CENACE dispatches the Sistema Eléctrico Nacional. CFE Transmisión and CFE Distribución own the wires. The Ley de la Industria Eléctrica, published in the DOF on 11 August 2014, reserves both functions to the state. A plant in Monterrey or Silao cannot buy grid balancing software, because it does not operate a grid.

What it does control is the generation sitting behind its own meter, the charge and discharge schedule of any battery next to it, and the 15 minute interval that sets its billed demand. Those three are where a forecasting model moves a number a finance director recognizes. The rest of the standard artificial intelligence pitch for renewable energy is somebody else's purchase order.

This article sorts the use cases by who in Mexico can buy them, then prices the two a load center owns against the CFE tariff line each one moves.

Who controls what in the Mexican system

Vendor material bundles four capabilities together: demand and generation forecasting, real time grid balancing, storage dispatch optimization, and distribution routing. In Mexico they have different owners.

System level forecasting and dispatch belong to CENACE, which operates the Mercado Eléctrico Mayorista and issues the instructions that decide which plants run. Network operation belongs to CFE Transmisión and CFE Distribución. Routing on the distribution network is theirs alone. A customer's transmission and distribution charges come from the tariff schedules the Comisión Reguladora de Energía approves, not from the path power takes to reach the site, so no routing algorithm lowers an industrial invoice.

Two capabilities are left. Forecasting a single site's own load and its own on-site output runs on private data and needs nobody's permission. Scheduling a battery behind the meter is a private decision bounded only by the interconnection contract. Those two belong in a capital request. The other two do not.

Where a forecast turns into money on a GDMTH bill

A forecast pays only where a tariff prices the difference between one hour and the next. For most industrial sites on CFE basic supply that tariff is GDMTH, and it prices the difference in two places.

The first is energy. GDMTH splits the day into base, intermedia and punta hours, with the ranges set by tariff region, season and day type in the schedules CFE publishes each month. Shifting a kWh out of punta into base is worth the spread between those two energy charges, and no more than that.

The second is demand, and that is usually where the money sits. Maximum demand is the average kilowatts over the 15 minute interval with the highest consumption in the billing period. For GDMTH, DIST and DIT, the capacity charge is assigned to the maximum demand coincident with the punta period. The distribution charge is applied to the lesser of two figures, the maximum demand recorded in the previous year and the maximum demand recorded in the billing month. The CRE set that criterion in acuerdo A/001/2018 of 1 February 2018 and recorded it in its Memoria Documental on final basic supply tariffs of 31 October 2018.

That is what makes the punta window the thing worth controlling. The capacity charge does not follow the plant's highest kilowatt of the month. It follows the highest kilowatt that lands inside punta. Both demand charges are billed on demand measured inside the billing month, and neither carries one bad interval forward on its own. A model that predicts a punta coincident peak twenty minutes ahead, so that load sheds or a battery discharges into it, holds down the kilowatt figure both charges are billed on in the month the peak forms. That is the calculation to put in front of a CFO, and it can be run on the site's own bills before anyone quotes software.

On-site generation is a permit question before it is an algorithm question

The generation has to exist and be lawful before any dispatch model matters. Article 17 of the Ley de la Industria Eléctrica exempts generation below 0.5 MW from requiring a CRE generation permit. It exempts nothing else. The site still needs an interconnection contract with CFE Distribución, municipal construction and land use approvals, and structural and electrical sign off.

The contract also fixes what surplus is worth. CRE's distributed generation rules, issued as RES/142/2017 and published in the DOF on 7 March 2017, provide three schemes. Under medición neta, surplus is credited against later consumption rather than paid in cash, and credit still unused after twelve months is liquidated at the local marginal price. Under facturación neta, injected energy is paid at the node's local marginal price, which sits well below the retail tariff. Under venta total, all output is sold and nothing is self consumed.

This bears directly on the forecasting question. Under medición neta an exported kWh is worth a retail credit only if the site later consumes it, so a model that maximizes export is solving for the wrong quantity. Shaping on-site solar against the plant's own load curve is the quantity worth solving for. The same logic decides whether a company is better served by on-site or offsite renewable supply, which are priced on completely different terms.

The installed base is not trivial. SENER's PRODESEN 2023-2037, published on 29 May 2023, records 2,613 MW of distributed generation in its base year and projects 11,442 MW by 2037. National installed capacity was 89,890 MW in 2022 and net consumption was 333,662 GWh, which SENER's planning scenario grows at 2.5% a year to 479,987 GWh in 2037.

