Regional Solar Irradiance Prediction In Vietnam Using Deep Learning

Abstract

The necessity of weather forecasting in general and solar power in specific using deep learning is completely reasonable for weather prediction, transportation planning, and the overall social human being. Accurate solar power provides the essential role of sophisticated forecasting tools in improving agricultural efficiency for farmers, forwarded information for logistics activities. Driven by the necessity to assess solar energy's potential and allocation within Vietnam, specifically focusing on Da Nang, this research offers critical insights into the feasibility and enhancement of solar power resources. The methodology leverages a Knowledge Distillation strategy to refine the prediction process by filtering the behavior from an elaborate teacher model to a lightweight student model. The data utilized in this study includes variables such as wind velocity, humidity, air temperature, Direct Normal Irradiance (DNI), Global Tilted Irradiance (GTI), and the direction of wind, enabling a comprehensive examination of their relationship with GHI. As the current conventional weather forecasting techniques face a numerous of difficulties such as: diverse of correlated features, huge volume, industrial impact, human impact. This approach allows for the efficient prediction of GHI, thus streamlining the feature analysis. The outcomes indicate that the Knowledge Distillation method, from the teacher to the student model, significantly outperforms a direct training approach on the student model, showcasing the method's effectiveness in boosting prediction accuracy. The forecasting metrics showed a significant result in predicting from two scenarios which is trained from teacher to student and other model is trained directly to student. This result indicates the importance of training model from distillation teacher.