Before you own the compute, model the power.
MicrogridModeler.com is live—and it gives remote-power teams a serious, browser-based way to size the system, inspect every hour, and ask what fails before hardware reaches the field.
- MicrogridModeler is live in the browser with no install or account required for the interactive modeler.
- It models PV, wind, batteries, diesel generators, converters, and optional grid service through chronological 8,760-hour dispatch.
- Its sizing search enforces supported constraints, compares dispatch controls, and exposes hourly operations, lifecycle economics, and resilience instead of returning one unexplained size.
- For remote AI, the useful question changes from “How big is the battery?” to “Which load is critical, what can fail, and how does the service recover?”
A resilient AI node is not a GPU attached to a battery. It is a power system with a compute workload.
That distinction matters. The electrical design has to survive nights, weather, generator starts, battery limits, maintenance, fuel constraints, and the occasional component that simply does not come back online. A nameplate spreadsheet can make the hardware add up. It cannot show whether the system survives Tuesday at 4:00 a.m. in the worst week of the year.
MicrogridModeler is the most interesting new tool we have seen for closing that gap early. It turns a system concept into a chronological, inspectable planning model without making the operator install a desktop application or hide assumptions inside a consultant's workbook.
The off-grid design loop, in one place
The product starts with the things that actually define a remote system: site, load, solar or wind resource, equipment limits, fuel, and operating policy. You can begin with one of 20 public reference projects or bring an 8,760-hour load CSV. The engine then runs the energy balance hour by hour and re-dispatches the system across the project life.
That chronology is the point. Annual energy totals can hide the sequence that kills an off-grid design: three poor solar days, a high overnight load, a battery already near its minimum state of charge, and a generator that cannot accept—or efficiently serve—the resulting duty cycle. MicrogridModeler keeps the sequence intact.
It also keeps the controls visible. Users can compare forecast-aware load following, cycle charging, diesel-first operation, and a maximum-state-of-charge strategy. Generator starts, low-load operation, wet-stacking hours, curtailment, battery state of charge, unmet load, and fuel use remain reviewable in the result.
For AI infrastructure, load shape is architecture
Remote compute has an unusual advantage: some work can move in time. Batch inference, indexing, model updates, and fine-tuning do not all need to run at the same priority. Cooling and networking add their own profiles. An operator can use those distinctions to shrink the expensive part of the power system—if the load model preserves them.
A practical first pass is to separate the always-on service from shiftable work, export a year of expected node demand, and upload that series. Then test a design with a conservative peak, real conversion losses, realistic battery limits, and enough fuel to cover the outage case. This builds on the same principle in our field note, “A Battery Is Not Redundancy”: the battery carries a transition, but resilience belongs to the complete service.
A design is not resilient because it has storage. It is resilient when the critical service survives a credible failure—and the model shows how.
It models failure, not only the sunny day
This is where MicrogridModeler becomes unusually relevant to Off Grid Labs readers. The tool can test outage survival from the battery state of charge reached in normal chronological dispatch. It reports the worst, average, and best outage-start survival, plus the probability of serving critical load for selected durations.
It can also disturb the equipment itself. Solar, wind, battery strings, and generator units can fail under deterministic or reproducibly seeded stochastic scenarios, remain unavailable for a modeled repair time, and return over an hourly recovery curve. The result includes energy not served, recovery, fleet outcomes, and four resilience measures derived from the NPS Microgrid Planner framework.
That does not make the forecast true. It makes the assumptions inspectable—which is far more useful than calling a system “redundant” because a second box appears on the one-line diagram.
The recommendation comes with receipts
The sizing workflow searches a declared coarse grid of PV, battery, and generator candidates, then refines promising designs jointly across equipment sizes and dispatch policies. Optional robust screening re-runs finalists against high load, renewable drought, and cost or fuel downside cases. Supported feasibility constraints are checked before candidates are ranked by lifecycle net present cost.
The wording matters: this is an exhaustive search of the declared sampled grid, followed by refinement—not a claim to have proven the best point in an infinite continuous design space. Search boundaries, local-gap diagnostics, and Pareto-efficient alternatives stay visible so an engineer can challenge the envelope.
Every run can produce a versioned JSON package with inputs, resource provenance, engine version, outputs, and a content fingerprint. Hourly dispatch and cash flows can be exported as CSV. Identical inputs reproduce identical deterministic results. That gives a remote-power team something much more valuable than a beautiful dashboard: a result another person can review.
MicrogridModeler is planning-grade chronological energy, resilience, and techno-economic analysis. It does not replace protection coordination, transient stability, detailed distribution power flow, site engineering, equipment selection, or controller hardware-in-the-loop certification. Use it to make early decisions sharper—and to make the detailed-engineering handoff better.
A useful first run for a remote AI node
- Build the real load series. Include compute, cooling, networking, storage, controls, and parasitic loads. Preserve workload schedules instead of flattening everything into one average.
- Define the service that cannot stop. Set the critical-load share or upload a dedicated critical-load series, then choose outage durations that match the site's logistics.
- Model an honest resource. Start with the labeled synthetic profile, then replace it with supported PVWatts, NSRDB, Wind Toolkit, or uploaded historical data as the project advances.
- Search a wide enough equipment envelope. Let the optimizer reject candidates that miss supported constraints, then inspect boundary warnings and the trade space—not only the recommended row.
- Run the dark test. Stress equipment failures, repair times, generator delay, and available outage fuel. Export the package and give it to the person who will challenge the assumptions.
Once the design has a credible load and resilience target, bring the resulting architecture back to the Off Grid Labs node configurator. One tool defines the compute you intend to own; the other helps pressure-test the power system that keeps it useful.
FAQ
What does MicrogridModeler model?
It performs planning-grade chronological techno-economic and resilience screening for systems combining solar PV, wind, battery storage, diesel generators, converters, and optional utility-grid import or export.
Can it evaluate an off-grid AI node?
Yes. Upload an 8,760-hour load profile for the compute node, define supported reliability and operating constraints, compare equipment sizes and dispatch policies, and inspect outage and component-failure results. The quality of the answer still depends on the quality of the load and resource assumptions.
Does it replace detailed electrical engineering?
No. It is a planning model for sizing, energy balance, economics, and resilience. Protection, transient, distribution, controls, civil, environmental, and interconnection work remain part of detailed engineering.
Model the power. Then own the compute.
MicrogridModeler is live now. Open a benchmark, change the system, inspect the hard hour, and export the evidence behind the answer.