PID-based CNN-LSTM for accuracy-boosted virtual sensor in battery thermal management system

Battery thermal management is essential to achieve good performance and a long battery system lifespan in electric vehicles and stationary applications. Such a thermal management system is dependent on temperature monitoring, which is frequently hampered by the limited sensor measurements. The virtual sensor is brought forward to overcome this physical restriction and provide broader access to the battery's temperature distribution. Through leveraging the combined convolutional neural network (CNN) and long short-term memory (LSTM) networks to extract both spatial and temporal information from the data, this paper proposes a novel virtual sensing platform. A PID compensator is included to offer auxiliary correction to the inputs and drive the prediction error to zero over time in a feedback loop. Off-line and online modes of this CNN-LSTM virtual sensor are considered. The network, which is trained off-line, will work with the PID compensator in the online mode with real-time sensor data. With the PID-based accuracy-boosted virtual sensor, the performance of the trained CNN-LSTM prediction on real-time data inputs is improved. Besides, this PID compensator reduces the number of hyper-parameters to be tuned. Based on control theory, the design of PID and its analysis are presented as well. With generated battery thermal data, numerical simulations show that the CNN-LSTM-PID virtual sensing framework can achieve the real-time prediction error reduction rate to 35.52% on average with 18.78% less online calculation time.

Automated summary

Battery thermal management is crucial for the performance and lifespan of electric vehicles and stationary applications. This paper proposes a novel virtual sensing platform that utilizes CNN and LSTM networks to extract spatial and temporal information for temperature monitoring. The PID compensator reduces the prediction error over time in a feedback loop. The virtual sensor works in both offline and online modes with real-time sensor data.