Skip to content

Energy and utilities

AI for keeping power reliable through the 15-minute market

An energy business breaks in a hundred small places: the one quarter-hour nomination that missed by 4 MW, a gearbox oil residual creeping upward in a 10-minute SCADA feed, the transformer replacement list you have to defend in front of EWRC. That is the span this page covers. Most of that work is data plumbing — timezone stamps, curtailment filtering, meter-to-transformer mapping — with a model at the end of it.

Book a callSee our projects →

30 minutes with the engineer who would do the work — not a salesperson. No obligation, and you keep whatever we work out on the call.

Control-room operator watching grid schematics across a wall of screens

The numbers this sector is run on

1.5–3%
Day-ahead load forecast MAPE, national seriesThe working band at TSO level. It degrades to 3–8% at a primary substation and 10–25% at a feeder or LV transformer, where the law of large numbers stops helping — quote a single number without the level and it means nothing. MAPE is fine on load, which never crosses zero; for price we report MAE and rMAE.
5–10%
Day-ahead PV forecast nRMSE, % of installed capacityOperational grade for plant-level day-ahead solar; published models report 6.3–8.2%. Satellite nowcasting in the 0–2 hour band roughly halves it, which is where the intraday euro now sits.
€149/MWh
Average daily day-ahead peak-to-trough spread, Bulgaria, 2026 YTDIBEX data via Capalo AI: €126 (2023), €183 (2024), €165 (2025), €149/MWh (2026 YTD). The headline spread is narrowing, but its shape is not: the midday discount went €5.56 (2024) → €11.63 (2025) → €23.45/MWh (2026 YTD). Value is migrating from the daily range into the intraday and midday legs, which is a different trade from the one a 2024 battery business case assumed.
≈1 / asset / month
Actionable alerts a condition-monitoring model may spendNot a benchmark — an operating constraint. Above roughly one actionable alert per asset per month, site teams stop reading them, which is how most utility analytics modules end up unused.

What we build

Energy analyst reading price and load charts on two monitors

Forecasting scored in euros, not in R²

Load, wind, solar and price predicted directly on the 15-minute grid rather than resampled from hourly models. ECMWF IFS HRES and ENS with ICON-D2 for the short range, bias-corrected per site, plus satellite nowcasting from MSG and MTG rapid scan inside four hours. Quantiles rather than point forecasts, so the desk can position against the expected sign of the imbalance price.

  • Direct 15-minute MTU models — never an hourly forecast interpolated four ways
  • Walk-forward recalibration and a full year of untouched out-of-sample
  • Price scored in MAE and rMAE against a naive benchmark, never MAPE — with 170 negative-price hours in 2025 and 166 in the first seven months of 2026, the denominator collapses at every near-zero hour and loses meaning once the sign flips. MAPE stays only where it is defensible: strictly positive load series
Field technician inspecting a substation cabinet with a tablet

Asset health with an agreed alert budget

Normal-behaviour models per asset and per component, trained only on healthy periods with curtailment, derating and status codes filtered out. Residuals tracked with an EWMA against a fleet-relative threshold, so one weather event does not fire forty alerts. Transformers scored from DGA, moisture, furan, bushing tan-delta and OLTC counts, with high-energy arcing handled as an immediate action rather than a ranking input.

  • Alert budget fixed up front, with every alert disposition reviewed for the first quarter
  • Confirmed alerts arrive in Maximo or SAP PM as work-order candidates with the residual plot
  • Retraining triggered by CMMS events — a component swap invalidates the model that watched it
Engineer checking a row of electricity meters in a utility cabinet

Meter and network data that closes the balance

AMI voltage and consumption correlated against transformer measurements to recover the true meter-to-transformer mapping and phase, returned to GIS as an auditable change set. VEE estimation replaced with per-meter models conditioned on temperature and day type, so estimates are not wrong on exactly the hot days, holidays and outage days that matter. Everything validated against the per-transformer energy balance.

  • Confidence-scored GIS corrections for review, never a silent overwrite
  • Technical and non-technical losses separated per secondary substation
  • Prosumer export treated as normal operation, not as an anomaly to be flagged
Two engineers reviewing a network schematic in a control room

An architecture your OT team will sign off

Nothing lands on the control network. The pattern is a historian replica in the IDMZ at Purdue level 3.5, fed through a unidirectional gateway or data diode, with every model downstream in the IT zone — under IEC 62443 a zone with a one-way conduit. Recommendations return as advisories inside the existing HMI, never as commands on a control path.

