State-of-charge estimation of IoT devices using genetic algorithm-optimized long short-term memory networks under real-world conditions

Accurate State of Charge (SOC) estimation is crucial for remaining energy estimation and remaining useful life prediction. While in application scenarios such as Electric Vehicles (EVs), accurate SOC estimation can help to plan journeys, reducing range anxiety; for WSN and IoT devices on the other hand, accurate SOC estimation may result in prolonged device lifetime. In this regard, most of the techniques consider battery terminal voltage, capacity, and even EMF to estimate the SOC, but mostly ignore the real-world conditions. To address this gap, a controlled experimental corpus covering three battery chemistries (Li-Ion, Li-Po, and Ni-MH), three constant discharge currents (30, 50, and 100 mA), and five ambient temperatures (5°C, 15°C, 25°C, 35°C, and 45°C) was assembled. Each condition was repeated three times, where cycles with anomalies were excluded to quantify reproducibility. A Long Short-Term Memory (LSTM) model tuned with a Genetic Algorithm (GA) was trained per chemistry using five-fold cross-validation, with optimization targeting hyperparameters (including layer depth, hidden units, dropout, learning rate, and batch size).The proposed model (LSTM-GA) provided an average low Mean Absolute Value (MAE) of 0.349, 0.390, and 0.412 for Li-Ion, Li-Po, and Ni-MH battery chemistry, respectively. The proposed technique also offers a complete adaptive, scalable code pipeline for model training and deployment. Finally, training cost and inference latency were quantified on NVIDIA Titan V and RTX 4070 GPUs (running TensorFlow 2.20 with CUDA 13), with latencies shown to be appropriate for embedded platform deployment.