The MADD methodology, in one battery, one week.

A grid-scale battery only reduces emissions if it shifts dirty generation to clean generation. How much it reduces emissions is a surprisingly subtle question — and the answer depends, hour by hour, on which specific power plant ramps down when the battery discharges, and which plant ramps up when the battery charges.

This page walks through exactly that calculation for a real Chilean battery: the AES Atacama BESS Stand-alone project (100 MW / 400 MWh / 4-hour duration). We follow the official methodology used by the Sherpas/Scientiis platform to estimate the project's 2025 emission reduction — a number eligible for issuance under Article 6.2 of the Paris Agreement as Internationally Transferred Mitigation Outcomes (ITMOs). All numbers are drawn from the project's audited 2025 dataset; no figures are illustrative.

↓ Annual summary (CSV) · ↓ Monthly breakdown (CSV) · ↓ Hourly anchor-week table (CSV)

§1 Why a battery can reduce emissions

Electricity is a commodity with two unusual properties. First, it is nearly impossible to store at grid scale — until recently. Second, the physics of the grid require that supply and demand match second by second. To make this work, every country runs a dispatch process: a single operator decides, every five minutes, which generators are turned on and how much each produces.

Chile's electric independent operator is the Coordinador Eléctrico Nacional (CEN). It dispatches generators by a rule called the merit order: cheapest first, most expensive last. The price of the last generator that has to turn on to meet demand sets the spot price (also called the marginal cost) for that hour, at that location.

This is why batteries can matter. If a battery charges when many cheap, clean generators are about to be wasted (curtailed solar at noon), and discharges when expensive, dirty generators are running (gas turbines at peak demand), the net effect is to move energy in time from clean hours to dirty hours. That is the source of every tonne of CO₂ a battery can credibly claim to have avoided.

The detail that matters: at any given hour, dispatch is set by economics, not by carbon emissions. The marginal generator is the one with the highest variable cost, not the highest emission factor. So the questions we have to answer rigorously are:

• Which specific generator's output gets reduced when the battery discharges?
• Which specific generator's output gets increased (or un-curtailed) when the battery charges?
• What are their emission factors, hour by hour?

This is not solved by using a single grid-average emission factor. The Chilean grid average in 2025 was about 0.32 tCO₂/MWh, but the specific unit that ramps down when this battery discharges is, as we'll see, almost always a fossil-fired unit with a much higher emission factor (0.4 to 0.9 tCO₂/MWh). Using the grid average would significantly under-state the project's contribution.

§2 The actors: who generates electricity in Chile

Chile's Sistema Eléctrico Nacional (SEN) is a single interconnected system stretching from Arica to Chiloé, with about 30 GW of installed capacity. Roughly half is dispatchable thermal generation (coal, natural-gas combined-cycle and gas turbines, diesel motors); the other half is non-dispatchable renewable (solar PV, wind, run-of-river hydro), with the balance from reservoir hydro that behaves part-way between the two.

In our anchor week (1–7 September 2025) the grid dispatched 1.65 TWh spread across 13 distinct technology classes:

The economics of the dispatch are reflected in each plant's variable cost (VC). Solar and wind have near-zero VC because their fuel — sunlight and air — is free. Reservoir hydro is also free at the margin (water has an opportunity cost the operator manages through a different mechanism). The expensive plants are the thermal ones: coal, gas, and diesel all burn fuel that has to be bought every hour. So in any given hour, the merit-order stack rises from "renewables first" at the bottom to "diesel motors last" at the top.

In the methodology, every plant carries a known per-MWh cost VCu,i (audited variable cost) and a known per-MWh emission factor EFu (audited technology-specific emission factor). The u indexes the generating unit, the i indexes the hour.

Technology glossary

Quick reference for the technology classes used throughout the page.

§3 One week in the life of the grid

To make the methodology tangible, we anchor everything that follows on a single typical week of 2025: the seven days from Monday, 1 September to Sunday, 7 September. We picked this week because all hours have full data, no extreme outages occurred, and the marginal-unit mix is representative of the rest of the year.

Below is the entire week, hour by hour, generation stacked by technology, with the system-wide generation-weighted spot price overlaid as a dotted line. The lightly shaded vertical bands mark the battery's fixed charging window (10–13h, in pale gold) and discharging window (19–22h, in burgundy).

Two visual rhythms tell the whole story. First, the solar bulge each day from about 09:00 to 17:00: a huge wave of yellow that pushes the thermal stack down. Second, the evening peak after sunset — the yellow disappears, and the merit order must turn on its costliest, dirtiest generators to cover the same demand. The spot price line follows this rhythm precisely: cheap during daylight, expensive at night.