  • Purdue and IEC 62443 zoning scoped before the modelling architecture, not after
  • Supplier security assessment expected under NIS2 and Delegated Regulation (EU) 2024/1366
  • Every decision reconstructable: inputs, model version, recommendation, and what the human did

Worked examples

These are designs, and the ranges we would contract against, drawn from published sector data — not Palamed results. The four systems we have actually delivered are on case studies.

Portfolio forecasting for a balancing group coordinator

Problem

Hundreds of PV plants, wind farms and C&I loads net into one schedule and settle at a single imbalance price per 15-minute ISP — the settlement period has been quarter-hourly here since 1 October 2022. One bad convective afternoon erases a month of retail margin. Since the 15-minute MTU went live on 1 October 2025 the trading grid finally matches the settlement grid, which removes the old hourly-shaping error and exposes every model that was only ever fitted on hourly data.

Approach

Gradient-boosted trees plus a quantile network per asset cluster, fed by ECMWF IFS HRES and ENS members with ICON-D2 for the short range, bias-corrected per site, joined to plant SCADA and AMI history. Predicted directly on the 15-minute grid rather than interpolated from hourly. Backtested walk-forward with recalibration, on the data vintages that existed at gate closure, and scored in €/MWh of settled imbalance rather than RMSE.

What we would target

Most of the gain comes from calling the fifteen worst quarter-hours in a month, not from average accuracy — roughly 0.5% of the 2,880 quarter-hours in a month, and we compute that concentration curve from your own settlement file on the first call, because it decides whether the work is worth doing at all. The by-product matters as much: P10/P90 bands the desk can size a deliberate long or short against, and a risk committee number that survives questioning.

Plant-level PV nowcasting for intraday re-trading

Problem

A 60 MW park bids day-ahead, then a cloud field arrives that no NWP resolved. The deviation shows on SCADA at 11:15, the intraday order book is already thin, and by the time a broker is called the price has moved. That is where the marginal euro now sits: IBEX intraday volume went from 3.32 TWh in 2024 to 5.60 TWh in 2025, up 68.6%.

Approach

Two tiers. An NWP-driven day-ahead model, and a 0–4 hour nowcast built from MSG and MTG rapid-scan cloud motion vectors and on-site irradiance, refreshed every 15 minutes. Clear-sky index modelling through PVLib, with soiling, temperature derating and inverter clipping modelled explicitly rather than absorbed into the residual. Output is pushed into the trading system as a signed delta-to-schedule — a suggested intraday volume, not a chart on a dashboard.

What we would target

Nowcast nRMSE roughly half the day-ahead value in the 0–2 hour band, converting into recovered value on the thirty to sixty days a year when convective cloud drives the largest deviations. We scope it as a target measured against your own day-ahead error on those same days — not as a Palamed result.

PV asset health: string underperformance, soiling and derating

Problem

Bulgaria added roughly 1,350 MW of solar in 2025 to reach about 5,910 MW, most of it young and watched through a portal that reports availability. Availability says nothing about the string that has been producing 4% under its neighbours since March — and with the midday discount widening, energy lost at noon is not worth what energy lost at 19:00 is worth.

Approach

Every string and inverter regressed against modelled expectation from plane-of-array irradiance, module temperature and clear-sky index, with soiling, shading and inverter clipping separated from irradiance instead of collapsed into one underperformance figure. Drift in IV-curve-derived parameters and step changes after maintenance tracked per device. Findings ranked by lost MWh multiplied by the price in the hour it was lost, so cleaning and repair schedules follow money rather than percentages.

What we would target

A work list a two-person O&M team can close in a week, with the euro figure beside each line, and a defensible answer to the question the owner actually asks: is this plant under-producing, or was that simply the weather we had?

Forecasting a newly liberalised household cohort

Problem

Household supply left the public-supplier arrangement from 1 July 2025, and thousands of residential profiles now sit inside your balancing group. They forecast at feeder-level accuracy — the 10–25% band, not the 1.5–3% one — and their interval reads now underpin settlement rather than an invoice.

Approach

Profile-class clustering conditioned on weather and calendar, then hierarchical reconciliation so the cohort forecast sums to the group nomination instead of fighting it. Estimated reads handled explicitly, because a VEE average is wrong on exactly the cold snaps and holidays that drive the cohort. GDPR scope for household interval data is settled before any modelling starts: what is pseudonymised, what stays aggregated, and what never leaves your infrastructure.

