Research and optimization of time series prediction model based on lightweight Transformer

Le résumé fourni par la source

Traditional Transformer-based models are difficult to meet actual deployment requirements due to their high computational complexity and large resource consumption. To this end, this paper proposes a lightweight Transformer time series prediction optimization model based on local attention mechanism, low-rank feature compression and structural simplification. Through experimental evaluation on the ETTh1 standard dataset, the designed model significantly reduces the number of parameters and inference time while maintaining prediction accuracy, with an inference speed increase of 43% and a model complexity reduction of more than 50%. In addition, ablation experiments verify the contribution of each design module to the overall performance improvement. The research results show that the lightweight Transformer can balance inference efficiency and resource friendliness while ensuring accuracy, providing effective support for the optimization of time series prediction systems in practical application scenarios.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.