"""Code for transforming EIA data that pertains to more than one EIA Form.
This module helps normalize EIA datasets and infers additonal connections
between EIA entities (i.e. utilities, plants, units, generators...). This
includes:
- compiling a master list of plant, utility, boiler, and generator IDs that
appear in any of the EIA 860 or 923 tables.
- inferring more complete boiler-generator associations.
- differentiating between static and time varying attributes associated with
the EIA entities, storing the static fields with the entity table, and the
variable fields in an annual table.
The boiler generator association inferrence (bga) takes the associations
provided by the EIA 860, and expands on it using several methods which can be
found in :func:`pudl.transform.eia._boiler_generator_assn`.
"""
import importlib.resources
from collections import namedtuple
from enum import StrEnum, auto
from typing import Literal
import networkx as nx
import numpy as np
import pandas as pd
import timezonefinder
from dagster import (
AssetIn,
AssetOut,
AssetsDefinition,
Field,
Output,
asset,
multi_asset,
)
import pudl
from pudl.helpers import convert_cols_dtypes
from pudl.metadata import PUDL_PACKAGE
from pudl.metadata.enums import APPROXIMATE_TIMEZONES
from pudl.metadata.fields import apply_pudl_dtypes, get_pudl_dtypes
from pudl.metadata.resources import ENTITIES
from pudl.settings import EiaSettings
[docs]
logger = pudl.logging_helpers.get_logger(__name__)
[docs]
TZ_FINDER = timezonefinder.TimezoneFinder()
"""A global TimezoneFinder to cache geographies in memory for faster access."""
[docs]
class EiaEntity(StrEnum):
"""Enum for the different types of EIA entities."""
[docs]
def find_timezone(*, lng=None, lat=None, state=None, strict=True):
"""Find the timezone associated with the a specified input location.
Note that this function requires named arguments. The names are lng, lat,
and state. lng and lat must be provided, but they may be NA. state isn't
required, and isn't used unless lng/lat are NA or timezonefinder can't find
a corresponding timezone.
Timezones based on states are imprecise, so it's far better to use lng/lat
if possible. If `strict` is True, state will not be used.
More on state-to-timezone conversion here:
https://en.wikipedia.org/wiki/List_of_time_offsets_by_U.S._state_and_territory
Args:
lng (int or float in [-180,180]): Longitude, in decimal degrees
lat (int or float in [-90, 90]): Latitude, in decimal degrees
state (str): Abbreviation for US state or Canadian province
strict (bool): Raise an error if no timezone is found?
Returns:
str: The timezone (as an IANA string) for that location.
Todo:
Update docstring.
"""
try:
tz = TZ_FINDER.timezone_at(lng=lng, lat=lat)
if tz is None: # Try harder
# Could change the search radius as well
tz = TZ_FINDER.closest_timezone_at(lng=lng, lat=lat)
# For some reason w/ Python 3.6 we get a ValueError here, but with
# Python 3.7 we get an OverflowError...
except (OverflowError, ValueError) as err:
# If we're being strict, only use lng/lat, not state
if strict:
raise ValueError(
f"Can't find timezone for: lng={lng}, lat={lat}, state={state}"
) from err
tz = APPROXIMATE_TIMEZONES.get(state, None)
return tz
[docs]
def occurrence_consistency(
entity_idx: list[str],
compiled_df: pd.DataFrame,
col: str,
cols_to_consit: list[str],
strictness: float = 0.7,
) -> pd.DataFrame:
"""Find the occurence of entities & the consistency of records.
We need to determine how consistent a reported value is in the records
across all of the years or tables that the value is being reported, so we
want to compile two key numbers: the number of occurances of the entity and
the number of occurances of each reported record for each entity. With that
information we can determine if the reported records are strict enough.
Args:
entity_idx: a list of the id(s) for the entity. Ex: for a plant
entity, the entity_idx is ['plant_id_eia']. For a generator entity,
the entity_idx is ['plant_id_eia', 'generator_id'].
compiled_df: a dataframe with every instance of the
column we are trying to harvest.
col: the column name of the column we are trying to harvest.
cols_to_consit: a list of the columns to determine consistency.
This either the [entity_id] or the [entity_id, 'report_date'],
depending on whether the entity is static or annual.
strictness: How consistent do you want the column records to
be? The default setting is .7 (so 70% of the records need to be
consistent in order to accept harvesting the record).
Returns:
A transformed version of compiled_df with NaNs removed and with new columns with
information about the consistency of the reported values.
"""
# select only the colums you want and drop the NaNs
# we want to drop the NaNs because
col_df = compiled_df[entity_idx + ["report_date", col]].copy()
if get_pudl_dtypes(group="eia")[col] == "string":
nan_str_mask = (col_df[col] == "nan").fillna(False)
col_df.loc[nan_str_mask, col] = pd.NA
col_df = col_df.dropna()
if len(col_df) == 0:
col_df[f"{col}_is_consistent"] = pd.NA
col_df[f"{col}_consistent_rate"] = pd.NA
col_df["entity_occurences"] = pd.NA
return col_df
# determine how many times each entity occurs in col_df
occur = (
col_df.assign(entity_occurences=1)
.groupby(by=cols_to_consit, observed=True)[["entity_occurences"]]
.count()
.reset_index()
)
# add the occurances into the main dataframe
col_df = col_df.merge(occur, on=cols_to_consit)
# determine how many instances of each of the records in col exist
consist_df = (
col_df.assign(record_occurences=1)
.groupby(by=cols_to_consit + [col], observed=True)[["record_occurences"]]
.count()
.reset_index()
)
# now in col_df we have # of times an entity occurred accross the tables
# and we are going to merge in the # of times each value occured for each
# entity record. When we merge the consistency in with the occurances, we
# can determine if the records are more than 70% consistent across the
# occurances of the entities.
col_df = col_df.merge(consist_df, how="outer")
# change all of the fully consistent records to True
col_df[f"{col}_consistent_rate"] = (
col_df["record_occurences"] / col_df["entity_occurences"]
)
col_df[f"{col}_is_consistent"] = col_df[f"{col}_consistent_rate"] > strictness
col_df = col_df.sort_values(f"{col}_consistent_rate")
return col_df
[docs]
def _lat_long(
dirty_df: pd.DataFrame,
clean_df: pd.DataFrame,
entity_id_df: pd.DataFrame,
entity_idx: list[str],
col: str,
cols_to_consit: list[str],
round_to: int = 2,
) -> pd.DataFrame:
"""Harvests more complete lat/long in special cases.