What we would target

The cohort’s contribution to settled imbalance stated separately from the C&I and generation legs, so you can price the household book rather than cross-subsidise it out of retail margin.

Battery co-optimisation across day-ahead, intraday and aFRR

Problem

A RESTORE-funded 50 MW / 150 MWh system has to decide every day how much capacity to hold back for aFRR and how much to leave free for arbitrage, while respecting state of charge, the cycle budget and its prequalification obligations. Spreadsheet dispatch leaves value on the table and occasionally breaks the reserve commitment. Since ESO joined PICASSO in February 2025, with cross-border aFRR energy exchange with Greece from March, the activation price the battery is bidding against is set by a European merit order, not a domestic one.

Approach

Stochastic MILP over a 48-hour horizon at 15-minute resolution, taking price quantiles from the forecasting stack as scenarios, with round-trip efficiency, C-rate limits, minimum bid size, the SOC headroom in both directions needed to deliver the contracted aFRR capacity for its full activation duration, the non-delivery penalty, and a cycle-degradation cost term encoded as constraints. Re-solved on every intraday price update; bid files emitted in the exchange format and reserve schedules in the TSO format. Shadow mode against the human trader for 60 days before any automation.

What we would target

Realistic uplift is single-digit to low-double-digit percent over a competent trader working from the same information — and it compounds, because the optimiser never forgets the state-of-charge constraint at 03:00 the way a tired desk does.

Wind drivetrain failure prediction from 10-minute SCADA

Problem

An unplanned gearbox replacement means crane mobilisation, three to six weeks of downtime and a six-figure invoice. The OEM alarm list fires constantly and the site manager has already learned to ignore it.

Approach

Normal-behaviour models per turbine and per component: gearbox oil, bearing and generator winding temperatures regressed on power, rotor and generator speed, nacelle and ambient temperature, trained on a healthy period with curtailment, derating and status-code windows filtered out. That filtering step is where most wind analytics quietly fails. Residuals smoothed with an EWMA and thresholded fleet-relatively, so a farm-wide weather event does not fire forty alerts, then routed into Maximo or SAP PM as work-order candidates with the supporting plot.

What we would target

One or two unplanned replacements per 50-turbine fleet per year converted into planned campaigns with a crane already booked. Retraining is tied to CMMS work-order events, because a gearbox swap invalidates the model that was watching it.

Transformer health index and DGA-driven capex ranking

Problem

Two hundred power transformers, average age past 35 years, and a budget that replaces four a year. The list is ranked by age, which predicts failure poorly, and the regulator wants a justification for every unit entering the RAB.

Approach

Online and offline DGA interpreted with the Duval Triangle and Pentagon under IEEE C57.104-2019 and IEC 60599, combined with moisture and furan for paper degradation, bushing tan-delta trends, OLTC operation counts and motor-current signatures, and thermal loading history from SCADA. Scored into a health index, then multiplied by a criticality score — customers downstream, N-1 exposure, energy not supplied if it fails. Hard rules stay outside the score: acetylene above roughly 1–2 ppm in a sealed unit is an action, not a ranking input.

What we would target

A replacement schedule with an evidence trail EWRC will accept, and usually a materially different top ten from the age-based list — typically two or three mid-life units under thermal stress that nobody had been watching.

LV observability: meter-to-transformer mapping and loss localisation

Problem

The GIS says which secondary substation a customer belongs to, and it is wrong for a meaningful share of connections; phase assignment is largely unknown. That makes loss allocation and hosting-capacity checks unreliable, and it caps everything the ADMS advertises for the low-voltage network.

Approach

Correlate 15-minute AMI voltage and consumption series against transformer-level measurements: voltage-profile correlation identifies the feeding transformer, correlation structure across the three phases identifies the phase. The method needs voltage registers at ≤60-minute granularity on a substantial share of the LV meters and measurement at the secondary substation; where those are absent we start from the MV/LV energy balance and last-gasp and PLC topology signals, and say up front what confidence that supports. Adjacent transformers on the same MV feeder are the known failure mode — we resolve those with energy-balance closure, not correlation alone, and flag them as unresolved rather than guessing. Reconcile against ArcGIS Utility Network or Smallworld and return a confidence-scored change set for review rather than silently overwriting records, then close the energy balance per transformer to separate technical from non-technical losses.

