Long sequence multivariate time series recurrent network prediction method based on attention mechanism

Abstract

The application relates to the field of artificial intelligence and a recurrent neural network, and discloses a long-sequence multivariate time-series recurrent network prediction method based on an attention mechanism, which comprises the following steps: segment embedding coding is performed on original multivariate time-series data; local attention weights are calculated through a sparse local window attention module to generate a weighted feature sequence; time-series dynamic modeling is performed by using a gated recurrent state memory unit to output a recurrent state sequence containing long-term dependence; and multi-scale state information is fused through a cross-layer feature aggregation decoder to generate a multi-step prediction result. The system comprises corresponding functional units. Through the fusion of local sparse attention and a gated recurrent structure, the application significantly reduces the calculation complexity, improves the prediction accuracy and reasoning efficiency, and realizes the collaborative optimization of calculation resources and performance.

Images