A Machine Learning System for Pressure-Based Wind Profiling—Proof-of-Concept Results

Wind profiling gives the wind speeds at different heights above the ground. In addition to its value for weather research and forecasting, wind profiling has important applications in industries such as aviation and wind energy. In aviation, wind profiles have the potential to improve airport safety and utilization by providing real-time information about wind conditions and wake-vortices over the runways. In wind-energy, wind profiles are used in the planning stages for site selection as well during operations for real-time control of wind turbines to maximize their output and protect them wind gust damage.

The traditional instruments for measuring wind profiles have been based on the vertical emission of strong microwave, laser, or acoustic signals. Not only can these instruments be large and expensive, their emissions can interfere with communications and navigation equipment (microwave profilers), can pose a health hazard (laser profilers), or can cause a nuisance (acoustic profilers). Moreover, they may not be able to operate in rainfall (laser and acoustic profilers). Rather than an instrument that transmits energy into the atmosphere to measure the winds, in this paper we explore the possibility of a passive approach to wind profiling. The approach involves inferring wind profiles from the low frequency sounds produced by the winds passing over an infrasound sensor located at or near ground level. The advantage of such a pressure-based wind profiler over microwave, laser, and acoustic wind profilers is that it would it have low life-cycle cost, it would work via “passive listening,” emitting no potentially interfering or hazardous signals, and it would have all-weather capability, able to profile the winds in “clear air” and precipitating conditions.

The principle behind our proposed pressure-based wind-profiling approach is that wind creates turbulent eddies as it blows over the ground. These turbulent eddies have a range of physical sizes and they move along at the speed of the wind. As the eddies move over a fixed sensor they create a number of local effects. One effect is a change in refractive index at the eddy boundaries. This is the source of the Bragg scatter exploited by microwave and laser wind-profilers. Another effect is a change in the local speed of sound at the eddy boundaries. This is the source of scatter exploited by acoustic wind profilers. The effect we propose to exploit are the pressure fluctuations produced by the eddies as they pass over a sensitive pressure sensor located on the ground. In particular, it is our hypothesis that the pressure fluctuations produced by different size eddies can be exploited to infer wind speeds at different heights above the ground: smaller eddies providing information about wind speeds close to the ground; larger eddies providing information about wind speeds higher up above the ground.

Towards a proof-of-concept our hypothesis, the approach we propose for inferring wind profiles from pressure samples is a machine-learning approach whereby we attempt to train a machine learner to learn mappings between the pressure time-series recorded by a pressure sensor at ground level and the wind speeds recorded by anemometers located different heights above ground level. We formulate this as a supervised machine-learning problem and explore different machine learning systems ranging from two-layer artificial neural networks to random forests to the convolutional neural networks that are currently so popular. After describing the apparatus that we deployed for data collection, the final paper will present our data analysis and machine learning results.