What we would target

A corrected connectivity model at a fraction of the cost of truck rolls, per-transformer loss figures the regulator can be shown, and an inspection list ranked by expected recovered kWh times probability — so the crew visits the big meters, not the most suspicious small ones.

Systems we work with

We integrate with what you already run. If a platform below is missing, tell us — the pattern usually transfers.

Control room, network model and GIS

  • Siemens Spectrum Power and Gridscale X
  • GE Vernova GridOS ADMS and e-terra
  • Hitachi Energy Network Manager
  • Schneider Electric EcoStruxure ADMS
  • Oracle Utilities Network Management System
  • Survalent SurvalentONE
  • Esri ArcGIS Utility Network
  • GE Smallworld and Schneider ArcFM

Metering, billing and customer systems

  • Oracle Utilities Meter Data Management
  • Itron IEE and Temetra
  • Siemens EnergyIP
  • Landis+Gyr Gridstream Connect
  • Iskraemeco and Kamstrup head-ends
  • SAP IS-U and S/4HANA Utilities
  • DLMS/COSEM and OBIS register mapping

Historians, asset performance and O&M

  • AVEVA PI SystemData Archive, Asset Framework, PI Vision
  • GE Proficy Historian and AspenTech InfoPlus.21
  • IBM Maximo and SAP PM
  • Hitachi Energy Lumada APM and Copperleaf C55
  • Power Factors Drive
  • Fluence Nispera
  • ONYX InSight
  • IEC 61850
  • IEC 60870-5-104
  • DNP3
  • ICCP/TASE.2
  • OPC UA

Markets, trading and weather

  • IBEX DAM
  • IDM
  • CMBC and the GO trading system
  • ESO EADschedule submission, nomination deadlines and the monthly settlement file layout a balancing group coordinator reconciles against
  • The market clock: SDAC day-ahead gate closure D-1 12:00 CET with results around 12:45, continuous intraday until minutes before delivery, settlement on the 15-minute ISPthe models here are built around those times, not around a nightly batch window
  • ENTSO-E Transparency Platform (treated as a research source, not a system of record)
  • ION Allegro
  • ION Openlink Endur
  • Trayport Joule
  • ECMWF IFS HRES and ENS
  • ECMWF AIFS
  • DWD ICON-EU and ICON-D2
  • EUMETSAT MSG and MTG rapid scan
  • CAMS aerosol

What we design against

10–30%
Lower imbalance cost per MWh of throughputThe published range for portfolio forecasting improvements at a balancing group coordinator, expressed as total settled imbalance cost divided by scheduled MWh. We scope it as a target, measured in shadow mode against your current forecaster on the same data vintages — not as a Palamed result. We also report imbalance volume as a share of scheduled volume, because the cost number can move on price alone.
About half
Nowcast PV error against the day-ahead model, 0–2 hours outSatellite nowcasting from MSG and MTG rapid scan roughly halves day-ahead nRMSE inside the 0–2 hour band in published work. The value shows up on the thirty to sixty days a year when convective cloud drives the largest deviations — which is where the marginal euro sits now that IBEX intraday volume has risen 68.6% in a year. A target we design against, measured on your own days.
25–40%
Confirmed-theft hit rate per inspection visitField programmes move from roughly one confirmed case in ten on rule-based lists into this band. The 95%+ accuracy numbers in the literature come from rebalanced academic datasets, not from truck rolls.
12–37 days
Warning before a drivetrain failurePublished work spans roughly 12–37 days on drivetrain components from 10-minute SCADA alone; a five-turbine bi-LSTM normal-behaviour study reports 32–37 days on gearbox, generator and generator-bearing faults, and about three months on a high-speed shaft when vibration is fused in. The target we design the alert budget around.

