<![CDATA[Lithium iron phosphate (LiFePO₄) batteries are widely utilized in renewable energy storage systems due to their high thermal stability and long cycle life. However, accurate battery capacity estimation remains a significant challenge, particularly for small-scale batteries with limited datasets. This study proposes a short-term capacity estimation framework based on a Temporal Convolutional Network (TCN) for a 4S LiFePO₄ 12.8 V 6 Ah battery pack under constant-current cycling conditions. A custom data acquisition (DAQ) system was developed using a Raspberry Pi Pico microcontroller integrated with ADS1115, INA226, and DS18B20 sensors to monitor voltage, current, and temperature in real time during charging and discharging experiments. A dataset consisting of 76 cycles was collected experimentally, from which 13 anomalous early-cycle samples were removed due to electrochemical stabilization phenomena, resulting in 63 effective samples. The TCN model was trained using a sliding-window approach with a sequence length of 5 timesteps, incorporating causal and dilated convolutions, residual connections, and dropout regularization. Evaluation on the testing dataset produced an RMSE of 0.02114 Ah, MAE of 0.01288 Ah, and MAPE of 0.827%, indicating high prediction accuracy with an average relative deviation below 1%. The negative value of −5.8753 was attributed to the statistical limitations of the metric when applied to small-scale datasets with low variance. The results demonstrate that the proposed TCN architecture is capable of learning short-term temporal degradation characteristics from limited battery cycling data with high relative accuracy.]]>
