Modeling decisions for EIA-anchored electricity disaggregation

[jvendries]

This discussion lists the modeling decisions for putting an EIA-anchored generation / transmission / distribution path into Cornerstone production. It replaces today’s 3-way split and mixed-units conversion (Discussion #85). Co-production cleanup is unchanged. The 405-sector canonical schema is unchanged (407 outputs still only when the existing electricity flag is on).

There are 15 decision points (D0–D14) in independent comments below, then a Phi note. Please comment in that order; earlier decisions constrain later ones. (A former D15 comment remains as a historical pointer: exports as a class were merged into D0.)

These are the intended production methods, not a diagnostics overlay. Code is not in this discussion; a separate implementation plan will cite these decisions.

Methods record (downloadable). eia_anchored_gtd_methods_record.md (raw download) is the living methods log for this replacement: settled design (guiding principles, D0–D14, Phi), the discussion log (why each decision was made, including options not chosen), and implementation notes. Use it as the source of truth if a comment and the log ever disagree. It is not the code implementation plan.

Guiding principles

1. MWh values in the IO model follow EIA's sales of electricity to ultimate customers by class (for example, Residential > Industrial + Direct Use in Table 2.2). We follow that class mix, not the published sales totals as-is. Those four classes share (eGRID − Table 2.14 export MWh). Physical exports are their own class on F04000 (Table 2.14), not Commercial. Inside each class, MWh still follow who already buys electricity in the IO table.
2. Total generation in the model equals national eGRID plant net generation (us_total_net_generation_mwh). GGL interconnect-loss helpers exist but are not added to this cap. EIA class shares only divide that total among industries and final demand. There is no 2017 eGRID, so the 2017 tables use an estimate (~4,039 TWh). The published 2024 model uses 2024 eGRID.
3. We keep each industry's and household's 2017 electricity bill in dollars. Splitting electricity into generation, transmission, and distribution only divides that existing bill; it does not change who pays how much.
4. US generation equals domestic use of generation. Imported electricity is extra. For 2017, that extra is close to EIA's reported imports (~0.96×).
5. When the generation industry uses generation, that use is counted in Industrial + Direct Use. Generation, transmission, and distribution do not buy each other's electricity products.
6. Whatever is left of a purchaser's electricity bill after generation is transmission and distribution. We do not move that leftover using EIA retail prices. If generation would cost more than the bill, water-fill the clipped generation $ onto others in the same class so leftover is not negative and class MWh still hit D0. Nibble that class (and eGRID) only if the class's total bills cannot cover p × class MWh.
7. How large generation dollars are, relative to leftover T&D, follows BEA's 2017 generation share of the electric power sector (~34%), including in the published later year. EIA decides who gets generation MWh; BEA decides only that overall dollar split.
8. Of leftover T&D dollars, transmission versus distribution stays the 2017 BEA mix (~6% / 94%), including after we scale to a later year. That split does not change the generation price.

We do not add new on/off switches for this method. Using manufacturing survey kWh inside Industrial is later work.

Summary

“Current production” is today’s live 3-way split plus mixed-units conversion (#85 as implemented).

