Accurately predicting thermal behavior is critically important in the real-world thermal management of supercapacitor modules with ultrahigh power and discharging current. In this work, an artificial intelligence approach based on the improved multiscale coupled electro-thermal model is employed for the first time to accurately predict the thermal behavior of a 350 F supercapacitor module under air-cooling conditions. Different from previous work that used commercial cells, the 350 F supercapacitors are fabricated from our proprietary pilot-scale production line. This approach provides a platform to precisely measure the structural parameters, electrical and thermal properties of electrodes and electrolytes (e.g., the temperature/current dependent equivalent series resistance and axial/radial thermal characteristics), which can improve the model for characterizing the irreversible heat generation and thermal transport processes. In particular, coupled with molecular dynamics simulations, the molecular origin of entropy is revealed via probing the atomic-level information (e.g., 1D/2D electric double-layer structure, electrical field/potential distributions, areal capacitance, and diffusion kinetics) to accurately predict the reversible heat generation. As a consequence, the deviation between our improved model and experimental results is substantially reduced to below 5%. A deep neural network based on the long short-term memory (LSTM) approach is trained to build a temperature database for practical supercapacitor modules under different operating conditions (including charging/discharging currents, cooling airflow rates, and cycle duration). This work demonstrates the potential of LSTM in predicting the thermal behavior, which can be broadly used for industry-relevant thermal management applications.

在具有超高功率和放电电流的超级电容器模块的实际热管理中，准确预测热行为至关重要。在这项工作中，首次使用基于改进的多尺度耦合电热模型的人工智能方法来准确预测350 F超级电容器模块在空冷条件下的热性能。与以前使用商用电池的工作不同，350 F超级电容器是由我们专有的中试规模生产线制造的。这种方法提供了一个平台，可以精确地测量电极和电解质的结构参数，电和热特性（例如，取决于温度/电流的等效串联电阻和轴向/径向热特性），这可以改善表征不可逆热产生和热传递过程的模型。特别是，结合分子动力学模拟，通过精确探测原子级信息（例如1D / 2D双电层结构，电场/电势分布，面电容和扩散动力学）来揭示熵的分子起源。预测可逆热量的产生。因此，我们改进的模型与实验结果之间的偏差已大大降低到5％以下。经过训练的基于长短期记忆（LSTM）方法的深度神经网络可以为不同操作条件（包括充电/放电电流，冷却气流速率和循环持续时间）下的实际超级电容器模块建立温度数据库。