Transformer Ridership Prediction

Climate-induced Transformer model for Bangkok MRT station ridership prediction

A deep learning pipeline that forecasts daily passenger in/out counts at individual MRT stations using a Transformer architecture, incorporating climate variables (temperature, rainfall, humidity) alongside historical ridership data.

Overview

Model	Transformer (encoder-only, per-station)
Target	Ridership in & out per MRT station
Input features	Historical ridership + climate variables
Sequence length	28 days → forecast 1 day ahead
Framework	PyTorch (CUDA-ready)

Features

• Per-station model training with checkpoint saving
• NaN-safe preprocessing and robust normalization options (standard, minmax, robust)
• Evaluation metrics: MAE, RMSE, R²
• Auto-generated prediction plots per station

Data

File	Description
Master.csv	Daily ridership + climate data per station
static.csv	Station metadata (line, location, etc.)

Usage

pip install torch pandas numpy matplotlib
python main.py

Set MODE = 0 for train + test, MODE = 1 for inference only (loads saved checkpoints).

Project Context

Part of the Bangkok Living Lab research under CUTI (Chulalongkorn University Transportation Institute), investigating climate resilience of urban transit systems.