Retrieving diurnal cycles from daily projections for impact studies

Impact studies, devoted to the energy demand in buildings or to the watershed hydrology, often need climate variable time series at the hourly time scale. However, climate projection outputs are mostly available at the daily time step, except for a few variables provided at the 3- or 6-hourly time step in some cases. In this paper, two ways of computing a diurnal cycle developed and used in impact studies at EDF/R&D are discussed.

The first approach has been defined to provide consistent hourly projections for four variables used to estimate the energy demand for heating/cooling of buildings at the 2050 horizon: temperature, wind speed, radiation and relative humidity at different geographical locations in France. The main idea is to use the distribution of the daily and geographical mean temperature over the whole French territory to identify an analogue day in the ERA5 reanalysis database. Then, for each variable and each location, the differences (for temperature) or ratios (for the other variables) between the daily mean and each hourly value are added to / multiplied by the daily mean value to assess local diurnal cycles. The approach is illustrated for 3 locations in France and its validation in terms of spatial and intervariable consistency is discussed, together with a highlight of the limitations and possible improvements.

The second approach uses a statistical bias correction method, namely the CDF-t (“Cumulative Distribution Function-transform”) approach (Michelangeli et al., 2009). The CDF-t was initially developed to spatially downscale and bias correct climate model projections by defining a correction function regarding the cumulative distribution functions of observed and modelled data over the reference and future time periods. In this work, it was adapted to temporally downscale climatic projections of surface thermal radiation downward from the daily to the 3-hourly time step for seven locations. It was based on daily series for the 1976-2065 time period and 3-hourly time series for 33 years scattered in the total time period. The CDF-t was applied as follow: 3-hourly time series are considered as “observations”. The 33 years displaying those data constitute the calibration time period, the remaining 57 years from the 1976-2065 time period are considered as the projection time period. Daily data are first roughly downscaled to the 3-hour time step by allocating the same daily values to all the 3-hour time steps across the 1976-2065 time period. Then the CDF-t method was applied to the 3-hour series considering the calibration and projection time periods. Results show satisfactory performances in terms on inter-annual and seasonal variability and cumulative distribution function. Mean annual bias is below 5% across the seven locations.

 

Reference:

Michelangeli, P.-A., Vrac, M., and Loukos, H. Probabilistic downscaling approaches: Application to wind cumulative distribution functions, Geophys. Res. Lett., 2009, 36, L11708, doi:10.1029/2009GL038401.