The battery's job is to surf those rhythms. Here is what AES Atacama BESS actually did across the same week:

Each day: 100 MWh charged between 10:00 and 13:00 (the curtailed-solar zone), and 93 MWh discharged between 20:00 and 23:00 (the post-sunset peak, when the gas plants would otherwise be running flat out). The 7 MWh missing on discharge are the round-trip efficiency losses — heat in the inverters and chemistry.

So our job, for every one of those 28 discharge hours and 28 charge hours in the week, is to identify the specific units on the other side of the trade. Let's zoom into one hour.

§4 Discharge: 20:00, Saturday 6 September 2025

This is the hero hour. The battery discharges its 93 MWh into the grid over the 20:00–21:00 window. Somewhere in the SEN, that 93 MWh of generation that would otherwise have come from some mix of dispatchable plants is no longer needed. The methodology's job is to identify exactly which plants.

It does this by sorting all dispatchable units by their economic distance to the marginal price, defined as DIFu,i = VCu,i − SPPu,i. The discharge merit order DORd sorts units by DIF in ascending order: the unit with the most-negative DIF — the cheapest one already running well below its busbar's spot price — sits at rank 1 and is the first the battery's 93 MWh would offset, followed by progressively more expensive units until the cap is reached.

This ordering is also the conservative one: by assuming the battery first offsets the cheapest units (which are typically the lowest- emission ones — biomass cogen, run-of-river, etc.), the methodology credits the project with the smallest avoided emissions consistent with the merit order. Any other ordering would yield more credits.

Below is exactly that stack for 20:00 on 6 September 2025. Each row is one generating unit; the faded bar shows how much energy it could have contributed at that hour (its available energy, AVE), and the solid colored portion shows how much of that energy was actually offset by the battery's 93 MWh discharge.

Walking the stack from rank 1 onward: the first four units sit deep "below water" with DIF between −150 and −116 USD/MWh — they were producing well under their busbars' spot price. All four (TER CMPC SANTA FE U1, TER NUEVA ALDEA U1, TER NUEVA ALDEA U2, TER CHOLGUAN U1) are biomass cogeneration units from the Chilean pulp-and-paper industry with EF = 0, so the battery offsetting their 67 MWh of output adds zero tCO₂. The fifth unit, PMGD TER ORAFTI U1 (cogen gas turbine, EF = 0.72), is the first dirty plant in the stack and contributes 5.08 × 0.72 = 3.66 tCO₂. Then more biomass (TER MASISA, TER LICANTEN, TER PLANTA DE ACIDO SULFURICO MEJILLONES) at zero, and a small landfill-gas motor (TER SANTA MARTA U1, 4.10 MWh, EF = 0.56) adding 2.30 tCO₂. The cumulative AVE crosses the 93 MWh cap inside the 10th unit, TER PETROPOWER U1 (petroleum-coke steam, EF = 0.99) — the dirtiest plant in the stack — of which only 8.49 MWh out of 56 are offset, contributing 8.41 tCO₂. This unit — outlined in burgundy in Figure 4 — is the marginal unit of the hour.

The total avoided emissions in this single hour are:

One hour, one specific battery, ten specific power plants identified by name. Notice how the conservative DIF-ascending ordering shapes the result: Petropower (EF ≈ 1) does enter the stack, but only at rank 10 and only with a partial slice — the rest of the 93 MWh has already been consumed by biomass and small thermal units higher in the order. Had we instead sorted by emission factor descending, Petropower would have received the full slice and the credited reduction would have been more than six times larger. The DIF-ascending rule is what keeps the methodology auditor-defensible.

Why sort by DIF, not by EF?

The merit order is set by economics, not by emissions. The operator dispatches by cost. If we sorted units by emission factor instead, we would identify the dirtiest plants as the marginal ones — which would be tempting (it gives a bigger answer) but wrong: the dirtiest plants are often not the ones the operator would choose to ramp down. They are often cheap to run (coal) or essential to the system's stability and must keep operating.

So DIF = VC − SPP captures the right idea: among the units actually producing at this hour, which are closest to being switched off? Sorting by DIF ascending puts the most-below-water units first — the ones the battery effectively replaces — and forces the methodology to credit their (typically low) emission factor rather than that of the dirtiest plant on the system.

A note on units dispatched for system security. Units with VC > SPP (positive DIF) are running because the operator forces them on for voltage support, reserves or transmission constraints — not for economics. The methodology never credits the battery with replacing them: in discharge, DIF-ascending pushes them past the 93 MWh cap; in charge, Eq24 excludes them (VC > SPP ⇒ EGEMc = 0). They are locked in by physical constraints, so the operator would not have moved them either way.

