Quantile Regression Ensemble Summer Year (QRESY)

Data collection method:

The Quantile Regression Ensemble Summer Year (QRESY) creation process starts by collecting hourly weather data over a long period. Typically, weather files attempt to be representative of periods around 20-40 years, and here we do the same, however much longer periods could be used. The existence, variables and quality of hourly weather data varies depending on the location. The variables usually include temperature, atmospheric pressure, cloud cover, wind speed and wind direction, precipitation, etc.

Data processing and preparation activities:

For the QRESY process, preprocessing of this data is required to ensure it contains no long sequences of missing data. If large amount of data are missing in any of the variables, the whole year is removed from the analysis. At this point it is also necessary to decide the target level of extreme weather to work with. That is, to fix the quantile level for the subsequent construction of the Quantile Regression models depending on the distance to the median (quantile 50, Q50). Running a Quantile Regression model for every year under analysis is an ``embarrassingly parallel'' problem, as it is straightforward to separate the problem into a number of parallel tasks and the code run on a parallel machine. The set of regressions is combined in a unique year of hourly data. This is done by endowing higher weights to quantiles away from Q50 for ensembles within upper quantiles. At the same time, it increases the importance of quantiles near to Q50 for combining lower quantiles. The idea being to focus on explaining critical phases of summer temperatures. Each ensemble is thereby made over the predictors of a number of regression models corresponding to each of the years in the database. The ensemble parameters can be tuned by cross-validation over random partitions of the data into training and test summer periods.