For all of the entities were there is not a consistent enough reported record for
latitude and longitude, this function reduces the precision of the reported lat/long
by rounding down the reported records in order to get more complete set of
consistent records.
Args:
dirty_df: dataframe with entity records having inconsistently reported lat/long.
clean_df: dataframe with entity records having consistently reported lat/long.
entity_id_df: a dataframe with a complete set of possible entity ids.
entity_idx: a list of the id(s) for the entity. Ex: for a plant entity, the
entity_idx is ['plant_id_eia']. For a generator entity, the entity_idx is
['plant_id_eia', 'generator_id'].
col: the column name of the column we are trying to harvest.
cols_to_consit: a list of the columns to determine consistency. This either the
[entity_id] or the [entity_id, 'report_date'], depending on whether the
entity is static or annual.
round_to: The number of decimal places we want to preserve while rounding down.
Returns:
DataFrame with all of the entity ids. Some will have harvested records from the
clean_df. Some will have harvested records that were found after rounding. Some
will have NaNs if no consistently reported records were found.
"""
# grab the dirty plant records, round and get a new consistency
ll_df = dirty_df.round(decimals={col: round_to})
logger.debug(f"Dirty {col} records: {len(ll_df)}")
ll_df["table"] = "special_case"
ll_df = occurrence_consistency(entity_idx, ll_df, col, cols_to_consit)
# grab the clean plants
ll_clean_df = clean_df.dropna()
# find the new clean plant records by selecting the True consistent records
ll_df = ll_df[ll_df[f"{col}_is_consistent"]].drop_duplicates(subset=entity_idx)
logger.debug(f"Clean {col} records: {len(ll_df)}")
# add the newly cleaned records
ll_clean_df = pd.concat([ll_clean_df, ll_df])
# merge onto the plants df w/ all plant ids
ll_clean_df = entity_id_df.merge(ll_clean_df, how="outer")
return ll_clean_df
[docs]
def _round_operating_date(
dirty_df: pd.DataFrame,
clean_df: pd.DataFrame,
entity_id_df: pd.DataFrame,
entity_idx: list[str],
col: str,
cols_to_consit: list[str],
group_by_freq: Literal["M", "Y"],
) -> pd.DataFrame:
"""Harvests operating dates by combining dates within the selected group_by_freq.
For all of the entities where there is not a consistent enough reported
operating date, this function reduces the precision of the reported operating date
by only keeping the last record when records are within the time bandwidth
(default of one year) of one another.
Args:
dirty_df: a dataframe with entity records that have inconsistently reported
operating dates.
clean_df: a dataframe with entity records that have consistently reported
operating dates.
entity_id_df: a dataframe with a complete set of possible entity ids
entity_idx: a list of the id(s) for the entity. Ex: for a plant entity, the
entity_idx is ['plant_id_eia']. For a generator entity, the entity_idx is
['plant_id_eia', 'generator_id'].
col: the column name of the column we are trying to harvest.
cols_to_consit: a list of the columns to determine consistency. This either the
[entity_id] or the [entity_id, 'report_date'], depending on whether the
entity is static or annual.
group_by_freq: Frequency to combine by ("M" for month, or "Y" for year)
Returns:
A dataframe with all of the entity ids. Some will have harvested records from
the clean_df. Some will have NA values if no consistently reported records were
found.
"""
# grab the dirty plant records, round and get a new consistency
op_df = dirty_df.assign(
operating_rounded=dirty_df[col].dt.to_period(group_by_freq).dt.to_timestamp()
)
logger.debug(f"Dirty {col} records: {len(op_df)}")
# Group all records within the same rounded time period and assign them the max
# value within that time period.
op_df[col] = op_df.groupby(cols_to_consit + ["operating_rounded"])[col].transform(
"max"
)
op_df["table"] = "special_case"
op_df = op_df.drop("operating_rounded", axis=1)
op_df = occurrence_consistency(entity_idx, op_df, col, cols_to_consit)
# grab the clean plants
op_clean_df = clean_df.dropna()
# find the new clean plant records by selecting the True consistent records
op_df = op_df[op_df[f"{col}_is_consistent"]].drop_duplicates(subset=entity_idx)
logger.info(f"Clean {col} records: {len(op_df)}")
logger.info(
f"Rescued rounded {col} for the following units ({entity_idx}): "
f"{sorted(op_df[entity_idx].apply(lambda row: '_'.join(row.to_numpy().astype(str)), axis=1))}"
)
# add the newly cleaned records
op_clean_df = pd.concat([op_clean_df, op_df])
# merge onto the plants df w/ all plant ids
op_clean_df = entity_id_df.merge(op_clean_df, how="outer")
return op_clean_df
[docs]
def _add_timezone(plants_entity: pd.DataFrame) -> pd.DataFrame:
"""Add plant IANA timezone based on lat/lon or state if lat/lon is unavailable.
Args:
plants_entity: Plant entity table, including columns named "latitude",
"longitude", and optionally "state"
Returns:
A DataFrame containing the same table, with a "timezone" column added.
Timezone may be missing if lat / lon is missing or invalid.
"""
plants_entity["timezone"] = plants_entity.apply(
lambda row: find_timezone(
lng=row["longitude"], lat=row["latitude"], state=row["state"], strict=False
),
axis=1,
)
return plants_entity
[docs]
def _add_additional_epacems_plants(plants_entity: pd.DataFrame) -> pd.DataFrame:
"""Adds the info for plants that have IDs in the CEMS data but not EIA data.
The columns loaded are plant_id_eia, plant_name, state, latitude, and
longitude. Note that a side effect will be resetting the index on
plants_entity, if onecexists. If that's a problem, modify the code below.
Note that some of these plants disappear from the CEMS before the
earliest EIA data PUDL processes, so if PUDL eventually ingests older
data, these may be redundant.
