A Joint Prediction Method for State of Charge and Life of Lithium Batteries Based on Deep Learning

Abstract

The invention discloses a method for jointly predicting the state of charge and life of a lithium battery based on deep learning, belonging to the technical field of lithium batteries. A full-life cycle charge-discharge cycle life experiment of the lithium battery under specific working conditions is carried out. The deep learning model Autoformer is trained through experimental data to obtain a joint prediction model for the state of charge and remaining service life of the lithium battery. Historical data is input into the joint prediction model to first obtain the predicted SOC and predicted voltage, and then the autocorrelation calculation is performed between the predicted voltage curve and the voltage curve of the first cycle of the lithium battery to obtain the autocorrelation coefficient. The autocorrelation coefficient is input into the remaining service life sub-prediction model to obtain the predicted value of the lithium battery capacity. The invention utilizes the excellent learning ability of Autoformer for ultra-long sequence data to extract the charge-discharge and capacity attenuation characteristics of the lithium battery from full-life cycle data, providing an accurate prediction means for the joint state estimation of the lithium battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium batteries, and specifically relates to a method for jointly predicting the state of charge and life of lithium batteries based on deep learning. Background Art

[0002] With the rapid development of new energy vehicles and energy storage systems, the prediction of the state of health (SOH) and remaining useful life (RUL) of lithium batteries has become a core challenge for battery management systems. In existing technologies, state of charge (SOC) and SOH / RUL prediction are usually modeled separately, including:

[0003] (1) SOC prediction based on electrochemical models: SOC is estimated by combining an equivalent circuit model (such as the Thevenin model) with a Kalman filter algorithm. However, parameter drift caused by battery aging (such as increased internal resistance and increased polarization) is not considered, and the error accumulation is significant after long-term use.

[0004] (2) Data-driven RUL prediction: The mainstream method uses LSTM or CNN to extract capacity decay features, but it is limited by the model's ability to remember long-term degradation sequences (>1000 cycles), making it difficult to capture the gradual-sudden coupling pattern before the capacity "dive";

[0005] (3) Multi-stage decomposition method: EEMD (ensemble empirical mode decomposition) or wavelet transform is used to decompose the voltage / capacity series into trend terms and fluctuation terms, which are then input into a shallow neural network (such as Elman) for prediction. However, the decomposition process relies on manual mode selection, and the dynamic correlation between SOC and capacity attenuation is not established. Summary of the Invention

[0006] In order to solve the problems of the existing technology in the joint prediction of lithium battery SOC and RUL, and the insufficient extraction of long-cycle degradation features, the present invention provides a method for joint prediction of lithium battery state of charge and life based on deep learning. The lithium battery is subjected to a full life cycle charge and discharge cycle life experiment under specific working conditions. The deep learning model Autoformer is trained with experimental data to obtain a joint prediction model of the lithium battery state of charge and remaining service life. The historical data is input into the joint prediction model to first obtain the predicted SOC and predicted voltage. The predicted voltage curve is then autocorrelated with the voltage curve of the first cycle of the lithium battery to obtain the autocorrelation coefficient. The autocorrelation coefficient is input into the remaining service life prediction model to obtain the predicted value of the lithium battery capacity. The present invention utilizes the excellent learning ability of Autoformer for ultra-long sequence data to extract the charge and discharge and capacity decay characteristics of the lithium battery from the full life cycle data, providing an accurate prediction means for the joint state estimation of the lithium battery.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A joint prediction method for the state of charge and life of a lithium battery based on deep learning, comprising the following steps:

[0009] Step 1: Conduct a charge-discharge cycle life experiment on the lithium battery under specified working conditions to obtain experimental data, including current, voltage, temperature, state of charge, and remaining capacity of the lithium battery at different cycle numbers in the full life cycle;

[0010] Step 2: Perform data normalization processing on the experimental data obtained in Step 1 to construct an input data set containing normalized parameters and derivative features;

[0011] Step 3: Based on the Autoformer-Transformer architecture, build a joint prediction model for the state of charge and remaining useful life of the lithium battery; the joint prediction model includes:

[0012] A state-of-charge sub-prediction model, which is built with Autoformer as the architecture and is used to predict the state of charge at future times under real-time working conditions;

[0013] A voltage sub-prediction model, which is built with Autoformer as the architecture and is used to start when the state of charge is lower than a certain value and predict the voltage curve from this moment to the state of charge being 0;

