Non-Stationary Bayesian Learning for Global Sustainability

Abstract

An increasingly warming planet calls for widespread use of sustainable energy sources like solar energy. To meet the rising energy demand, the focus of state-of-the-art solar energy models on local predictions is no longer sufficient as it only leads to local optimization of solar resources. Hence, a new class of models is needed that can provide a global response towards sustainability. In this paper, therefore, we propose a new approach that models cloud movement as a multilayer network and then performs parameter learning on it to generate short-term predictions of cloud fraction/solar irradiance simultaneously at a large number of locations. These learned parameters capture the spatio-temporal interdependencies of solar energy which can allow power-grid operators and policy-makers at different locations to know who impacts the solar energy of whom. Our results indicate a Root Mean Square Error (RMSE) of 8-18% in one-hour cloud fraction prediction. Finally, using our network approach, we show that the cloud movement likely follows a power law distribution, an important domain knowledge discovery that may be useful for future models. A major consequence of our approach is that it can enable power-grid operators/policy-makers to see beyond the local boundaries of their respective geographical locations.