EV Battery State-of-Charge Estimation: Coulomb Counting, OCV Lookup and Kalman Filtering on Public Data

SOC-Bench is a declared-protocol benchmark of three state-of-charge (SOC) estimation algorithms for electric-vehicle lithium-ion batteries, executed entirely on public, citable laboratory datasets from the NASA Prognostics Center of Excellence (PCoE) and the CALCE Battery Group at the University of Maryland.

SOC is the quantity every battery management system must report, and India's EV fleet — dominated by two- and three-wheelers — is extremely sensitive to range anxiety and premature cell ageing, both of which trace back to SOC error. Yet student projects typically demonstrate one estimator on one dataset and declare victory. This study instead implements three estimators in one framework — Coulomb counting with periodic OCV re-anchoring, OCV-lookup with resistive-drop correction, and an Extended Kalman Filter (EKF) on a first-order equivalent-circuit model — and scores them on the same cells, the same current profiles (including CALCE's dynamic DST, FUDS and US06 drive-profile tests) and the same metrics.

The evaluation protocol is frozen before any estimator is scored, in the spirit of 22RMI16: "truth" SOC is defined by OCV-anchored capacity determination, three cold-start SOC-error trials (±10 %, ±20 %) test convergence, RMSE and post-settling maximum error are pre-declared, and a parameter-sensitivity sweep (±20 % on R0, R1, C1) answers how much identification quality the EKF actually needs.

The deliverable is a paper-style study with a reproducible Python pipeline — the honest comparison an M.Tech viva can defend, with every number traceable to a public dataset and a fixed random seed.

EV battery management systems must report state-of-charge accurately under drive-cycle loads, cold starts with unknown initial SOC, and imperfect model parameters, yet published student-level work overwhelmingly demonstrates a single estimator on a single dataset with a single run — no cold-start trials, no parameter sensitivity, no declared truth definition — making accuracy claims unverifiable and comparisons between papers meaningless. There is no open, reproducible benchmark that implements Coulomb counting, OCV lookup and Extended Kalman Filtering in one framework, scores them on the same public NASA PCoE and CALCE cells under identical dynamic drive profiles, and quantifies (a) steady-state accuracy, (b) convergence from deliberately wrong initial SOC, and (c) sensitivity to equivalent-circuit parameter error. This dissertation fills that gap with a frozen evaluation protocol, a reproducible Python pipeline, and results tables that regenerate from public data — giving the Indian EV community an honest baseline for the estimator decision every BMS designer faces.

Abstract

This dissertation benchmarks three state-of-charge estimation algorithms for EV lithium-ion batteries on public datasets: Coulomb counting with periodic OCV re-anchoring, open-circuit-voltage lookup with resistive-drop correction, and an Extended Kalman Filter on a first-order equivalent-circuit model. Data come from the NASA PCoE and CALCE battery archives, whose DST/FUDS/US06 profiles emulate realistic drive-cycle loading. A pre-declared protocol defines truth SOC, evaluation windows, metrics (RMSE, post-settling maximum error, convergence time) and cold-start error trials, so every reported number regenerates from source data and fixed seeds. The EKF is expected to win under dynamic loads and large initial errors; the study quantifies the margin, the cost in parameter-identification effort, and the failure modes of the simpler estimators.

Introduction

An electric-vehicle battery management system owns one question — how much energy is left? — and answers it with a SOC estimate that must stay within a few percent across temperatures, ageing and aggressive load transients. Over-estimating SOC strands the rider; under-estimating it wastes usable capacity and feeds the range anxiety that Indian two- and three-wheeler buyers report most. The algorithms are classical, but their relative merits depend on data quality, model identification and initial-condition error — exactly what student projects rarely control.

Existing system and literature gap

The SOC literature is enormous, but accessible published comparisons either use private data, compare estimators under different model assumptions, or report a single favourable run without cold-start trials or parameter sensitivity. Industry BMS documentation rarely quantifies expected accuracy under drive-cycle loads. What is missing is a like-for-like benchmark on fully public, citable data with a declared protocol — same cells, truth definition, metrics and initial-error trials — so that any reviewer can regenerate every figure. This study provides that for three estimator families spanning the practical spectrum.