A subtle point: how ties are broken

Sometimes two units have the same DIF. The methodology then leans on EF as a tiebreaker, and the direction of that tiebreaker is set differently on each side — always in the direction that reduces the project's credited reduction. On the discharge side, ties break by EF ascending: among equally-cheap units, the cleanest is "displaced" first, crediting the project with the smallest avoided emissions consistent with the tie. On the charge side, ties break by EF descending: among equally-eligible units, the dirtiest is treated as the one whose output rose, charging the project with the largest induced emissions consistent with the tie. Same conservativeness principle, just pointed opposite ways.

§5 Charge: 12:00, Saturday 6 September 2025

Now the symmetric question: when the battery charged 100 MWh between 10:00 and 14:00, whose output went up to supply it? Charging adds to demand, so whatever generator was on the margin between dispatch and not dispatch has to produce a little more. The methodology screens candidates with Eq24: only units that pass VCu,i ≤ SPPu,i are economically eligible (their variable cost is at or below the spot price at their busbar), and everything else has EGEMc = 0 and never enters the merit order. Concretely, at 12:00 on 6 Sep, 322 of the 1187 units in the SEN are filtered out before sorting; the 829 that remain are the universe the battery could plausibly draw from.

The charge merit order DORc then sorts those eligible units by DIF in descending order — the unit closest to the margin (least below water) first, the deepest-curtailed last. The conservative direction here is the opposite of discharge: starting with the highest-DIF eligible unit (which on a high-curtailment hour is a unit sitting right at DIF = 0) and breaking ties by EF descending assigns charging to the dirtiest plausible source first, maximising induced emissions and minimising the project's net credit. Below is the stack for the hero charge hour:

This is the most important chart on the page. Even with the conservative dirtiest-first tie-break, every unit the battery actually reaches is a renewable: wind (PE ATACAMA), small run-of-river hydro (HP GUAYACAN, HP RIO HUASCO), and solar PV (PFV LLANO DE LLAMPOS, PMGD PFV PAMA, PMGD PFV TAMBO REAL, PMGD PFV SANTA CECILIA, PMGD PFV LOMAS COLORADAS, PFV PILAR LOS AMARILLOS, and PFV JAVIERA — the marginal). Their emission factors are all zero. So the induced emissions from charging in this hour are:

This is not a coincidence of one cherry-picked hour. It is a structural feature of the Chilean grid in 2025: during midday hours, the SEN regularly curtails 5–15% of its solar and wind generation simply because the demand is not enough to dispatch all the solar/wind units, and partly because the transmission system can't deliver it where it is needed. The battery, by soaking up 100 MWh in that window, is partially un-doing that curtailment — taking energy that would otherwise have been thrown away. The "induced" emissions of the charging side are therefore close to zero across the whole year:

TGHGc = 1,405 tCO₂, just 2.1% of TGHGd.

This asymmetry is what makes BESS so effective in this grid — and it is exactly the kind of fact that an hour-by-hour, unit-by-unit methodology can capture, but a single global emission factor cannot.

Why does charging displace renewables, not thermal? Because at midday the system is over-generating: there is more renewable supply available than demand, and the operator has to throttle renewables to keep the grid balanced. When the battery adds 100 MWh of demand, the operator doesn't crank up a gas turbine — it simply lets the renewables run a bit more.

§6 Where the numbers come from

Before reading the equations, it is worth knowing what feeds them. Every input variable in the calculation is monitored at hourly granularity by Chile's Coordinador Eléctrico Nacional (CEN), which is the statutory grid operator and the single source of truth for dispatch, prices, and generation. The methodology consumes eight monitored quantities directly from CEN's public datasets:

• UGEN — hourly generation per unit (audited every five minutes by the CEN's SCADA) · source portal
• SPB — hourly spot price at each injection busbar (the CEN's official marginal-cost publication) · source portal
• VC — audited variable cost per unit and per hour · source portal
• CURT — documented curtailment per unit and per hour · source portal
• CR — hourly capacity availability factor · source portal
• EF — emission factor per unit (audited annually)
• Bc, Bd — battery hourly charge and discharge (metered at the project's substation)

The platform downloads these files monthly through an automated Python pipeline that reproduces a human analyst's steps on the CEN portal, one by one. The pipeline preserves the original files alongside cleaned, standardized tables — every value remains traceable back to its source file, the original-name fragment, and a confidence level on the unit identification. A step-by-step walkthrough of the manual download procedure for each of the five hourly variables is available in this explanatory video.

