Tharunp2431/weather-forecasting

🌤️ Weather Prediction with Hybrid Deep Learning Models

MSc Deep Learning Applications (CMP-L016) — Project #28 Author: Tharun Bisai

Overview

This project investigates hybrid deep learning architectures for short-term temperature forecasting using the Jena Climate dataset (2009–2022). We compare LSTM, TCN, TCN-LSTM hybrid, and stacking ensemble approaches.

Models

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run Training (Colab)

Upload notebooks/Tharun_ML2.ipynb to Google Colab and run all cells.

3. Run Evaluation

Upload notebooks/Tharun_ML3.ipynb to Colab (after M2 finishes).

4. Launch Dashboard

🌍 Live Cloud Deployment: 👉 [weather-tcn-forecasting.streamlit.app](https://weather-tcn-forecasting.streamlit.app)

To run legally/locally:

streamlit run app.py

Project Structure

├── app.py                     # Streamlit web dashboard
├── requirements.txt           # Python dependencies
├── notebooks/
│   ├── Tharun_ML1.ipynb       # Data exploration
│   ├── Tharun_ML2.ipynb       # Model training
│   └── Tharun_ML3.ipynb       # Evaluation & analysis
├── docs/
│   └── Tharunbisai_FinalReport.md  # IEEE report
├── data/raw/                  # Dataset (not tracked)
├── outputs/
│   ├── figures/               # Generated plots
│   ├── models/                # Saved weights (.pt)
│   └── results/               # JSON results
└── src/models/hybrid.py       # Model class definitions

Dataset

Jena Climate Dataset — Max Planck Institute for Biogeochemistry 14 meteorological features, hourly resolution, 2009–2022

Results

• —Best model: Stacking Ensemble (MSE ~0.0043)
• —Best standalone: LSTM (MSE ~0.0044)
• —Fastest training: TCN (24 epochs vs 84 for LSTM)

Tech Stack

Python 3.10+ • PyTorch • Streamlit • Plotly • scikit-learn

License

Academic use only — MSc coursework submission.