7.9.26. Battery State of Charge (SOC) Estimation

Estimate the state of charge of a lithium-ion cell from voltage, current, temperature and accumulated capacity on the MSPM0G5187 microcontroller with its integrated NPU. New in 1.5.0.

Example folder: examples/bms_soc_estimation (GitHub)

7.9.26.1. Overview

Task: timeseries regression (generic_timeseries_regression)
Application: battery management systems (BMS), prevention of over-charge and over-discharge, remaining runtime estimation, battery life extension
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder)
Model: REGR_20k_NPU (about 20,000 parameters, INT8 NPU quantization)
Target: SOC

Traditional coulomb counting and open-circuit-voltage lookup drift over time and are sensitive to sensor noise and battery aging. A short window of raw voltage, current and temperature is not enough either, because SOC depends on the charge and discharge history. The model therefore also receives the accumulated charge (capacity_Ah) as an input.

7.9.26.2. The Dataset

The dataset is the public LG 18650HG2 Li-ion battery data (LG_18650HG2_Liion_Battery_Data), charge and discharge cycles logged at fixed intervals, split into 4491 files. Each file provides four input variables in this column order:

voltage_V
current_A
temperature_C
capacity_Ah

The target is SOC. The config downloads https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/LG_18650HG2_Liion_Battery_Data.zip. The layout is the regression layout (files/ plus annotations/), see Regression Dataset Format.

7.9.26.3. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/bms_soc_estimation/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\bms_soc_estimation\config_MSPM0.yaml

7.9.26.4. Configuration

common:
  target_module: timeseries
  task_type: generic_timeseries_regression
  target_device: MSPM0G5187
  run_name: '{date-time}/{model_name}'
dataset:
  dataset_name: LG_18650HG2_Liion_Battery_Data
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/LG_18650HG2_Liion_Battery_Data.zip
data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 16
  stride_size: 0.5
  variables: 4
training:
  enable: true
  model_name: REGR_20k_NPU
  quantization: 2
  num_gpus: 1
  training_epochs: 50
  batch_size: 128
  learning_rate: 0.001
testing: {}
compilation: {}

quantization: 2 selects INT8 NPU-optimized quantization, and compilation with the TI Neural Network Compiler generates mod.a and tvmgen_default.h for the MSPM0G5187 NPU. num_gpus: 1 expects a GPU; use 0 for CPU.

7.9.26.5. Results

The example readme reports the following for REGR_20k_NPU on the full dataset (best run):

Metric	Value
Test RMSE	2.56% SOC (0.0256)
Test R2	0.99
Best-epoch R2 during training	0.987
Flash on MSPM0G5187	26.39 KB (code 2.43 KB, read-only data 21.52 KB, read-write data 2.45 KB)

The readme sets a stretch target of RMSE below 2% SOC with R2 above 0.99.

Note

The readme results were measured with an earlier release. In the automated 1.5.0 config test run config_MSPM0.yaml failed while loading the dataset (the annotation list file was not found, IndexError in load_list), so these results are unverified in 1.5.0.

7.9.26.6. Output Location

Results are written to ./data/projects/LG_18650HG2_Liion_Battery_Data/run/<date-time>/REGR_20k_NPU/. The device files are compilation/artifacts/mod.a, compilation/artifacts/tvmgen_default.h and the golden vectors in training/quantization/golden_vectors/.

7.9.26.7. Running on Device

Use the MSPM0 SDK (2.08.00 or later) with Code Composer Studio and follow Deploying Regression Models from ModelMaker to Device and Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check). For the MSPM0G5187 NPU see Supported Devices.