QUANTIZED LSTM FOR PREDICTIVE THERMAL MONITORING OF TEMPERATURE OF LITHIUM-ION BATTERY

The large memory requirements and low inference speed renders the deployment of Long short-term memory (LSTM) impractical for predictive monitoring in battery thermal management system. Thus, this study proposes quantization aware training, where the weights and biases were quantized using ternary quantization scheme and while the activation function were quantized using 8-, 6- and 4-bit level schemes. The various quantization strategies were evaluated with statistical analysis such as mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) and train and validated with oxford battery dataset. In addition, the various bit level schemes were compared in terms of performance.As 4-bit level level offers lowest memory footprint. This space saving is done where 24-bit = 16 values are assigned for the activation unit and unable to cover the whole spectrum compared to full precision, thus results in large mean accuracy loss. On the other hand, 6-bit level quantization proves to be better performance with relatively smaller memory footprint compared to 8-bit level quantization.Thus, 6-bit level quantization is suitable for low power edge application for battery management system.