🌞 Weather-Net (Solar Energy Prediction)

This project builds a deep learning model to forecast hourly solar energy output (W/m²) using weather features from NASA's POWER API. It includes a FastAPI server for making real-time predictions using geographic coordinates. It has a Latency of 35 ms

---

📦 Features

• Predict next hour’s solar irradiance (ALLSKY_SFC_SW_DWN)
• Uses historical hourly weather data (past 24h)
• FastAPI endpoints for real-time predictions and recent actual data
• Model built using a spike-aware hybrid CNN-LSTM-Attention architecture (SpikeAwareHybrid)

---

🧠 Model Architecture

• Input: 24 hourly sequences of 14 weather features
• Model: SpikeAwareHybrid (CNN + LSTM + Multihead Attention)
  • Multi-scale CNN feature extraction (trend + spike detection)
  • Bidirectional LSTM for temporal dependencies
  • Multihead Attention for context
  • Global mean/max pooling for robust spike-aware prediction
• Output: Next hour solar energy prediction (in W/m²)

---

📷 Images

• Actual Vs Predicted
  
• Average Hourly Solar Prediction
  
• On Live Data Never Seen Before
  

---

🚇 Metrics

• RMSE Error = 118.63706223892275
• MAE = 97.60
• R2 socre = 0.9136192701253345

---

📁 Project Structure

weather_net/
├── final/
│   └── dataset.py
├── models_code/
│   ├── model_5.py             # Spike-aware hybrid CNN + LSTM + Attention architecture
│   ├── model_3.py             # LSTM-Attention model
│   └── other models…
├── scalers/
│   ├── feature_scaler_seq24.pkl
│   ├── feature_scaler_seq12.pkl
├── models/
│   ├── spike_aware_q95_seq24.pth
│   ├── spike_aware_q95_seq12.pth
│   ├── version2.pth
│   └── other model weights…
├── training_codes/
│   ├── training_5.py
│   ├── training_6.py
├── final_fetch.py             # Preprocessing & Weatherbit integration
├── main.py                    # FastAPI server
├── pred.py                    # CLI prediction script
├── test_prediction_plot.py    # Local 7-day evaluation script
├── requirements.txt
└── readme.md

Descriptions:

• final/ — Data handling and sequencing logic (PyTorch Dataset class)
• models_code/ — All model architectures (CNN, LSTM, Transformer, hybrids, etc.)
• scalers/ — Saved feature scalers used during training/inference
• models/ — Saved PyTorch weights for trained model versions
• training_codes/ — Python scripts used to train models (with spike-aware logic)
• final_fetch.py — Weatherbit API fetching + feature harmonization (TOA, clipping, night masking)
• main.py — FastAPI server providing live prediction and recent data endpoints
• pred.py — CLI for predicting the next hour's value from console
• test_prediction_plot.py — Script to compare predictions vs real for the past 7 days and plot results

---

🚀 Running the API Server

1. Install dependencies:

   pip install -r requirements.txt

2. Start FastAPI server:

   uvicorn main:app --reload

3. Test API in browser:
   Open: http://127.0.0.1:8000/docs

---

📥 API Usage

1️⃣ Predict Next Hour Solar Irradiance

Endpoint:

GET /predict_current

Test URL (Delhi, India): http://127.0.0.1:8000/predict_current?lat=28.6139&lon=77.2090

Response Example:

{
  "latitude": 28.6139,
  "longitude": 77.2090,
  "timestamp_predicted_ist": "2025-08-02 14:00:00",
  "predicted_next_hour_wm2": 378.65
}

---

2️⃣ Get Last 7 Hours Actual Solar Data

Endpoint:

GET /last_7hours_real

Test URL (Delhi, India): http://127.0.0.1:8000/last_7hours_real?lat=28.6139&lon=77.2090

Response Example:

{
  "latitude": 28.6139,
  "longitude": 77.2090,
  "data": [
    {"timestamp_ist": "2025-08-02 07:00:00", "ghi_wm2": 512.34},
    {"timestamp_ist": "2025-08-02 08:00:00", "ghi_wm2": 498.21},
    ...
  ]
}

---

🛰️ Data Source

Training Data

• Source: NASA POWER API
• Features used:
  • Solar Radiation & Irradiance:
    ALLSKY_SFC_SW_DIFF, ALLSKY_SFC_SW_DNI, TOA_SW_DWN, ALLSKY_SFC_SW_DWN, ALLSKY_SFC_LW_DWN
  • Weather Parameters:
    RH2M (Relative Humidity at 2m), QV2M (Specific Humidity at 2m), PS (Surface Pressure), WS2M (Wind Speed at 2m), CLOUD_AMT (Cloud Amount)
  • Temperature:
    T2M (Temperature at 2m)
  • Cyclic Time Features:
    hour_sin, hour_cos, month_sin, month_cos

---

Prediction Data

• Source: Open-Meteo API
• Features used: Same as the training data.
• Note: Two features not provided by the API were calculated in code:
  1. Relative Humidity at 2m (RH2M)
  2. Top of atmosphere shortwave downward radiation (TOA_SW_DWN)

---

🛠 Tech Stack

• Language: Python 3.10+
• Framework: FastAPI, PyTorch
• Data Fetching: NASA POWER API, Weatherbit API
• Model: SpikeAwareHybrid (CNN + LSTM + Attention)

---

📄 License

MIT License © 2025

---