WeatherNet: Nowcasting Net Radiation at the Edge

In addition to natural processes such as photosynthesis and evapotranspiration, net radiation affects industrial applications such as photovoltaic energy management and solar thermal collection. We propose a deep learning approach for nowcasting net radiation within subhourly and intrahour horizons to better understand and control processes influenced by net radiation. Specifically, we developed a deep-learning-based CNN-LSTM, named WeatherNet, that combines multiple local ground-based cameras and weather sensor data to predict net radiation. Unlike previous methodologies, our approach involves images from three different cameras: a sky-facing RGB camera, a horizon-facing RGB camera, and a horizon-facing forward-looking infrared camera. Further, WeatherNet was designed to run "at the edge" using the Waggle edge computing framework to reduce the data bandwidth, improve the latency of predictions, and eliminate centralized data collection. With our proposed dataset and model, WeatherNet, we present a novel methodology using relatively inexpensive equipment for nowcasting net radiation precisely between a 15- and 90-minute horizon.

I. Premise

The task of weather forecasting in a short horizon (such as a 24-hour window) is called nowcasting [1]. Intrahour nowcasting can provide needed information to optimize industrial applications such as photovoltaic energy management and solar thermal collection and can provide crucial alerts about unexpected dangerous weather changes. Nowcasting is traditionally performed by using radar data and Numerical Weather Prediction (NWP) models [1]. However, NWP models perform poorly at intrahour nowcasting because of the limits of using radar data, such as high computational expense, spatial resolution, and infrequent availability of observations [2]. We posit that these methods can be improved by supplementing them with local sensor data.