#	Decision	This method	vs current production
D0	Purchaser rows	Four Table 2.2 classes: shares of Total End Use × (eGRID − Table 2.14 exports); F04000 is the Exports class (Table 2.14 MWh); F05000 is not a class; within class ∝ electricity $; p numerator = 2017 UGO gen share of 221100 (same share after year scaling)	Compensating w_row forces the UGO G/T/D dollar mix on every purchaser. No EIA class MWh. Mixed units then warp class MWh via Table 2.4. F04000 sits in Commercial.
D1	Use 3×3	U[G,G] = generation industry’s slice of Industrial+Direct Use; off-diagonals 0; no Table 8.3 stacked on those MWh	Diagonal split of the intersection using Table 8.3 Purchased Power + T/D weights.
D2	Make	Make-last: diagonal shares follow Use+Y G/T/D totals	Make is split first from UGO305 (~34% / 4% / 62%).
D3	Columns + VA	Same recipe (fuels → gen, VA residual), but column sizes from Make-last; spill other non-fuel if VA_G would go negative	Column sizes from UGO. Negative VA is warned and kept; no non-fuel spill.
D4	Mixed units	c_col = eGRID / q_$; c_row = 1/p (flat). Table 2.4 out. T&D stay dollars. Domestic Use+Y MWh = eGRID (class nibble is D8)	c_row varies by EIA class (∝ 1 / Table 2.4). Households get ~38% of Residential generation MWh.
D5	eGRID minus EIA end use	No extra Use-row cell. Remaining gap = eGRID − export MWh − Total End Use, smeared on the four Table 2.2 classes. F04000 holds Table 2.14 only (D0)	No eGRID class targets, so no explicit gap to park.
D6	Year scaling	Keep per-child UGO GO growth as an intermediate, then re-apply D0 at the model year (end of derive_cornerstone_Aq_scaled, 3-way flag) and Make-last. Published q follows that year’s EIA MWh mix. D8 bills from the pre-1a row sum	Per-child UGO GO growth lasts on electricity q and A rows (after summary "22").
D7	E and B	Keep production placement: combustion on generation, SF₆ on transmission, B_gen /= c_col. Split GHG-year parent GO with P5 q shares, not 2017 V. Do not overwrite 2017 V	Same E placement and no gas-type fallback. B’s x today follows UGO Make V.
D8	Leftover T&D $	Keep each purchaser’s 221100 $; leftover = bill − generation $. Water-fill within class; nibble a class only if its bills cannot cover p × class MWh. Split leftover with 2017 UGO T/(T+D). Uimp 100% generation; F04000 is a D8 purchaser (Exports class from D0). Do not reassign leftover using Table 2.4	No leftover construct. Every purchaser gets the same G/T/D mix. Bills still sum to 221100.
D9	Flags	No new flags. Existing 3-way-split and mixed-units flags run this method. Reallocation unchanged	Same three flag names. Code behind 3-way and mixed units changes. Old vs new is a committed freeze, not a fourth flag.
D10	Self-use 3×3	T_dom = Udom[221100,221100]; Udom[G,G] from D1 clipped to T_dom; remainder on Udom[T,T] / Udom[D,D]; Uimp intersection → Uimp[G,G] only; off-diagonals 0	Whole intersection split by Table 8.3 diagonal weights.
D11	Imports vs eGRID	Domestic generation Use+Y MWh = eGRID = q. Extra import MWh = abs(Y[221100, F05000]) / p. Imported 221100 is generation only	Mixed units already set domestic row MWh = eGRID then convert Aimp. Imports still get the UGO/w_row G/T/D mix, not 100% generation.
D12	2017 eGRID proxy	2017-chain eGRID = eGRID 2016 × (EIA Table 3.1 2017 / 2016) = 4,039 TWh. After D6, real eGRID at model_base_year	Today’s mixed units already use real eGRID at model_base_year (canonical 2024). The 2017 3-way is monetary and does not call eGRID. 2018 was the diagnostics CF proxy, not live production.
D13	Negative Use/Y cells	Clip to 0 only when forming within-class dollar shares. Live cells unchanged. F05000 and F04000 out of Table 2.2 weights per D0	No class-share allocator, so no class-share clip. Equal-price split can pass negatives through.
D14	Leftover T vs D year	Freeze 2017 UGO T/(T+D) (~5.92% / 94.08%) on the 2017 chain and after D6. Does not enter p. D8 already uses this freeze; this comment is the year-choice record	No leftover T vs D. After scaling, T vs D of electricity q follows UGO GO growth (model-year mix).
Phi	Electricity children	When the 3-way flag is on, Phi = 1 on 221110 / 221121 / 221122 (same as USEEIO Phoebe on 221100). Leftover T&D is D8, not a Phi haircut	Aggregate 221100 already has Phi = 1 in Phoebe. Children get 1 today only via reindex fill.

[jvendries]

D0. Purchaser generation / T&D package

Implementation (intended): electricity_disaggregation.py (commodity rows + Y); Y wiring in derive_disagg_Ytot_with_trade / disaggregate_electricity_commodity_row_in_y in cornerstone_disagg_pipeline.py. Drop compensating w_row. Table 2.14 loader (epa_02_14): fifth class Exports, sole member F04000. End-use map: F04000 → Exports, not Commercial. F05000 is not a D0 class.

In today’s 3-way split, compensating row weights restore BEA UGO305 G/T/D dollar Use+Y totals after a Table 8.3 3×3. Mixed units then put class retail gaps into generation-row conversion factors (c_row ∝ 1 / EIA Table 2.4). Class generation MWh therefore do not match EIA (households ~38% of Residential; Transportation ~11×). Putting F04000 in Commercial made export MWh = bill / p (~73 TWh), which is not EIA Table 2.14 exports (~14 TWh).

Forcing BEA GO row-dollar totals, EIA class MWh, and one national generation price p at once is inconsistent. We drop BEA GO as a Use-row constraint. Class-specific generation prices were rejected in favor of one national p.

Five purchaser classes. Four Table 2.2 classes (Residential, Commercial, Industrial + Direct Use, Transportation) plus Exports (F04000 only). F05000 is not a class (imports extra; detail in D11).

Class MWh. Four Table 2.2 classes get shares of Total End Use times (eGRID − Table 2.14 Canada+Mexico export MWh) — not EIA sales MWh as published, and not shares × full eGRID. F04000 generation MWh = that Table 2.14 total, same year as the eGRID in force (D12 on the 2017 chain; model year after D6). If EPA lags model_base_year, use the latest 2.14 year and log it. Within each Table 2.2 class, MWh still follow electricity-purchase dollars. Households get the scaled Residential total. The generation industry stays in the Industrial+Direct Use dollar pool. The remaining eGRID − export MWh − Total End Use gap is spread over the four Table 2.2 classes; scaled class MWh are not EIA delivered consumption. F04000 holds Table 2.14 only, not the rest of the gap.

Keep the F04000 dollar bill. Leftover on that column is D8 (one-purchaser class: nibble exports only; do not raid Commercial). Do not put F04000 100% on generation.

p numerator is the 2017 UGO generation share of 221100 Use+Y (~$156.6 B, ~34% of ~$458 B). Apply the same 2017 share to the inflated electricity Use+Y total at the model year. EIA MWh set who gets generation; BEA sets only the gen-vs-T&D dollar level. Not the model-year UGO generation share (2024 ≈ 30.6%). Allocated MWh (four classes + F04000) sum to eGRID, so generation Use+Y $ = p × eGRID.

There is no EIA series in the current extract that is “generation GO $.” UGO is required; missing UGO is an error. Table 8.3 Production share (~87% gen) is not a p backup (fuel-heavy opex, not GO).

