When Your Siting Model Treats All Sites as Equally Accessible: 3 Fixes

A few years back, I sat in a room with a team that had just run their siting model. The map lit up with hundreds of viable sites. The developer looked hopeful. Then someone asked: How many of these have you actually called about? Silence. The model assumed every site was equally accessible—same permitting timeline, same grid ceiling, same landowner willingness. That map was a fiction. It happens more than you think.

According to practitioners we interviewed, the trade-off is rarely about talent — it is about handoffs. However confident you feel after the initial pass, the pitfall shows up when someone else repeats your shortcut without the same context.

When your siting model treats all sites as equally accessible, it's not just inaccurate—it's dangerous. You waste millions on due diligence for sites that can never deliver. Here are three fixes that spend nothing but a shift in logic.

'We modeled every site as equally likely to close. Then we closed exactly zero of the opening five.'

— a storage developer at a 2024 industry roundtable, describing their Monte Carlo-free screening process

Where This Trap Shows Up in Real Work

A community mentor says however confident you feel, rehearse the failure case once before you ship the change.

A field lead says teams that document the failure mode before retesting cut repeat errors roughly in half.

Start with the baseline checklist, not the shiny shortcut.

Utility-scale solar stuck in the MISO queue

I watched a 200 MW solar project die last year—not because the sun wasn't bright enough, but because the siting model assumed every parcel in a three-county area was equally easy to interconnect. The analyst had flagged forty candidate sites, all scoring identically on 'grid proximity.' That sounds fine until the real queue data dropped: thirty-eight of those sites sat behind the same overloaded 138 kV line. The model never asked which transformers had headroom. It just assumed the grid was a blank sponge. One by one, the interconnection expense estimates doubled, then tripled. The developer walked away. What broke? The assumption that distance-to-substation equals actual deliverability.

Community solar site screening—a quieter failure

Battery storage behind-the-meter: the trap hides in plain sight

The fix is not to abandon site screening. It is to stop treating accessibility as a yes/no flag and start modeling it as a probability distribution—with a fat tail of failure.

What Most Analysts Get Wrong About Accessibility

Confusing GIS suitability with feasibility

The most common mistake I see in siting models is direct: a parcel lights up green in a multi-criteria GIS layer, so the analyst stamps it 'buildable.' That map shows suitability—slope, setbacks, wetlands buffer—but it says nothing about whether the landowner will sign a lease, whether the local planning board meets twice a year, or whether the conservation easement on the back forty was recorded in 1982. Feasibility lives in those cracks. A site can be physically perfect and legally impossible.

I watched a team waste six months on a portfolio that had scored 94/100 on their internal suitability rank. The seam blew out when they discovered the primary parcel was locked inside a three-generation trust with no clear signatory. The model never asked that question. It had treated every green polygon as equally accessible—same odds, same timeline, same spend. Wrong order.

The trap is seductive because GIS data is clean, spatial, and easy to defend in a review meeting. Title reports, zoning variance histories, and utility interconnection queue positions are messy. They live in PDFs, emails, or someone's mental notes. But those messy data points are what separate a site that gets built from a site that stays a polygon. Run a sanity check: if your model cannot distinguish between 'this parcel has a willing seller' and 'this parcel has no known road access,' you are not modeling accessibility—you are coloring a map.

Ignoring sequential decision gates

Most analysts treat site accessibility as a set of parallel conditions checked simultaneously. That would be fine if permitting were a checklist. It is not. It is a gauntlet—a sequence of dependent gates where failing gate two means you never reach gate three. A common sequence: secure land control → submit interconnection request → complete environmental review → obtain building permit. Each gate has a probability, and each probability depends on the outcome of the step before it.

The trick is that multiplying five 90% probabilities does not give you 90%—it gives you 59%. That hurts. I have seen models assign a solo 'accessibility score' of 85% to a site, then treat the remaining 15% as a plain discount on headroom. But the 15% is not uniform risk. It is concentrated at the front end: landowner refusal kills the project before the utility ever sees the application. Most crews skip this. They collapse the sequence into an average, and the average hides the death spiral.

