TariffKit

An offline electricity tariff engine for pricing, billing, and energy-system integrations. The first data provider supports PG&E residential rate plans under NEM 3.0 / the Net Billing Tariff (NBT); the package identity is deliberately not tied to one utility or tariff program.

Three schedules are vendored: E-ELEC (Electric Home), E-TOU-C (Time-of-Use, peak 4–9 p.m. every day), and EV2-A (Home Charging).

Under NBT your export credit is not a time-of-use schedule: it is an hourly Avoided Cost Calculator value that swings from about $0.06/kWh at midday to about $1.19/kWh on an August evening. Knowing what a kWh is worth right now, and what it will be worth over the next two days, is the input to every useful solar-and-battery dispatch decision.

from tariffkit import RateEngine

engine = RateEngine()
point = engine.price_now()

print(point.import_price.total)  # $/kWh to draw from the grid
print(point.export_price.total)  # $/kWh earned by exporting
print(point.spread)  # positive => exporting beats self-consuming

curve = engine.forecast(hours=48)
for hour in curve.best_export_hours(3):
    print(hour.start, hour.export_price.total)

Why it works offline

PG&E publishes 20 years of hourly export rates per vintage, as CPUC Resolution E-5301 requires (roughly 40 MB of CSV per vintage). But that file is a lossless expansion of a 576-cell matrix per year (12 months × 2 day types × 24 hours) per component. tariffkit collapses it at build time, verifying losslessness cell by cell, so the entire five-vintage dataset ships inside the wheel at 268 KiB and every lookup is a few list indexes.

The retail side is similar: these schedules' period boundaries are identical every day of the week including holidays and do not shift by season, so an import price is fully determined by (season, hour).

Nothing here touches the network at runtime.

Documentation

Works with

Prices are published in the shapes these already read, via either the custom component or the MQTT publisher — no template plumbing on your side. The two surfaces differ for EMHASS and Predbat: MQTT always publishes their attributes, while the custom component asks for a window on demand and keeps Predbat opt-in.

See docs/home-assistant.md and docs/mqtt.md for setup of each.

Install

pip install tariffkit              # core, zero dependencies
pip install 'tariffkit[mqtt]'      # + MQTT publisher with Home Assistant discovery
pip install 'tariffkit[web]'       # + FastAPI service
pip install 'tariffkit[secrets]'   # + OS keyring credential storage
pip install 'tariffkit[statements]' # + reading local PG&E statement PDFs
pip install 'tariffkit[all]'

The Home Assistant custom component is also available through HACS as a custom repository while its default-store submission is reviewed:

CLI

tariffkit now                          # current import/export price
tariffkit forecast --hours 48          # the upcoming curve
tariffkit forecast --format json       # machine-readable
tariffkit mqtt --broker 192.168.1.100  # publish hourly, with HA discovery
tariffkit serve                        # REST API on :8000
tariffkit bill intervals.csv           # compute a cycle from meter data
tariffkit info                         # which data is loaded, and from where
tariffkit account init home            # track a service agreement's history
tariffkit account source home show ha  # inspect profile grid-import/export entities

Configuration

Defaults target a PG&E-bundled residential customer. Point it at your own service agreement via ~/.config/tariffkit/config.toml:

supplier = "bundled"              # or "cca"
interconnection_year = 2026       # selects the NBT vintage and ACC Plus row
pto_date = "2026-06-03"           # starts the nine-year rate lock
acc_plus_segment = "residential"
base_services_charge_tier = 3

TARIFFKIT_* environment variables override any file setting. Long-lived PG&E, Home Assistant, InfluxDB, and MQTT credentials can be stored outside that file with tariffkit credentials set; see Configuration.

What the numbers include

• Import price is the marginal per-kWh cost: generation + distribution for the season and period, plus the flat riders. The AB 205 Base Services Charge is a fixed $/day amount and is deliberately excluded: folding it into a $/kWh figure would corrupt any marginal dispatch decision. Read it separately via engine.daily_fixed_charge().
• Export credit is the generation component plus the delivery component, plus your ACC Plus adder. Values past your nine-year lock are still returned but flagged locked=False; PG&E publishes them for illustration only.

Edge cases in PG&E's published data

Found by round-tripping the vendored matrices against all 1.75 million source rows. The library handles each; they are documented because they are surprising.

• The autumn DST hour. The fall-back day has 25 real hours but only 24 rate labels. PG&E gives the repeated 01:00 PST the HS2 label, so it is priced as 2am. Pricing it by wall-clock hour returns the wrong value for that hour.
• Far-future drift. From 2036 onward PG&E's own hour labels stop tracking Pacific daylight time, and NBT25/26/00 duplicate some holidays onto the following day. Each vintage records the last year verified exact; prices past it are returned with exact=False. Every year inside a nine-year rate lock is exact, and those years are already published as illustrative only.
• Vintages disagree about holidays in those same late years, so the holiday calendar used for a lookup is the one embedded in that vintage's own file.
• Only the June 2026 tariff sheets are vendored, for each of the three schedules. Earlier timestamps raise rather than silently back-dating current rates onto an older billing period.
• E-TOU-C's baseline credit is not in the marginal price. It applies to the first N kWh of a cycle, which is a quantity rather than a time, so price_at reports it as baseline_credit and the billing engine applies it.

CCA customers

If a Community Choice Aggregator supplies your generation, PG&E still delivers, and under NEM 3.0 you receive only the delivery component of the export credit from PG&E; generation compensation comes from the CCA.

An MCE rate card is vendored (generation by season/period, the Cost Relief Credit, Deep Green premium, and the 10% Solar Bonus Credit):

supplier = "cca"

[cca]
name = "MCE"
rate_card = "mce"
pcia_rate = 0.03476                # $/kWh, from your bill
franchise_fee_surcharge = 0.00042  # $/kWh, from your bill

For other CCAs, supply generation_rates directly; see docs/configuration.md. Until generation rates and a franchise fee are configured, CCA mode returns delivery-only prices flagged complete = False rather than quietly understating your rates.

Note that a CCA customer's PCIA is a charge, while a bundled customer's is a credit, so CCA service can cost several cents per kWh more on import.

Keeping rates current

Every vendored dataset is regenerated from the document that publishes it, by tools.regen. Nothing is hand-transcribed and nothing is hand-edited.

python -m tools.regen                              # rebuild every dataset
python -m tools.regen --check                      # exit 1 if a publisher moved
python -m tools.regen tariff --for-date 2025-12-15 # rebuild a superseded vintage
python -m tools.regen.export --download            # the 843 MB export-rate archive

Export matrices come from PG&E's CSV archive, which is large enough to have its own entry point; everything else comes from a published PDF. Nothing is written unless the rendered file survives being read back by the library code that will consume it, so a generator that drifts from the schema fails instead of shipping.

A weekly CI job runs --check, so a rate change surfaces as a failing build rather than as silent drift. Export files are updated by October 1 of any year the CPUC adopts a new Avoided Cost Calculator; retail rates change more often, via advice letters — three times in the first half of 2026 alone — which is why a superseded vintage can be rebuilt from the filing that adopted it. See docs/data.md.

Data sources

See docs/data.md for the source of every vendored table and why OpenEI's URDB is deliberately not used.

License

MIT