Storage dispatch is the real optimization problem behind the meter

A battery behind an industrial meter has one job in Mexico that pays predictably. It holds the 15 minute peak down. Solar alone cannot do that, because the punta window in most tariff regions runs into the evening, after output has fallen away. Storage moves energy across that gap. It also absorbs the short excursions that large motors and compressors create, and those excursions are what set maximum demand in the first place. The economics of energy storage in Mexico turn on that single billing mechanism more than on any technology claim.

The scheduling problem is genuinely hard, which is why it is the one place a trained model earns a license fee. The battery has to weigh how much charge to hold in reserve for a peak that may never arrive against the value of discharging into the punta energy charge now. Recharging has to happen without creating a new peak of its own. Cycle life is finite, so every avoided kilowatt carries a degradation cost.

One warning about the arbitrage story that usually accompanies this. A CFE basic supply customer does not face an hourly price. It faces three time bands and a demand charge, and that is the whole of it. Hourly local marginal prices reach only a market participant or a usuario calificado buying from a qualified supplier on an indexed contract. Registration as a usuario calificado requires 1 MW of maximum demand at the centro de carga over the previous twelve months. Direct participation in the MEM requires 5 MW plus 20 GWh a year under SENER's Manual de Registro y Acreditación de Participantes del Mercado, published in the DOF on 15 July 2016. Below those thresholds, a pitch built on hourly price arbitrage describes a market the site is not in.

What the software needs before it can do anything

Three data conditions decide whether any of this is buyable this year. The site needs interval metering at the main incomer rather than a monthly reading. It needs twelve clean months of CFE billing history, because the distribution demand charge is bounded by the maximum demand of the previous year. It needs submetering on the largest motors, chillers and compressors, because a peak the model cannot attribute is a peak it cannot pre-empt.

Sites that fail those conditions should spend the first tranche of capital on metering and monitoring rather than on software. A dispatch model trained on monthly totals has nothing to learn from.

What would make this read wrong

Two things. The first is tariff design. Every argument here rests on the CFE charge structure holding, and the capacity and distribution demand rules are administrative acts the CRE can amend. If the CRE changed the demand base for either charge, the case for peak shaving would move the same month.

The second is that power quality can cost more than the peak, and it is cheaper to fix. The CRE reissued the Código de Red as RES/550/2021, published in the DOF on 31 December 2021 and in force from 1 January 2022. Its power factor requirement, a band of 0.95 lagging to 1.0 measured in five minute intervals, is not enforceable until 8 April 2026. The older CFE surcharge bites today. Under the SHCP acuerdo published in the DOF on 31 October 2000, a billing period average power factor below 90% lagging triggers a surcharge applied to the invoice amount, capped at 120%, with the corresponding bonus capped at 2.5%. Capacitors and reactive compensation cost a fraction of a dispatch platform and usually pay back first.

Find out which AI use cases your site can actually buy

Send twelve months of CFE bills, your tariff class, your interval meter data if you have it, and a one line list of your largest motors, chillers and compressors. Mexico Energy Partners will review which line of your invoice a forecasting or dispatch tool would move, and which of the two would pay back first. If neither would move anything, we will tell you that.

Sources

• Congreso de la Unión, Ley de la Industria Eléctrica, DOF 11 August 2014. Article 17 generation permit exemption below 0.5 MW, and the assignment of system operation, transmission and distribution.
• Comisión Reguladora de Energía, RES/142/2017, disposiciones administrativas de carácter general y modelos de contrato para generación distribuida, DOF 7 March 2017.
• Comisión Reguladora de Energía, Tarifas Finales del Suministro Básico, Memoria Documental, 31 October 2018. Capacity charge assigned to demand coincident with the punta period for GDMTH, DIST and DIT. Acuerdo A/001/2018 of 1 February 2018 set the distribution charge on the lesser of the maximum demand of the previous year and the maximum demand of the billing month.
• SENER, Programa de Desarrollo del Sistema Eléctrico Nacional 2023-2037, published 29 May 2023. Installed capacity, net consumption, demand growth scenario and distributed generation projection.
• Comisión Reguladora de Energía, RES/550/2021, Código de Red, DOF 31 December 2021, in force 1 January 2022.
• SHCP, acuerdo que autoriza el ajuste a las tarifas para suministro y venta de energía eléctrica, DOF 31 October 2000. Power factor surcharge and bonus.
• SENER, Manual de Registro y Acreditación de Participantes del Mercado, DOF 15 July 2016. Thresholds for direct participation in the MEM.