The set of additional plants is every plant that appears in the hourly CEMS
data (1995-2017) that never appears in the EIA 923 or 860 data (2009-2017
for EIA 923, 2011-2017 for EIA 860).
Args:
plants_entity: The plant entity table to which we will append additional plants.
Returns:
The same plants_entity table, with the addition of some missing EPA CEMS plants.
"""
# Add the plant IDs that are missing and update the values for the others
# The data we're reading is a CSV in pudl/metadata/
# SQL would call this whole process an upsert
# See also: https://github.com/pandas-dev/pandas/issues/22812
cems_df = pd.read_csv(
importlib.resources.files("pudl.package_data.epacems")
/ "additional_epacems_plants.csv",
index_col=["plant_id_eia"],
usecols=["plant_id_eia", "plant_name_eia", "state", "latitude", "longitude"],
)
plants_entity = plants_entity.set_index("plant_id_eia")
cems_unmatched = cems_df.loc[~cems_df.index.isin(plants_entity.index)]
# update will replace columns and index values that add rows or affect
# non-matching columns. It also requires an index, so we set and reset the
# index as necessary. Also, it only works in-place, so we can't chain.
plants_entity.update(cems_df, overwrite=True)
return pd.concat([plants_entity, cems_unmatched]).reset_index()
[docs]
def _compile_all_entity_records(
entity: EiaEntity, clean_dfs: dict[str, pd.DataFrame]
) -> pd.DataFrame:
"""Compile all of the entity records from each table they appear in.
Comb through each of the dataframes in clean_dfs to pull out every instance of the
entity id.
"""
# we know these columns must be in the dfs
id_cols = ENTITIES[entity.value]["id_cols"]
static_cols = ENTITIES[entity.value]["static_cols"]
annual_cols = ENTITIES[entity.value]["annual_cols"]
# A dictionary of columns representing additional data to be harvested,
# whose names should map to an ID, static, or annual column name.
mapped_schemas = ENTITIES[entity.value].get("mapped_schemas")
base_cols = id_cols + ["report_date"]
# empty list for dfs to be added to for each table below
dfs = []
# for each df in the dict of transformed dfs
for table_name, transformed_df in clean_dfs.items():
# inside of main() we are going to be adding items into
# clean_dfs with the name 'annual'. We don't want to harvest
# from our newly harvested tables.
# if the df contains the desired columns the grab those columns
if set(base_cols).issubset(transformed_df.columns):
logger.debug(f" {table_name}...")
# create a copy of the df to muck with
df = transformed_df.copy()
# we know these columns must be in the dfs
cols = []
# check whether the columns are in the specific table
for column in static_cols + annual_cols:
if column in df.columns:
cols.append(column)
df = df[(base_cols + cols)]
df = df.dropna(subset=id_cols)
# add a column with the table name so we know its origin
df["table"] = table_name
dfs.append(df)
# check if there are columns that should be renamed and harvested
# as an additional table
# for map_col_dict in mapped_schemas: iterate through map_cols_dict
if mapped_schemas:
for i, map_cols_dict in enumerate(mapped_schemas):
base_cols_to_add = set(base_cols) - set(map_cols_dict.values())
if base_cols_to_add.union(set(map_cols_dict.keys())).issubset(
transformed_df.columns
):
mapped_df = transformed_df[
list(base_cols_to_add) + list(map_cols_dict.keys())
]
mapped_df = mapped_df.rename(columns=map_cols_dict)
mapped_df = mapped_df.dropna(subset=id_cols)
mapped_df["table"] = table_name + f"_mapped_{i}"
dfs.append(mapped_df)
# add those records to the compilation
compiled_df = pd.concat(dfs, axis=0, ignore_index=True, sort=True)
# strip the month and day from the date so we can have annual records
compiled_df["report_date"] = compiled_df["report_date"].dt.year
# convert the year back into a date_time object
year = compiled_df["report_date"]
compiled_df["report_date"] = pd.to_datetime({"year": year, "month": 1, "day": 1})
logger.debug(" Casting harvested IDs to correct data types")
# most columns become objects (ack!), so assign types
compiled_df = apply_pudl_dtypes(compiled_df, group="eia")
return compiled_df
[docs]
def _manage_strictness(col: str, eia860m: bool) -> float:
"""Manage the strictness level for each column.
Args:
col: name of column
eia860m: if True, ETL is attempting to include year-to-date EIA 860M data.
"""
strictness_default = 0.7
# the longitude column is very different in the ytd 860M data (it appears
# to have an additional decimal point) bc it shows up in the generator
# table but it is a plant level data point, it mucks up the consistency
strictness_cols = {
"plant_name_eia": 0,
"utility_name_eia": 0,
"longitude": 0 if eia860m else 0.7,
"prime_mover_code": 0,
}
return strictness_cols.get(col, strictness_default)
[docs]
def harvest_entity_tables( # noqa: C901
entity: EiaEntity,
clean_dfs: dict[str, pd.DataFrame],
eia_settings: EiaSettings,
debug: bool = False,
) -> tuple:
"""Compile consistent records for various entities.
For each entity (plants, generators, boilers, utilties), this function
finds all the harvestable columns from any table that they show up
in. It then determines how consistent the records are and keeps the values
that are mostly consistent. It compiles those consistent records into
one normalized table.
There are a few things to note here. First being that we are not expecting
the outcome here to be perfect! We choose to pull the most consistent
record as reported across all the EIA tables and years, but we also
required a "strictness" level of 70% (this is currently a hard coded
argument for :func:`occurrence_consistency`). That means at least 70% of the
records must be the same for us to use that value. So if values for an
entity haven't been reported 70% consistently, then it will show up as a
null value. We built in the ability to add special cases for columns where
we want to apply a different method to, but the only ones we added was for
latitude and longitude because they are by far the dirtiest.
We have determined which columns should be considered "static" or "annual".
These can be found in constants in the `entities` dictionary. Static means
That is should not change over time. Annual means there is annual
variablity. This distinction was made in part by testing the consistency
and in part by an understanding of how the entities and columns relate in
the real world.
Args:
entity: One of: plants, generators, boilers, or utilties
clean_dfs: A dictionary of table names (keys) and clean dfs (values).
eia860m: if True, the etl run is attempting to include year-to-date updated from
EIA 860M.
debug: if True, log when columns are inconsistent, but don't raise an error.
