Tyler McCandless


Machine Learning Techniques for Renewable Energy and Wildfire Prediction

What HOMEPAGE GR Meteo Colloquium
When Jan 16, 2019
from 03:30 pm to 04:30 pm
Where 112 Walker Building, John Cahir Auditorium
Contact Name George Young
Contact email
Add event to calendar vCal

McCandless 2019

The National Center for Atmospheric Research (NCAR) has been a leader in many areas of atmospheric science for applications of artificial intelligence (AI) to environmental science problems. The research in AI at NCAR began in the 1990’s, first from advanced applications of combined statistical methods, expert systems, and genetic algorithms to the full complexity of modern AI including machine learning and deep learning.  This talk will detail three projects at NCAR’s Research Applications Laboratory utilizing machine learning for improved short-range solar power prediction, wind power conversion and wildland fire decision support systems. 

Energy grid systems operators and utilities require accurate solar power forecasts to effectively balance the supply and demand of electricity due to the inherent variability in solar power production. The objective of a forecasting system is to model the actual relationships between the predictors and the predictand, and in the case of solar power forecasting, the relationship between the predictors and the predictand is frequently non-linear, especially in the case of significant cloud cover changes. One AI method uses an unsupervised approach to specifically classify regimes (k-means clustering) and then trains a supervised technique (artificial neural network) separately on each regime, which is correlated to utilizing direct relationships as described in causality theory. This research quantifies the value in a regime-dependent method predicts short-range (15-min to 180-min) solar power generation, which has inherent weather regimes, or cloud types, that have varying levels of predictability and non-linear relationships between predictors and predictand.  

Similar to solar power, wind power is a variable generation resource and therefore requires accurate forecasts for integration into the electric grid. Generally, the wind speed is forecast for a wind plant and the forecasted wind speed is converted to power to provide an estimate of the expected generating capacity of the plant. The average wind speed forecast for the plant is a function of the underlying meteorological phenomena being predicted; however, the wind speed for each turbine at the farm is also a function of the local terrain and the array orientation. Conversion algorithms that assume an average wind speed for the plant, i.e. the super turbine power conversion, assume that the effects of the local terrain and array orientation are insignificant in producing variability in the wind speeds across the turbines at the farm. Here, this research quantifies the differences resulting from Jensen’s Inequality in converting wind speed to power at the turbine-level compared to a super turbine power conversion for a hypothetical wind farm as from empirical data. This research shows that machine learning algorithms can be trained to over-come the power conversion differences caused by Jensen’s Inequality to accurately convert wind speed to wind power for the total power generated at the wind farm. 

Successful wildland fire decision support systems depend on accurate predictions of wildland fire spread and the fuel moisture content (FMC) is one of the critical drivers influencing the rate of a wildland fire’s growth. However, FMC is a sparsely and infrequently measured surface variable compared to most atmospheric variables. A high-resolution, gridded, real-time FMC data set does not currently exist for assimilation into operational wildland fire prediction systems. This research utilizes surface observations of live and dead FMC to train machine-learning models to estimate FMC based on satellite observations and surface characteristics. Random forests and gradient boosted regression machine learning algorithms are applied across the entire spatial grid for an initial test domain of Colorado, populated by Moderate Resolution Imaging Spectrometer (MODIS) Terra and Aqua reflectances as predictors, to achieve a gridded, real-time FMC dataset.