End-to-end wind power modelling: developing physics-informed machine learning models for atmospheric fluid dynamics

到本世纪末，人类活动产生的温室气体排放必须接近于零，才能稳定气候。全球能源供应部门是这些排放的最大单一来源（约35%），因此其深度脱碳势在必行。风力发电是最有前途的技术之一，因为它具有广泛的可用性、高通用性和成熟的供应链。管理风能的可变性和间歇性是一个关键的挑战，这需要更好的风能预测。一般来说，风力发电传统上有两种预测方法：1)通过代表第一原理的物理方程，或2)通过数据驱动的统计模型，将历史风力发电与相关的气象变量（如风速、风向、气温和压力）结合起来。这两种方法在地形和位置的通用性以及准确性方面都受到限制，因此需要新的方法。该研究项目旨在将物理系统和数据驱动的机器学习建模这两个领域结合到一个建模框架中，以产生有关风电预测和生产规律的基本知识。为了实现这一目标，该研究计划将专注于开发：1)用于物理信息神经网络（PINN）的新型深度学习架构；2)新的输入数据特征选择优化算法。综合起来，这些努力将对大气流体流动和动力学产生新的、基本的见解。通过将不完全确定的风速物理方程与数值天气预报（NWP）模型产生的大数据以及从气象塔测量的数据相结合，提出的方法将实现更可靠的风力预测。据推测，新的建模框架将在预测精度方面提供一个重大飞跃，从而可以从根本上实现风电生产在电力系统中的更高集成度。这项研究还将通过对风力建模的深入研究，对流体动力学的基本模型产生见解。为将流体动力学集成到深度学习神经网络中而开发的总体框架也将适用于广泛的机械、航空航天、生物医学和化学工程问题。

By the end of the century, greenhouse gas emissions from human activities must approach near-zero in order to stabilize the climate. The global energy supply sector is the largest single source of these emissions (approximately 35%), rendering its deep decarbonization an imperative. Wind power is one of the most promising technologies to contribute to this, given its extensive availability, high versatility, and mature supply chain. Managing the variability and intermittency of wind power is a key challenge-one that demands better wind power prediction. Broadly, wind power has traditionally been predicted using two approaches: 1) through physical equations that represent first principles, or 2) through data-driven statistical models that combine historical wind power production with associated meteorological variables like wind speed, wind direction, air temperature and pressure. Both of these approaches are limited in their generalizability across terrain and location, as well as their accuracy, hence the need for new methods. This research program seeks to combine two domains-physical systems and data-driven machine learning modeling-into one modeling framework, in order to generate fundamental knowledge into the laws governing wind power forecasting and production. To achieve this goal, the research program will focus on developing: 1) new deep learning architectures for physics-informed neural networks (PINN); and 2) new optimization algorithms for input data feature selection. Combined, these efforts will generate new, fundamental insights into atmospheric fluid flow and dynamics. The proposed approach would enable more reliable wind power prediction, by combining what is known from incompletely specified physical equations of wind speed with big data generated from numerical weather prediction (NWP) models and measured from meteorological towers. It is hypothesized that the new modeling framework will provide a significant leap forward in predictive accuracy that could enable radically higher integration of wind power production in the power system. This research would also generate insights into the fundamental models of fluid dynamics through the intensive study of wind power modeling. The general framework developed for integrating fluid dynamics into deep learning neural networks will also be applicable to a wide range of mechanical, aerospace, biomedical and chemical engineering problems.