Returns:
entity_df (the harvested entity table), annual_df (the annual entity table),
col_dfs (a dictionary of dataframes, one per harvested column, with information)
about their consistency and the values which were harvested)
Raises:
AssertionError: If the consistency of any record value is <90% (when
debug=False)
Todo:
* Return to role of debug.
* Determine what to do with null records
* Determine how to treat mostly static records
"""
# Do some final cleanup and assign appropriate types:
clean_dfs = {
name: convert_cols_dtypes(df, data_source="eia")
for name, df in clean_dfs.items()
}
if entity == EiaEntity.UTILITIES:
# Remove location columns that are associated with plants, not utilities:
for table, df in clean_dfs.items():
if "plant_id_eia" in df.columns:
plant_location_cols = [
"street_address",
"city",
"state",
"zip_code",
]
logger.info(f"Removing {plant_location_cols} from {table} table.")
clean_dfs[table] = df.drop(columns=plant_location_cols, errors="ignore")
# we know these columns must be in the dfs
id_cols = ENTITIES[entity.value]["id_cols"]
static_cols = ENTITIES[entity.value]["static_cols"]
annual_cols = ENTITIES[entity.value]["annual_cols"]
logger.debug(" compiling plants for entity tables from:")
compiled_df = _compile_all_entity_records(entity, clean_dfs)
# compile annual ids
annual_id_df = compiled_df[["report_date"] + id_cols].copy().drop_duplicates()
annual_id_df = annual_id_df.sort_values(["report_date"] + id_cols, ascending=False)
# create the annual and entity dfs
entity_id_df = annual_id_df.drop(["report_date"], axis=1).drop_duplicates(
subset=id_cols
)
entity_df = entity_id_df.copy()
annual_df = annual_id_df.copy()
special_case_cols = {
"latitude": [_lat_long, 1],
"longitude": [_lat_long, 1],
"generator_operating_date": [_round_operating_date, "Y"],
}
consistency = pd.DataFrame(
columns=["column", "consistent_ratio", "wrongos", "total"]
)
col_dfs = {}
# determine how many times each of the columns occur
for col in static_cols + annual_cols:
if col in annual_cols:
cols_to_consit = id_cols + ["report_date"]
if col in static_cols:
cols_to_consit = id_cols
strictness = _manage_strictness(col, eia_settings.eia860.eia860m)
col_df = occurrence_consistency(
id_cols, compiled_df, col, cols_to_consit, strictness=strictness
)
# pull the correct values out of the df and merge w/ the plant ids
col_correct_df = col_df[col_df[f"{col}_is_consistent"]].drop_duplicates(
subset=(cols_to_consit + [f"{col}_is_consistent"])
)
# we need this to be an empty df w/ columns bc we are going to use it
if col_correct_df.empty:
col_correct_df = pd.DataFrame(columns=col_df.columns)
if col in static_cols:
clean_df = entity_id_df.merge(col_correct_df, on=id_cols, how="left")
clean_df = clean_df[id_cols + [col]]
entity_df = entity_df.merge(clean_df, on=id_cols)
if col in annual_cols:
clean_df = annual_id_df.merge(
col_correct_df, on=(id_cols + ["report_date"]), how="left"
)
clean_df = clean_df[id_cols + ["report_date", col]]
annual_df = annual_df.merge(clean_df, on=(id_cols + ["report_date"]))
# get the still dirty records by using the cleaned ids w/null values
# we need the plants that have no 'correct' value so
# we can't just use the col_df records when the consistency is not True
dirty_df = col_df.merge(clean_df[clean_df[col].isnull()][id_cols])
if col in special_case_cols:
clean_df = special_case_cols[col][0](
dirty_df,
clean_df,
entity_id_df,
id_cols,
col,
cols_to_consit,
special_case_cols[col][1],
)
if col in static_cols:
clean_df = clean_df[id_cols + [col]]
entity_df = entity_df.drop(columns=[col]).merge(clean_df, on=id_cols)
elif col in annual_cols:
raise AssertionError(
"Method currenty not configured to work with annual values."
)
if debug:
col_dfs[col] = col_df
# this next section is used to print and test whether the harvested
# records are consistent enough
total = len(col_df.drop_duplicates(subset=cols_to_consit))
# if the total is 0, the ratio will error, so assign null values.
if total == 0:
ratio = np.nan
wrongos = np.nan
logger.debug(f" Zero records found for {col}")
if total > 0:
ratio = (
len(
col_df[(col_df[f"{col}_is_consistent"])].drop_duplicates(
subset=cols_to_consit
)
)
/ total
)
wrongos = (1 - ratio) * total
logger.debug(
f" Ratio: {ratio:.3} "
f"Wrongos: {wrongos:.5} "
f"Total: {total} {col}"
)
if ratio < 0.9:
if debug:
logger.error(f"{col} has low consistency: {ratio:.3}.")
else:
raise AssertionError(
f"Harvesting of {col} is too inconsistent at {ratio:.3}."
)
# add to a small df to be used in order to print out the ratio of
# consistent records
consistency = pd.concat(
[
consistency,
pd.DataFrame(
{
"column": [col],
"consistent_ratio": [ratio],
"wrongos": [wrongos],
"total": [total],
}
),
],
ignore_index=True,
)
mcs = consistency["consistent_ratio"].mean()
logger.info(f"Average consistency of static {entity.value} values is {mcs:.2%}")
if entity == EiaEntity.PLANTS:
# Post-processing specific to the plants entity tables
entity_df = _add_additional_epacems_plants(entity_df).pipe(_add_timezone)
annual_df = fillna_balancing_authority_codes_via_names(annual_df).pipe(
fix_balancing_authority_codes_with_state, plants_entity=entity_df
)
return entity_df, annual_df, col_dfs
@asset(
ins={
table_name: AssetIn()
for table_name in [
"_core_eia860__boiler_generator_assn",
"_core_eia923__generation",
"_core_eia860__generators",
"_core_eia923__boiler_fuel",
]
},
config_schema={
"debug": Field(
bool,
default_value=False,
description=(
"If True, debugging columns will be added to boiler_generator_assn."