Not chosen for exports: parking the eGRID − end-use gap on F04000; scaling classes by (eGRID − X) but leaving F04000 inside Commercial; giving F04000 both a 2.14 target and a Commercial $ share; using F04000 / p as EIA exports.

Dollar toy. Electricity Use+Y $100; Steel $40, shop $30, households $30; generation-dollar share 34% → generation $34, leftover $66.

Leftover T&D dollars are D8 (each purchaser’s bill minus generation $). Make-last is D2.

Modeling implications: Generation output and domestic generation use both equal eGRID. Household generation MWh will exceed published Residential sales. All-in ¢/kWh will not equal Table 2.4. Commercial MWh get closer to Table 2.2 Commercial. Leftover T&D on the export column is D8 uniformity, not a physical claim that exported kWh travel US distribution. 2017 bills look slack (F04000 / p ≫ 2.14).

Resolution: Four Table 2.2 class MWh = EIA Total End Use shares × (eGRID − Table 2.14 export MWh); Industrial includes Direct Use; F04000 is the Exports class (generation MWh = Table 2.14 Canada + Mexico); F05000 is not a D0 class; p numerator = 2017 UGO generation share of 221100 Use+Y, including at the model year. Drop w_row.

[jvendries]

D1. Use 3×3 and generation self-use

Implementation (intended): disaggregate_use_intersection() in electricity_disaggregation.py. Off-diagonals stay 0. U[G,G] comes from the purchaser-row builder (D0), not Table 8.3. Remainder of the self-use cell is D10.

The 3×3 of electricity industries buying electricity commodities decides (1) whether T/D total EFs pick up generation combustion through L (off-diagonal purchases) and (2) whether extra generation MWh sit outside the D0 class buckets.

Not putting generation emissions into T/D EFs is the more important EF choice, so off-diagonals stay 0. Direct Use is not a separate cell: it is inside the Industrial+Direct Use share of (eGRID − export MWh). U[G,G] is the generation industry’s dollar-weight slice of that bucket × p. Stacking Table 8.3 generation dollars on top was rejected because D0’s four-class targets plus F04000 already sum to eGRID.

The generation industry column has no T&D leftover (D8). T/D diagonal cells take the rest of T_dom = Udom[221100, 221100] (D10) and stay in dollars, so they do not add generation MWh. Imported self-use is Uimp[G,G] only (D11).

Modeling implications: T/D total EFs do not inherit generation’s direct intensity through this block. U[G,G] MWh is not EIA Direct Use (~141 TWh) and not Table 8.3 / p. Most Direct Use lands on non-utility Industrial columns via dollar weights. A vertical chain (D buys T, T buys G) is not this method; leftover T&D is a dollar markup on the same purchaser (D8).

Resolution: Off-diagonals = 0. U[G,G] = generation industry’s slice of Industrial+Direct Use. Do not add Table 8.3 generation $ on top. T/D diagonals take the remainder of domestic self-use (T_dom; D10). Imported self-use is Uimp[G,G] (D11).

[jvendries]

D2. Make table — Make-last

Implementation (intended): Reverse today’s order. Build commodity rows + Y (and the Use 3×3) first, then disaggregate_make_intersection from Use+Y G/T/D row-total shares instead of build_electricity_disagg_go_weights(). Off-diagonals stay 0. D6 re-runs Make-last after model-year D0.

After co-production cleanup, electricity Make is a diagonal 3×3, so commodity output q and industry gross output x share one mix. Today that mix is BEA UGO305 (~34.2% / 3.9% / 61.9%). D0/D1 set Use+Y from EIA class MWh plus leftover T&D, not from BEA GO. If Make stayed UGO, T/D use would not match q.

Make-last: after D0/D1/D8 produce generation / T / D commodity-row totals (intermediate Use plus Y), set the three Make diagonal cells so each child’s share of aggregate 221100 Make equals that child’s share of Use+Y. Then q follows use (up to imports). Because Make is diagonal, industry x gets the same mix.

This is not copying the Use 3×3 into Make. U[G,G] is generation self-use (a slice of Industrial). V[G,G] is the generation industry’s entire output of the generation commodity — almost all of q_G. Make-last copies row totals, not that self-use cell.

Modeling implications: BEA’s 34/4/62 is no longer who produces G/T/D. UGO is still used for the generation-dollar share (D0) and leftover T vs D (D14). The published CF Make-last percentages (35.8 / 3.8 / 60.4) were sales-only gen dollars and are not the production mix.

Resolution: Make inherits the Use+Y G/T/D split. Do not use BEA GO shares for Make.

[jvendries]

D3. Industry columns + value added

Implementation (intended): disaggregate_use_industry_columns takes Make-last weights instead of UGO. After computing VA_G, if negative, rescale generation’s non-fuel, non-electricity Use rows down (remainder to T/D by Make-last T/D shares) until residual is 0; if other rows hit 0 and residual is still negative, warn and keep the negative VA (same as today).

Make-last sets how large each G/T/D industry is (x_k). Each column still has to buy non-electricity inputs and have VA so that inputs + VA = x_k. D1 already settled electricity purchases in those columns (generation buys generation only; T/D do not buy generation). D3 is everything else: fuels, other intermediates, VA.

Default: same recipe as production, Make-last weights instead of BEA GO. Fuels → generation (100%). Other non-electricity rows ∝ Make-last shares. x_k from Make-last. VA residual. Column still balances.

