Predictive pretrained transformer (PPT) for real-time.pptx

Liderazgo

1. Predictive pretrained transformer (PPT) for real-time battery health diagnostics

2. Highlights •Developed a predictive pretrained Transformer (PPT) for real-time battery health diagnosis. •The multi-model fusion PPT for  transfer learning  across diverse battery types and conditions. •Improved capture of long-range dependencies with multi-head probsparse self-attention. •Achieved substantial computational cost savings and enhanced generalization and scalability. •The bespoke  neural networks  achieve second-level health diagnostics for battery entire life cycle.

3. Abstract Modeling and forecasting the evolution of  battery  systems involve complex interactions across physical, chemical, and  electrochemical processes , influenced by diverse usage demands and dynamic operational patterns. In this study, we developed a predictive pre-trained Transformer (PPT) model equipped with 1,871,114 parameters that enhance identification of both short-term and long-term patterns in time-series data. This is achieved through the integration of  convolutional layers  and probabilistic sparse self-attention mechanisms, which collectively enhance prediction accuracy and efficiency in diagnosing  battery  health. Moreover, the customized hybrid-model fusion supports parallel computing and employs  transfer learning , reducing computational costs while enhancing scalability and adaptability. Consequently, this allows for precise real-time health estimations across various battery cycles. We validated this method using a public dataset of 203 commercial  lithium iron phosphate  (LFP)/graphite batteries charged at rates ranging from 1 C  to 8 C . By using only partial charge data—from an 80 %  state of charge  to the maximum charging voltage (3.6 V for LFP batteries, 4.2 V for ternary batteries)—and avoiding complex feature engineering, error metrics were achieved below 0.3 % for  root mean square error  (RMSE), weighted mean absolute percentage error (WMAPE), and  mean absolute error  (MAE), with an R 2  of 98.9 %. The generalization capabilities were further demonstrated across 36 different testing protocols, encompassing 23,480 cycles throughout the entire life cycle, with a total inference time of 9.88 s during the testing phases. Further experiments on 30 nickel cobalt  aluminum  (NCA) batteries and 36  nickel cobalt manganese  (NCM) batteries, across different battery types and operational scenarios, resulted in RMSE, WMAPE, and MAE all below 0.9 %, with R 2  values of 94.1 % and 94.4 %, respectively. These findings highlight the potential of our customized deep transfer  neural networks  to enhance diagnostic accuracy, accelerate training, and improve generalization in real-time applications.

4. Introduction Lithium-ion batteries are widely used as primary energy storage devices due to their high energy density, high power density, strong environmental adaptability, and low self-discharge characteristics [[1], [2], [3], [4]]. As lithium-ion battery technology continues to mature, significant cost reductions are expected [5,6], driven primarily by advancements in manufacturing processes, economies of scale, and widespread adoption in electric vehicles [7,8] and energy storage applications [9]. The ongoing improvements in battery chemistry, such as higher energy densities and longer cycle life, contribute to more efficient use of materials and resources, further driving down costs. However, their performance gradually declines over time during storage and usage due to their electrochemical properties [10]. Additionally, under complex working conditions such as dynamic cycling, operational conditions, temperature/thermal effects, and other environmental factors, the rate of battery aging accelerates [[11], [12], [13]]. Regardless of battery design, environmental impacts and dynamic cycling expedite battery aging, hindering optimal performance throughout its lifespan [14]. Therefore, accurately monitoring battery degradation and assessing whether its performance meets expected levels is crucial. The state of health (SOH) is a common metric for evaluating battery degradation, reflecting the current energy storage and power supply capability relative to its initial state at the beginning of its lifecycle [15,16]. To maintain the safety and reliability of aging batteries, developing high-precision health diagnostic techniques is essential [17,18].

5. To accurately and reliably estimate the SOH of batteries, researchers have proposed various methods, including direct measurement, model-based approaches, and data-driven techniques [19]. Direct measurement methods are used in laboratory environments to measure capacity and internal resistance directly, such as ampere-hour counting, resistance/impedance methods, electrochemical impedance spectroscopy (EIS), and open circuit voltage methods [[20], [21], [22], [23]]. Although these methods are easy to deploy, their estimation accuracy relies on the precision of the measurement techniques and may not be sufficiently reliable in practical applications. Therefore, they are primarily suitable for offline SOH estimation in laboratory settings [24,25]. In contrast, model-based approaches treat lithium-ion batteries as electrochemical or electrical models to simulate the dynamic systems within batteries and track the degradation process [26]. However, accurate SOH estimation requires a thorough understanding of the degradation processes, which is challenging due to the complexity of battery electrochemical dynamics [27]. Moreover, the accuracy of model-based estimation depends heavily on the precision of battery modeling, making it difficult to balance SOH estimation accuracy and computational complexity in practical scenarios. Conversely, data-driven approaches, especially those based on machine learning, are pivotal in fields such as scientific data analysis, intelligent systems, cognitive science, and artificial intelligence [28,29]. These methods excel in representing and manipulating uncertainty within models and predictions [30]. These intelligent learning methodologies have become powerful tools across a wide range of applications, including classification, regression, anomaly detection, and space- or time-dependent state prediction. They can be applied in supervised, unsupervised, semi-supervised, or self-supervised modes, particularly within the context of Industry 4.0 [31].