Improving Satellite Cloud Thickness Characterization by means of Ceilometer Data

As PV technology evolves, more sophisticated methods of solar irradiance forecasting and ex-post estimation are needed for PV power forecasting and monitoring purposes, especially regarding the different components ( Direct, Global and Diffuse) and their spectral distribution. To achieve an accurate prediction across the spectrum for each component, radiative transfer models (RTMs) are often used; however, they are based on a vertical characterization of the atmosphere, and particularly of cloud thickness, which  traditionally is retrieved as a fixed thickness values for each different cloud type (under the assumption that same-type clouds have the same thickness) or from NWP models.

In this field, a relatively new and quite promising technology that helps in better characterizing the clouds is the ceilometer, a ground-based lidar instrument designed for vertical profiling offering high-resolution assessments of cloud properties. RSE S.p.A. installed one such instrument (Vaisala CL61) in Milan (Italy) in June 2023. The goal of this early-stage study is to exploit the ceilometer’s cloud base measurements and satellite’s cloud top height estimations to train a machine learning model to derive the information on the vertical thickness. We aim to use  the same model to obtain experiment-based estimates of cloud thickness over the entire Italian domain, to be used in RSE’s choice RTM (libRadtran).

The starting dataset to train the model consists of:

• Backscatter and depolarization measurements from the Milan-based ceilometer, from which it is straightforward to obtain the cloud base height. The data has a 1-minute time frequency and a 5m vertical resolution over a 10-month period.
• Meteosat Second Generation satellite data, from which cloud position, cloud type and cloud top height have been inferred. The data has a 15-minute time frequency and a 4 km spatial resolution (parallax and data acquisition time have been accounted for).

From the dataset, it is straightforward to calculate the thickness of single-layer clouds as the difference between the cloud top (satellite) and cloud base (ceilometer). The cloud thickness has therefore been used as predictand in a random forest model (RF), using the cloud type and the cloud top height as predictors. Other meteorological variables can also be added as predictors.

The method’s accuracy has been tested indirectly by comparing the irradiance components calculated using libRadtran fed with the RF derived cloud thickness with respect to the experimental ones, registered in different Italian sites.