Backup if default VA_G < 0: spill other non-fuel intermediates off generation onto T/D until VA_G = 0. Fuels stay 100% on generation. U[G,G] is not spilled.

If the backup is not enough (fuels + U[G,G] > x_G even with other non-fuel on generation at 0): warn and write the negative VA, same as today’s production. No further guardrails: do not reduce fuels on generation, do not raise x_G, do not shrink U[G,G].

Modeling implications: Fuels stay on the industry that burns them. The only repair is moving non-fuel intermediates, which does not leak fuel-chain EFs onto T/D. Negative generation VA remains possible in the pathological case. T&D-as-markup columns (VA absorbs leftover) were not chosen.

Resolution: Option 1 with Make-last weights. Spill other non-fuel if VA_G would go negative; else warn and keep negative VA as today.

[jvendries]

D4. Mixed units

Implementation (intended): Replace electricity_class_row_factors (Table 2.4 / λ) with a constant 1/p. Keep electricity_output_factor as eGRID / q_$. Rewrite mixed-units tests that currently assert Industrial c_row > Residential c_row. eGRID year is D12 on the 2017 chain and D6 at the model year.

Mixed units convert generation from dollars to MWh and leave T&D in dollars. Today’s class-varying c_row ∝ 1 / Table 2.4 would undo D0 class MWh. If Make-last q_$ ≠ generation Use+Y dollars, a single factor (c_row = c_col) makes output = eGRID but Use+Y MWh ≠ eGRID.

Option 2: eGRID constrains generation on Make (q / x via c_col = eGRID / q_221110_$) and on Use+Y (c_row[j] = 1/p for every purchaser). Table 2.4 does not enter mixed units. B generation is still divided by c_col. If q_$ ≠ Use+Y $, the self-use A cell is scaled by c_row / c_col ≠ 1. Keep today’s two y_nab getters: published / Aq_scaled y_nab stays dollars; mixed-units y_nab is hybrid (generation in MWh) so y = q − A q matches mixed A/q. Do not force mixed-units y_nab[221110] to dollars. Imports use the same generation-row conversion as domestic (D11).

Modeling implications: Class generation MWh from D0 survive mixed units. We give up one shared conversion factor. D6’s Make-last after re-anchor makes q_$ ≈ Use+Y $ at the model year (up to imports), so the two-factor case is mostly imports.

Resolution: c_col = eGRID / q_$; c_row = 1/p (flat). Table 2.4 out of conversion. Domestic Use+Y MWh = eGRID (the class-level nibble, if any, is D8).

[jvendries]

D5. eGRID minus EIA end use

eGRID exceeds EIA Total End Use. That gap is generation that is not EIA end use (losses, plant use, and the part of trade not peeled off as D0 exports). The diagnostics counterfactual left the gap off the Use row. D0 already smears eGRID − export MWh − Total End Use across the four Table 2.2 classes. F04000 holds Table 2.14 only (D0). D4 then makes Make output and Use+Y both eGRID in MWh, so there is no MWh hole left to park. Leftover dollars for T&D are D8.

Illustration: 2018 eGRID 4,168 TWh minus EIA Total End Use ~3,864 TWh ≈ 304 TWh (CF working table, before peeling exports onto F04000). Production 2017-chain uses D12 4,039 TWh minus Table 2.14 2017 exports on F04000. The placement rule does not change: no dummy cell.

Modeling implications: Scaled four-class MWh are not EIA delivered consumption. There is no explicit losses / plant-use generation cell. Exports are on F04000 via D0.

Resolution: No extra Use-row cell. The remaining gap exists only as D0 scale-up on the four Table 2.2 classes. Direct Use stays inside Industrial’s share. Exports are D0 (F04000 = Table 2.14).

[jvendries]

D6. Year scaling

Implementation (intended): Leave rescale_electricity_children_to_detail_GO_growth_A / _q as they are (per-child UGO GO growth after summary "22"). That hook is before commodity-PI inflation. Re-apply D0 at the end of derive_cornerstone_Aq_scaled (3-way flag, not mixed-units-only): D8 bills from the pre-1a electricity-row sum, then PI; four-class MWh from model-year EIA 2.2 × (eGRID − 2.14 exports); F04000 = that year’s Table 2.14 (D0). Do not spread F04000 from 2017 Y column shares. Extra import MWh from the scaled imports vector (D11). Rebuild electricity A rows / q (Make-last in A/q) and Adom columns. Do not overwrite 2017 V. Then D4.

The 3-way split is built on 2017 detail tables. The published model is a later year (e.g. 2024).

Today, after disaggregation, all 407 sectors get BEA summary-sector growth. Electricity children sit in Utilities "22". Production then applies a per-child GO-growth correction (code Decision 7 in year-scaling; not this discussion’s D7): G/T/D grow at different UGO rates (2017→2022 illustration: generation ~1.62×, transmission ~1.29×, distribution ~1.33× vs Utilities ~1.43×).

If we only inflate 2017 D0 dollars, 2024 class MWh is the 2017 EIA mix grown, not the 2024 Residential/Industrial mix. D0 as a model-year identity wants EIA 2.2 shares × (eGRID − 2.14 exports) at the model year, plus that year’s Table 2.14 on F04000.

Ruled out: split-last (2017 3-way must still go through the same summary inflation as flag-off). Scale-only with no EIA re-anchor.