Across the eight monitored parameters and the year 2025, the final audited tables ("TRF tables") reach 99.7% coverage over all hours, units, and busbars. The remaining 0.3% corresponds to units recently commissioned (with sparse historical data) or to busbars that the CEN itself publishes inconsistently. Sensitivity tests show those gaps do not move RED by more than ±0.2%.

The unit-identity layer — which is the harder problem, because the same physical plant can appear under different names in different CEN publications — uses a four-step matching cascade (exact match → name similarity → fuzzy logic on multiple attributes → human-reviewed AI suggestion). Every cross-reference carries an explicit confidence score, and benchmark sheets maintained by the project's reference electrical engineer were used to validate the matches.

The end result: a tabular dataset where every hour, every unit, and every variable is traceable back to its CEN source, with an audit trail for any imputation or harmonization step. That dataset is what the next section's 26 equations operate on.

§7 The math, in 26 equations

The full chain is 26 equations. They are listed below in compact form. Every variable name is the canonical name used in the MADD methodology document (Section 5.7) and in the Python implementation that produced this page. The equations are not approximations of the methodology — they are the methodology.

Inputs · hourly monitored (every hour, every unit)

The five hourly time-series the platform consumes from the CEN's public datasets. Each value is timestamped, traceable to its source file, and paired with a confidence level on the unit identification.

Hourly monitored inputs · anchor week (CSV)

Inputs · annually monitored (per-unit, fixed within the year)

Static parameters that the platform refreshes once per audited year. They describe each unit's identity, its place in the grid, and its physical limits.

Annual per-unit parameters · 2025 (CSV)

Inputs · supplied by the project's meter (validated against public data)

The battery's own measurements, taken at the project's substation by the metering chain installed for the issuance. These values are independently cross-checked against the CEN's published injection data for the same node.

Pre-computations

The spot price at each unit is just the busbar price where that unit injects:

The economic distance to the spot price:

The energy that could have been dispatched but wasn't:

For the charge side, the eligible generation is the unit's existing output if the price test holds:

Discharge stack

Sort all units by DIF ascending → DORdk,i (Eq. 13). Compute cumulative dispatchable energy along the sorted stack (AEADd, Eq. 10) and cumulative emissions (ACHGEd, Eq. 11, where each unit contributes CHGEd = AVE × EF, Eq. 12).

Find the index LPFdi of the last unit fully dispatched without exceeding the battery's discharge (Eq. 9). Then:

Charge stack

Symmetric, but sort descending by DIF, use EGEMc instead of AVE, and clip the marginal-unit term at zero (Eq. 17):

Top-level reduction

Sum the hourly results over the year (Eq. 5 and Eq. 15):

The net GHG balance:

The observed system emissions (sum of UGEN × EF across the year) (Eq. 2) plus that balance gives the counterfactual baseline: what emissions would have been without the battery (Eq. 3):

And finally:

For AES Atacama BESS Stand-alone, 2025: RED = 65,743 tCO₂. That single number is the output of running this whole calculation against the 8 760 hours of 2025, with 1 187 units, across 13 technology classes. Every intermediate quantity — every LPFdi, every GHGFdi, every DORdk,i — is preserved as an audit checkpoint.

§8 From one hour to one week

So far we have done the calculation for two specific hours. The methodology repeats exactly this procedure for every one of the week's 168 hours. Most of those hours (140 of them) lie outside the battery's operating windows, so Bdi = Bci = 0 and they contribute nothing. The remaining 28 charge hours and 28 discharge hours produce the picture below.

The negative side of the axis is almost empty. In our week, the battery induced essentially no emissions while charging, because the renewables being un-curtailed all have EF = 0. Each of the 28 evening discharges, by contrast, displaces tens of tonnes of CO₂ — generally a single dirty gas turbine bears the brunt.

Stacking that all up, the weekly RED accumulates almost linearly:

§9 From one week to one year

Repeating the same procedure for every hour of 2025 — all 8 760 of them — gives the year's RED of 65,743 tCO₂. The seasonal pattern is mild but visible: winter months (June–August in the southern hemisphere) show slightly higher mitigation because the gas turbines run harder, summer months slightly less because solar curtailment is lower.

An interesting derived view: which technologies bore the brunt of the avoided emissions? Or equivalently — which kinds of generators did this battery most often displace? In 2025:

Open-cycle gas turbines supplied the most marginal hours, with combined-cycle gas units and diesel engines making up most of the rest. Coal-fired steam units appear comparatively less often as the marginal unit because in 2025 they are running mid-merit — cheaper than gas but with positive output even at low spot prices.

