marmossburg/logistic-regression

EV Charging Logistic Regression

PyTorch logistic regression predicting whether to charge an EV based on battery level and electricity price.

Training script: logistic_regression.py Data generator: generate_dataset.py

Dataset

Synthetic EV charging dataset (300 samples, linearly separable by design).

Feature	Range	Description
battery_percent	0.0 – 100.0	Battery state of charge (%)
electricity_price	0.050 – 0.500	Price per kWh ($)
charge_now	{0, 1}	Target: 1 = charge now

Class balance: ~40% positive (charge_now=1)

Decision boundary (generative):

score = -0.08 * battery_percent - 3.0 * electricity_price + N(0, 0.8)
charge_now = 1 if score > -4.0 else 0

• Lower battery → more likely to charge
• Lower price → more likely to charge
• Linear boundary (no interaction terms)

Training

Setting	Value
Epochs	200
Batch size	64
Optimizer	SGD
Learning rate	0.01
Loss	BCEWithLogitsLoss
Train/val split	80/20 (random)
Normalization	Z-score (fit on train)
Device	5060ti

Checkpointing: Best model by validation loss saved to logistic_best.pt

Results (Final Epoch 200)

Metric	Value
Train Loss	0.3025
Val Loss	0.2731
Accuracy	93.33%
Precision	89.66%
Recall	96.30%
F1	0.9286

Peak accuracy (epoch 102): 95.00% (F1: 0.9412)

• Validation loss continued decreasing after epoch 102 (0.3485 → 0.2731)
• Accuracy/F1 plateaued — model gained confidence without changing decisions

Files

File	Description
logistic_best.pt	Best model weights + metadata (feature_cols, target_col, X_mean, X_std)
logistic_latest.pt	Final epoch weights
logistic_regression.py	Training script
generate_dataset.py	Data generator (linear boundary)
figures/logistic_learning_curve.png	Loss/accuracy curves
figures/logistic_decision_boundary.png	Decision boundary visualization

Usage

import torch
from pathlib import Path
import sys
sys.path.insert(0, str(Path(__file__).parent.parent))
from _july_2.logistic_regression import LogisticRegressionModel

# Load checkpoint
ckpt = torch.load("logistic_best.pt", map_location="cpu")

# Recreate model
model = LogisticRegressionModel(len(ckpt["feature_cols"]))
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

# Prepare input (2 features: battery_percent, electricity_price)
X_raw = torch.tensor([[20.0, 0.15], [80.0, 0.40]], dtype=torch.float32)

# Normalize using training stats
X_norm = (X_raw - ckpt["X_mean"]) / ckpt["X_std"]

# Predict
with torch.no_grad():
    logits = model(X_norm)
    probs = torch.sigmoid(logits)
    preds = (probs >= 0.5).int()

print(f"Probabilities: {probs.squeeze().tolist()}")
print(f"Predictions: {preds.squeeze().tolist()}")
# [low battery, low price] -> prob] -> ~0.9 (charge)
# [high battery, high price] -> ~0.1 (don't charge)

Inference API (for HF Spaces)

The checkpoint contains everything needed:

ckpt = torch.load("logistic_best.pt")
feature_cols = ckpt["feature_cols"]      # ["battery_percent", "electricity_price"]
target_col = ckpt["target_col"]          # "charge_now"
X_mean, X_std = ckpt["X_mean"], ckpt["X_std"]

Limitations

• Synthetic data — not trained on real EV charging behavior
• Linear boundary — cannot capture price sensitivity that varies with battery level
• Small dataset — 300 samples, 240 train / 60 test
• No temporal/contextual features — time of day, trip distance, user preference ignored

Citation

@misc{ev-charging-logreg-2026,
  title={EV Charging Logistic Regression},
  author={marmossburg},
  year={2026},
  url={https://huggingface.co/...}
}