Demand allocation based on population at census tract level population.
Can take input demand from arbitrary geometries (HIFLD, EIA 861 service territories, etc.)
Currently checking our aggregations against the monthly sales reported to EIA 861 by customer class.
At the state level, there's a 99% correlation between the reported EIA 861 sales, and our state-by-state allocations. The total reported demand/sales are different by a constant factor of about 12% in the two datasets.
Have mostly cleaned the time-series data from FERC 714, converting all times into UTC + appropriate timezones, identified and fixed data reporting errors (non-standard timezone codes, demand reported as a negative rather than positive number, etc.)
Generalized method for applying location dependent weightings to the demand allocation, so that we can use state, regional, or other variables to improve the match between our demand allocation and the target inputs.
Unexpected things that have come up
We need to generate our own planning area geometries, since none of FERC, EIA, or HIFLD appear to have them.
About 88% of total reported demand is associated with one of the HIFLD geometries in 2018. This drops gradually to 58% in earlier years as utility IDs and/or service areas changed, so there's a substantial lack of coverage, even in the current year.
Generally using the methodology outlined in Auffhammer et al.
What we're working on now / Decision points:
Generating historical geometries using EIA 861 service territory counties
Re-aggregating the allocated demand up to larger areas using the full time series, not just monthly or annual totals.
Trying to understand systematic under / over allocation of demand in some states, and whether it corresponds to relative importance of different customer classes in this states, which planning areas are involved, region of the US, other variables.
Categorization and quantification of remaining outlier demand values in demand time series.
Imputation of appropriate values where demand time series is missing data or contains bad data. Generally following the process outlined by Tyler Ruggles in his recent work on EIA 930 data.
What we've gotten done from the original scope.
Unexpected things that have come up
What we're working on now / Decision points: