Wind Resource Data for Southeast Asia using a Hybrid Numerical Weather Prediction with Machine Learning Super Resolution Approach

This hybrid Numerical Weather Prediction + Generative Adversarial Networks (GANs) paradigm was used to generate 15-year high-resolution wind, temperature, and pressure data from January 2007 through December 2021 at multiple hub heights over Southeast Asia. The resolutions of this novel data set are 3-km horizontally and 15-minute temporally. The WRF model setup involved selecting optimal boundary conditions and physics options for simulations based on comparisons to tower observations in Bangladesh for a 21-day period during February 2016. Low-resolution (9-km grid) WRF simulations were performed over Southeast Asia, and the domains for these simulations were stitched together to provide a single final domain. The GANs were trained on coarsened WIND Toolkit data, and the stitched domain was enhanced by 3x along each spatial dimension and 4x along the temporal dimension. The hybrid model substantially reduces computational costs while producing high accuracy wind data appropriate for modeling and site evaluation. The hybrid approach using GANs downscaling to 3-km 15-minute resolution provides a 16x speedup when compared to carrying out WRF simulations at 3-km 15-minute resolution directly. Python code developed through this project for feature engineering, data handling, model training, and inference is publicly available on GitHub as the Super Resolution for Renewable Resource Data (sup3r) package. Validation of the WRF model involved comparing simulated wind speeds to tower observations in Bangladesh. The results showed that the WRF model performed well in simulating wind speeds, with a centered root mean square error (cRMSE) of 0.5 m/s. The GANs were also validated against WRF simulations, and the results showed that the GANs were able to downscale the WRF simulations to 3-km 15-minute resolution with high accuracy. The generated wind data can serve as essential inputs to many wind energy analysis tools and models and benefit a broad range of stakeholders working across Southeast Asia to deploy wind energy, including developers, policymakers, governments, utilities, system operators, technical institutes, and energy consultants. The hybrid model and the methods described in this report can be used to generate high-resolution wind data for other regions and can be applied to other renewable energy resources. The wind data generated through this project can be used to evaluate the potential of wind energy in Southeast Asia and to identify areas with high wind resources. The data can also be used to develop wind energy policies and to support the development of wind energy projects. The hybrid model and the methods described in this report can be used to generate high-resolution wind data for other regions and can be applied to other renewable energy resources. The project team used a combination of WRF simulations and GANs to generate high-resolution wind data for Southeast Asia. The WRF simulations were used to generate low-resolution wind data, which was then used to train the GANs. The GANs were then used to downscale the WRF simulations to 3-km 15-minute resolution. The resulting wind data was validated against tower observations in Bangladesh and against WRF simulations. The project team also developed a Python package called sup3r, which is publicly available on GitHub. The sup3r package includes code for feature engineering, data handling, model training, and inference. The package can be used to generate high-resolution wind data for other regions and can be applied to other renewable energy resources.