🌦️ Weather Forecasting with Time Series Models

Multi-step temperature & precipitation forecasting using classical and deep learning approaches — SARIMA, Prophet, and LSTM.

🎯 Project Overview

This project explores time series forecasting applied to historical weather data, comparing statistical and neural network models to predict future temperature and precipitation values across multiple horizons (1-day, 7-day, 30-day).

Main Goal: Build and evaluate forecasting pipelines that can be adapted to any weather station dataset, with structured model comparison and reproducible experiments.

---

🏆 Models & Results

Model	MAE (temp °C)	RMSE	Notes
Baseline (naive)	~3.1	~4.2	Prior-day carry-forward
SARIMA	~1.8	~2.5	Seasonal decomposition
Prophet	~1.6	~2.3	Holiday & trend effects
LSTM (stacked)	~1.4	~2.0	Best on longer horizons

---

🔍 What Was Built

1. Data Pipeline

• Ingestion and cleaning of raw weather station data
• Handling of missing values (interpolation + flagging)
• Feature extraction: lag features, rolling statistics, cyclical encoding of date components

2. Classical Models

• SARIMA: Grid search over (p,d,q)(P,D,Q,s) with AIC selection
• Prophet: Additive model with weekly/yearly seasonality and changepoint detection

3. Deep Learning — LSTM

• Stacked LSTM with dropout regularization
• Sliding-window sequence generation for multi-step forecasting
• Early stopping + learning rate scheduling

4. Evaluation Framework

• Walk-forward validation (no data leakage)
• Metrics: MAE, RMSE, MAPE
• Residual analysis and forecast visualisations

---

📁 Project Structure

weather-forecasting-timeseries/
├── data/
│   ├── raw/            # Original weather station CSVs
│   └── processed/      # Cleaned & feature-engineered datasets
├── models/             # Serialised model artefacts (.pkl / .h5)
├── notebooks/
│   ├── 01_eda.ipynb
│   ├── 02_sarima_prophet.ipynb
│   └── 03_lstm.ipynb
├── src/
│   ├── data_loader.py
│   ├── features.py
│   ├── models/
│   │   ├── sarima_model.py
│   │   ├── prophet_model.py
│   │   └── lstm_model.py
│   └── evaluate.py
├── reports/            # Plots and summary reports
├── requirements.txt
└── project_structure_run.py

---

⚙️ Quickstart

# 1. Clone and set up environment
git clone https://github.com/pelabdang/weather-forecasting-timeseries.git
cd weather-forecasting-timeseries
python -m venv .venv && source .venv/bin/activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Initialise project structure
python project_structure_run.py

# 4. Run notebooks in order
jupyter lab notebooks/

---

🛠️ Tech Stack

---

📊 Key Takeaways

• LSTM outperforms classical models on longer forecast horizons (7–30 days)
• Prophet handles trend shifts and seasonality with minimal tuning
• Walk-forward validation is essential — standard train/test splits overestimate model performance on time series

---

📧 Contact: Angelo Pelisson · GitHub