Chosen: keep 1a (per-child GO growth) as an intermediate, then re-apply D0, then Make-last again. 1a does not last on electricity q or purchaser rows. Prefer that year’s EIA MWh mix over BEA gen GO growth on published output.

D6 does not keep the 2024 UGO generation share. 1a would temporarily change gen’s share of q; D0 and Make-last overwrite it. What lasts: (1) who gets generation MWh — that year’s EIA 2.2 shares × (eGRID − 2.14) plus Table 2.14 on F04000; (2) how large generation $ are vs leftover — 2017 generation-dollar share × inflated 221100.

Toy. 2017 Use+Y and q both $100 (generation $34). Summary "22" 1.43× → $143. 1a (2017→2022 illustration) would make q about $55 gen / $5 T / $83 D. D0 re-apply might land generation Use+Y at $50; Make-last sets generation q to $50, not $55.

Modeling implications: The inflation mechanism is unchanged when the electricity flag is on. Published electricity q follows EIA use mix, not lasting 1.62× vs 1.29×/1.33×. Electricity-column purchases of non-electricity commodities follow the post-Make-last x.

Resolution: 1a as intermediate; re-apply D0 at the model year; re-run Make-last so published q follows EIA MWh mix. p numerator stays the 2017 generation-dollar share.

[jvendries]

D7. Emissions E and B

Implementation (intended): Leave eGRID FBS mapping and apply_electricity_unit_conversion_to_B as they are. Do not overwrite 2017 V. When splitting GHG-year parent 221100 GO across G/T/D (distribute_electricity_aggregate_x_using_v_row_shares), use P5 published q shares from derive_cornerstone_Aq_scaled().scaled_q, not 2017 V row shares. Vnorm stays on 2017 V. Missing eGRID inventory remains an error (no gas-type fallback), per #85.

Direct emissions E are an industry fact (who combusts). Intensities B = E / x (then mixed units divide generation B by c_col). D is column sums of B. N is D through the supply chain (B L). D0/D1 already decided T/D do not buy generation, so T/D total EFs do not pick up generation combustion through the electricity 3×3. Who does inherit generation’s D is whoever buys generation MWh (EIA class mix).

Keep production placement:

• Plant combustion (non-SF₆) → generation 221110
• SF₆ → transmission 221121
• Distribution 221122 ≈ 0 direct E
• Fuels in the Use table go to the generation column (D3)
• When B uses E-year GO, aggregate 221100 x is expanded with P5 q shares (published Make-last after model-year D0), not 2017 V row shares and not BEA GO
• Mixed units: B[:, 221110] /= c_col

Leftover T&D can be a large $ industry. Putting combustion E there would make T/D look carbon-intensive because they are a markup, not because they burn fuel.

Modeling implications: D0 changes purchaser N (who buys gen MWh). T/D D stay near zero except SF₆ on T. Generation x (and thus D_gen dilution) follows published Make-last q after P5, not 2017 V and not UGO.

Resolution: Keep production E/B placement. Do not put combustion on leftover T&D. Do not use a BEA-GO denominator for B.

[jvendries]

D8. Leftover T&D dollars, clip, map, Y

Implementation (intended): For each non-electricity purchaser, domestic bill = Udom + Y cell of 221100 (except F05000). Do not include Uimp in the D8 bill — write all Uimp 221100 onto the generation row with no leftover. Water-fill gen $ within class so gen ≤ bill and class totals hit D0 when class bills allow. F04000 is a one-purchaser Exports class (D0): gen $ = p × Table 2.14 MWh; leftover = bill − gen; nibble that class only (do not raid Commercial). td = bill − gen. Split td with 2017 UGO T/(T+D) (~5.92% / 94.08%), including after D6. Write Use rows and Y. Drop w_row. Keep today’s end-use map (households = Residential; electricity children = Industrial). In P2, set END_USE_MAPPING_REVIEW_STATUS to adopted for EIA-anchored G/T/D (no longer DRAFT). F05000 stays out of D0 class pools; F04000 maps to Exports (D0). Published y_nab stays dollars; mixed-units y_nab is hybrid (D4). Electricity 3×3: D1 U[G,G] plus D10 remainder on U[T,T] / U[D,D].

Generation dollars are already set (class MWh × one generation price). What remains is leftover T&D on each purchaser, and whether that changes that purchaser’s electricity bill. Changing bills to follow EIA retail-price gaps would move $ off industries onto households and force a VA repair (or break x = inputs + VA).

Option 3 (leave industry bills alone, put the leftover pattern only in Y) was ruled out — it would not model industry electricity use more accurately. Option 2 (change bills, VA absorbs) hits an EIA leftover-dollar pattern but rewrites sector VA; when $ move onto households, national industry VA rises.

Option 1 (chosen): keep each purchaser’s electricity $ (the old 221100 cell). Generation $ = class MWh × p, allocated within class by electricity-purchase dollars. Leftover on that purchaser is the rest of their bill. Do not spread leftover using EIA Table 2.4. Class leftover ¢/kWh is (class IO $ − class gen $) / class MWh, not Table 2.4 minus p.

If generation $ would exceed that purchaser’s bill, water-fill the clipped $ onto others in the same class so leftover is not negative and class MWh still hit D0. Do not cut generation $ to match an EIA retail bill. Nibble that class only if class bills < p × class MWh. On 2017 live IO this does not bind.

