Turbulence-aware inflow feature engineering for physics-guided surrogate modeling of wind turbine dynamics

Physics-based aero-servo-elastic simulators are indispensable for wind turbine design and control but become prohibitively expensive when thousands of operating conditions must be evaluated. Data-driven surrogates offer a faster alternative, yet often ignore microscale turbulence and provide limited physical interpretability. This study proposes PIFENet, a turbulence-aware, physics-guided inflow feature-engineering framework for real-time prediction of wind turbine dynamic responses. Three-dimensional turbulent inflow fields are transformed into a low-dimensional, information-dense feature set that combines general flow statistics with turbine-centric structured features constructed from spatial weighting matrices. A minimum redundancy–maximum relevance algorithm is then used to select task-oriented features, which are fed into a multi-task neural network to simultaneously predict blade-root and tower-base loads, tower-top displacement, and generator power. Comprehensive OpenFAST simulations for the NREL 5 MW turbine across diverse wind speeds, turbulence intensities, and yaw conditions show that PIFENet achieves R2 values above 0.94 for all targets while faithfully reproducing both periodic and stochastic dynamics, with residual autocorrelation substantially reduced. The results demonstrate that physics-guided inflow feature engineering is an effective, modular route to robust surrogate models for wind turbine dynamics.