Spatiotemporal information conversion machine for time-series forecasting

Making predictions in a robust way is difficult task only based on the observed data of nonlinear system. In this work, neural network computing framework, spatiotemporal information conversion machine (STICM), was developed to efficiently and accurately render multistep-ahead prediction time series by employing spatial-temporal (STI) transformation. STICM combines advantages both STI equation temporal convolutional network, which maps high-dimensional/spatial future values target variable, thus naturally providing variable. From variables, also infers causal factors variable sense Granger causality, are turn selected as effective spatial improve robustness time-series. The successfully applied benchmark systems real-world datasets, all show superior performance prediction, even when were perturbed noise. theoretical computational viewpoints, has great potential practical applications artificial intelligence (AI) or model-free method data, opens new explore high-dimensional dynamical manner for learning.