[0014] An autocorrelation coefficient calculation sub-module, which is used to calculate the autocorrelation coefficient between the predicted voltage curve output by the voltage sub-prediction model and the corresponding voltage curve of the first cycle, and input the calculated autocorrelation coefficient into the remaining useful life sub-prediction model;

[0015] A remaining useful life sub-prediction model, which is built with a CNN-Transformer hybrid architecture and is used to predict the capacity decay state in the next several cycles according to the autocorrelation coefficient calculated by the autocorrelation coefficient calculation sub-module;

[0016] Step 4: Input the processed feature data into the joint prediction model built in Step 3 for training and optimization to obtain an optimized joint prediction model;

[0017] Step 5: Input historical data into the optimized joint prediction model in Step 4 to obtain prediction results for the state of charge and remaining life.

[0018] Further, the Step 2 includes:

[0019] S21. Normalize the current, voltage, state of charge, and remaining capacity data obtained from the experiment in Step 1 to obtain normalized parameters;

[0020] S22. Divide the time into six scales of year, month, day, hour, minute, and second according to the timestamp of the data, and use the five periodic features of month, day, hour, minute, and second as the time position encoding of the training data; construct an input data set containing normalization parameters and derivative features.

[0021] Further, the input of the state of charge sub-prediction model includes the current, voltage, current SOC, and cycle number features during the battery cycle process, and the output is the predicted value of the state of charge at a future moment.

[0022] Further, the construction principle of the state of charge sub-prediction model is as follows:

[0023] Utilize the ability of the Autoformer deep learning model to decouple periodic ultra-long time series data into trend terms and seasonal terms, and capture the changes in battery charge and discharge characteristics caused by battery aging and the periodic changes in SOC during the cycle.

[0024] The construction process of the state of charge sub-prediction model is as follows:

[0025] Record the first k historical data of current, voltage, cycle number, and SOC as and input it into the Autoformer encoder. Capture the period-based dependencies of the sequence through the autocorrelation mechanism, and aggregate similar subsequences from the underlying periods; splice the current, voltage, cycle number, and SOC data from the (k - l)-th to the k-th time and p zeros as the input of the encoder, denoted as , where l is the amount of labeled data and p is the amount of predicted data; record the true SOC data from the k-th time to the (k + l)-th time as as the label, and calculate the loss with the output of Autoformer to obtain the loss of this training and backpropagate to optimize the model parameters; obtain the state of charge sub-prediction model after multiple trainings.

[0026] Further, the input of the voltage sub-prediction model is the current, voltage, SOC, cycle number, and remaining capacity features of each cycle, and the output feature is the predicted value of the voltage at a future moment.

[0027] Further, the construction principle of the voltage sub-prediction model is as follows:

[0028] The Autoformer deep learning model decouples the ultra-long sequence data with periodic characteristics into trend components and seasonal components through the time domain decomposition mechanism, and respectively models the drift of charge and discharge characteristics induced by battery aging and the periodic pattern of voltage fluctuations during the cycle; through the joint characterization of the two trends, multi-scale prediction of battery degradation and cycle dynamic evolution is achieved.

[0029] Further, the calculation method of the autocorrelation coefficient is as follows:

[0030] Align the predicted voltage curve output by the voltage sub-prediction model with the voltage curve of the first cycle through SOC, and make the two voltage curves have the same length and be aligned in time through normalizing the time axis and linear interpolation. Then splice the two voltage curves within the same SOC change value together, and set the lag k to half of the total length of the curve, that is , and the calculation formula for the autocorrelation coefficient Acf is as follows:

[0031] where , , is the i-th voltage value of the spliced voltage curve, is the i + k-th voltage value of the spliced voltage curve, and N is the total amount of spliced voltage data.

[0032] Further, the input features of the remaining useful life sub-prediction model include the number of cycles, the remaining capacity, and the voltage autocorrelation coefficient output by the autocorrelation coefficient calculation sub-module, and the output is the predicted value of the future capacity decay trajectory.

[0033] Further, the construction principle of the remaining useful life sub-prediction model is as follows:

[0034] The CNN module extracts short-term mutation features in the capacity fluctuation through local convolution operations, and the Transformer module uses the global self-attention mechanism to model the long-range dependence relationship between the number of cycles, the autocorrelation coefficient, and the capacity decay;

[0035] The construction process of the remaining useful life sub-prediction model is as follows:

[0036] Embed the number of cycles and the voltage autocorrelation coefficient of the historical k cycles into the encoder through CNN, and capture the global dependence relationship through the self-attention module; use the capacity offset by 1 cycle as the decoder input, and establish a non-linear relationship between the number of cycles, the voltage autocorrelation coefficient, and the cycle capacity after self-attention and cross-attention, and output the capacity prediction value.