Regulation and standards in scope

  • Commission Regulation (EU) 2017/2195 (EBGL) — standard balancing products, merit-order activation and the European platforms (PICASSO for aFRR, MARI for mFRR, IGCC for netting) — with Commission Regulation (EU) 2015/1222 (CACM) as the basis for the 15-minute Market Time Unit that went live across SDAC, and on IBEX, for delivery day 1 October 2025. Bulgaria had settled on a 15-minute ISP since 1 October 2022, so the MTU change closed a three-year mismatch: every hourly-shaped bid template, block product and internal report keyed on hour-ending had to be rebuilt, and hourly models resampled instead of retrained now carry a hidden error.
  • Commission Regulation (EU) No 543/2013 on submission and publication of data in electricity markets — Bulgaria has been assessed as having material gaps here, including late publication of imbalance prices. That sets the cadence of any shadow-mode scoring: we score against the ESO settlement file when it lands, not against a Transparency Platform pull, and we say up front what that does to the evaluation window.
  • REMIT, Regulation (EU) 1227/2011 as amended by (EU) 2024/1106 — any component that generates or submits bids is an algorithmic trading system: inside-information disclosure via UMM, the manipulation prohibition, registration of persons professionally arranging transactions, and order and trade record-keeping. We build the audit trail to that standard and the desk retains submission authority.
  • Commission Delegated Regulation (EU) 2024/1366 (Network Code on Cybersecurity), applying since 2 July 2025, stacked on Directive (EU) 2022/2555 (NIS2). Both reach suppliers: expect a supplier security assessment during onboarding and budget calendar time for it rather than discovering it at contract stage.
  • Regulation (EU) 2024/1689 (AI Act) — Annex III point 2 classifies AI used in the management and operation of electricity supply as high-risk, with duties on risk management, data governance, logging, human oversight and technical documentation. Decision support that leaves the operator in control is a different posture from closed-loop control, and the distinction is documented on day one.
  • The Bulgarian Energy Act, the Electricity Trading Rules and ПУЕЕС (the transmission system operation rules), with Правила за управление на електроразпределителните мрежи where a DSO is in scope — balancing group coordinator licensing and the settlement mechanics ESO applies on a 15-minute ISP with a single imbalance price (deficit price = surplus price), in force since the move away from the strict two-price system. EWRC/KEVR decisions have historically set the upward-regulation marginal price at the hourly IBEX day-ahead price plus a fixed adder of BGN 100/MWh (≈ €51.13/MWh at the fixed 1.95583 rate) and floored downward regulation at zero. Single pricing plus that asymmetry is what makes a deliberate, sized position defensible rather than reckless — and it moves a BRP’s economics more than most modelling choices.
  • Amendments to the Electricity Trading Rules effective 1 July 2025 removed the public-supplier arrangement for households. Thousands of residential profiles moved inside balancing groups, and household interval reads became a settlement input rather than a billing one. Regulation (EU) 2023/2854 (Data Act, applicable since 12 September 2025) and the GDPR then govern how those reads may be shared and modelled — household consumption at interval resolution is personal data.
  • IEC 62443, ISO/IEC 27001 and ISO 55000/55001 — the frameworks a utility procurement team will ask about — alongside euro adoption on 1 January 2026 at the fixed 1.95583 BGN/EUR, which every historical price series, tariff and settlement record has to carry as documented lineage or backtests break silently.

Most of it is data plumbing, with a model at the end of it.

What you are probably thinking

Our data is not clean enough for this.

Correct, and that is the project. Expect roughly the first third of the engagement on tag dictionaries, timezone normalisation and the duplicated hour in late October, curtailment and status-code filtering, and reconciling GIS against meter reality. Anyone who tells you their model runs out of the box on your historian has not seen a real utility tag namespace.

We are not putting anything on the OT network.

Nothing needs to be. A historian replica sits in the IDMZ at Purdue level 3.5, fed through a data diode or unidirectional gateway, and every model runs downstream in the IT zone — under IEC 62443 that is a zone with a one-way conduit. The model gets a copy, never a control path. Where a recommendation has to reach an operator, it arrives as an advisory in the HMI they already use.

The last vendor’s analytics module sits unused. It fired 200 alerts a week.

Alarm fatigue is the default failure mode of condition monitoring, not an unlucky outcome. We threshold on fleet-relative residuals instead of absolute limits, suppress correlated fleet-wide events, and agree an alert budget before we build anything — roughly one actionable alert per asset per month — with every alert disposition reviewed together for the first quarter.

How do we know this beats the forecaster we already pay for?

Shadow mode. The model runs alongside your incumbent on live data for a defined window, scored in €/MWh of imbalance cost with a Diebold-Mariano test on the error series. If it does not beat persistence and the incumbent by a statistically significant margin, that is a finding and you keep your supplier. When a backtest looks too good, the cause is usually known: revised ENTSO-E values that did not exist at gate closure, an actuals column leaking into features, or capacity that changed mid-series.

Bulgaria is too small and too odd for tools calibrated somewhere else.

