EnergyMap Modeler technical masterclass: from Data Shop evidence to engineering results

This is where a power-system tutorial should start: with the result we must be able to defend. In this 168-hour zonal demonstration, Modeler reports 26,573.4 gigawatt-hours served, ₹106.2 billion of modeled system cost, 10.7 million tonnes of carbon dioxide, zero modeled curtailment and unserved energy, and an average modeled price of ₹3,996.1 per megawatt-hour.

Those values are not the conclusion. They are the top of an evidence stack. The carbon and energy totals imply roughly 403 grams of CO2 per kilowatt-hour. The price is about ₹4.00 per kilowatt-hour before network charges, losses, taxes or a procurement contract. Zero unserved energy in one week does not prove adequacy under another weather year, outage or load shape.

By the end, we will build the data contract, preserve the counterfactual, validate the energy balance, read hourly dispatch, test a 30 megawatt four-hour battery, and examine a 150 megawatt Dholera data-centre case. The point is not to produce one impressive number. The point is to know exactly why a number moved and what evidence could reverse the decision.

Before touching the model, define the counterfactual. A good question changes one controlled input and names the metrics that could change the decision. For a BESS screen, that might be: what changes when a 30 megawatt, four-hour battery is placed at one candidate bus? For a data centre: what changes when 150 megawatts of constant grid draw is added near Dholera, with and without storage?

Write the non-goals beside the question. This is a nodal economic-dispatch screen over the displayed network and horizon. Unless the run contract explicitly says otherwise, it is not CERC SCED, AC load flow, N-1 security proof, a retail tariff, a price forecast or utility approval.

Choose metrics before solving: energy balance, unserved energy, production cost, local marginal-price proxy, curtailment, total and marginal emissions, storage state of charge, and loading on the affected corridors. Pre-committing to the scorecard prevents cherry-picking after the solve.

Open Atlas Data Shop and read the card as a data contract. Do not select by title alone. We need the dataset identifier, source, unit, lineage, geography, time resolution, coverage and supported destination. Structural datasets can create a model. A time series or calibration signal normally attaches to an existing compatible model.

For the solved demand example, select Gujarat, 1 April 2025 through 1 April 2026, hourly resolution and a new-model target. The receipt must preserve that normalized scope. A complete non-leap Indian financial year has 8,760 hourly timestamps. We expect monotonic time, no duplicate timestamps, no unexplained gaps, no negative demand and one declared timezone.

Energy is the first quantitative cross-check. Sum hourly megawatts times one hour and divide by one thousand to obtain gigawatt-hours. Reconcile the materialized total against the source aggregate within 0.1 percent or explain the transformation. Then inspect peak, mean, load factor, p95 and the largest ramps. A plausible annual total can still hide a shifted peak or duplicate hour.

Now attach the IEX-derived price band. Its unit is a zero-to-one index. Verify every value remains inside that interval after joining it to the demand clock. This is relative calibration evidence, not raw rupees per megawatt-hour, not a tariff and not a nodal market price. Save the overlap count, excluded hours and both ingest receipts with the study manifest.

Preserve the base before changing anything. Record the network, scenario, formulation, solver, horizon, snapshot stride and weights. Lint the model, resolve errors and document accepted warnings. A completed solver job proves execution; it does not prove the data or formulation answer the project question.

Validate the physical balance before reading economics. At every snapshot, generation plus imports plus storage discharge must equal demand plus exports plus storage charge plus losses, within the recorded numerical tolerance. When snapshots are weighted, multiply power by duration before aggregating energy. Megawatts summed without hours are not megawatt-hours.

If unserved energy is non-zero, find the hour, bus and constraint. If cost is unexpectedly low, check whether a shortfall slack or missing constraint created the saving. If a scenario has a different number of snapshots from the base, stop. A comparison with different horizons is not a counterfactual.

Read results in a fixed order: solver and balance, reliability, cost and prices, emissions and curtailment, then asset-level constraints. Report absolute base and scenario values before percentages. A percentage is unstable when the base is near zero.

For cost, identify which generators, starts, fuels or scarcity hours changed. The system objective is not the same thing as a customer bill or project NPV. For price, distinguish a modeled nodal marginal value from observed IEX products, balancing settlement and a delivered tariff.

For carbon, show both total tonnes and intensity. Then explain the marginal mechanism: did the new load call coal, gas, hydro, imports or previously curtailed renewable generation? Annual renewable share cannot answer which generator served the critical hour.

Finally, connect every headline to an asset. Open the map or result table, name the bus or corridor and show the loading change. System cost can fall while a local price or line loading rises. That is not a contradiction; it can be the distributional effect of congestion.