Domain. No leftover on the generation industry column (D1). Uimp is 100% generation (not a D8 bill). F04000 is a D8 purchaser (Exports class from D0). Extra import MWh = |F05000| / p is D11. The year freeze on T/(T+D) is recorded as D14; D8 already uses 2017.

Dollar toy. Electricity Use+Y $100; Steel $40, shop $30, households $30. Generation $ stay $34. EIA shares put about 29% / 35% / 36% of generation on Steel / shop / HH, so generation $ ≈ $10 / $12 / $12.

• Option 1: bills stay $40 / $30 / $30. Leftover ≈ $30 / $18 / $18. Steel’s other inputs and VA do not move.
• Option 2 (not chosen): leftover might be $9 / $24 / $33, so all-in bills become $19 / $36 / $45. Steel’s electricity bill falls $21; households’ rises $15. Steel x is unchanged, so unless VA rises $21 the column does not balance.

Modeling implications: Industry columns stay balanced without moving VA. End-use classes still differ in generation vs leftover mix via D0. Leftover $ by class will not follow Table 2.4 price gaps. All-in ¢/kWh will not equal Table 2.4. A class-level nibble of D0/eGRID MWh remains possible if that class's bills cannot cover p × class MWh (not observed on 2017 IO).

Resolution: Keep each purchaser’s electricity $. Leftover = bill − generation $. Water-fill within class; nibble a class only if its bills cannot cover p × class MWh. Do not reassign leftover $ using Table 2.4. Split leftover with 2017 UGO T/(T+D). Uimp is not a D8 bill. F04000 is a D8 purchaser (D0 Exports class).

[jvendries]

D9. Config flags

Implementation (intended): No new flags in usa_config.py. implement_electricity_reallocation stays as it is (co-production cleanup unchanged). implement_electricity_disaggregation is this new 3-way split. implement_electricity_mixed_units is the new generation conversion. Existing YAML files keep the same flag names; only the code behind them changes.

P0 first (on today’s code, before replacing the 3-way): run today’s 3-way and mixed-units waterfall configs and write a committed snapshot under bedrock/analysis/electricity_disagg_eia/output/baseline_current_production/. Canonical v0.3 (all electricity flags off) is unchanged. Reallocation-only configs are unchanged. Production 3-way-split and mixed-units tests are rewritten in the same PR as the replacement (not skipped). Analysis d_85 stays historical.

Today there are three electricity switches: reallocation, 3-way split, mixed units. A fourth switch would keep the old 3-way around for comparison but adds config surface and waterfall states.

Modeling implications: Old vs new in one checkout is a committed freeze of today’s production, not a live dual path. Waterfall config names stay: footing → co-production cleanup → 3-way split (this method) → mixed units (this conversion).

Resolution: Do not add flags. Replace 3-way-split and mixed-units implementations in place. Compare to the P0 freeze, not to a second production flag.

[jvendries]

D10. Electricity self-use 3×3

Implementation (intended): After D0 writes Udom[G,G], set Udom[T,T] and Udom[D,D] from T_dom − Udom[G,G] with 2017 UGO T/(T+D) (D14). Write Uimp[221100,221100] onto Uimp[G,G] only (D11). Do not write off-diagonals. Same rule after D6 / P5 (T_dom = collapsed domestic electricity-industry bill; Aimp intersection 100% generation).

D8 leftover is retail bill minus generation $ for Steel / shop / HH. Self-use is an intersection cell. Do not unbundle it with leftover language. Place the cell on the diagonal.

Aims: (1) do not lose dollars of old U[221100, 221100]; (2) do not lose dollars of 221100 x or q; (3) generation MWh = eGRID; (4) no Use 3×3 off-diagonals.

• D1 only (U[G,G], T/T = D/D = 0) hits (3) and (4) but drops the remainder of the cell, so Use+Y and Make-last q/x fall.
• Plain diagonal split U[k,k] = w_k × T hits (1), (2), and (4) but U[G,G] = w_G × T is not D1’s purchaser slice, so generation Use+Y $ ≠ p × eGRID.

The construction that hits all four:

• Domestic intersection: T_dom = Udom[221100, 221100]. Udom[G,G] = min(D1 purchaser slice, T_dom). Remainder on Udom[T,T] and Udom[D,D] with 2017 UGO T/(T+D). Those cells stay dollars.
• Imported intersection (D11): Uimp[221100, 221100] writes entirely to Uimp[G,G]. No leftover T/D on imported self-use. Do not put imported $ into T_dom.
• Clip D1 against T_dom only, not T_dom + Uimp. Off-diagonals 0.

If D1 U[G,G] would exceed T_dom, clip to T_dom and put the clipped generation $ on other Industrial purchasers in that class (same water-fill idea as D8, inside Industrial). Compute that inside the shared purchaser builder so Use and Y see the same result.

Dollar toy. Use+Y $100, domestic self-use T_dom $8, D1 U[G,G] ≈ $2. Write $2 / $0.4 / $5.6 on Udom G/T/D diagonals. Imported self-use, if any, sits on Uimp[G,G] only. Steel / shop / HH bills unchanged. Generation $ still $34.

Modeling implications: Circular T/T and D/D self-use is ordinary IO; it does not put generation emissions on T/D through this block. T and D industries buy their own commodities. U[G,G] is still D1’s Industrial slice, not w_G × T. Import MWh sit on top of eGRID.

