Informers for turbulent time series data forecast

Long-sequence time-series forecasting requires deep learning models with high predictive capacity to capture long-range dependencies between inputs and outputs effectively. This study presents a methodology for forecasting pressure time series in shock-wave, turbulent boundary layer interaction flows. Pressure signals were extracted below the λ-shock foot for six deformed rigid panel surface cases, where the low-frequency unsteadiness of the shock-boundary layer interaction is most prominent. The Informer model demonstrated superior performance in accurately predicting the pressure signals. Comparative numerical experiments revealed that the Informer model generally outperformed the Transformer, as indicated by lower root mean square errors and a more accurate power spectrum. The Informer effectively resolved the low-frequency unsteadiness of the λ-shock foot and better matched the ground truth's low- to mid-frequency power content. The forecasted pressure signals accuracy remained robust across all six rigid surface deformation cases, though subtle yet noticeable discrepancies still manifested. The accuracy of the Informer forecasted pressure time series was heavily dependent on the forecasting time step size. A step size of four provided a closer match to the ground truth in a deterministic manner, while a step size of eight achieved a better agreement in a stochastic sense. Larger time step sizes resulted in a gradual decline in accuracy.