license: cc-by-4.0
tags:
- electric-vehicles
- telemetry
- data-engineering
- medallion-architecture
- databricks
- data-quality
- olap
size_categories:
- 100K<n<1M
configs:
- config_name: messy_stream_2024
data_files: 1_car_telemetry_2024.csv
- config_name: charging_sessions
data_files: 2_charging_sessions.csv
- config_name: grid_carbon
data_files: 3_grid_carbon_hourly.csv
- config_name: archive_rare_user
data_files: archive_2020_2024_rare_user.csv
- config_name: archive_daily_user
data_files: archive_2020_2024_daily_user.csv
- config_name: archive_moderate_user
data_files: archive_2020_2024_moderate_user.csv
- config_name: archive_heavy_user
data_files: archive_2020_2024_heavy_user.csv
⚡ EV Fleet Sensor Data — Databricks Medallion Case Study
Raw source CSVs for an end-to-end Databricks medallion architecture (Bronze → Silver → Gold) case study: a deliberately messy EV sensor stream is cleaned, quality-gated, and distilled into decision-grade insights — battery health ranking, DTC risk, charger siting, and a clean-charging window.
- Question data:
1_car_telemetry_2024.csv— 35,363 rows of 2024 telemetry with planted faults - Answer key: the four
archive_2020_2024_*.csvfiles — a clean ~175K-row sensor archive (2020–2024) - The messy 2024 stream was derived from the archive (faults planted, seed = 7), so every fault is verifiable. The two sources intentionally overlap in 2024 (stream = question, archive = ground truth).
Dataset Structure
🔴 Feed 1 — Car telemetry (the messy stream)
1_car_telemetry_2024.csv — hourly readings, 4 vehicles, calendar year 2024.
| Column | Type | Description |
|---|---|---|
timestamp |
datetime | Hour of reading |
vehicle_id |
string | EV-101 … EV-104 |
user_profile |
string | rare_user / daily_user / moderate_user / heavy_user |
district |
string | Home district: University, Downtown, Harbor, Industrial |
SOC |
float | State of charge, % |
SOH |
float | State of health (battery), % |
Charging_Cycles |
int | Lifetime charge cycles |
charging_now |
int | 0/1 — currently charging |
Battery_Temp |
float | Battery temperature, °C |
Motor_RPM |
int | Motor speed |
Motor_Torque |
float | Motor torque, Nm |
Motor_Temp |
float | Motor temperature, °C |
Brake_Pad_Wear |
float | Brake pad wear, % |
Charging_Voltage |
float | Charging voltage, V |
Tire_Pressure |
float | Tire pressure — ⚠️ planted unit faults (kPa vs PSI) |
DTC |
int | 1 = diagnostic trouble code active |
grid_carbon_g_per_kwh |
float | Grid carbon intensity at that hour |
Planted fault census: exact duplicates (419, to be dropped), kPa→PSI unit errors (fix), out-of-bounds values (fix + flag), missing readings (flag). Silver-layer policy: only duplicates die — the row always survives; the faulty value is nulled.
🔌 Feed 2 — Charging sessions
2_charging_sessions.csv — 484 completed sessions.
| Column | Description |
|---|---|
session_id, station_id |
Identifiers |
district |
Where the session happened (drives charger-siting analysis) |
start_time, end_time |
Session window |
energy_kwh_source_analog, power_kw_analog |
Energy delivered / power draw |
soc_start_pct, soc_end_pct |
Battery % before → after |
battery_kwh, vehicle_model |
Vehicle pack size and model |
station_temp_c |
Station temperature |
status |
Session status |
⚡ Feed 3 — Grid carbon intensity
3_grid_carbon_hourly.csv — 168 rows (7 days × 24 h): date, hour, carbon_g_per_kwh.
Midnight peaks near 911 gCO₂/kWh; midday (12:00) drops to ~319.
🟢 The clean archive (ground truth)
Four files, one per user profile, each with 43,800 rows (24 h × 365 d × 5 yrs, 2020–2024):
archive_2020_2024_{rare,daily,moderate,heavy}_user.csv — same sensor columns as the stream
(leading index column, no timestamps/district).
| File | Vehicle | Profile |
|---|---|---|
archive_2020_2024_rare_user.csv |
Renault Zoe (52 kWh) | EV-101 |
archive_2020_2024_daily_user.csv |
Hyundai Ioniq 5 (77 kWh) | EV-103 |
archive_2020_2024_moderate_user.csv |
VW ID.4 (82 kWh) | EV-102 |
archive_2020_2024_heavy_user.csv |
Ford F-150 Lightning (131 kWh) | EV-104 |
Example decisions this data supports
Transparent, verifiable arithmetic (no black-box model):
- Risk score =
soh_decay_pct_per_year × 20 + dtc_rate_pct × 10, thresholds 60 (act) / 30 (watch) → EV-102 scores 54.1 (decay 1.958 %/yr + 654 DTC events) → service first - Charger siting =
total_co2_g ÷ vehicle_days→ Industrial district worst at 43,311 gCO₂/vehicle-day across 86 sessions → site the next charger there - Charging window = cleanest vs dirtiest hour → shift charging to ~12:00: ~319 vs ~911 gCO₂/kWh (2.86× cleaner, ~592 g saved per kWh shifted)
How it was used
The files are ingested as-is into a Databricks Free Edition lakehouse:
Bronze (raw, immutable) → Silver (cleaned, quality gate: 35,363 − 419 dupes = 34,944 rows)
→ Gold (6 decision tables: soh_trend, dtc_risk_rank, co2_per_vehicle_day,
grid_carbon_by_district_hour, fleet_worry_list, final_decision)
Usage
from huggingface_hub import hf_hub_download
import pandas as pd
path = hf_hub_download(repo_id="jeyanthiii/sample_ev_data_csv",
filename="1_car_telemetry_2024.csv",
repo_type="dataset")
df = pd.read_csv(path)
Notes
- Faults are synthetic and intentional (teaching/demo dataset); the archive is the reference truth.
- 2024 rows exist in both stream and archive by design — that overlap is what makes the quality gate provable.
- Local backup of this folder is kept at
~/Documents/work/temp/cl_demo/data/.
License
CC-BY-4.0 — free to use with attribution.