the formula: The capacity state at time k; This is a neural network function.
[0046] Furthermore, the Transformer neural network employs a position encoder to process sequence order information; Position encoders use sine and cosine functions of different frequencies to encode position.
[0047] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for assessing the state of capacity of lithium batteries based on a multi-timescale dual neural network, characterized in that, include: Measure and acquire the operating voltage and current data of the lithium battery; The operating voltage and current data are input into an LSTM neural network, and the state of charge of the lithium battery is calculated by the LSTM neural network. Based on the state of charge and the voltage and current data, multi-time-scale charge features and multi-time-scale capacity features are extracted based on multi-time-scale characteristics. The multi-timescale charge characteristics and multi-timescale capacity characteristics are input into a Transformer neural network, and the capacity state of the lithium battery is calculated by the Transformer neural network.
2. The method according to claim 1, characterized in that, The process of measuring and acquiring the operating voltage and current data of a lithium battery includes: The voltage and current data of the lithium battery during charging and discharging are acquired in real time using a data acquisition card with a sampling frequency of 10Hz.
3. The method according to claim 1, characterized in that, The process of obtaining the state of charge (SOC) of a lithium battery includes: The sampled voltage and sampled current at time k are input into a neural network with 5 layers and 128 LSTM units per layer. The state of charge at time k is obtained through the calculation function of the neural network.
4. The method according to claim 3, characterized in that, The LSTM unit controls the flow of information through input gates, forget gates, and output gates, and uses sigmoid and tanh functions as activation functions; the sigmoid function transforms the input features to a non-linear range of 0 to 1, and the tanh function transforms the input features to a non-linear range of -1 to 1.
5. The method according to claim 1, characterized in that, The process of acquiring the multi-timescale charge characteristics includes: Based on the difference between the state of charge at time k and the state of charge at time kn, the multi-timescale charge characteristics at time k are obtained, where n is the cumulative timescale constant.
6. The method according to claim 1, characterized in that, The process of obtaining the multi-timescale capacity features includes: Based on the cumulative current over multiple sampling periods prior to time k, the multi-timescale capacity characteristics at time k are obtained, where the sampling period is m.
7. The method according to claim 1, characterized in that, The process of obtaining the state of capacity of a lithium battery includes: The multi-timescale charge features and multi-timescale capacity features are input into a Transformer neural network containing a multi-head attention mechanism, and the capacity state at time k is obtained by calculating the function of the neural network.
8. The method according to claim 7, characterized in that, The multi-head attention mechanism is implemented in the following ways: After linearly transforming the query matrix Q, key matrix K, and value matrix V through the weight matrix, the scaled dot product attention is calculated. The outputs of multiple attention heads are then concatenated and linearly transformed to obtain the multi-head attention output.
9. The method according to claim 7, characterized in that, The Transformer neural network uses a position encoder to process sequence order information; The position encoder uses sine and cosine functions of different frequencies to encode the position.