The increasing global demand for clean and sustainable energy has made accurate solar radiance prediction vital for enhancing solar energy utilization. This project proposes a novel deep learning architecture that integrates Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and an attention mechanism to forecast real-time sky radiance for optimal solar panel angle adjustment. The hybrid model leverages CNN to extract local spatial features, BiLSTM to capture complex temporal dependencies, and an attention layer to emphasize critical time steps that contribute most significantly to prediction accuracy. The study uses a real-world dataset from the DKA Solar Center, comprising 32,687 samples with features including radiation, temperature, humidity, wind direction, and speed. Comprehensive feature engineering, including temporal and astronomical data transformation, was conducted to enhance input quality. Data normalization using standard scaler was applied to ensure consistency across variable magnitudes. The model architecture employs a dual-channel Conv1D structure with different kernel sizes to simultaneously capture short- and mid-term temporal patterns. Following feature fusion and BiLSTM processing, a time-based soft attention mechanism is introduced to weigh sequence importance dynamically. The system was trained and evaluated using MAE, MSE, RMSE, and R² metrics. Results demonstrate strong performance with high prediction accuracy and generalization, particularly for short-term forecasting. Additionally, Explainable AI (XAI) via LIME was used to interpret model outputs, revealing the dominance of recent radiation values and pressure features in prediction efficacy. A user-friendly GUI was also developed, allowing real-time model training, multi-step prediction (1–12 steps), and interactive visual analysis of attention weights, activation maps, and prediction distributions. This study contributes a robust, interpretable, and deployable AI-based framework for solar radiance forecasting, offering tangible benefits to energy sector professionals, policymakers, and sustainability-focused researchers.
The project successfully demonstrates the potential of a hybrid deep learning framework to accurately predict sky radiance using real-time weather and satellite data. By combining CNN for spatial feature extraction and BiLSTM for temporal sequence modeling, enhanced by a focused attention mechanism, the model achieved outstanding accuracy with an R² of 0.982. The application facilitates solar panel angle optimization, maximizing energy efficiency and contributing to smarter renewable energy deployment.
Despite high accuracy, the model's predictive performance is constrained under extreme weather conditions or regions with limited data. The system also demands significant computational power for real-time processing, which may hinder deployment in remote or under-resourced areas.
Future directions include integrating reinforcement learning for adaptive panel control, enhancing robustness to missing data using generative models, and expanding the model’s usability through mobile-based applications. Incorporating drone-based sky imagery could further refine spatial feature extraction.
Yang Kaijia (Alex), majoring in Computer Science and Artificial Intelligence at Chengdu University of Technology - Oxford Brookes College, led this project with a passion for renewable energy and smart automation. His dedication to solving climate-related challenges with AI has inspired ongoing research into smart grid systems.