Configure 30 megawatts, four hours and one site. That is 120 megawatt-hours of nameplate energy. With 92 percent charging and 92 percent discharging efficiency, round-trip efficiency is 84.64 percent. If charging energy costs ₹3,000 per megawatt-hour, the loss-only discharge hurdle is about ₹3,544. The ₹544 spread still excludes degradation, auxiliary load, fees and network charges.

Inspect charge, discharge and state of charge on the same clock as demand and price. State of charge must remain between zero and 120 megawatt-hours, power within 30 megawatts, and a cyclic model must finish where it started. Otherwise the horizon can create value by consuming stored energy it never replaces.

The recorded Dholera benchmark is deliberately uncomfortable. Storage reduced unserved energy from 3.58 to 2.94 gigawatt-hours, about 18 percent, and lowered modeled operating cost by about 3.6 percent, roughly ₹19 lakh over the week. But marginal emissions increased from 563 to 607 grams per kilowatt-hour and local corridor peak loading rose from 22.5 to 24.4 percent.

That mixed result is the analytical payoff. The battery improved reliability and cost but was not automatically cleaner and did not reduce every network stress. A project decision must state which objective matters, then test power, duration, efficiency, weather, price and constraint sensitivities before connecting dispatch value to capex or financing.

First define 150 megawatts. Is it IT load, facility grid draw, sanctioned demand or contracted supply? If it is facility draw and PUE is 1.30, IT load is about 115.4 megawatts and facility overhead about 34.6. A data-centre interconnection and clean-energy plan must use the power the grid actually sees.

A flat 150 megawatts over 168 hours consumes 25.20 gigawatt-hours. At 100 percent annual load factor it becomes 1.314 terawatt-hours. A reported 51 percent annual green share is 670.14 gigawatt-hours, but annual matching is not 24-by-7 carbon-free energy and does not prove local deliverability.

Load the named Gujarat detailed network and verify the Dholera EHV yard and surrounding corridors. The workflow solves the untouched base, clones it, adds one named 150 megawatt load, solves the impact case and persists the pair. The recorded native-hourly no-battery screen produced about ₹2,148 per megawatt-hour of incremental modeled operating cost, 3.58 gigawatt-hours of unserved energy, 563 grams per kilowatt-hour of marginal emissions and 22.5 percent peak loading on the local corridor.

Run the battery case from the same untouched base. Then align facility load, unserved energy, local price, marginal emissions, corridor loading, battery power and state of charge on one clock. The important question is whether the worst reliability, price, carbon and loading hours coincide—or whether improving one objective worsens another.

Do not collapse power cost into one unsupported number. Keep eight ledgers: compute-to-grid volume, bulk energy, balancing, green supply, network delivery, flexibility or storage, carbon, and policy plus project-capital gaps.

For market exposure, assign explicit shares to PPA or RTC supply, DAM, RTM and imbalance. Stress five and ten percent imbalance volumes with median and p95 spreads. Do not multiply all annual energy by a p95 price. Add losses, transmission, distribution, taxes, green premium, bay and reinforcement as separate sourced or TBD inputs.

Policy analysis uses the same discipline. A statewide 7.5 gigawatt ambition is not a single-bus scenario. A 51 percent annual green requirement is not hourly clean power. A support envelope is not a project entitlement until the applicable notified terms and eligibility are verified.

An AI agent should begin read-only. Ask it to restate the active project, model, scenario, component counts, modeled horizon and latest result. If it cannot identify context, do not grant mutation or compute authority.

A mutation prompt must name the target, value, unit, scenario, action, comparison and required evidence. Review the typed action chip and read the effective value back in Tables. For a workflow, verify its arguments, result IDs, solver settings and horizon rather than trusting the prose summary.

Then challenge the conclusion. Ask which single plausible input is most likely to reverse it, define a low/base/high range and specify the smallest run that could falsify the recommendation. Confidence should increase because the result survived a targeted test, not because the agent sounded certain.

The handoff is compact but complete: project, model, scenario, source and Data Shop receipts, result IDs, timestamps, solver and formulation, modeled horizon and weights, assumptions, absolute values, deltas, benchmark calculations, limitations and the next falsification test.

Electrical engineers should be able to find the constrained asset. Environmental engineers should be able to trace marginal carbon. Energy economists should see the counterfactual and excluded cash flows. Policy makers should see which result supports which claim. Traders should see whether a value is observed, derived, forecast or modeled.

That is the standard for Modeler: not one magic answer, but an evidence chain that can be rerun when the data, site, asset, policy or commercial assumption changes.