Solar Panel Electroluminescence (EL) Image Analysis

Project Overview

This project implements an ensemble machine learning system for analyzing electroluminescence (EL) images of solar panels to predict power loss and assess panel health. The system combines a physics-based model with a deep learning ResNet model to provide accurate predictions of solar panel performance.

Dataset

This project uses the ELPV dataset, released under the CC BY 4.0 License. Please download it directly from the official source.

We do not host the dataset files in this repository to respect bandwidth, license, and size limitations.

Project Structure

h-proj/
├── scripts/
│   ├── run_inference.sh      # Main inference script
│   ├── inference.py          # Python inference implementation
│   └── physics_ensemble_final.ipynb  # Training notebook
├── models/
│   └── best_ensemble_model.pth       # Trained ensemble model 
├── data/
│   ├── data.csv              # Dataset metadata (720 samples)
│   ├── processed.zip         # Compressed processed images (322MB)
│   └── processed/            # Processed EL images (719 images)
└── results/
    └── image_9_results.txt   # Sample inference results

Key Features

Ensemble Model Architecture

• Physics Model: Analyzes inactive pixel percentage using threshold-based calculations
• ResNet Model: Modified ResNet18 architecture adapted for grayscale EL images
• Meta Learner: Linear combination layer that optimally weights both model predictions

Model Components

Physics Model

• Calculates inactive pixel percentage using configurable threshold
• Applies linear regression: pred = 1.5752 * inactive_pct - 0.6471
• Provides interpretable physics-based predictions

ResNet Model

• Modified ResNet18 architecture for single-channel (grayscale) input
• First layer adapted to handle grayscale images
• Outputs power prediction values

Ensemble Model

• Combines physics and ResNet predictions using learned weights
• Provides final ensemble prediction for power loss assessment

Dataset

• Size: 720 samples with corresponding EL images
• Images: 719 processed PNG files (grayscale EL images)
• Metadata: CSV file containing:
  • Peak power measurements
  • Nominal power values
  • Pressure readings
  • Excitation class
  • Module type and instance
  • Power group classifications
  • Cross-validation fold assignments

Usage

Prerequisites

• Python 3.x with PyTorch
• Conda environment named "solar"
• CUDA-compatible GPU (optional but recommended)

Running Inference

1. Configure the script: Edit scripts/run_inference.sh to set your parameters:

   # Model Settings
   MODEL_PATH="/home/.../models/best_ensemble_model.pth"
   USE_GPU=true
   
   # Image Processing
   IMAGE_SIZE=224
   THRESHOLD=0.2
   
   # Input/Output
   INPUT_FILE="/home/.../data/processed/image_9.png"
   OUTPUT_DIR="results"

2. Run inference:

   cd /home/.../
   chmod +x scripts/run_inference.sh
   ./scripts/run_inference.sh

Direct Python Usage

python scripts/inference.py \
    --image /path/to/el/image.png \
    --model /home/.../models/best_ensemble_model.pth \
    --output results/ \
    --gpu true \
    --size 224 \
    --threshold 0.2

Output Format

The system generates comprehensive analysis results including:

• Physics Model Prediction: Physics-based power prediction
• ResNet Model Prediction: Deep learning model prediction
• Ensemble Prediction: Final combined prediction
• Power Loss: Calculated as 1 - ensemble_prediction
• Relative Power: Direct ensemble prediction value
• Inactive Pixel %: Percentage of inactive pixels detected

Sample Output

Physics Model Prediction: 0.9007
ResNet Model Prediction: 1.3446
Ensemble Prediction: 0.8933
Power Loss: 0.1067
Relative Power: 0.8933
Inactive Pixel %: 0.9007

Configuration Parameters

Parameter	Description	Default Value
IMAGE_SIZE	Input image resolution	224
THRESHOLD	Inactive pixel threshold	0.2
USE_GPU	Enable GPU acceleration	true
CONDA_ENV	Conda environment name	solar

Model Performance

The ensemble model combines the interpretability of physics-based analysis with the pattern recognition capabilities of deep learning:

• Physics Model: Provides interpretable predictions based on inactive pixel analysis
• ResNet Model: Captures complex visual patterns in EL images
• Ensemble: Optimally combines both approaches for improved accuracy

File Descriptions

Scripts

• run_inference.sh: Main bash script for running inference with configuration
• inference.py: Python implementation of the inference pipeline
• physics_ensemble_final.ipynb: Jupyter notebook containing training and model development

Models

• best_ensemble_model.pth: Trained ensemble model weights (43MB)

Data

• data.csv: Dataset metadata with power measurements and cross-validation folds
• processed/: Directory containing 719 processed EL images
• processed.zip: Compressed archive of processed images

Results

• <image_number>_results.txt: Sample inference results showing model predictions

Technical Details

Model Architecture

• Input: Grayscale EL images (224x224 pixels)
• Physics Model: Threshold-based inactive pixel analysis
• ResNet Model: Modified ResNet18 with grayscale adaptation
• Output: Power prediction values (0-1 range)

Dependencies

• PyTorch
• torchvision
• PIL (Pillow)
• numpy
• argparse

License

This project is for research and development purposes. Please ensure compliance with relevant data usage agreements when working with solar panel EL images.

Contact

For questions or issues related to this solar panel EL image analysis system, please refer to the project documentation or contact the development team.

Credits

Codebase Creator: Harini Iyar Project: Solar Panel Electroluminescence (EL) Image Analysis
Date: July 2025