Predicting high-demand EV charging sessions at workplace facilities before they occur.

Workplace EV charging drives significant electricity demand charges — costs determined not by total energy consumed, but by the highest 15-minute interval of consumption in a billing period. When employees plug in simultaneously during morning arrival (8–10am), the resulting demand spike can dominate the monthly electricity bill.

High-demand sessions — the top 25% by energy delivered (above 13.7 kWh) — account for 50.3% of all energy consumed despite being only one quarter of all sessions. The challenge: these sessions need to be identified at plug-in time, before charging begins, so that interventions (scheduling nudges, rate adjustments, managed charging) can be deployed in time to matter.

A Random Forest classifier trained on the NREL Workplace Charging Dataset that flags high-demand sessions at the moment of plug-in, using only information available before charging starts.

Two agentic POC systems built on top of the model:

The classification threshold was lowered from 0.50 to 0.40 to prioritise recall — missing a high-demand session costs money in demand charges; a false alarm costs only an unnecessary nudge.

NREL Workplace Charging Dataset — publicly available from the National Renewable Energy Laboratory.

• Source: https://data.openei.org/submissions/4538
• Size: 40,179 sessions after cleaning (from 40,979 raw)
• Period: November 2016 — October 2021
• Sites: Single workplace facility, 141 charging stations

peak-charge/
│
├── README.md
│
├── data/
│   ├── sessions.csv              # Cleaned session data (200-session replay sample)
│   └── all_sessions.csv          # Full cleaned dataset for model maintenance
│
├── models/
│   ├── rf_model.pkl              # Trained Random Forest classifier
│   ├── rf_model_backup.pkl       # Auto-backup before each model swap
│   ├── feature_cols.pkl          # Ordered feature list
│   └── baseline_auc.pkl          # Current baseline AUC for drift detection
│
├── agents/
│   ├── session_monitoring_agent.py    # Problem 1 — real-time session flagging
│   └── model_maintenance_agent.py    # Problem 4 — drift detection + retraining
│
├── logs/
│   ├── session_log.csv           # Per-session scoring decisions
│   └── drift_log.jsonl           # Model health check audit trail
│
└── notebooks/
    └── peak_charge_analysis.ipynb    # Full EDA, modelling, and evaluation

pip install pandas numpy scikit-learn joblib matplotlib seaborn

Note: Model files (*.pkl) are not committed to this repo. Run the notebook end-to-end to generate them locally before running the agents.

Open notebooks/peak_charge_analysis.ipynb and run all cells. This produces rf_model.pkl, feature_cols.pkl, baseline_auc.pkl, sessions.csv, and all_sessions.csv.

# Default — replay 200 sessions at threshold 0.40
python agents/session_monitoring_agent.py
 
# Replay all sessions
python agents/session_monitoring_agent.py --n 6696
 
# Slower replay for demos
python agents/session_monitoring_agent.py --delay 0.1
 
# Custom threshold
python agents/session_monitoring_agent.py --threshold 0.50

Sample output:

#0071  Driver 16      Stn 07B   kWh_req  24.0  Score 0.800  → HIGH  ✓
 
  [HIGH DEMAND ALERT]
  Driver:   16
  Station:  07B
  Time:     2016-11-14 07:00:00
  kWh req:  24.0   Miles req: 80
  Score:    0.800  (threshold: 0.4)
  Action → Notify driver / flag for managed charging

# Normal run — check model health across 6 time windows
python agents/model_maintenance_agent.py
 
# Simulate concept drift to see full retraining cycle
python agents/model_maintenance_agent.py --simulate-drift
 
# Custom evaluation window and drift threshold
python agents/model_maintenance_agent.py --window 90 --threshold 0.84

All features are constructed from information available at plug-in time — no post-session data is used as input.

q75 = df['energy_charged'].quantile(0.75)   # = 13.72 kWh
df['high_demand'] = (df['energy_charged'] > q75).astype(int)
# Class distribution: 75% low-demand, 25% high-demand

Random Forest selected as final model — higher last-fold AUC (0.87 vs 0.84), better recall on high-demand class (0.83 vs 0.78), and more robust to nonlinear feature interactions.

TimeSeriesSplit with 5 folds — training always on the past, testing always on the future. Prevents data leakage from future sessions into training.

Fold 1: train=6,699   test=6,696   (earliest data)
Fold 2: train=13,395  test=6,696
Fold 3: train=20,091  test=6,696
Fold 4: train=26,787  test=6,696
Fold 5: train=33,483  test=6,696   (most recent data)

Tesla Model 3 sessions are high-demand 85% of the time. Prius Prime, Audi A3 E-Tron, and Ford Fusion (PHEVs with small batteries) are near 0%. Vehicle type is a strong proxy for battery capacity and therefore demand potential.

Sessions under paid charging (afterPaid=1) show only marginally higher demand rates than free sessions (27% vs 24%). Pricing regime is not a meaningful predictor.

This project demonstrates how prediction alone is insufficient — acting on predictions at scale requires agentic systems. Four problems and their agent-based solutions:

• Temporal drift: The dataset spans 2016–2021. The vehicle fleet has changed significantly since then — Tesla/Bolt share has grown, increasing average battery sizes and shifting the demand distribution.
• Single-site data: All sessions come from one facility. Station-level features (station_encoded) may not generalise to other sites.
• kwh_requested caveat: This is a strong predictor (r=0.456 with target) but also highly correlated with actual energy delivered. While it is genuinely available at plug-in time, it may partially reflect self-fulfilling driver behaviour.
• No real charger integration: Both agents use CSV replay as a simulated event stream. Production deployment would require an OCPP webhook or charger management system API.

This project was developed with AI assistance using Claude (Anthropic) as a pair-programming and analysis tool — a workflow sometimes referred to as vibe coding.

Claude was used throughout the project lifecycle: exploring and cleaning the dataset, engineering features, selecting and evaluating models, writing the agentic POC scripts, and drafting documentation. All outputs were reviewed, validated, and tested by the author at each step.

Guardrails kept in mind throughout:

• No data leakage — features were explicitly restricted to information available at plug-in time; session outcomes were never used as predictors
• Time-aware validation — TimeSeriesSplit was used throughout to ensure the model was always evaluated on future data, never on data it had already seen
• Human-in-the-loop review — every model decision, threshold choice, and agent behaviour was examined and confirmed before being accepted
• Honest evaluation — metrics were computed on a held-out temporal fold, not on training data; limitations are documented explicitly
• Reproducibility — all random seeds are fixed; the full pipeline from raw data to trained model is contained in a single notebook

AI assistance accelerated development but did not replace critical thinking about the problem, the data, or the validity of results.

Dataset provided by the National Renewable Energy Laboratory (NREL) via the Open Energy Data Initiative (OEDI).

A. Burrell, N. Rustagi, et al. Workplace EV Charging Dataset. NREL, 2021. https://data.openei.org/submissions/4538