Power economics: anticipating the peak and building the VPP

≈ 4 min read · updated Aug 2026

The entire power system is sized for a single hour: the most demanding one of the year. Anticipating that peak with machine learning, and deciding in time what to shift or curtail, pays off in infrastructure, contracts and investment.

Power economics · machine learning
The system is sized for the peak, and that peak can be anticipated
How to read it: above, a day in the system, sized for its highest hour; below, the model that anticipates that peak days in advance
00:00 06:00 12:00 18:00 24:00 MW system capacity the evening peak a shrinking margin The proposed model learns from history and calls the peak days in advance −30 d −15 d today +7 d +14 d peak daily alert: high peak likely recommendation: shift flexible load
Day's demand (load curve) System capacity Peak forecast Observed daily peaks Model error band
  • Deep learning · time series
  • Gradient boosting
  • Anomaly detection
  • Predictive maintenance
  • Multi-agent RL · lowest-cost dispatch
In simple terms: the peak follows patterns of hour, temperature and calendar, and that is exactly what a model can learn. Anticipating it days in advance means deciding with time to spare instead of reacting on the fly. This is a diagram of the method: real models are trained on each system's own operational data.

Energy and power are not the same thing. Energy is what you consume over time, the kWh on the bill; power is how much you need at the same moment, and the entire system is sized for that moment: the peak.

Anticipating the peak, deciding in time

In Paraguay’s hydroelectric system, operated by ANDE, machine learning earns its keep on four fronts:

  • Peak demand: time-series models by hour and by area.
  • Hydrology and generation: how much water will arrive and how much energy there will be.
  • Losses: anomaly models that separate technical losses from non-technical ones.
  • Predictive maintenance: the signals equipment sends before it fails.

Each forecast drives an hourly decision. In the design we propose, the model ranks the options by cost and the decision stays with the operator:

  1. Shift your own flexible load: cooling, pumping, furnaces that can wait.
  2. Curtail what your own operation can spare while the peak lasts.
  3. Dispatch backup or pay for expensive capacity, only as a last resort.

Virtual power plants (VPPs): responding to the peak

Virtual power plant · VPP
Many small resources, one plant at the moment of peak
How to read it: on the left, the resources coordinated by software · on the right, what they do to the curve
rooftop solar batteries flexible loads generators forecast + decision one plant 12:00 19:00 · peak 24:00 the peak, trimmed result: a lower peak, without adding new generation
Peak without a VPP Peak with a VPP What the VPP avoids
  • Conceptual diagram · no data
  • Forecast + cost-order dispatch
  • No regulatory framework in Paraguay
In simple terms: the software dispatches resources that already exist in cost order, following the forecast. This is a conceptual diagram.

What you can build today, without waiting for the regulatory framework that is still missing, is the brain: forecasting and dispatch over your own load pay off on their own.

The data: your hourly load curve, the tariff schedule and each process’s flexibility with its opportunity cost.

The technique: deep learning forecasts the peak and multi-agent reinforcement learning learns, first in simulation, when each asset should shift or dispatch.

The decision: an hourly dispatch plan for the week, with the estimated capacity savings.

In that design, the agents learn to execute the ladder above better; the bar to beat is classic optimization (MILP).

Who this matters for and where to start

The structural backdrop is the end of the power surplus: while power is abundant, getting the peak wrong is cheap; once the surplus runs out, that same mistake shows up in contracts and in construction.

That end has an estimated date, and it is close. The capacity available to Paraguay across Itaipú, Yacyretá and Acaray is around 8,670 MW (ABC Color, 2025), the system’s peak already reached 5,752 MW in January 2026 (ABC Color, 2026) and consumption grew 12.5% in 2025 (ANDE, 2026) and 18.3% in the first half of 2026 (Agencia IP, 2026). With demand growing at 9% per year, industry groups and analysts place the crossing between 2030 and 2032 (La Nación, 2026).

Projection · end of the surplus
The crossing year: when demand reaches the available capacity
How to read it: the green line is the system's annual peak, the cyan one is today's available capacity · where they cross, the new mix starts closing the gap
2025 2027 2029 2031 2033 MW 10,000 8,000 6,000 solar + batteries small hydro biomass Itaipú + Yacyretá + Acaray: 8,670 MW actual peak: 5,752 MW · Jan 2026 projected at 9% per year the surplus runs out: 2030-2032
Annual peak (observed and projected at 9%) Available capacity: Itaipú + Yacyretá + Acaray Proposed mix: solar + batteries, small hydro, biomass
  • Illustrative projection on cited sources
  • 2030-2032 crossing · ABC Color 2025, La Nación 2026
  • The mix is a study proposal, not an approved plan
In simple terms: the peak already reached 5,752 MW in January 2026 and the available capacity is around 8,670 MW. At 9% per year the crossing lands between 2030 and 2032, per the cited sources. The colored layers show what each source in the mix we propose to study would contribute: the exact share is calculated with the actual load curve.

The gap that opens after the crossing is closed with a mix, and each source contributes something different: solar with batteries generates during the day and shifts stored energy to the evening peak; small hydro plants (PCH) add firm capacity, small and close to consumption, with 22 sites already identified by the Parque Tecnológico Itaipú (La Nación, 2025); and biomass is the only one of the three that can be dispatched at any hour. What share of each one makes sense is exactly the question the model we propose is designed to answer with each zone’s actual load curve.

  • Industrial users: the peak sets your contracted capacity; modeling your own load curve is real money in the supply contract.
  • Generation investors: demand and hydrology forecasts carry the business case (legal framework at BACN).
  • Grid operators and large consumers: forecasting, alerts and maintenance, in the order your own data marks as most expensive to ignore.

Two concrete projects, starting from your data and a diagnosis:

  • Forecasting models across the four fronts above, as a dashboard with alerts.
  • Power economics analysis: where to connect a large load or a project, and at what risk.

How to measure it: a dashboard comparing each week’s forecast peak against the actual one, and what that error costs in contracted capacity.

This study is published to be argued with. If you read the end of the surplus differently, if you have a load curve that is costing you money in capacity charges, or if you are deciding where to connect a large load or a generation project, write to us and we will put it to the test against your actual data.

Assess your connection risk →

Initial diagnosis at no cost · we reply within 24 business hours