Sky Image Classification Based on Transfer Learning Approaches.

Area of Science:

• Computer Science
• Renewable Energy Systems
• Atmospheric Science

Background:

• Accurate forecasting of photovoltaic (PV) energy generation is crucial for grid stability and management.
• Cloudy conditions significantly impact local solar power output, posing a challenge for real-time energy management.
• Real-time sky condition assessment is vital for optimizing standalone PV system operations and managing energy consumption versus generation.

Purpose of the Study:

• To evaluate the effectiveness of deep learning models for classifying sky images to predict local weather conditions.
• To identify the most accurate convolutional neural network (CNN) architectures for sky image classification in the context of renewable energy forecasting.
• To leverage transfer learning (TL) techniques to enhance the performance of sky image classification models.

Main Methods:

• Utilized convolutional neural networks (CNNs) and transfer learning (TL) for sky image classification.
• Tested various architectures from the EfficientNet family and two ResNet models.
• Applied cross-validation methods to rigorously assess model performance across different experimental setups.

Main Results:

• Achieved a mean accuracy of 98.09% using EfficientNetV2-B1 and EfficientNetV2-B2 models.
• Demonstrated the high efficacy of the employed CNN architectures for classifying sky images.
• Identified specific EfficientNet models as yielding the best performance in sky condition classification.

Conclusions:

• The study confirms the significant potential of deep learning, particularly CNNs and TL, for accurate sky image classification.
• EfficientNetV2-B1 and EfficientNetV2-B2 models are highly effective for real-time sky condition assessment, improving renewable energy forecasting.
• Accurate sky image classification can enhance decision-making in renewable energy system operations, optimizing energy generation and consumption.