Reference

Glossary

Every term used on this site, in plain language. Dotted-underlined words elsewhere on the site link back here.

Basics

State of charge (SOC): How full the battery is, from 0 % (empty) to 100 % (full) — the EV equivalent of a fuel gauge.Unlike a fuel tank, SOC cannot be measured directly. It has to be estimated from what a battery management system can measure: current, voltage and temperature. Getting it wrong means either stranding the driver or leaving usable range on the table.
SOC estimator: An algorithm that turns measured current, voltage and temperature into an SOC estimate, once per second.Estimators range from simple current integration (Coulomb counting) to model-based filters and neural networks. On this site an estimator is a MATLAB function Model(X, z) that is called for every 1 Hz sample.
Cell: A single battery unit; a Tesla Model 3 pack contains 2,976 of the 2170 cells tested here.The dataset is measured on individual cylindrical 2170 cells (21 mm diameter, 70 mm long). Pack-level behaviour is derived by scaling the vehicle power demand down to one cell.
Battery management system (BMS): The on-board electronics that monitor a battery and run the SOC estimator in a real vehicle.

Metrics

RMSE: Root-mean-square error between estimated and true SOC over a cycle, in % SOC. The main accuracy metric; lower is better.RMSE penalises large errors more than small ones, so a model that is mostly right but occasionally far off scores worse than one with small steady errors.
MAE: Mean absolute error — the average size of the SOC error over a cycle, in % SOC.
Max error (MAXE): The single worst instantaneous SOC error in a cycle. Shows how badly a model can be wrong even if it is usually accurate.
Weighted error: The leaderboard score: a weighted average of the test-case RMSE values, so hard conditions and robustness count as much as easy ones. Lower is better.Every test type (blinded cell, charging, loads, drive-cycle type, temperature, initial-SOC error, sensor offset) carries weight 0.1, split across its sub-cases; 'all cells' is weighted 0 because it is already contained in the others. Weights sum to 1, so the score reads like an RMSE in % SOC.
Complexity: A 1–10 classification of how much computation a model needs in the evaluator (1 = trivial coulomb counter, 10 = extreme).
Test case: A named subset of the blinded cycles (e.g. '−20 °C' or 'charging') whose RMSE values are averaged into one number.
Initial SOC error: A robustness test: the estimator starts while the true SOC is 90, 60 or 30 % instead of 100 %, as if the vehicle woke from an unknown state.
Current sensor offset: A robustness test: a constant ±0.1 A or ±0.3 A error is added to the measured current, which makes pure coulomb counting drift.

Tests & data

Drive cycle: A standard speed-vs-time profile (e.g. UDDS = city driving, HWFET = highway) converted into the current a cell would see.UDDS (urban), HWFET (highway), LA92 (aggressive urban) and US06 (high-speed, high-acceleration) are US EPA certification cycles. HWCUST and HWGRADE are custom 100–130 km/h highway cycles, the latter with up to 10 % grades to mimic mountain passes. Cycles are repeated until the cell can no longer deliver 60 kW at pack level.
Blinded data: Test data that is never released. Models are scored on it so they cannot be tuned to the answers.One whole cell (m448) and the standard drive cycles for every cell are withheld. This is what makes the leaderboard a fair comparison between methods rather than a measure of who over-fitted the test set.
Open data: The part of the dataset you can download and train or parameterise on: characterization tests plus reordered and custom drive cycles for three cells.
HPPC test: Hybrid Pulse Power Characterization: short current pulses at many SOC levels used to measure a cell's internal resistance and open-circuit voltage.Ten-second discharge and charge pulses of several magnitudes are applied at 100, 95, 90, 80 … 5 % SOC after a one-hour rest. The voltage response gives resistance and dynamic parameters for equivalent-circuit models; the rested voltage gives the OCV–SOC curve.
Open-circuit voltage (OCV): The cell voltage after resting with no current; it maps almost one-to-one to SOC, which many estimators exploit.
C-rate: Current relative to capacity: 1C fully discharges a 4.5 Ah cell in one hour, C/20 takes twenty hours.
Payload / HVAC cases: Each cell was cycled with a different vehicle load: 80, 448 or 1000 kg of payload, with cabin heating/cooling on or off, so the data spans light to heavy use.

Methods

Coulomb counting: Integrating current over time to track charge in and out. Simple and fast, but drifts with sensor offset and needs a known starting SOC.
Kalman filter (EKF/UKF): A model-based estimator that blends a battery model's prediction with voltage measurements, correcting drift. EKF and UKF handle the non-linear voltage–SOC relationship in different ways.
Equivalent-circuit model (ECM): A battery represented as a voltage source plus resistors and capacitors; its parameters come from HPPC data and feed Kalman-filter estimators.
Neural-network estimators (FNN, LSTM, GRU, Transformer): Data-driven models trained on the open data to map measurements to SOC. Recurrent types (LSTM, GRU) keep memory of past samples; FNNs use averaged inputs instead.