Proposed work

• Parse the NASA PCoE and CALCE battery archives into a documented, versioned store; verify dataset integrity against published test descriptions.
• Identify a first-order (R0–R1–C1) equivalent-circuit model per cell by non-linear least squares; report parameter fits and residuals.
• Implement the three estimators behind one interface, each consuming only what a real BMS would see — terminal current and voltage.
• Evaluate on the dynamic cycles under the declared protocol: three initial-SOC error trials per cycle, pre-declared metrics, EKF consistency checks and a ±20 % parameter-perturbation sweep.
• Report per-cycle and aggregate results with residual diagnostics, and an operational recommendation for a low-cost Indian EV BMS.

Feasibility

• Technical: all estimators run in Python on a 16 GB laptop (FilterPy for the EKF); the full benchmark takes minutes per cell.
• Data: both sources are public and citable (NASA PCoE repository; CALCE open battery data with required citation), removing the private-database objection entirely.
• Economic: zero licence cost; no hardware or cell cycling is required.
• Operational (research): the protocol is scripted and frozen; every results table regenerates with two commands, which is the reproducibility evidence M.Tech panels increasingly demand.

In scope

• Cell-level estimation on public datasets: NASA PCoE 18650 cells and CALCE CS2/CX2 prismatic cells, including their dynamic DST/FUDS/US06 drive-profile tests and OCV/capacity characterisation runs.
• Three estimator families: Coulomb counting (with OCV re-anchoring), OCV lookup (with R0 correction), EKF on a first-order equivalent-circuit model.
• Declared protocol: truth SOC from OCV-anchored capacity, cold-start error trials, pre-declared metrics, EKF consistency checks, parameter-sensitivity sweep.
• Room-temperature data only, as published in the source datasets.

Out of scope

• Pack-level estimation, cell balancing and thermal modelling beyond what the datasets provide.
• Ageing/health (SOH) prediction and remaining-useful-life forecasting — separate research questions.
• Hardware prototyping or embedded implementation on a real BMS microcontroller.
• New estimator design (e.g. unscented/particle filters appear only as future scope); the contribution is the benchmark, not a novel filter.

The work follows a quantitative, comparative-experimental research design in six phases over 22 weeks, with the evaluation protocol frozen at the end of Phase 2 — before any estimator is scored.

Validation inside the study: synthetic replay tests (run the identified model forward, confirm each estimator recovers the planted SOC), EKF consistency checks (NIS within chi-square bounds), coulomb-counting sign/direction unit tests, and a full re-run reproducing all tables from fixed seeds.

The pipeline keeps data, model identification and estimation strictly separated, so the benchmark cannot accidentally train and test on the same artefact.

flowchart TD
  A["NASA PCoE + CALCE archives\ncurrent, voltage, temperature"] --> B["Parser and QC\nintegrity vs test descriptions"]
  B --> C[("Versioned data store\nper-cell time series")]
  C --> D["OCV-SOC curve + capacity\ncharacterisation per cell"]
  C --> E["ECM identification\nR0, R1, C1 by least squares"]
  D --> F["Truth-SOC builder\nOCV-anchored"]
  E --> G["Estimator interface\nsees only current + voltage"]
  F --> H["Scoring layer\nRMSE, max error, convergence"]
  G --> H
  H --> I["Cold-start trials\n±10 %, ±20 % initial error"]
  H --> J["Parameter sweep\n±20 % on R0, R1, C1"]
  I --> K["Results tables and figures"]
  J --> K

Three design decisions carry the viva. First, the estimator interface exposes only terminal current and voltage — never truth SOC, never the identified parameters of the cell under test unless the estimator legitimately needs them — so information leakage is structurally impossible and a unit test enforces it. Second, truth SOC is defined once, from OCV-anchored capacity determination on the characterisation runs, and the same definition scores every estimator; a benchmark with a per-estimator truth would be meaningless. Third, the scoring layer is the only place metrics are computed, from paired error series with declared settling windows; the cold-start and parameter-sweep runs are re-parameterisations of the same harness with recorded configurations, so the protocol in the dissertation is quoted verbatim from code rather than re-typed.

• Add an unscented or particle filter as a fourth estimator and test whether nonlinearity actually costs accuracy at cell level.
• Repeat the benchmark on temperature-segmented data (the CALCE storage/temperature tests) to study cold-weather estimator degradation.
• Extend to state-of-health co-estimation (dual Kalman filter) on the same public cells.
• Study LFP chemistry explicitly: its flat OCV plateau breaks OCV lookup, making the EKF's model-based advantage largest — quantify it.
• Port the winning estimator configuration to an embedded target and measure real-time cost per cycle for a low-cost Indian two-wheeler BMS.