CFD-guided memory-enhanced LSTM predicts leeward flow of railway windproof structures

Under the protection of windproof structures, wind speed along high-speed railway (HSR) lines decreases while turbulence intensity increases, particularly in transition segments between different windproof structure types. Considering the challenges of placing trackside sensors and the lengthy computational time required for numerical simulations, we propose a novel predictor that integrates computational fluid dynamics (CFD) with neural network (NN) methodologies to estimate internal wind speed signals in real-time based on external measurements from the windproof structures. This architecture combines an echo state network with a dynamic reservoir (dESN) as a feature extractor and a memory-enhanced long short-term memory (eLSTM) network as the output module. It is characterized by a dataset from an experimentally validated delayed detached eddy simulation (DDES)-based CFD model. Applied to a segment of the Lanzhou-Xinjiang railway crossing the Baili windy zone, the predictor achieves over 85% accuracy in estimating in-plane wind speeds and vertical wind profiles, outperforming conventional methods. Additionally, it demonstrates effectiveness under extreme conditions, including exceedingly high incoming flow speeds, absence of downstream measurement points, and sparse or distant sensor arrangements.