SBIR Phase I: Proximate Wind Forecasts: A New Machine Learning Approach to Increasing Wind Energy Production

这个小型企业创新研究(SBIR)第一阶段项目的更广泛/商业影响将是展示以非常低的成本将现有风电场的风能产量增加(2%)的潜力。这项技术将分布在地形上的网络化气压传感器与人工智能/机器学习(AI/ML)相结合，将使风力发电场运营商能够提前发出即将到来的风和阵风的警报，从而先发制人地调整叶片螺距和涡轮机偏航等设置。这些调整将导致更多的风能生产和更少的涡轮机损坏。这项技术将显著增加能源收入，降低成本。2022年，美国风电场生产了380太瓦时(TWh)的能源。如果只为现有电厂的一半提供服务，这项技术每年可以额外产生3.8太瓦时的可再生能源和超过1.5亿美元的美国风能销售收入。在竞争激烈的风电行业，这些收入可以极大地提高运营利润率，并有助于加快行业和清洁能源就业的增长。使用政府的排放数据，这一部署还将在20年内避免24亿吨二氧化碳(GTCO2)。这个小型企业创新研究(SBIR)第一阶段项目将展示如何通过将新的传感器模式-分布式压力传感器-与新的机器学习(ML)模型相结合，在未来10-600秒内测量和预测风。压力传感器比风传感器(如多普勒激光雷达)便宜得多，但将压力传感器的数据处理成风的预测是复杂的。手工编码统计模型来根据地面压力测量来预测涡轮机高度风是不可能的。取而代之的是，人们可以依靠学习的ML模型来做出这些预测。以前的研究使用ML来模拟区域或全球尺度的天气，但该项目是第一个为适用于风电场运营的小得多、要求更高的尺度创建模型，并针对对风电场运营商重要的指标进行优化。由于ML模型还没有被直接开发用于在这种时空尺度上组合气压和风数据，该项目将结合基于注意力的模型(如变形金刚)的进步和尊重物理先验的模型的进步(如哈密尔顿神经网络)，并将导致一种新的传感形式，它将比以前在这个价格点可能实现的更准确。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project will be to demonstrate the potential to increase (by 2%) wind-energy production from existing wind farms at very low cost. Combining networked, air-pressure sensors distributed on the landscape with artificial intelligence/machine learning (AI/ML), the technology will empower wind farm operators with advance alerts of oncoming winds and gusts to preemptively adjust settings like blade pitch and turbine yaw. These adjustments will result in more wind energy production and less turbine damage. This technology will significantly increase energy revenues and decrease costs. In 2022, US wind farms produced 380 terawatt hours (TWh) of energy. If serving just half of existing plants, this technology could yield an additional 3.8 TWh of renewable energy and over $150 million to US wind energy sales annually. In the competitive wind industry, these revenues can greatly increase operating margins and help accelerate the growth of the industry and clean energy jobs. Using government emissions figures, this deployment would also avert 2.4 gigatons of carbon dioxide (GTCO2) over 20 years. This wind alert technology could also benefit solar tracker safety and increase safety at aerial vehicle ports and lift-crane operations.This Small Business Innovation Research (SBIR) Phase I project will show how wind can be measured and predicted 10–600 seconds in the future by combining a new sensor modality — distributed pressure sensors — with new machine learning (ML) models. Pressure sensors are far cheaper than wind sensors (e.g., Doppler LIDAR), but processing data from pressure sensors into predictions of the wind is complex. It is impossible to hand-code statistical models to predict turbine-height wind from ground-level pressure measurements. Instead, one may rely on learned ML models to make these predictions. Previous studies have used ML to model weather on regional or global scales, but this project is the first to create models for the much smaller and more demanding scales applicable to wind farm operation and to optimize for metrics important to wind farm operators. Because ML models have not yet been developed directly for combined pressure and wind data at this spatial and temporal scale, this project will combine advances in attention-based models (like Transformers) with advances in models that respect physical priors (like Hamiltonian Neural Networks) and will lead to a new form of sensing which will be far more accurate than was previously possible at this price point.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.