Unsupervised machine learning as a method to identify patterns in regional climate downscaling reanalyses

Regional climate downscaling reanalyses are used in a wide range of applications including wind-energy and pollution assessments, event planning, and epidemiological studies. A key component of these assessments are the atmospheric conditions traditionally obtained from climatological averages, which at best do not fully represent the range of possible conditions and at worst may actually be misleading with respect to actual atmospheric conditions. Using the daily climate record may be too time consuming and/or too costly. A more ideal solution is to examine select cases that represent the probability distribution of the climate record.

To improve the range of atmospheric conditions used in consequence assessments, we use an unsupervised machine learning technique based upon Kohonen neural networks (Self Organizing Maps, SOMs) to identify the atmospheric patterns in the historical record. The historical record is generated with a high-resolution limited area atmospheric model that re-analyzes the past 30 years based upon global data and local observations. The SOM is trained using daily patterns of multivariate output from the atmospheric model re-analysis. A good SOM analysis typically overestimates the total number of patterns resulting in many patterns that bear close similarity to each other, which can make interpretation of the results difficult. In order to overcome this limitation we apply an optimization stage based upon hierarchical trees to identify the minimum number of unique atmospheric patterns. We can then provide the analyst with a reasonable number of cases, along with the frequency of occurrence, that represent the full range of possible atmospheric conditions.