Solar energy prediction plays a key role in optimizing the integration of renewable energy into the grid. However, the complexity and variability of solar radiation patterns pose significant challenges. Traditional prediction methods often face difficulties in achieving high accuracy, particularly when trying to capture the complex spatial and temporal dependencies within solar data. Although the advancement of machine learning has enhanced the predictive ability, the inaccuracy of solar motion vector prediction remains a persistent problem. This project utilizes deep learning to enhance prediction accuracy by developing a new hybrid model, known as BiLSTM, combined with a dual-path architecture. This innovative model combines the advantages of BiLSTM in capturing time dependencies, enabling it to focus on both spatial and temporal granularities simultaneously while maintaining the computational efficiency of accurate solar motion vector predictions. The proposed model has demonstrated outstanding prediction. The performance is specifically manifested as a Test Loss as low as 0.0016, a Test Mean Absolute Error (Test MAE) of 0.0233, and a Test Root Mean Square Error (Test RMSE) of 0.0301. It indicates that the error between the predicted value and the actual value of the model is extremely small. Although the Test Mean Absolute Percentage Error (Test MAPE) reached 2198.8198, the Test Determination Coefficient (Test R²) of the model was as high as 0.982. This indicates that the model can accurately capture the underlying patterns of the data, and the reliability of the prediction and the fitting effect are extremely excellent, highlighting its potential in practical applications.
The hybrid deep learning framework demonstrated exceptional performance in solar energy prediction by learning complex cloud movement and solar irradiance dependencies. Its ability to process both spatial and temporal data efficiently makes it a reliable model for real-world grid applications and smart solar farm management.