Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA


Solar radiation nowcasting through advanced CNN model integrated with ResNet structure

Lei Chen 1, Yangluxi Li 1, Hu Du 1, Yukun Lai 2
1 Welsh school of Architecture, Cardiff University, United Kingdom
2 School of Computer Science and Informatics, Cardiff University, United Kingdom

Abstract: Although a range of solar radiation forecasting methods have been developed for predicting photovoltaic generation, only a few of them focus on solar radiation forecasting for building energy demand. From the perspective of building performance modelling, solar radiation forecasting needs to meet several critical requirements including high spatial resolution (1m-2km) and high temporal resolution (5-60mins), the accurate value of Direct Normal Irradiance (DNI) and Diffuse Horizontal Irradiance (DHI), which differs the requirement for predicting photovoltaic generation. As the geometric sum of DNI and DHI, accurate prediction of Global Horizontal Irradiance (GHI) with high spatialtemporal resolution tends to be the prerequisite for precise prediction of DNI and DHI. This research aims to construct a hybrid nowcasting model to predict GHI in high spatial-temporal resolution. In this article, the authors adopt an advanced Convolutional Neural Network (CNN) model with Residual Neural Network (ResNet) structure to identify the cloud image information and predict the GHI at 10 minutes intervals merely using cloud images captured by a ground-based sky camera. On this basis, several ResNet structures are compared to achieve the optimal nowcasting model for GHI. The results present that the ResNet structure can efficiently capture the cloud information and the ResNet152 achieves better performance than other alternative structures on the nowcasting of GHI. Finally, the authors discussed the calculation of synchronous DNI and DHI using the predictive GHI and Dirint model, and the application of DNI and DHI as the input for the simulation of building energy management. Key Contributions • Classifying solar radiation forecasting methods and reviewing recent representative articles. • Proposing a hybrid method using CNN-ResNet models to nowcast solar irradiance. • Only using ground-based cloud images as input to predict solar irradiation. • Comparing several ResNet structures to achieve the optimal nowcasting model for GHI. Practical Implications This study makes it possible for relevant researchers to utilise predictive solar radiation data in a convenient manner, taking only cloud images as input instead of other difficult-to-obtain weather data.
Keywords: Solar Irradiance Nowcasting, Global Horizontal Irradiance, High Spatial-Temporal Resolution
Pages: 1959 - 1966