Embedded artificial intelligence in battery management systems: Pruning and quantization for efficient state-of-charge and state-of-health estimation

Online battery management requires accurate state-of-charge (SOC) and state-of-health (SOH) estimates within strict memory and timing budgets. To address these constraints, this work investigates how stateful long short-term memory (LSTM) networks with multi-layer perceptron (MLP) heads can be optimized for embedded artificial intelligence (AI), using structured pruning and recurrent-weight quantization to reduce memory and computational costs while retaining estimation accuracy on an STM32H753ZI microcontroller. Using the open Lithium Iron Phosphate (LFP) Cycle Ageing Dataset , full-precision models provide the accuracy reference against which the compressed variants are evaluated. Long-horizon replay yields mean absolute error (MAE) values of 2.3–2.8 percentage points for SOC and 0.9–1.5 for SOH, showing that compression largely preserves accuracy. The evaluated 30% pruning configuration reduces flash by 37.2% and 43.3% in the available firmware builds and inference time by 42.9% and 44.0% in the recorded hardware-in-the-loop benchmarks for SOC and SOH, respectively. The analytical complexity model relates computational savings to pruning fraction and network dimensions. Quantization reduces flash by 49.7% and 55.9% relative to the respective full-precision models, but raises runtime in the mixed-precision implementation. Together, the study provides a reproducible reference for resource-efficient embedded battery-state estimation.