Resolution: Keep the whole self-use cell. T_dom = Udom[221100,221100]; Udom[G,G] from D1 clipped to T_dom; remainder on Udom[T,T] / Udom[D,D]; Uimp intersection → Uimp[G,G] only; off-diagonals 0.

[jvendries]

D11. Imports vs the eGRID cap

Implementation (intended): D0/D8 eGRID allocation on domestic bills only. Place import 221100 entirely on the generation row (Uimp and the F05000 generation cell). Convert at 1/p. National check: import generation MWh ≈ abs(Y[221100, F05000]) / p. Intermediate Uimp is the industry slice of F05000, not additional.

eGRID is US plant net generation (domestic output). BEA also has imported electricity. D0/D4 said generation Use+Y MWh = eGRID = q. If “Use+Y” includes imports, imported MWh are taken out of the eGRID cap. If “Use+Y” is domestic only, then q = domestic use = eGRID and import MWh sit on top. You cannot have all three of: q = eGRID, total (domestic+import) use = eGRID, and import MWh extra.

Today’s mixed units already pick the domestic identity: preserve domestic row MWh = eGRID, then apply the same c_row to Aimp. D5’s eGRID − end-use gap is EIA losses / plant use / trade smeared into class MWh; it is not the BEA import row.

Option 1: domestic generation Use+Y MWh = eGRID = q. Extra import MWh = abs(Y[221100, F05000]) / p. Imported 221100 is the generation commodity (wholesale), not a US bundled retail bill. Do not apply the UGO generation-dollar share or D8 leftover to imports.

Bedrock eGRID on this path is plant net generation only (us_total_net_generation_mwh); GGL losses are not added. It has no national import/export series. EIA Table 2.14 (already extracted as epa_02_14) does. On 2017 IO dollars with the CF 2018 eGRID proxy, abs(F05000) / p ≈ 65.1 TWh vs EIA 2018 imports 58.3 TWh (1.12×) and EIA 2017 imports 65.7 TWh (0.99×). The 2018 gap is mostly year. After D12, 2017-chain p ≈ $38.56/MWh and abs(F05000) / p ≈ 63.0 TWh vs EIA 2017 65.7 TWh (0.96). Exports do not use this import recipe (F04000 / p vs EIA exports). D0 assigns Table 2.14 export MWh to F04000 inside eGRID.

Modeling implications: Output = domestic use, not all use. Four-class MWh shares × (eGRID − exports) describe domestic ultimate-customer use of US generation. Export MWh are D0. Imports are extra supply on the generation row.

Resolution: Domestic Use+Y MWh = eGRID = q. Extra import MWh = abs(F05000) / p. Imported 221100 is generation only.

[jvendries]

D12. 2017 eGRID proxy

Implementation (intended): Helper egrid_mwh_for_io_year(2017) = us_total_net_generation_mwh(2016) × EIA Table 3.1 total MWh 2017 / EIA Table 3.1 total MWh 2016. Other years: real eGRID. Plant net generation only; do not add GGL losses. D0/D4/D11 on the 2017 chain use this. After D6, us_total_net_generation_mwh(model_base_year) — the same call today’s mixed units already make. Table 2.14 export MWh (D0) uses the same year as that eGRID; if EPA lags, latest 2.14 year and log it.

Current production: mixed units in electricity_conversion_factors call us_total_net_generation_mwh(cfg.model_base_year). Canonical v0.3 is model_base_year=2024, so 2024 eGRID. There is no 2017 fallback; a missing inventory year raises InventoryNotAvailableError. The 2017 3-way split is monetary and does not call eGRID. 2018 as a 2017 proxy was the diagnostics CF, not live production.

D0/D4/D11 use eGRID MWh as domestic generation output. The 3-way split is built on 2017 detail tables. Stewi eGRID inventories in bedrock are 2014, 2016, 2018–2024 — no 2017. D6 already re-applies D0 at the model year with that year’s real eGRID, so this proxy does not set published q on canonical v0.3. D12 is the 2017-chain cap (needed once D0 writes MWh on the 2017 tables).

Option 4: estimate 2017 eGRID from the EIA 2016→2017 trend:

eGRID_2017 = eGRID_2016 × (EIA Table 3.1 2017 / EIA Table 3.1 2016)

Table 3.1 is all-sector net generation (3.1.A fossil/nuclear/pumped storage + 3.1.B renewables, including estimated small-scale solar). Assumption: the eGRID/EIA coverage ratio is stable from 2016 to 2017.

Calculation (bedrock, 2026-08-22):

• EIA 3.1 2016 = 4,096.387 TWh; 2017 = 4,059.433 TWh; ratio = 0.990979 (−0.90%)
• eGRID 2016 = 4,075.323 TWh
• eGRID_2017 = 4,038.559 TWh (report 4,039 TWh)
• Same coverage gap vs EIA as 2016 (0.99486); 0.5% below EIA 2017

Not chosen: raw eGRID 2018 (4,168 TWh, the CF proxy, 2.7% above EIA 2017); raw eGRID 2016 (misses the 2016→2017 dip); raw EIA 3.1 2017 (not eGRID coverage); 2018-backward twin (4,019 TWh).

Knock-on for D11 on the 2017 chain: p ≈ $38.56/MWh (was $37.36 with 2018 eGRID). abs(F05000) / p ≈ 63.0 TWh vs EIA Table 2.14 2017 imports 65.7 TWh (0.96).