),
),
},
required_resource_keys={"dataset_settings"},
io_manager_key="pudl_io_manager",
)
[docs]
def core_eia860__assn_boiler_generator(context, **clean_dfs) -> pd.DataFrame:
"""Creates a set of more complete boiler generator associations.
Creates a unique ``unit_id_pudl`` for each collection of boilers and generators
within a plant that have ever been associated with each other, based on the boiler
generator associations reported in EIA860. Unfortunately, this information is not
complete for years before 2014, as the gas turbine portion of combined cycle power
plants in those earlier years were not reporting their fuel consumption, or
existence as part of the plants.
For years 2014 and on, EIA860 contains a ``unit_id_eia`` value, allowing the
combined cycle plant compoents to be associated with each other. For many plants not
listed in the reported boiler generator associations, it is nonetheless possible to
associate boilers and generators on a one-to-one basis, as they use identical
strings to describe the units.
In the end, between the reported BGA table, the string matching, and the
``unit_id_eia`` values, it's possible to create a nearly complete mapping of the
generation units, at least for 2014 and later.
Args:
clean_dfs: a dictionary of clean EIA dataframes that have passed through the
early transform steps.
Returns:
A dataframe containing the boiler generator associations.
Raises:
AssertionError: If the boiler - generator association graphs are not bi-partite,
meaning generators only connect to boilers, and boilers only connect to
generators.
AssertionError: If all boilers do not end up with the same unit_id each year.
AssertionError: If all generators do not end up with the same unit_id each year.
"""
debug = context.op_config["debug"]
eia_settings = context.resources.dataset_settings.eia
# Do some final data formatting and assign appropriate types:
clean_dfs = {
table_name: (
convert_cols_dtypes(df, data_source="eia")
.pipe(_restrict_years, eia_settings)
.pipe(PUDL_PACKAGE.encode)
)
for table_name, df in clean_dfs.items()
}
# compile and scrub all the parts
logger.info("Inferring complete EIA boiler-generator associations.")
logger.debug(f"{clean_dfs.keys()=}")
# grab the core_eia923__monthly_generation table, group annually, generate a new tag
gen_eia923 = clean_dfs["_core_eia923__generation"]
gen_eia923 = (
gen_eia923.set_index(pd.DatetimeIndex(gen_eia923.report_date))
.groupby([pd.Grouper(freq="YS"), "plant_id_eia", "generator_id"])
.net_generation_mwh.sum()
.reset_index()
.assign(missing_from_923=False)
)
# compile all of the generators
gens = pd.merge(
gen_eia923,
clean_dfs["_core_eia860__generators"],
on=["plant_id_eia", "report_date", "generator_id"],
how="outer",
)
gens = gens[
[
"plant_id_eia",
"report_date",
"generator_id",
"unit_id_eia",
"net_generation_mwh",
"missing_from_923",
]
].drop_duplicates()
# create the beginning of a bga compilation w/ the generators as the
# background
bga_compiled_1 = pd.merge(
gens,
clean_dfs["_core_eia860__boiler_generator_assn"],
on=["plant_id_eia", "generator_id", "report_date"],
how="outer",
)
# Create a set of bga's that are linked, directly from bga8
bga_assn = bga_compiled_1[bga_compiled_1["boiler_id"].notnull()].copy()
bga_assn.loc[:, "bga_source"] = "eia860_org"
# Create a set of bga's that were not linked directly through bga8
bga_unassn = bga_compiled_1[bga_compiled_1["boiler_id"].isnull()].copy()
bga_unassn = bga_unassn.drop(["boiler_id"], axis=1)
# Side note: there are only 6 generators that appear in bga8 that don't
# apear in gens9 or gens8 (must uncomment-out the og_tag creation above)
# bga_compiled_1[bga_compiled_1['og_tag'].isnull()]
bf_eia923 = clean_dfs["_core_eia923__boiler_fuel"].assign(
total_heat_content_mmbtu=lambda x: x.fuel_consumed_units * x.fuel_mmbtu_per_unit
)
bf_eia923 = (
bf_eia923.set_index(pd.DatetimeIndex(bf_eia923.report_date))
.groupby([pd.Grouper(freq="YS"), "plant_id_eia", "boiler_id"])
.agg({"total_heat_content_mmbtu": pudl.helpers.sum_na})
.reset_index()
.drop_duplicates(subset=["plant_id_eia", "report_date", "boiler_id"])
)
# Create a list of boilers that were not in bga8
bf9_bga = bf_eia923.merge(
bga_compiled_1,
on=["plant_id_eia", "boiler_id", "report_date"],
how="outer",
indicator=True,
)
bf9_not_in_bga = bf9_bga[bf9_bga["_merge"] == "left_only"]
bf9_not_in_bga = bf9_not_in_bga.drop(["_merge"], axis=1)
# Match the unassociated generators with unassociated boilers
# This method is assuming that some the strings of the generators and the
# boilers are the same
bga_unassn = bga_unassn.merge(
bf9_not_in_bga[["plant_id_eia", "boiler_id", "report_date"]],
how="left",
left_on=["report_date", "plant_id_eia", "generator_id"],
right_on=["report_date", "plant_id_eia", "boiler_id"],
)
bga_unassn["bga_source"] = np.where(
bga_unassn.boiler_id.notnull(), "string_assn", pd.NA
)
bga_compiled_2 = pd.concat([bga_assn, bga_unassn]).fillna(
{"missing_from_923": True}
)
# Connect the gens and boilers in units
bga_compiled_units = bga_compiled_2.loc[bga_compiled_2["unit_id_eia"].notnull()]
bga_gen_units = bga_compiled_units.drop(["boiler_id"], axis=1)
bga_boil_units = bga_compiled_units[
["plant_id_eia", "report_date", "boiler_id", "unit_id_eia"]
].copy()
bga_boil_units = bga_boil_units.dropna(subset=["boiler_id"])
# merge the units with the boilers
bga_unit_compilation = bga_gen_units.merge(
bga_boil_units,
how="outer",
on=["plant_id_eia", "report_date", "unit_id_eia"],
indicator=True,
)
