Uncertainty-Aware Weather Forecasting System

Live Data Anchored, 7-Day Forecasting with Machine Learning & React

📌 Project Overview

This project is a machine learning–based weather forecasting system designed to predict the next 7 days of temperature using:

Historical learning Live weather anchoring Uncertainty estimation

Unlike traditional student projects that rely on static datasets or single-point predictions, this system:

Anchors forecasts to real-time weather data via Open-Meteo Uses recursive multi-step forecasting to predict future days Explicitly models prediction uncertainty using residual error analysis Presents results through a decoupled FastAPI backend and a modern, glassmorphic React/Tailwind UI

The project is suitable for academic research, capstone evaluation, and demonstrating applied ML system design.

🎯 Problem Statement

Most beginner weather forecasting projects suffer from:

Reliance on static datasets that do not reflect real-world conditions Providing single deterministic predictions without communicating reliability Exposing internal ML features (e.g., lag variables) directly to users Lack of system-level design (monolithic scripts instead of separated API and UI layers)

This project addresses these issues by building a realistic, interpretable, and modular forecasting system that combines machine learning with live data integration and uncertainty-aware predictions.

🧠 Key Concepts & Methodology 1️⃣ Historical Learning with Machine Learning Core model: XGBoost Regressor Trained on historical weather data Uses lag-based features (previous 7 days’ temperatures) to capture temporal dependencies 2️⃣ Live Weather Anchoring (Open-Meteo API) Backend fetches: Temperature Wind speed Humidity Cloud cover These values act as Day 0 input for forecasting Ensures predictions reflect current real-world conditions

⚠️ Live data is used for anchoring, not retraining

3️⃣ Recursive Multi-Step Forecasting (7-Day Prediction) Predict Day 1 Feed prediction back to predict Day 2 Repeat recursively for 7 days

This is a standard and academically valid time-series approach.

4️⃣ Uncertainty-Aware Forecasting & Expanded Metrics

Instead of single-point predictions:

Residual errors are analyzed Standard deviation is computed A 95% confidence interval is generated Uncertainty is dynamically calibrated based on deviation from current conditions

Additionally:

ML predictions are merged with 8-day API forecasts Provides:

1. Wind Speed
2. Cloud Cover
3. Precipitation

🏗️ System Architecture Open-Meteo API │ ▼ Live Weather Anchoring │ ▼ XGBoost ML Model (Historical) │ ▼ Recursive 7-Day Forecast │ ▼ Uncertainty Estimation │ ▼ FastAPI Backend ──► React / Tailwind Frontend

🧩 Project Structure weather-forecasting/ │ ├── app/ │ └── main.py # FastAPI backend (prediction API & Data Merging) │ ├── frontend/ # React / Vite application │ ├── src/ │ │ ├── App.jsx # Main UI Component with Expandable Accordions │ │ ├── index.css # Tailwind v4 Configuration │ │ └── main.jsx │ └── package.json │ ├── scripts/ │ ├── train_model.py # Model training + uncertainty computation │ └── fetch_current_weather.py
│ ├── data/ │ └── weather.csv # Historical training dataset │ ├── models/ │ ├── xgb_model.pkl # Trained ML model │ ├── scaler.pkl # Feature scaler │ └── uncertainty.pkl # Residual-based uncertainty value │ ├── results/ │ ├── metrics.txt │ └── actual_vs_pred.csv │ ├── requirements.txt # Python dependencies └── README.md

🚀 How to Run the Project 1️⃣ Python Backend Setup pip install -r requirements.txt pip install fastapi uvicorn pydantic requests python scripts/train_model.py uvicorn app.main:app --reload API: http://localhost:8000 Docs: http://localhost:8000/docs

2️⃣ React Frontend Setup cd frontend npm install npm run dev

UI: http://localhost:5173 📊 Output Example Day 3 (Wednesday)

Predicted Temperature: 14.8 °C Uncertainty: ±2.1 °C Confidence Range: 12.7 – 16.9 °C

Expanded view:

Wind: 12 km/h Clouds: 45% Precipitation: 0.0 mm

📈 Evaluation Metrics Mean Absolute Error (MAE) Root Mean Square Error (RMSE)

Stored in: results/metrics.txt

⚠️ Limitations

1. Model does not retrain automatically
2. Uncertainty reflects model confidence, not real-world guarantees
3. Recursive forecasting may accumulate error

🔮 Future Enhancements

1. Multi-variable forecasting (humidity, wind via ML)
2. Deep learning models (LSTM / GRU)
3. Automated retraining (CI/CD pipelines)
4. Cloud deployment (AWS / GCP / Vercel)