§10 Why this matters — vs. using one number for the whole grid

The simplest possible methodology would be: take the grid's annual-average emission factor (about 0.32 tCO₂/MWh in 2025), multiply by the battery's net energy throughput, declare that the RED. Why isn't that enough?

Because the units that the battery actually displaces are nothing like the average. Discharge happens in the late evening, when the grid's marginal unit is essentially always a fossil-fired plant. Charge happens at midday, when the marginal unit is essentially always a renewable. The hour-by-hour marginal emission factor swings between near-zero and near-1.0 tCO₂/MWh:

If we had used the system-average EF for both charge and discharge, the answer would have under-estimated RED by ~40%. The unit-level methodology captures the project's real grid-shaping effect, not an averaged abstraction of it.

Compared to the CDM Tool 007

The closest standardized methodology in the carbon-credit ecosystem is the CDM's Tool to calculate the emission factor for an electricity system (commonly called Tool 007). Tool 007 produces a single annual emission factor for an entire grid, blending a build-margin and an operating-margin component. For the Chilean SEN in 2025, Tool 007 would yield an annual average somewhere near 0.4 tCO₂/MWh — within the same order of magnitude as the unit-level marginal EF we observe in this methodology, on average. The two approaches converge in the annual mean.

They diverge sharply in the hourly dimension. Tool 007 cannot distinguish between a 2 a.m. discharge displacing a gas turbine and a midday charge absorbing curtailed solar. The unit-level methodology documented here can — and that distinction is precisely what makes a BESS-specific RED defensible. For projects where the battery's value proposition is grid-shaping (every BESS), Tool 007 mis-credits the project's actual contribution. For projects where the operation is flat (a baseload generator), the two methods would agree.

One-to-one with the Python implementation

Each equation in §7 corresponds to exactly one function in the project's open Python module shared/madd_engine/equations.py — same name, same signature, same intermediate variables. The 26 equations of MADD §5.7 are implemented in 26 named functions:

The module is pure math: no I/O, no database calls, no global state. It consumes pandas DataFrames as inputs and returns pandas DataFrames as outputs. Every intermediate quantity — every LPFdi, every CHGEdk,i — is preserved as a row in those DataFrames. An auditor can re-run any equation in isolation against the project's CSV exports and verify that the function output matches the documented value, without needing to read or run Python beyond the call site.

All hourly intermediate tables for the anchor week of this page are downloadable via the buttons on each chart card. The same tables are available for any week or month of any audited year.

Conservativeness baked into every choice

At every place where the methodology has discretion, the choice deliberately reduces the project's credit. Some examples:

• The dispatchability flag DPP classifies geothermal as non-dispatchable, even though it is thermally generated, because its output cannot in practice be ramped down to absorb the battery's discharge. The effect: lower TGHGd.
• The maximum-capacity parameter MAXC is estimated from the historical maximum of UGEN, not from the authorized nameplate. The effect: lower available headroom, lower TGHGd.
• Tie-breakers in the discharge merit order favor the cleaner unit first; in the charge merit order, the dirtier unit first. The combined effect: both reduce RED.
• When a unit's published variable cost is temporarily unavailable for a given hour, the engine applies a deterministic three-step fallback — unit historical median, then technology-class minimum, then system-wide minimum. Each step further into the fallback tends to under-state the unit's true cost and therefore its likelihood of being on the margin. The fallback is flagged in the dataset so an operator can review it; the effect on RED is, again, always negative.

None of these choices are visible in the final number, but each of them is one reason an auditor can sign off the project's claim. The complete list of decisions, each with its per-decision sensitivity to RED, is maintained internally as part of the calculation engine's audit trail and is available on request.

§11 Bottom line

A 100-MW / 400-MWh battery in northern Chile, in 2025, avoided 65,743 tonnes of CO₂ — equivalent to taking about 14,292 U.S. passenger cars off the road for a year, by the EPA's accounting. That number is built from 8,760 hourly calculations, each of which identifies the specific power plants on both sides of the trade.

The methodology is auditable end-to-end: every variable feeding every equation is either a directly monitored measurement (with a metering chain that the auditor can validate) or a value taken from a public audited dataset (the Coordinador's published technical and economic parameters). The Python implementation that produced the numbers on this page is a one-to-one mapping of the 26 equations in the methodology document, with every intermediate dataframe preserved as an inspection checkpoint. The numbers shown here are the same numbers the auditor will see when verifying the project's emission reduction claim.

For deeper questions about specific units, decision boundaries, or the relationship between this methodology and CDM AM0026, see the formal methodology document MADD Sherpas BESS GFM Programme (Section 5, pages 39–66) and the associated data architecture report.