[0037] Further, step five includes: inputting historical data into the optimized joint prediction model, first obtaining the predicted SOC and predicted voltage, then calculating the autocorrelation between the predicted voltage curve and the voltage curve of the first cycle of the lithium battery through the autocorrelation coefficient calculation sub-module to obtain the autocorrelation coefficient, and inputting the autocorrelation coefficient into the remaining useful life sub-prediction model to obtain the predicted value of the lithium battery capacity.

[0038] The present invention has the following beneficial effects:

[0039] The present invention utilizes the ability of Autoformer to capture the characteristics of ultra-long time-series data, extracts the charge-discharge characteristics and capacity attenuation characteristics of lithium batteries from the full-life cycle data, innovatively proposes using the voltage autocorrelation coefficient as the training feature, synchronously analyzes the coupling relationship between charge-discharge characteristics and capacity attenuation through Autoformer ultra-long sequence modeling, realizes the joint prediction of SOC-RUL and early fault warning, and breaks through the technical bottlenecks of traditional methods in long-cycle modeling and dynamic coupling analysis. Description of the Drawings

[0040] Figure 1 is a schematic diagram of the overall process of a method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to the present invention.

[0041] Figure 2 is a comparison chart of the normalized voltage autocorrelation coefficient and the battery capacity in an embodiment of the present invention.

[0042] Figure 3 is a schematic diagram of the principle of a joint prediction model for the state of charge and remaining useful life of the lithium battery according to the present invention.

[0043] Figure 4 is a schematic diagram of the state-of-charge sub-prediction model and voltage sub-prediction model of the lithium battery in an embodiment of the present invention.

[0044] Figure 5 is a schematic diagram of the remaining useful life sub-prediction model of the lithium battery in an embodiment of the present invention.

[0045] Figure 6 is a comparison chart of single-step capacity prediction with and without autocorrelation coefficient in the input features in an embodiment of the present invention.

[0046] Figure 7 is a comparison chart of ten-step capacity prediction with and without autocorrelation coefficient in the input features in an embodiment of the present invention. Detailed Embodiments

[0047] The present invention will be further explained below in conjunction with embodiments and drawings.

[0048] In this embodiment, the data of graphite / lithium batteries is taken as an example, and a joint prediction model for the state of charge and remaining useful life of lithium batteries is built and trained to describe the present invention comprehensively and in detail. This method is not limited to graphite / lithium batteries and is applicable to all commercial and self-made lithium batteries.

[0049] This embodiment is a method for jointly predicting the state of charge and life of a lithium battery based on deep learning. As Figure 1 shown, it mainly includes five steps:

[0050] Step 1: Conduct charge-discharge cycle life experiments on lithium batteries under specified working conditions to obtain data such as current, voltage, temperature, state of charge, and remaining capacity at different cycle numbers during the full life cycle of the lithium battery, as shown in Table 1;

[0051] Table 1. Charge-discharge cycle life experiment data

[0052] month day hour minute s Number of cycles Charge and discharge state Current / A Voltage / V SOC Capacity / Ah 8 8 0 49 5 1 0 0 4.1962 1 1.4527 8 8 0 49 15 1 -1 -0.749 4.1115 1 1.4527 8 8 0 49 25 1 -1 -0.749 4.1031 0.998 1.4527 8 8 0 49 35 1 -1 -0.749 4.0962 0.997 1.4527 8 8 0 49 45 1 -1 -0.749 4.0905 0.995 1.4527

[0053] Step 2: Perform data standardization processing on the experimental data obtained in Step 1. Normalize data such as current, voltage, state of charge, and remaining capacity to construct an input data set containing standardized parameters and derivative features;

[0054] Specifically, Step 2 includes:

[0055] S21. Normalize the data such as current, voltage, state of charge, and remaining service life obtained from the experiments in Step 1, so that data of different scales are scaled to the interval [0,1], facilitating the model to learn the features of data between different scales;

[0056] The data obtained from the charge-discharge cycle life experiments are generally current, voltage, temperature, cycle number, SOC, and battery capacity after each cycle with timestamps. For current, voltage, SOC, and remaining capacity, the data of different scales are unified through normalization calculation to avoid the situation that the model can only learn the changes of a certain feature due to the overly large change range of a certain feature. The normalization calculation formula is as follows:

[0057] In the formula, is the normalized data, is the input data, is the minimum value of the input data, is the maximum value in the input data.

