A novel spatiotemporal Fourier neural operator for dynamic wake prediction

Accurate and efficient modeling of wake dynamics is essential for power optimization and load mitigation in wind farms. Analytical wake models and numerical simulation methods are either limited to steady-state scenarios or computationally expensive. Recent advances in deep learning surrogates have shown promise in accelerating dynamic wake prediction, yet most struggle to maintain accuracy over long-term horizons and to generalize across multiple different parameter combinations, including inflow wind speeds, turbulence intensities, and yaw angles, which are critical for effective wake control to optimize wind farm performance. To address these challenges, we propose the spatiotemporal Fourier neural operator (ST-FNO), a novel physics-guided operator learning framework that reformulates dynamic wake prediction as a parameterized PDE-solving problem. ST-FNO jointly models spatial wake characteristics and long-term temporal dependencies through frequency disentanglement, frequency-specific enhancement, and progressive feature fusion with a Transformer backbone. Extensive experiments on a DWM-generated dataset covering diverse inflow and control conditions demonstrate that ST-FNO achieves superior short- and long-term accuracy, strong generalization, and millisecond-level inference efficiency in parameterized dynamic wake prediction task. Comprehensive ablation studies and hyperparameter sensitivity analyses, which are important yet rarely reported in prior work, further confirm the necessity and robustness of the proposed design. To the best of our knowledge, this is one of the first attempts to introduce neural operator learning into dynamic wake prediction.