Parallel Adaptive Window Dynamic Time Warping Attention Transformer Network Model for Time Series Forecasting

Time series forecasting, a fundamental task in data analysis, holds significant application value across various sectors, including finance, energy, and transportation. This paper introduces an innovative Parallel Adaptive Window SDTWNet model, which integrates the Spring Dynamic Time Warping (DTW) algorithm with the attention mechanism of the transformer architecture. By leveraging a CUDA-implemented banded window constraint and a diagonal filling strategy, the model decomposes the dynamic programming matrix computation of DTW into parallelizable diagonal operations, effectively mitigating the serial bottleneck inherent in traditional DTW. Furthermore, the model dynamically adjusts the window radius at each time step through learnable parameters, enabling it to adaptively expand or contract the alignment search range based on the local characteristics of the sequence. This capability significantly enhances alignment accuracy and improves generalization. Compared to traditional DTW-based networks, the Parallel Adaptive Window SDTWNet notably reduces algorithmic complexity, leading to superior computational efficiency. Experimental results demonstrate that the proposed model achieves predictive performance comparable to state-of-the-art methods, highlighting its potential for practical applications.