Spatiotemporal Autoencoder-Based Framework for GHI Prediction Using Visible Satellite Imagery

Project Description

This project focuses on short-term Global Horizontal Irradiance (GHI) forecasting using visible-band satellite imagery from the INSAT-3D database. The model uses a spatiotemporal autoencoder integrated with ConvLSTM layers to capture cloud movement and predict GHI for the next 2 hours at 30-minute intervals. It combines:

• Physics-based clear-sky irradiance modeling
• Cloud mask generation and cloud index computation
• Deep learning for temporal pattern recognition

Features & Methodology Summary

• Area of Interest (AOI) based processing: Focuses only on the specified geographic region to reduce computation time.

• Cloud detection: Uses autoencoder reconstruction error to identify cloud coverage in visible satellite imagery.

• Cloud index calculation: Quantifies cloud attenuation effect on GHI.

• Clear-sky GHI estimation: Uses PVLib's Solis model.

• Spatiotemporal deep learning model:

  • Encoder for extracting spatial features
  • ConvLSTM layers for capturing temporal patterns
  • Decoder for reconstructing future frames

• GHI prediction: Predicts 4 future steps (2 hours) from 6 past observations (3 hours).

How to Run the Code

1. Clone the Repository

2. Install Dependencies pip install -r requirements.txt

3. Prepare Input Data

   • Satellite Data: INSAT-3D visible-band imagery in .h5 format.
   • Cloud Mask Files: . npy binary mask files for each timestamp.
   • Place your input files in a data/ folder inside the project directory.

4. Running the Notebook

   • Open the Jupyter Notebook: jupyter notebook GHI_Forecasting_final_code.ipynb

   • Run all cells in order.

5. AOI and File Creation During Execution

   • At the start, you define your Area of Interest (AOI) with latitude, longitude, and altitude in the code.

   • The script:

     • Clips satellite imagery to your AOI.
     • Creates cloud mask arrays (. npy files).
     • Computes Cloud Index (CI) for each timestamp.
     • Generates clear-sky GHI values using PVLib.
     • Prepares time series sequences for training and validation.

   • After training, the model creates:

     • Saved Model Weights in .h5 format.
     • Prediction Results in .csv format.
     • Visualization Plots (actual vs predicted GHI).

Dependencies & Requirements

Python 3.8+ and the following packages:

• numpy
• pandas
• matplotlib
• h5py
• tensorflow
• scikit-learn
• pvlib
• jupyter
• opencv-python
• seaborn

Authors & Credits

• Lakshmi P “Vellore Institute of Technology, Chennai"
• Kalpalathika N “Vellore Institute of Technology, Chennai"
• Jumin Salih “Vellore Institute of Technology, Chennai"
• Nihal Siddiqi “Vellore Institute of Technology, Chennai"
• Bhuvaneswari A “Assistant Professor, Vellore Institute of Technology, Chennai"

Research Paper: Enhanced Short-Term GHI Prediction Using Cloud-Aware Deep Autoencoder Network “IEEE, ICSCSA 2025.