Introducing NILM Metadata: a schema for energy datasets and prior knowledge about appliances
One of our aims with the open-source energy disaggregation toolkit NILMTK is to make it easy to import any of the 10+ NILM datasets currently available. One of the pain points when writing a NILMTK importer for a new dataset is that each dataset uses a different metadata schema and, sometimes, there simply is no metadata associated with some datasets. At best, this means that we have to manually map from the dataset's appliance names to NILMTK's standard appliance names. At worse, it means that it's impossible to unambiguously import the dataset (did that channel really only include the fridge? It sure looks like there are other appliances on there. What's the wiring hierarchy between the mains meter, the circuit meters and the appliance meters? Does that channel record active power or apparent power? What pre-processing has already been applied? etc. etc.)
NILM Metadata is a draft proposal for a schema and controlled vocabulary for describing energy datasets. It also stores a set of 'central metadata'. This includes simple stuff like the mapping from appliance type name (e.g. 'fridge') to a set of categories (e.g. ['cold', 'large']) and the mapping from country name to nominal mains voltages for that country. But we also try to capture probabilistic prior information about variables such as the on-duration of each appliance, the on-power of each appliance, which appliances tend to be used with which other appliances etc. The 'central metadata' uses a simple but powerful inheritance mechanism. For example, a 'fridge' is a type of 'cold appliance' and inherits properties from 'cold appliance'. This inheritance allows us to minimise redundancy when specifying the common metadata, and it also allows us to propagate prior knowledge down the inheritance tree. For example, say we have no prior knowledge about the times-of-day that a plasma TV is used. But we know that a plasma TV is a subtype of a television, so we can use our prior knowledge about televisions.
NILM Metadata is on github, and has a fair amount of documentation (including a step-by-step guide to describe the REDD dataset using NILM Metadata), and a research paper.
I would dearly love for NILM Metadata to become a community project and for folks to modify and extend it to their needs. It really is just a draft at the moment - I'm sure there's lots missing from the current proposal. Please go ahead and fork the project on github and submit a pull request with your changes (here's a guide from github on forking). Or add ideas to the issue queue.
Some questions on the evolution of NILM Metadata:
- Perhaps the 'central metadata' should be stored in a semantic wiki rather than a set of YAML files to make it really easy for people to edit, and for machines to query it.
- How best to represent inferred models of appliances? I've made a start at defining an ApplianceModel schema but it needs more thought
- How best to represent event-based datasets? (Mario Bergés and I discussed this at the NILM Workshop in Austin (his BLUED dataset is event-based)). Perhaps each ApplianceModel could enumerate every state for every appliance and then these states could be referenced from the event-based dataset? How to define a set of states for a complex appliance like a washing machine or a computer? Maybe we would have multiple ways to define states... e.g. the set of states that a human operator would care about (wash, spin, dry etc).