Advice on model to use #2014
Unanswered
davidshumway
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Predicting future weather from weather stations.
Each weather station records 0-24 daily observations with at most one observation per hour. For each observation, there may be 1 or more characteristics measured, such as temperature, wind speed, wind direction, humidity, sky cover (e.g. "clear/overcast"), and solar radiation. Weather history is available for an extended timespan (perhaps 20 years is sufficient?). There are 15,000 stations worldwide. Each station has a lat/lon location.
The goal is to predict all features for all stations in one or more future timesteps. For example, if data is through 10am, then make a prediction for all features and stations at 11am.
One idea is to model the observations as features. Then use the spatiotemporal models suggested in the timeseries demos: (https://stellargraph.readthedocs.io/en/stable/demos/time-series/index.html).
Alternatively, data could be modeled as a knowledge graph for use with DistMult or ComplEx models. In this case, there would be nodes for
Stations
andObservations
, and relations for various characteristics such as temperature and wind speed.Observations
could be tied toStations
by afromStation
relation, and vice versa through ahasObservation
relation. Finally, each observed characteristic (temp., wind speed, ...) could also be a node type, e.g.Observation-01 -> hasTemp -> Temp1
, whereTemp1
is a node labeled by the observed temperature (e.g.32
).Beta Was this translation helpful? Give feedback.
All reactions