A time series forecasting model for high-resolution compressor aerodynamic noise prediction using nonlinear autoregressive neural networks

In the prediction of compressor aerodynamic noise, traditional transient computational fluid dynamics (CFD) simulations can provide detailed data on flow fields and noise sources. However, these methods face significant challenges in practical applications due to the requirement for very small computational timesteps and extensive computational resources. These limitations significantly impact the ability to deeply analyze and understand noise sources. To address these issues, this study introduces a Nonlinear Autoregressive Neural Network (NARNN) based time series prediction model, which can effectively utilize limited numerical simulation data to predict longer time series of sound pressure. This approach initially uses 0.02 seconds of numerical simulation sound pressure data to train the NARNN model, which can then extend the sound pressure data to longer durations, such as 0.08 seconds, 0.2 seconds, and 0.5 seconds. This method does not require additional computational resources and can significantly enhance the frequency resolution and prediction accuracy of the noise spectrum. To further optimize the prediction accuracy and applicability of the model, this study also explores the impact of using smaller training datasets (such as 0.015 seconds and 0.01 seconds) on model performance, aiming to identify the minimum dataset required for achieving high precision predictions. By comparing with experimental data and traditional numerical simulation results, the NARNN model was confirmed to keep the error within 3 dB of the Overall Sound Pressure Level (OASPL), demonstrating good consistency. These results not only validate the effectiveness of the NARNN model in reducing computational demands while maintaining high frequency resolution and accurate peak prediction capabilities but also provide a viable solution for computational resource constraints faced in turbomachinery design.