A novel framework for temporal super-resolution of wind in urban energy applications

Timely and precise access to the temporal evolution of planetary boundary layer wind is crucial for urban wind energy scheduling and management. Nevertheless, the expensive costs of predicting high temporal resolution turbulence using physical models hinder engineering applications. Deep learning techniques have become a promising alternative to numerical methods, whereas current studies on the high temporal resolution wind reconstruction remain scarce. This study proposes a novel framework incorporating Sparse Window-based Attention for cost-effective super-resolution of turbulence fields. It allows the customization of attention sparsity by modifying the stride value. Relative Physical-informed Loss is proposed to guarantee the physical plausibility of the generated wind fields. Compared to Window-based Attention, the proposed attention significantly reduces computational costs and improves the inference efficiency. Even though increasing interpolated wind snapshots slightly reduces performance, the model still reconstructs the wind field's structure. Evaluations of statistical metrics, turbulence features, power spectra, and coherence functions exhibit strong physical consistency of the reconstructed wind fields. Meanwhile, it shortens training time by 32.92 % and decreases computational energy consumption by 32.89 %. Larger strides further reduce energy consumption but at the cost of decreased performance, highlighting a trade-off between accuracy and efficiency.