It is odd, and that is the argument for local calibration rather than against the work. The EWRC upward-regulation adder over the day-ahead price, downward regulation floored at zero, cold reserve as a separate product, prequalification that has favoured dispatchable generation, 3.4 GW and roughly 8.6 GWh of grid batteries reported by ESO in May 2026 — about two and a half hours of duration — against 5.9 GW of solar, with 9.7 GWh of RESTORE awards still working through commissioning, and a currency changeover that breaks every historical series — a model fitted on German data will be confidently wrong here.

Who carries it when the model is wrong at the intraday gate?

The trader, always. The model produces a distribution and a suggested position; the desk holds the mandate and the submission. Every decision is reconstructable — inputs, model version, recommendation, and what the human did — which is also what REMIT record-keeping requires.

We are regulated. How do we recover this cost?

Frame it against the KPIs EWRC already prices: SAIDI, SAIFI, energy not supplied, the technological-cost allowance, and capex deferred by health-index-based replacement. A ranked replacement list with an evidence trail is easier to defend in a price review than an age-based one.

When we are the wrong choice

  • Closed-loop control. We do not write protection settings, automatic switching logic or anything that can trip a breaker — that belongs to your protection engineers and a type-tested vendor. Our output advises; your procedure decides.
  • Sites without history. Under roughly two years of usable 10-minute SCADA or AMI intervals, a normal-behaviour model has nothing to learn a healthy baseline from, and you are better served by physics-based checks and better instrumentation first.
  • Replacing your ADMS or MDM. Those are two-to-four-year programmes whose critical path is network model preparation and GIS, CIS and work-management interfaces, not analytics. We work alongside that prime contractor, never instead of one.

Questions we get asked

We already buy a forecast from a vendor. What would you add?

Often the vendor forecast is fine and the loss is downstream — in how it is nominated, in an hourly shape still being applied to a 15-minute settlement grid, or in a schedule nobody adjusts intraday. We measure your settled imbalance against a reconstruction of what the same information could have produced, and if the gap is small we say so. That answer is cheaper for you than a modelling project.

Our SCADA and historian are isolated. How would you get data at all?

Through a replica, not a tunnel. A copy of the historian lands in the IDMZ across a unidirectional gateway, exports run on a schedule your OT team controls, and no Palamed component ever holds a route into the control network. For a proof of value, a one-off export of two to three years of 10-minute SCADA and AMI intervals is usually enough to start.

Everything we have is in BGN up to 1 January 2026. Does that break our history?

It breaks it silently if nobody handles it, which is the dangerous case. Every price series, tariff, contract and settlement record needs an explicit conversion at the fixed 1.95583 BGN/EUR with the conversion recorded as lineage, not applied in place. Mixed-currency series are a common source of a backtest that looks better than the market ever was.

We coordinate a balancing group. Can you tell us which member caused the imbalance?

Yes, and it is usually the highest-return work on this page. Per-member, per-quarter-hour deviation is reconstructed from metered values against nominations, then the group’s settled imbalance is allocated by marginal contribution — a member deviating against the group’s net position reduces cost and should see that in the invoice. In the published accounts of this method, groups improve within two settlement cycles — before any model gets better.

Is this a high-risk AI system under the AI Act?

If it is used in the management or operation of electricity supply, Annex III point 2 says yes. In practice the posture matters: decision support that leaves a dispatcher or trader in control carries different obligations from closed-loop control, and we write that distinction into the design documentation at the start rather than arguing it during an audit.

What do we own at handover, and can our own people retrain it?

The pipeline, the feature definitions, the retraining runbook and the evaluation harness, in your repository and running on your infrastructure. Handover includes training a named person on your team to retrain and re-score the model. If nobody on your side can do that at the end, the engagement failed regardless of what the metrics said.

Keep reading

AI advisory and readinessBusiness efficiency automationHow we work: stages, durations, kill criteria
Warm light ribbons on a dark field

Book a 30-minute forecast and imbalance review

Bring one month of settled imbalance and your current forecaster’s output. On the call we will tell you whether the cost sits in the weather, the model, or the schedule you nominated — and which of the three is worth paying to fix. If it is worth going further, the next step is a fixed-fee two-week diagnostic, not a proposal.

Book a callSee our projects →

You talk to the engineer who would do the work — and if your incumbent forecast is already close to the achievable limit, we will say that too. There is nothing here to sell you.