A0466
Title: Multi-output Gaussian process based on neural kernel learning and its prediction applications
Authors: Hsiang-Ling Hsu - National University of Kaohsiung (Taiwan) [presenting]
Abstract: The basic concepts of multivariate normal distribution, kernels, non-parametric models and conditional probability are essential to build a Gaussian process. The flexibility and uncertainty measure inherent in predictions makes the Gaussian process regression models widely applicable to various fields. However, Gaussian process regression is based on the similarity of the sample function space and independently builds each predictive model but ignores the correlation between time points. On the other hand, the prediction results of the Gaussian process are influenced by the structure assumptions of the covariance function. Hence, a multi-output Gaussian process regression model attempts are proposed to combine different kernel functions through a neural kernel network framework to acquire the complicated correlation patterns between data for predictions. The electric load data of Kyushu in Japan and the apparent temperature of Lingya, Kaohsiung City, Taiwan, are analyzed based on the proposed method with two kinds of evaluation indexes, the root mean square error (RMSE) and mean absolute percentage error (MAPE), to measure the prediction performances. Compared with persistence, seasonal autoregressive integrated moving average (SARIMA) and traditional Gaussian process regression (GPR), the experimental results show that the proposed method possesses low MAPE and RMSE in forecasting the electricity load and apparent temperature 2020.