Modeling implications: The 2017 chain does not bit-match the published CF’s 2018 proxy. The estimate is documented and reproducible from on-disk EIA 3.1 + eGRID 2016.

Resolution: 2017-chain eGRID = eGRID 2016 × (EIA 3.1 2017 / EIA 3.1 2016) = 4,039 TWh. Model year still uses that year’s real eGRID (D6).

[jvendries]

D13. Negative Use/Y cells

Implementation (intended): clip(lower=0) on the dollar series used for within-class shares (same as the diagnostics CF). Drop F05000 and F04000 from Table 2.2 class weights (D0 class map). Map F04000 to Exports (D0). Leave live Use/Y values unchanged.

D0 splits each Table 2.2 class’s MWh among purchasers in proportion to their electricity dollars. A negative dollar cell then wants negative MWh. That is a different clip from D8 (water-fill / nibble on a positive bill).

On 2017 BEA detail 221100, industry Use has no negative cells. The only negative Y cell is F05000 = −$2.431 B. After expansion or year scaling, small inventory/scrap negatives can still appear.

Resolution: Clip to 0 only when forming within-class dollar shares. No purchaser gets negative MWh. Do not rewrite live Use/Y cells to 0. Exclude F05000 and F04000 from Table 2.2 class weights (D0: imports extra; Exports is its own class). On a rare negative industry bill: generation $ = 0, leftover stays that negative cell (D8 leftover ≥ 0 does not bind there). F05000 is not a D8 purchaser. F04000 is a D8 purchaser.

Modeling implications: Matches the CF weight clip. Keeps q = Utot + Y. Does not double-count D11. Leftover may be negative on quirk cells.

[jvendries]

D14. Leftover T vs D year

Implementation (intended): T/(T+D) from 2017 UGO in build_electricity_disagg_go_weights() (or equivalent), including after D6 re-apply. Do not read model_base_year UGO for this split. Same ratio for D8 leftover and D10 self-use remainder.

D8 leftover on each (positive) purchaser is bill minus generation $. That leftover pool is split transmission vs distribution with UGO T/(T+D), not the full G/T/D mix. D10 uses the same T/(T+D) for the self-use remainder. D0 already froze the generation-dollar share at the 2017 UGO ratio even after year scaling. D14 is whether T/(T+D) is also frozen at 2017.

UGO T/(T+D) barely moves: 2017 5.92% / 94.08%; 2024 6.03% / 93.97%. Generation’s share of G+T+D does move in UGO (2017 34.2% → 2024 30.6%). That is not the published D6 mix (1a does not last).

p = (2017 generation-dollar share × 221100 Use+Y $) / eGRID. T/(T+D) is not in that formula. Model-year T/(T+D) would also leave p unchanged. What would change p is using the model-year generation share of 221100, which D0 already ruled out.

Using 2024 T/(T+D) would mix UGO years: 2017 for gen-vs-leftover, 2024 for leftover T vs D. Using 2024 UGO for generation share as well would reopen D0.

Toy. Leftover $66. 2017 split → T $3.91, D $62.09. 2024 split (not used) → T $3.98, D $62.02. Generation $ still $34; p unchanged.

Modeling implications: Leftover T vs D stays on the same 2017 UGO structure as the generation-dollar share. The year-to-year T/(T+D) gap is ~0.1–0.3 percentage points.

Resolution: Freeze 2017 UGO T/(T+D) (~5.92% / 94.08%) on the 2017 chain and after D6.

[jvendries]

D15 (historical). Exports as a D0 class — merged into D0

This comment is not a later override. Physical exports as their own class (F04000 = EIA Table 2.14 Canada + Mexico; four Table 2.2 classes share eGRID minus that MWh) is part of D0. Read D0 for the live identity.

Settled 2026-08-22, then folded into D0 so Discussion #88’s “earlier decisions constrain later ones” holds. Permalink kept so older links still resolve.

Resolution: See D0. Do not treat this comment as a sixteenth independent decision.

[jvendries]

Phi on disaggregated electricity

Implementation (intended): When implement_electricity_disaggregation is on, phi_for_sectors sets 221110 / 221121 / 221122 to 1.0 (producer price = purchaser price). Do not rely only on reindex(..., fill_value=1.0). Other commodities keep derive_phi_cornerstone_usa_at_year. No new flag. Do not error if 3-way and cornerstone_industry_avg_margins are both on. P6: assert those three Phis are 1 on a 407 index.

D8 leftover T&D is a producer-price split of each purchaser’s old 221100 bill. Phi is a different conversion: purchaser-price N = producer-price N × Phi. Canonical v0.3 has margins on and electricity off. The electricity waterfall has margins off. 2025_usa_cornerstone_full_model_electricity_disaggregation already has both flags on.

USEEIO Phoebe 221100/US is exactly 1.0 for every year 1997–2024. USEEIO does not put electric T&D into Phi on aggregate electric power. After a 3-way split the children are missing from the 405-sector Phi list, so they already get 1 via fill — set it explicitly. Mixed units: generation N is per MWh, so a dollar-margin Phi ≠ 1 on 221110 would be the wrong units.

Modeling implications: Leftover T&D stays in the producer-price Use table (D8). Phi = 1 on G/T/D does not add a second markup on electricity’s own N. Steel can have Phi ≠ 1 while electricity has Phi = 1.

Resolution: Phi = 1 on the three electricity children when the 3-way flag is on, matching Phoebe on 221100. Leftover T&D is D8, not a Phi haircut.