Research on Dynamic System Resource Prediction Based on Transformer
Time series of dynamic system resources typically exhibit complex long- and short-term dependencies, posing significant challenges for accurate multi-step forecasting. To address the limitations of conventional single-architecture models in capturing both global patterns and local dynamics, this paper proposes an enhanced multivariate forecasting model, PatchTST_SLG(Serial LSTM-GRU). This model utilizes the Patch Time Series Transformer (PatchTST) as the backbone to extract global dependencies from long sequences and serially integrates Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) modules at the output to smooth prediction trajectories and correct local errors. Experiments on the Abilene dataset demonstrate that the proposed model significantly outperforms baseline models, including LSTM, GRU, and the original PatchTST, achieving substantial reductions in MAE and RMSE. The results confirm the effectiveness of this serial structure and its significant contribution to improving predictive stability.