# label the bga_source
bga_unit_compilation.loc[
bga_unit_compilation["bga_source"].isnull(), "bga_source"
] = "unit_connection"
bga_unit_compilation = bga_unit_compilation.drop(["_merge"], axis=1)
bga_non_units = bga_compiled_2[bga_compiled_2["unit_id_eia"].isnull()]
# combine the unit compilation and the non units
bga_compiled_3 = pd.concat([bga_non_units, bga_unit_compilation])
bga_compiled_3 = bga_compiled_3[
[
"plant_id_eia",
"report_date",
"generator_id",
"boiler_id",
"unit_id_eia",
"bga_source",
"boiler_generator_assn_type_code",
"steam_plant_type_code",
"net_generation_mwh",
"missing_from_923",
"data_maturity",
]
]
# label plants that have 'bad' generator records (generators that have MWhs
# in gens9 but don't have connected boilers) create a df with just the bad
# plants by searching for the 'bad' generators
bad_plants = bga_compiled_3[
(bga_compiled_3["boiler_id"].isnull())
& (bga_compiled_3["net_generation_mwh"] > 0)
].drop_duplicates(subset=["plant_id_eia", "report_date"])
bad_plants = bad_plants[["plant_id_eia", "report_date"]]
# merge the 'bad' plants back into the larger frame
bga_compiled_3 = bga_compiled_3.merge(
bad_plants, how="outer", on=["plant_id_eia", "report_date"], indicator=True
)
# use the indicator to create labels
bga_compiled_3["plant_w_bad_generator"] = np.where(
bga_compiled_3["_merge"] == "both", True, False
)
# Note: At least one gen has reported MWh in 923, but could not be
# programmatically mapped to a boiler
# we don't need this one anymore
bga_compiled_3 = bga_compiled_3.drop(["_merge"], axis=1)
# create a label for generators that are unmapped but in 923
bga_compiled_3["unmapped_but_in_923"] = np.where(
(bga_compiled_3.boiler_id.isnull())
& ~bga_compiled_3.missing_from_923
& (bga_compiled_3.net_generation_mwh == 0),
True,
False,
)
# create a label for generators that are unmapped
bga_compiled_3["unmapped"] = np.where(
bga_compiled_3.boiler_id.isnull(), True, False
)
bga_out = bga_compiled_3.drop("net_generation_mwh", axis=1)
bga_out.loc[bga_out.unit_id_eia.isnull(), "unit_id_eia"] = pd.NA
bga_for_nx = bga_out[
["plant_id_eia", "report_date", "generator_id", "boiler_id", "unit_id_eia"]
]
# If there's no boiler... there's no boiler-generator association
bga_for_nx = bga_for_nx.dropna(subset=["boiler_id"]).drop_duplicates()
# Need boiler & generator specific ID strings, or they look like
# the same node to NX
bga_for_nx["generators"] = (
"p"
+ bga_for_nx.plant_id_eia.astype(int).astype(str)
+ "_g"
+ bga_for_nx.generator_id.astype(pd.StringDtype())
)
bga_for_nx["boilers"] = (
"p"
+ bga_for_nx.plant_id_eia.astype(int).astype(str)
+ "_b"
+ bga_for_nx.boiler_id.astype(pd.StringDtype())
)
# dataframe to accumulate the unit_ids in
bga_w_units = pd.DataFrame()
# We want to start our unit_id counter anew for each plant:
for pid in bga_for_nx.plant_id_eia.unique():
bga_byplant = bga_for_nx[bga_for_nx.plant_id_eia == pid].sort_values(
["generators", "boilers"]
)
# Create a graph from the dataframe of boilers and generators. It's a
# multi-graph, meaning the same nodes can be connected by more than one
# edge -- this allows us to preserve multiple years worth of boiler
# generator association information for later inspection if need be:
bga_graph = nx.from_pandas_edgelist(
bga_byplant,
source="generators",
target="boilers",
edge_attr=True,
create_using=nx.MultiGraph(),
)
# Each connected sub-graph is a generation unit:
gen_units = [
bga_graph.subgraph(c).copy() for c in nx.connected_components(bga_graph)
]
# Assign a unit_id to each subgraph, and extract edges into a dataframe
for unit_id, unit in zip(range(len(gen_units)), gen_units, strict=True):
# All the boiler-generator association graphs should be bi-partite,
# meaning generators only connect to boilers, and boilers only
# connect to generators.
if not nx.algorithms.bipartite.is_bipartite(unit):
raise AssertionError(
f"Non-bipartite generation unit graph found."
f"plant_id_eia={pid}, unit_id_pudl={unit_id}."
)
nx.set_edge_attributes(unit, name="unit_id_pudl", values=unit_id + 1)
new_unit_df = nx.to_pandas_edgelist(unit)
bga_w_units = pd.concat([bga_w_units, new_unit_df])
bga_w_units = bga_w_units.drop(["source", "target"], axis=1)
# Check whether the PUDL unit_id values we've inferred conflict with
# the unit_id_eia values that were reported to EIA. Are there any PUDL
# unit_id values that have more than 1 EIA unit_id_eia within them?
bga_unit_id_eia_counts = (
bga_w_units.groupby(["plant_id_eia", "unit_id_pudl"])["unit_id_eia"]
.nunique()
.to_frame()
.reset_index()
)
bga_unit_id_eia_counts = bga_unit_id_eia_counts.rename(
columns={"unit_id_eia": "unit_id_eia_count"}
)
bga_unit_id_eia_counts = pd.merge(
bga_w_units, bga_unit_id_eia_counts, on=["plant_id_eia", "unit_id_pudl"]
)
too_many_codes = bga_unit_id_eia_counts[
bga_unit_id_eia_counts.unit_id_eia_count > 1
]
too_many_codes = (
too_many_codes[too_many_codes.unit_id_eia.notna()]
.groupby(["plant_id_eia", "unit_id_pudl"])["unit_id_eia"]
.unique()
)
for row in too_many_codes.items():
logger.warning(
f"Multiple EIA unit codes:"
f"plant_id_eia={row[0][0]}, "
f"unit_id_pudl={row[0][1]}, "
f"unit_id_eia={row[1]}"
)
bga_w_units = bga_w_units.drop("unit_id_eia", axis=1)
# These assertions test that all boilers and generators ended up in the
# same unit_id across all the years of reporting:
pgu_gb = bga_w_units.groupby(["plant_id_eia", "generator_id"])["unit_id_pudl"]
if pgu_gb.nunique().gt(1).any():
logger.error("Inconsistent inter-annual plant-generator-units!")
