🚆 Train Ridership Forecasting

Predicting daily train ridership and allocating the required number of trains per station using machine learning — with a comparative analysis of ElasticNet regression and LSTM models across a dataset spanning 2019–2022, including the COVID-19 pandemic period.

---

📌 Project Overview

Rail operators need to know how many trains to deploy at each station and time period to avoid both overcrowding and resource waste. This project builds an end-to-end pipeline that:

1. Explores ridership patterns and the structural impact of COVID-19
2. Engineers time-series features (lag, rolling mean, cyclical encoding)
3. Trains and tunes two distinct model families
4. Compares their performance at both global and per-station levels
5. Outputs concrete train allocation recommendations

---

📂 Repository Structure

Train-Ridership-Forecasting/
│
├── data/
│   ├── Ridership.csv                            # Raw dataset (2019–2022)
│   ├── elasticnet_allocation_summary.csv        # ElasticNet train allocation output
│   └── lstm_allocation_summary.csv             # LSTM train allocation output
│
├── models/
│   ├── ElasticNet_model/
│   │   ├── elasticnet_model.pkl                 # Trained ElasticNet model
│   │   └── preprocessor.pkl                    # Fitted ColumnTransformer
│   └── LSTM_model/
│       ├── lstm_model.h5                        # Trained LSTM model
│       ├── scaler_X.pkl                         # Feature scaler (MinMaxScaler)
│       └── scaler_y.pkl                         # Target scaler (MinMaxScaler)
│
├── notebooks/
│   ├── EDA.ipynb                                # Exploratory Data Analysis
│   ├── ElasticNet_Ridership_Forecasting.ipynb   # ElasticNet pipeline
│   ├── LSTM.ipynb                               # LSTM pipeline
│   └── Model_Comparison.ipynb                  # Side-by-side model comparison
│
├── README.md
└── requirements.txt

---

📊 Dataset

Column	Type	Description
Year	int	Calendar year (2019–2022)
Month	str	Month name
Day	int	Day of month
Week Number	int	ISO week number
Corridor	str	Train route corridor (7 unique)
Workday	str	y = workday, n = weekend/holiday
Station	str	Station identifier (45 unique)
Period	str	Time-of-day segment (5 unique: AM Peak, Midday, PM Peak, Evening, Weekend/Holiday)
Ridership	int	Target — number of passengers
N_trains	int	Number of trains deployed
Covid19	int	Binary flag: 1 = pandemic period

Key statistics:

• 64,369 observations · No missing values · No duplicate rows
• Ridership ranges from 0 to 26,798 (strongly right-skewed)
• COVID-19 reduces average ridership by ~73% (from ~1,676 to ~448)

---

🔧 Feature Engineering

All features are computed on the training set only and applied to the test set to prevent data leakage.

Feature	Description
month_sin / month_cos	Cyclical month encoding (preserves Dec→Jan continuity)
day_of_week / is_weekend	Day-level calendar features
lag_1/2/3_period	Ridership 1–3 steps ago (per Station × Period group)
rolling_mean_3/7_period	Rolling averages shifted by 1 to avoid leakage
Station_enc	Target encoding of Station by mean ridership (train-only)

---

🤖 Models

ElasticNet Regression

Combines L1 (Lasso) and L2 (Ridge) regularization — well-suited for datasets with correlated lag and rolling features.

• Hyperparameter tuning via GridSearchCV with TimeSeriesSplit(n_splits=5)
• Search space: alpha ∈ [0.001 … 100], l1_ratio ∈ [0.1, 0.5, 0.9]
• Preprocessing: OneHotEncoder for categoricals + StandardScaler for numerics
• Low-frequency stations (< 30 records) are grouped into an Other category

LSTM (Long Short-Term Memory)

Deep learning model for sequential pattern learning.

