BOSSWAVE interface on top of Clipper
Model calls for thermal model (Python):
# start at midnight
normal_schedule = [
# midnight - 8:00am
(50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90), (50, 90),(50, 90),
# 8:00am - 4:00pm
(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),(70, 74),
# 4:00pm - 6:00pm
(70, 74),(70, 74),(70, 74),(70, 74),
# 6:00pm - 12:00am
(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90),(50, 90)
]
params = {
'zone': 'http://buildsys.org/ontologies/ciee#CentralZone',
'date': '2018-02-06 00:00:00 UTC',
'schedule': normal_schedule
}
import requests
import json
resp = requests.post('http://localhost:1337/<model name>/predict', data=json.dumps({'input': params}))
print resp.json()['output']For each model, we keep the following metadata:
- model name
- model version (combination of name, version is unique and immutable)
- predictive range + resolution (list of days, date ranges that the model can predict)
- possibly support other definitions of what are the 'good' ranges for this model
- possibly a function that gets called?
model.can_predict(start date, end date, resolution) => {true, false}
- training data streams
- the URIs or UUIDs or Brick model names of the streams used to train this model
- can be used to determine if user can access this model (based on the URIs and BOSSWAVE DoTs)
- optional
- training data range
- the range of data this model was trained on
- optional
Models are defined as a Python closure running through Clipper. This gives us a very flexible execution model. These models could be a simple evaluation such as using a weight vector, a more involved iteration (such as MPC under some schedule), or even a call against an external API (like weather underground -- we'd likely use a caching layer here too to avoid hitting the API too many times).
It should be super simple to deploy a model from a piece of Python code; likely just a simple wrapper for Clipper (make the call idempotent, or automatically increment the version number).
Query contains:
- model type: (consumption, thermal, occupancy, weather)
- model context: (zone, room, etc)
- prediction range: start, end timestamp e.g.
2018-01-30 00:00:00 PST - 2018-01-31 00:00:00 PST - prediction granularity:
30min,1h
Idea is that you could use these parameters to have the model server choose the model.
Optional parameters:
- data to use to train the model
- entity to execute the model as a given entity
- model name (invoke a model directly)