[0058] S22. Divide the time into 6 scales according to year, month, day, hour, minute, and second according to the timestamps of the data. Use the 5 periodic features of month, day, hour, minute, and second as the time position encoding of the training data; construct an input data set containing standardized parameters and derivative features.

[0059] During the implementation of this embodiment, in order to verify the non-linear relationship between the voltage autocorrelation coefficient and the capacity attenuation curve, the voltage curve of the first cycle experiment is used as the benchmark, and the voltage curves of each subsequent cycle are calculated for the autocorrelation coefficient with the voltage curve of the first cycle. An autocorrelation coefficient can be obtained for each cycle. After calculating and normalizing the autocorrelation coefficients of all cycles, it can be seen that there is a certain non-linear relationship between the autocorrelation coefficient and the capacity attenuation curve, as Figure 2 shown.

[0060] Step 3: Based on the Autoformer-Transformer architecture, build a joint prediction model for the state of charge (SOC) and remaining useful life (RUL) of a lithium battery. The joint prediction model synchronously integrates an SOC sub-prediction model, a voltage sub-prediction model, a remaining life sub-prediction model, and an autocorrelation coefficient calculation module. Since the historical m-cycle data and the historical k-time data are time-series data of different time scales, the state parameters and the remaining useful life (RUL) are associated and modeled by sharing the time-series feature layer. As Figure 3 shown, the joint prediction model includes:

[0061] (1) The SOC sub-prediction model, that is, the SOC sub-prediction model, which is built with Autoformer as the architecture. As Figure 4 shown, the input of the SOC sub-prediction model includes features such as current, voltage, current SOC, and cycle number during the battery cycle, and the output is the predicted value of the SOC at a future time. The SOC prediction model is used for second-level data prediction, that is, for predicting the SOC at a future time under real-time working conditions.

[0062] The Autoformer deep learning model can decouple periodic ultra-long time-series data into trend terms and seasonal terms, and capture the changes in battery charge and discharge characteristics caused by battery aging and the periodic changes in SOC during cycling. The model architecture is as shown in the attached Figure 3 specification. The model records the first k historical data of current, voltage, cycle number, and SOC as and inputs it into the Autoformer encoder. Through the autocorrelation mechanism, the sequence-based periodic dependence relationship is captured, and similar subsequences are aggregated from the underlying cycles. The data of current, voltage, cycle number, and SOC from the (k - l)-th to the k-th time and p times of 0 data are concatenated as the input of the encoder, denoted as , where l is the amount of labeled data and p is the amount of predicted data. The true SOC data from the k-th to the (k + l)-th time is denoted as as the label, and the loss is calculated with the output of Autoformer to obtain the loss of this training and backpropagate to optimize the model parameters. After sufficient training, the SOC sub-prediction model is obtained.

[0063] (2) The voltage sub-prediction model, which is built with Autoformer as the architecture. As Figure 4 shown, the input of the voltage prediction model is features such as current, voltage, SOC, cycle number, and remaining capacity for each cycle, and the output feature is the predicted value of the voltage at a future time. The voltage prediction model is also used for second-level data prediction and is activated when the SOC is lower than a certain value to predict the voltage curve from this moment to when the SOC is 0.

[0064] The Autoformer deep learning model decouples ultra-long sequence data with periodic characteristics into trend components and seasonal components through a time-domain decomposition mechanism, and separately models the charge-discharge characteristic drift induced by battery aging and the periodic pattern of voltage fluctuations during cycling. This architecture realizes multi-scale prediction of battery degradation and cyclic dynamic evolution by jointly characterizing the two trends. In the data input process, except that the input features are changed to current, voltage, number of cycles, SOC, and remaining capacity, the rest is the same as described in the state of charge prediction model. Taking the capacity of the lithium battery as the prediction feature of voltage can increase the robustness of the model and improve the prediction accuracy of the model at different numbers of cycles.

