A hierarchical tensor error correction GRU model

Abstract

Massive and complex longitudinal data processing of extended target-effective observations highlights a major challenge: existing methods are concerned with the difficulty of improving prediction robustness and ensuring relatively low computing costs. This paper proposes a novel hierarchical tensor processing error correction gated recurrent unit-effective observation prediction lower- and upper-bound estimation ensemble model. First, we introduce a continuous-time tensor value function to integrate time and context information and create a dynamic variable coefficient embedding layer of continuous-time tensor decomposition, including time granularity, observation, Gaussian component, and context information through a B-spline function approximation. The main advantages of this method include accurate normalized distance squared, effective threshold delimitation, and effective prediction boundary estimation. Second, to solve the error interference caused by the physical factors of multivariate observations, the lower and upper interval boundaries of prediction interval (PI) are deeply integrated into the proposed tensor model through the error factor correction method. The experimental results for aviation datasets from Central and Southern China demonstrate that the proposed model is better than other benchmark models in terms of PI output quality, prediction performance, and computing resource consumption.