Nonlinear dynamic systems modeling using Gaussian processes: Predicting ionospheric total electron content over South Africa

Two different implementations of Gaussian process (GP) models are proposed to estimate the vertical total electron content (TEC) from dual frequency Global Positioning System (GPS) measurements. The model falseness of GP and neural network models are compared using daily GPS TEC data from Sutherland, South Africa, and it is shown that the proposed GP models exhibit superior model falseness. The GP approach has several advantages over previously developed neural network approaches, which include seamless incorporation of prior knowledge, a theoretically principled method for determining the much smaller number of free model parameters, the provision of estimates of the model uncertainty, and a more intuitive interpretability of the model. Key Points Gaussian processes are proposed as an attractive alternative to neural networks Model falseness should not be used as the only metric for empirical models