Forecast Journal fetches 12-hour temperature forecasts from the AccuWeather API and logs them in a structured PostgreSQL database hosted on NeonDB. The system allows analysis of prediction accuracy over time.
With the newly implemented frontend dashboard, users can now visualize trends and download historical data for further analysis.
✅ Weather Data Collection: Automated fetching of forecasts at regular intervals.
✅ Database Storage: Structured storage of forecasted vs. actual temperatures.
✅ Logging System: Tracks API and database transactions in a structured format.
✅ Frontend Dashboard: Visual representation of forecast accuracy trends.
✅ Downloadable Data: Users can retrieve historical forecast accuracy data.
✅ ML Prediction Model: Refines temperature forecasts using temporal patterns.
The frontend, built with React + TypeScript, provides interactive data visualizations. It includes:
- 📈 Time Series Chart – Predicted vs. Actual temperatures.
- 📊 Error Distribution Histogram – Frequency of forecast errors.
- 🔥 Error Heatmap – Displays error magnitude over time.
Hosting:
- Frontend (React) – Hosted on Vercel
- Backend API (Flask) – Hosted on Render
The ML model has been implemented at TempCastRNN, a transformer-based time series model that:
- 📊 Processes 12×12 forecast matrices leveraging data from Forecast Journal.
- 📈 Uses temporal attention to learn patterns in forecast changes over time.
- 🔄 Corrects inaccuracies in 12-hour forecasts with low average error (~1.7°C).
- 🧠 Implements causal masking to ensure predictions use only available data.
- 📉 Demonstrates improved accuracy through time-aware deep learning techniques.
Recent evaluation shows the model achieves an MAE of 1.698°C and RMSE of 2.017°C, effectively refining traditional weather forecasts.
The next step is improving the ML model to refine predictions further. The goal is to:
- Collect more data points for training.
- Enhance the model's architecture for better accuracy.
📁 API Docs – Details on how forecasts are fetched.
📁 Database Docs – Explanation of the storage schema.
📁 Scripts Docs – Background automation tasks.
📁 Frontend Docs – UI and API integration details.
- Clone the repository:
git clone https://github.com/your-username/forecast-journal.git
- Install required dependencies:
pip install -r requirements.txt
- Set up the database:
- Create a NeonDB account and configure the connection details.
- Use the provided schema (
schema.sql) to initialize the database.
- Run the forecast retrieval script:
python fetch_forecast.py
Contributions are welcome! Feel free to open an issue for suggestions or improvements. 🚀