Cloud motion is predictable on a 10 minute timescale (and potentially longer with more sophisticated models), and the LSST scheduler has the opportunity to use such predictions when generating its schedules. Without predicted cloud data, the scheduler may plan out an optimal path, only to have it disrupted by an encroaching cloud that could have been avoided more effectively. To solve this problem, we have created this cloud prediction model, which uses streaming cloud data to predict cloud propagation.
We model the sky as 2D cartesian sheet with clouds that all move together at a constant velocity. We calculate the velocity with pairwise cloud map comparisons spaced apart in time, and we use a median filter to reduce noise.
The CloudMap class represents a map of the clouds at a particular time. A series of CloudMap instances are managed by a CloudServer instance. The CloudServer class processes streaming cloud map data that is passed in as it becomes available. It also provides a prediction method that returns a predicted cloud map based on the data currently available in the CloudServer.
The Cloudmap objects compute the 2D FFT of the cloud sheet, and velocities are computed by cross-correlating pairs of Cloudmaps.
During full operations, an all-sky camera will be periodically imaging the whole sky in order to track clouds. Every time an image is taken, a HEALPix representation of the image will be sent to the cloud prediction code. There, it needs to be converted to a CloudMap using the CloudMap.fromHpix method. The resulting CloudMap needs to be posted to a running CloudServer instance using the CloudServer.postCloudMap function. Then, whenever the scheduler needs a cloud prediction, it can call CloudServer.predCloudMap(mjd).
Alternatively, under the publish/subscribe architecture, whatever code is hosting the cloud prediction code can periodically call CloudServer.predCloudMap and then publish the result.
Sample data is available at lsst-dev01.ncsa.illinois.edu:/datasets/public_html/sim-data/cloudy_night_example