pbu_gb = bga_w_units.groupby(["plant_id_eia", "boiler_id"])["unit_id_pudl"]
if pbu_gb.nunique().gt(1).any():
logger.error("Inconsistent inter-annual plant-boiler-units!")
bga_w_units = (
bga_w_units.drop("report_date", axis=1)
.loc[:, ["plant_id_eia", "unit_id_pudl", "generator_id", "boiler_id"]]
.drop_duplicates()
)
bga_out = (
pd.merge(
left=bga_out,
right=bga_w_units,
how="left",
on=["plant_id_eia", "generator_id", "boiler_id"],
)
.astype({"unit_id_pudl": pd.Int64Dtype()})
.pipe(apply_pudl_dtypes, group="eia")
)
# If we're NOT debugging, drop additional forensic information and bad BGAs
if not debug:
bga_out = (
bga_out[
~bga_out.missing_from_923
& ~bga_out.plant_w_bad_generator
& ~bga_out.unmapped_but_in_923
& ~bga_out.unmapped
]
.drop(
[
"missing_from_923",
"plant_w_bad_generator",
"unmapped_but_in_923",
"unmapped",
],
axis=1,
)
.drop_duplicates(
subset=["report_date", "plant_id_eia", "boiler_id", "generator_id"]
)
)
return bga_out
[docs]
def _restrict_years(
df: pd.DataFrame,
eia_settings: EiaSettings | None = None,
) -> pd.DataFrame:
"""Restricts eia years for boiler generator association."""
if eia_settings is None:
eia_settings = EiaSettings()
bga_years = set(eia_settings.eia860.years) & set(eia_settings.eia923.years)
df = df[df.report_date.dt.year.isin(bga_years)]
return df
[docs]
def map_balancing_authority_names_to_codes(df: pd.DataFrame) -> pd.DataFrame:
"""Build a map of the BA names to their most frequently associated BA codes.
We know there are some inconsistent pairings of codes and names so we grab the most
consistently reported combo, making the assumption that the most consistent pairing
is most likely to be the correct.
Args:
df: a data table with columns ``balancing_authority_code_eia`` and
``balancing_authority_name_eia``
Returns:
a table with a unique index of ``balancing_authority_name_eia`` and a column of
``balancing_authority_code``.
"""
return (
# count the unquie combos of BA code and name's.
df.assign(count=1)
.groupby(
by=["balancing_authority_name_eia", "balancing_authority_code_eia"],
observed=True,
)[["count"]]
.count()
.reset_index()
# then sort so the most common is at the top.
.sort_values(by=["count"], ascending=False)
# then drop duplicates on the BA name
.drop_duplicates(["balancing_authority_name_eia"])
.set_index("balancing_authority_name_eia")
.drop(columns=["count"])
)
[docs]
def fillna_balancing_authority_codes_via_names(df: pd.DataFrame) -> pd.DataFrame:
"""Fill null balancing authority (BA) codes via a map of the BA names to codes.
There are a handful of missing ``balancing_authority_code_eia``'s that are easy to
map given the balancing_authority_name_eia. This function fills in null BA codes
using the BA names. The map ofo the BA names to codes is generated via
:func:`map_balancing_authority_names_to_codes`.
Args:
df: a data table with columns ``balancing_authority_code_eia`` and
``balancing_authority_name_eia``
"""
pre_len = len(df[df.balancing_authority_code_eia.notnull()])
# Identify the most common mapping from a BA name to a BA code:
ba_name_to_code_map = map_balancing_authority_names_to_codes(df)
null_ba_code_mask = (
df.balancing_authority_code_eia.isnull()
& df.balancing_authority_name_eia.notnull()
& df.balancing_authority_name_eia.isin(ba_name_to_code_map.index)
)
# For each row with a missing BA code, identify the likely code based on its
# associated BA name. Here the argument to map() is a Series containing
# balancing_authority_code that's indexed by balancing_authority_name.
ba_codes = df.loc[null_ba_code_mask, "balancing_authority_name_eia"].map(
ba_name_to_code_map.balancing_authority_code_eia
)
# Fill in the null BA codes
df.loc[null_ba_code_mask, "balancing_authority_code_eia"] = ba_codes
post_len = len(df[df.balancing_authority_code_eia.notnull()])
logger.info(f"filled {post_len - pre_len} balancing authority codes using names.")
return df
[docs]
def fix_balancing_authority_codes_with_state(
plants: pd.DataFrame, plants_entity: pd.DataFrame
) -> pd.DataFrame:
"""Fix selective balancing_authority_code_eia's based on states.
There are some known errors in the ``balancing_authority_code_eia`` column that we
can identify and fix based on the state where the plant is located. Where we update
the ``balancing_authority_code_eia`` column, we also update the
``balancing_authority_name_eia`` column using the name generated by
:func:`map_balancing_authority_names_to_codes`.
This function should only be applied post-:func:`harvest_entity_tables`. The
``state`` column is a "static" entity column so the first step in this function is
merging the static and annually varying plants together. Then we fix known errors in
the BA codes:
* reported PACE, but state is OR or CA, code should be PACW
* reported PACW, but state is UT, code should be PACE
Args:
plants: annually harvested plant table with columns: ``plant_id_eia``,
``report_date`` and ``balancing_authority_code_eia``.
plants_entity: static harvested plant table with columns: ``plant_id_eia`` and
``state``.
Returns:
plants table that has the same set of columns and rows, with cleaned
``balancing_authority_code_eia`` column and an updated corresponding
``balancing_authority_name_eia`` column.
