Abstract

This research addresses a critical challenge in solar energy integration: the accurate prediction of power output fluctuations caused by cloud shadows. As photovoltaic installations continue to expand globally, the intermittent nature of solar resources presents significant obstacles for grid operators attempting to maintain stability and optimize power distribution. This study introduces a novel hybrid CNN-LSTM deep learning framework that analyzes sky imagery to track cloud movements and predict their impact on solar irradiance levels. The methodology involves preprocessing satellite imagery, extracting spatial features using convolutional neural networks, and leveraging Long Short-Term Memory networks to capture temporal dependencies in cloud patterns. Training was conducted using the "Solar Energy Forecasting using Low-Res Sky Images" dataset collected at Wollongong, Australia, comprising comprehensive sky imagery paired with corresponding PV power measurements. The proposed hybrid model achieved impressive performance metrics with Mean Squared Error of 0.0032, Mean Absolute Error of 0.042, and Mean Absolute Percentage Error of 17.48%. These results demonstrate substantial improvements over standalone approaches, with the CNN model yielding MAE of 0.6109 and the LSTM model producing MAE of 1.0986. Despite encountering overfitting challenges during implementation, which were addressed through regularization techniques, the hybrid architecture effectively captures both spatial features in cloud formations and temporal patterns in their movement. This research contributes valuable insights into applying deep learning techniques for solar energy forecasting and offers a promising solution for enhancing the reliability and efficiency of solar power systems, ultimately supporting broader renewable energy adoption and grid stability.

Project Overview

This project, led by Dong Linrui (Link), focuses on optimizing short-term solar energy prediction using cloud shadow mapping from sky images. The model integrates CNN, BiLSTM, and an attention mechanism to enhance the forecasting of irradiance patterns affected by cloud dynamics.

Key Contributions

• Introduced a hybrid deep learning model using CNN + BiLSTM + Attention for solar irradiance prediction.
• Utilized real-world sky imagery and weather data to model cloud motion vectors for forecasting solar output.
• Deployed an interactive dashboard allowing users to upload data, visualize predictions, and interpret attention maps.
• Achieved high accuracy and robustness as evidenced by RMSE, MAE, and R² metrics compared to traditional LSTM/CNN approaches.

Conclusion

The model effectively captures spatiotemporal features of cloud motion and their influence on solar irradiance. The use of attention mechanisms provides interpretable outputs, making it suitable for real-time energy management and forecasting applications in smart grid infrastructure.

Limitation

The model's accuracy is constrained by the availability and resolution of real-time sky images and the variability of atmospheric conditions in different regions. Hardware requirements for real-time processing and latency in satellite image feeds may also affect deployment in remote areas.

Future Work

• Integrate higher-resolution satellite feeds for improved shadow mapping granularity.
• Expand model to include multi-day forecasting capabilities using ensemble learning approaches.
• Deploy on edge devices with model pruning for real-time use in solar farms across developing regions.
• Collaborate with meteorological agencies to test model performance in live grid scenarios.