[0065] (3) The autocorrelation coefficient calculation sub-module, which is used to calculate the autocorrelation coefficient between the predicted voltage curve output by the voltage sub-prediction model and the corresponding voltage curve of the first cycle, and input the calculated autocorrelation coefficient into the remaining useful life sub-prediction model for future RUL prediction. The calculation method of the autocorrelation coefficient Acf is as follows:

[0066] Align the predicted voltage curve output by the voltage sub-prediction model with the voltage curve of the first cycle through SOC, and make the two voltage curves have the same length and be aligned in time through normalizing the time axis and linear interpolation processing. Then splice the two voltage curves within the same SOC change value together, and set the lag k to half of the total length of the curve, that is The following calculation formula can be obtained:

[0067] Where is the i-th voltage value of the spliced voltage curve, is the i + k-th voltage value of the spliced voltage curve, and N is the total amount of spliced voltage data.

[0068] (4) The remaining useful life sub-prediction model, which is built with a CNN-Transformer hybrid architecture, as shown in Appendix Figure 5 The input features include the number of cycles, the remaining capacity, and the autocorrelation coefficient of the lithium battery voltage (Acf) output by the autocorrelation coefficient calculation sub-module, and the output is the predicted value of the future capacity decay trajectory. The remaining useful life prediction model is used for data prediction between different cycles. After the voltage prediction model outputs the autocorrelation coefficient, it predicts the capacity decay state in the next several cycles.

[0069] The CNN module extracts short-term mutation features in capacity fluctuations through local convolution operations, while the Transformer module utilizes the global self-attention mechanism to model the long-range dependencies among the number of cycles, autocorrelation coefficient, and capacity decay. The hybrid architecture model embeds the number of cycles and voltage autocorrelation coefficient of the historical k cycles into the encoder through CNN, and captures global dependencies through the self-attention module; takes the capacity offset by 1 cycle as the decoder input, and establishes a non-linear relationship among the number of cycles, voltage autocorrelation coefficient, and cycle capacity through self-attention and cross-attention, and outputs the capacity prediction value. Compared with Autoformer that relies on the fixed trend / cycle assumption of sequence decomposition, this architecture can capture complex patterns of non-periodic decay and avoid the strong dependence on the steady-state time series prior of Autoformer through the synergistic effect of CNN-Transformer, thus significantly improving the adaptability to the non-stationary degradation process of lithium batteries.

[0070] Step 4: Input the processed feature data into the joint prediction model established in Step 3 for training, and iteratively update the weight parameters through the backpropagation algorithm to enable the joint prediction model to synchronously learn the dynamic change law of SOC and the capacity decay trend, and finally obtain an optimized joint prediction model.

[0071] Step 5: Input the historical data into the optimized joint prediction model in Step 4 to obtain the prediction results of the state of charge and remaining useful life.

[0072] According to the number of cycles and battery capacity of the input historical m cycles, and the current, voltage, and SOC data at historical k moments, use the joint prediction model to predict the RUL of the next n cycles and the SOC of the next l moments, and achieve the joint prediction of SOC and RUL.

Claims

1. A method for joint prediction of lithium battery state of charge and life based on deep learning, characterized in that: The following steps are involved: Step 1: Conduct a charge-discharge cycle life test on a lithium battery under specified operating conditions to obtain experimental data, including the current, voltage, temperature, state of charge, and remaining capacity of the lithium battery at different cycle times throughout its life cycle; Step 2: Perform data normalization on the experimental data obtained in step 1 to construct an input data set containing normalization parameters and derived features; Step 3: Based on the Autoformer-Transformer architecture, a joint prediction model for the state of charge and remaining service life of the lithium battery is built; the joint prediction model includes: The state-of-charge prediction model, built on the Autoformer architecture, is used to predict the state of charge at future times under real-time operating conditions. The voltage sub-prediction model, built on the Autoformer architecture, is used to start when the state of charge drops below a certain value and predict the voltage curve from that moment to the state of charge reaching 0; An autocorrelation coefficient calculation submodule, which is used to calculate the autocorrelation coefficient between the predicted voltage curve output by the voltage sub-prediction model and the corresponding voltage curve of the first cycle, and input the calculated autocorrelation coefficient into the remaining service life sub-prediction model; A remaining useful life sub-prediction model, built using a CNN-Transformer hybrid architecture, is used to predict the capacity decay state for several future cycles based on the autocorrelation coefficient calculated by the autocorrelation coefficient calculation sub-module; Step 4: Input the processed feature data into the joint prediction model built in step 3 for training and optimization to obtain an optimized joint prediction model; Step 5: Input the historical data into the joint prediction model optimized in step 4 to obtain the prediction results of state of charge and remaining life.

2. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The second step includes: S21. The current, voltage, state of charge, and remaining capacity data obtained in step 1 of the experiment are normalized to obtain standardized parameters; S22. Divide the time into six scales according to the data's timestamp: year, month, day, hour, minute, and second. Use the five periodic features of month, day, hour, minute, and second as the temporal position encoding for the training data. Construct an input dataset containing standardized parameters and derived features.

3. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The state of charge sub-prediction model input includes the current, voltage, current SOC and cycle number characteristics of the battery cycle process, and the output is the state of charge prediction value at a future time.

4. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 3, characterized in that: The construction principle of the state of charge prediction model is as follows: Leveraging the Autoformer deep learning model’s ability to decouple periodic time series data into trend and seasonal terms, it captures changes in battery charge and discharge characteristics due to battery aging and periodic changes in SOC during cycling. The process of building the state of charge prediction model is as follows: The previous k historical data of current, voltage, number of cycles, and SOC are recorded as x enc And input the Autoformer encoder, which captures the cycle-based dependencies of the sequence through the autocorrelation mechanism and aggregates similar subsequences from the underlying cycle; The current, voltage, number of cycles, SOC data from kl to k and the zero data of p are spliced as the input of the decoder, denoted as x dec , where l is the amount of label data and p is the amount of predicted data; the real SOC data from the kth to the k+lth time is recorded as SOC true As a label, and the output SOC of Autoformer predict Perform loss calculation to obtain the loss of this training and backpropagate to optimize the model parameters; after multiple training sessions, the state of charge prediction model is obtained.

5. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The voltage sub-prediction model inputs the current, voltage, SOC, number of cycles, and remaining capacity characteristics of each cycle, and the output characteristics are the voltage prediction value at a future moment.

6. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 5, characterized in that: The construction principle of the voltage quantum prediction model is: The Autoformer deep learning model decouples sequence data with periodic characteristics into trend components and seasonal components through a time domain decomposition mechanism, modeling the charge and discharge characteristic drift induced by battery aging and the periodic pattern of voltage fluctuation during the cycle respectively; by jointly characterizing the two trends, multi-scale prediction of battery degradation and cycle dynamic evolution is achieved.

7. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The calculation method of the autocorrelation coefficient is: The predicted voltage curve output by the voltage sub-prediction model is aligned with the voltage curve of the first cycle by SOC. The two sets of voltage curves are made to have the same length and aligned in time by normalizing the time axis and linear interpolation. Then, the two sets of voltage curves within the same SOC change value are spliced together, and the hysteresis k is set to half of the total length of the curve, that is, The calculation formula for the autocorrelation coefficient Acf is as follows: in, V i is the i-th voltage value of the voltage curve after splicing, V i+k is the i+kth voltage value of the voltage curve after splicing, and N is the total amount of voltage data after splicing.

8. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The input features of the remaining service life sub-prediction model include the number of cycles, the remaining capacity, and the voltage autocorrelation coefficient output by the autocorrelation coefficient calculation sub-module, and the output is the predicted value of the future capacity attenuation trajectory.

9. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 8, characterized in that: The construction principle of the remaining useful life sub-prediction model is as follows: The CNN module extracts short-term mutation features in capacity fluctuations through local convolution operations, while the Transformer module uses a global self-attention mechanism to model the long-range dependency between the number of cycles, autocorrelation coefficient, and capacity decay. The process of building the remaining useful life sub-prediction model is as follows: The number of cycles and voltage autocorrelation coefficients of the historical k cycles are embedded in CNN and input into the encoder, and the global dependency is captured through the self-attention module; the capacity offset by 1 cycle is used as the input of the decoder, and after self-attention and cross-attention, a nonlinear relationship between the number of cycles, voltage autocorrelation coefficient and cycle capacity is established, and the capacity prediction value is output.

10. The method for jointly predicting the state of charge and life of a lithium battery based on deep learning according to claim 1, characterized in that: The step five includes: inputting historical data into the optimized joint prediction model to first obtain the predicted SOC and predicted voltage, and then performing autocorrelation calculation on the predicted voltage curve and the voltage curve of the first cycle of the lithium battery through the autocorrelation coefficient calculation submodule to obtain the autocorrelation coefficient, and inputting the autocorrelation coefficient into the remaining service life sub-prediction model to obtain the predicted value of the lithium battery capacity.

Images