• Sequences built per station (window = 7 timesteps) to avoid cross-station contamination
• Architecture: 2× stacked LSTM → Dropout(0.2) → BatchNorm → Dense(16) → Dense(1)
• Training: EarlyStopping(patience=10) + ReduceLROnPlateau(patience=5)
• Scaling: MinMaxScaler for both features and target
• Low-frequency stations (< 30 records) are dropped entirely — too few records to build reliable sequences
• Original row indices are tracked inside create_sequences_per_station to prevent length-mismatch errors when aligning predictions back to the DataFrame

---

📈 Model Comparison

Four comparison dimensions are evaluated in Model_Comparison.ipynb:

#	Method	What it measures
1	Global Metrics	RMSE, MAE, R² on the shared aligned test set
2	Visual Comparison	Time series overlay, scatter plots, residual distributions
3	Per-Station RMSE	Winner at each individual station + delta chart
4	Train Allocation	Over/under-allocation rates, heatmaps per Station × Period

Results

Criterion	ElasticNet	LSTM	Winner
Global RMSE	433.82	563.71	✅ ElasticNet
Global MAE	275.28	374.11	✅ ElasticNet
Global R²	0.4847	0.1299	✅ ElasticNet
Per-Station RMSE wins	21 / 23	2 / 23	✅ ElasticNet
Train Allocation MAE	0.431	0.572	✅ ElasticNet

ElasticNet wins across all 5 criteria.

Why ElasticNet outperforms LSTM on this dataset

• Rich feature engineering — lag and rolling features hand the model pre-digested temporal patterns, removing LSTM's main advantage
• Moderate data volume per station — LSTM needs large sequence counts to generalize; splitting by station limits this significantly
• Near-linear relationships — ridership correlates strongly and linearly with Covid19, Period, and Station, which ElasticNet captures efficiently
• ElasticNet's L1+L2 regularization controls overfitting effectively at this scale, while LSTM tends to overfit with limited per-station sequences

---

🚉 Train Allocation Logic

required_trains = ceil(predicted_ridership / capacity)
required_trains = max(required_trains, min_trains)

Default parameters: capacity = 600 passengers/train, min_trains = 1

---

▶️ How to Run

1. Install dependencies

pip install -r requirements.txt

2. Run notebooks in order

1. EDA.ipynb
2. ElasticNet_Ridership_Forecasting.ipynb   ← saves models/ElasticNet_model/
3. LSTM.ipynb                               ← saves models/LSTM_model/
4. Model_Comparison.ipynb                  ← loads both saved models

⚠️ Model_Comparison.ipynb depends on saved artifacts from steps 2 and 3. Run them first.

3. Loading the LSTM model

from tensorflow.keras.models import load_model
from tensorflow.keras.metrics import MeanSquaredError
import joblib

lstm_model = load_model(
    '../models/LSTM_model/lstm_model.h5',
    custom_objects={'mse': MeanSquaredError()}
)
scaler_X = joblib.load('../models/LSTM_model/scaler_X.pkl')
scaler_y = joblib.load('../models/LSTM_model/scaler_y.pkl')

---

🛠️ Tech Stack

Tool	Version	Role
Python	3.10+	Core language
pandas / numpy	—	Data manipulation
scikit-learn	—	ElasticNet, preprocessing, evaluation
TensorFlow / Keras	—	LSTM model
matplotlib / seaborn	—	Visualization
joblib	—	Model serialization

---

📋 Key Design Decisions

Decision	Rationale
Time-based split at 2022-01-01	Respects temporal ordering — no future data leaks into training
Outlier clipping via IQR	Uses train statistics only; applied identically to test set
Per-station sequences for LSTM	Prevents cross-station contamination in sliding windows
Index tracking in create_sequences_per_station	Avoids ValueError from length mismatch when aligning predictions
custom_objects={'mse': MeanSquaredError()}	Resolves Keras 3 deserialization error when loading .h5 models
Station dropped (not grouped) in LSTM	Too few records produce unreliable sequences; dropping is cleaner

---

👤 Author

Ali Sarafraz
GitHub: @Ali-sarafraz