"""
# Identify the most common mapping from a BA name to a BA code:
ba_name_to_code_map = map_balancing_authority_names_to_codes(plants)
ba_name_to_code_map = ba_name_to_code_map.reset_index()
# Prior to 2013, there are no BA codes or names. Running a pre-2013 subset of data
# through the transform will thus return an empty ba_name_to_code_map.
if (
not ba_name_to_code_map.empty
and plants.balancing_authority_code_eia.isin(["PACW", "PACE"]).any()
):
logger.info("Spot fixing incorrect PACW/PACE BA codes and names.")
plants = plants.merge(
plants_entity[["plant_id_eia", "state"]], # only merge in state, drop later
on=["plant_id_eia"],
how="left",
validate="m:1",
)
BACodeFix = namedtuple(
"BACodeFix", ["ba_code_found", "ba_code_fix", "ba_name_fix", "states"]
)
fixes = [
BACodeFix(
"PACE",
"PACW",
ba_name_to_code_map.loc[
ba_name_to_code_map.balancing_authority_code_eia == "PACW",
"balancing_authority_name_eia",
].tolist()[0],
["OR", "CA"],
),
BACodeFix(
"PACW",
"PACE",
ba_name_to_code_map.loc[
ba_name_to_code_map.balancing_authority_code_eia == "PACE",
"balancing_authority_name_eia",
].tolist()[0],
["UT"],
),
]
for fix in fixes:
plants.loc[
(plants.balancing_authority_code_eia == fix.ba_code_found)
& (plants.state.isin(fix.states)),
["balancing_authority_code_eia", "balancing_authority_name_eia"],
] = [fix.ba_code_fix, fix.ba_name_fix]
plants = plants.drop(columns=["state"])
return plants
[docs]
def harvested_entity_asset_factory(
entity: EiaEntity, io_manager_key: str | None = None
) -> AssetsDefinition:
"""Create an asset definition for the harvested entity tables."""
harvestable_assets = (
"_core_eia860__boiler_cooling",
"_core_eia860__boiler_emissions_control_equipment_assn",
"_core_eia923__boiler_fuel",
"_core_eia860__boiler_generator_assn",
"_core_eia860__boiler_stack_flue",
"_core_eia860__boilers",
"_core_eia923__coalmine",
"_core_eia860__emissions_control_equipment",
"_core_eia923__energy_storage",
"_core_eia923__fuel_receipts_costs",
"_core_eia923__generation",
"_core_eia923__generation_fuel",
"_core_eia923__generation_fuel_nuclear",
"_core_eia860__generators",
"_core_eia860__generators_energy_storage",
"_core_eia860__generators_wind",
"_core_eia860__generators_solar",
"_core_eia860__ownership",
"_core_eia860__plants",
"_core_eia860__utilities",
)
@multi_asset(
ins={table_name: AssetIn() for table_name in harvestable_assets},
outs={
f"core_eia__entity_{entity.value}": AssetOut(io_manager_key=io_manager_key),
f"core_eia860__scd_{entity.value}": AssetOut(io_manager_key=io_manager_key),
},
config_schema={
"debug": Field(
bool,
default_value=False,
description=(
"If True, allow inconsistent values in harvested columns and "
"produce additional debugging output."
),
),
},
required_resource_keys={"dataset_settings"},
name=f"harvested_{entity.value}_eia",
)
def harvested_entity(context, **clean_dfs):
"""Harvesting IDs & consistent static attributes for EIA entity."""
logger.info(f"Harvesting IDs & consistent static attributes for EIA {entity}")
eia_settings = context.resources.dataset_settings.eia
debug = context.op_config["debug"]
clean_dfs = {
df_name: PUDL_PACKAGE.encode(clean_dfs[df_name]) for df_name in clean_dfs
}
entity_df, annual_df, _col_dfs = harvest_entity_tables(
entity, clean_dfs, debug=debug, eia_settings=eia_settings
)
return (
Output(output_name=f"core_eia__entity_{entity.value}", value=entity_df),
Output(output_name=f"core_eia860__scd_{entity.value}", value=annual_df),
)
return harvested_entity
[docs]
harvested_entities = [
harvested_entity_asset_factory(entity, io_manager_key="pudl_io_manager")
for entity in EiaEntity
]
[docs]
def finished_eia_asset_factory(
table_name: str, _core_table_name: str, io_manager_key: str | None = None
) -> AssetsDefinition:
"""An asset factory for finished EIA tables.
Args:
table_name: the name of the harvest table.
_core_table_name: the name of the unharvested input table
io_manager_key: the name of the IO Manager of the final asset.
Returns:
A harvested EIA asset.
"""
@asset(
ins={_core_table_name: AssetIn()},
name=table_name,
io_manager_key=io_manager_key,
)
def finished_eia_asset(**kwargs) -> pd.DataFrame:
"""Enforce PUDL DB schema on a cleaned EIA dataframe."""
res = PUDL_PACKAGE.get_resource(table_name)
return (
PUDL_PACKAGE.encode(kwargs[_core_table_name])
.pipe(convert_cols_dtypes, data_source="eia")
.pipe(res.enforce_schema)
)
return finished_eia_asset
[docs]
finished_eia_assets = [
finished_eia_asset_factory(
table_name, _core_table_name, io_manager_key="pudl_io_manager"
)
for table_name, _core_table_name in {
"core_eia923__monthly_boiler_fuel": "_core_eia923__boiler_fuel",
"core_eia923__entity_coalmine": "_core_eia923__coalmine",
"core_eia923__monthly_energy_storage": "_core_eia923__energy_storage",
"core_eia923__monthly_fuel_receipts_costs": "_core_eia923__fuel_receipts_costs",
"core_eia923__monthly_generation": "_core_eia923__generation",
"core_eia923__monthly_generation_fuel": "_core_eia923__generation_fuel",
"core_eia923__monthly_generation_fuel_nuclear": "_core_eia923__generation_fuel_nuclear",
"core_eia860__scd_ownership": "_core_eia860__ownership",
"core_eia860__scd_emissions_control_equipment": "_core_eia860__emissions_control_equipment",
"core_eia860__assn_yearly_boiler_emissions_control_equipment": "_core_eia860__boiler_emissions_control_equipment_assn",
"core_eia860__assn_boiler_cooling": "_core_eia860__boiler_cooling",
"core_eia860__assn_boiler_stack_flue": "_core_eia860__boiler_stack_flue",
"core_eia860__scd_generators_wind": "_core_eia860__generators_wind",
"core_eia860__scd_generators_solar": "_core_eia860__generators_solar",
"core_eia860__scd_generators_energy_storage": "_core_eia860__generators_energy_storage",
}.items()
]