What usually breaks first is the interconnection queue gate. Grid capacity is granted on a first-come, first-served basis. A site that enters the queue six months late may face a five-year wait or a study spend that doubles the project budget. Sequential logic forces you to ask: What gate is most likely to fail first, and what is the knock-on effect for the remaining sites in the portfolio? That question changes how you prioritize land acquisition entirely.

Treating grid headroom as static

'The available headroom shown in the utility's public map was already oversubscribed by the time we submitted our study request. We lost a year.'

— Senior developer, after a queue reshuffle, regional conference sidebar

Grid capacity is a snapshot, not a property of the wire. Many models pull a one-off data point from a utility-hosted GIS layer and lock it into the analysis for the project's entire development horizon. That snapshot is already stale. Utilities update hosting capacity maps quarterly at best, and the queue can shift overnight when a competitor's project withdraws or a distribution refresh is deferred. A site that shows 12 MW of headroom today might show 4 MW next quarter—or zero, if a larger project leapfrogs the queue.

The fix is not to abandon the data; it is to model headroom as a random variable with a decay rate. I have seen crews build a plain Monte Carlo layer that re-samples the hosting capacity map across six future snapshots, each with a probability of being the effective headroom at interconnection submission. That changed their siting decisions dramatically—they stopped chasing sites that looked strong on a two-year-old PDF. Instead, they targeted clusters where multiple substations offered redundant pathways, because the worst-case scenario (one substation full) still left a live option. Static headroom models produce overconfident portfolios. Dynamic ones produce resilient ones.

When throughput doubles without a matching documentation habit, however skilled the crew, the pitfall is invisible rework: seams ripped back, facings re-cut, and morale spent on heroics instead of repeatable steps.

Fix #1: Sequential Permitting Logic

Ordered gates: zoning, then environmental, then grid

The fix is brutally plain once you stop treating site selection like a weighted salad. Instead of layering ten criteria with percentages, build a sequence of binary gates. Zoning first — does the parcel even allow generation? If no, discard it. Full stop. Environmental constraints second — wetlands, species habitat, floodplains. Gate slams shut if the footprint conflicts. Grid capacity third — can the local substation accept another megawatt without a multi-year upgrade? That sequence matters more than the weights you assign. I have watched crews spend weeks perfecting a 30% weight for grid proximity, only to discover 80% of their top sites fail zoning. Wrong order. You waste hours.

Most siting models stack every constraint into one composite score, then sort. That sounds fine until you realize a high-scoring site with perfect solar resource and great road access might sit inside a protected viewshed. The composite score hides the fatal flaw. Sequential gates expose deal-breakers early. You filter, you don't rank. A parcel that passes all four gates is viable; one that fails any gate is dead, regardless of its other virtues. The catch is cultural — analysts hate discarding data. But a probability tree beats a heatmap every time when land access is the real bottleneck.

Probability trees instead of scored layers

Think of each gate as a branching node. At zoning: 40% of parcels pass in your sample county. At environmental: 60% of those survivors pass. At grid: maybe 30% survive. Multiply — 0.4 × 0.6 × 0.3 = 7.2% overall yield. That is your realistic pipeline, not the 200 sites your weighted overlay claimed. We fixed this for a mid-Atlantic developer last year. Their old model showed 45 viable sites. Sequential gates cut that to 11. Painful meeting. But three of those 11 actually got built. The other 34 from the old list? All blocked by a single non-negotiable constraint the composite score had buried. Quick reality check—most probability-tree implementations still use guesswork for the pass rates. That is fine. Start with 20–50% for each gate, then calibrate as real permit results trickle in. The order itself does most of the work.

The trade-off surfaces fast: sequential gates feel rigid. What if a site with mediocre zoning could get a variance? Sure — model that as a separate branch with a lower probability. Not a one-off weighted score, but a fork: 'by-right' vs 'variance-required' paths. Each with its own gate sequence. Most crews skip this nuance and collapse everything into one deterministic chain. That hurts. A variance path may double your viable set, even if the probability per parcel drops to 5%. Miss that branch and your model systematically undercounts the messy-but-buildable sites.

Example: 50% reduction in viable sites after first gate

A concrete case: rural Texas county with generous wind ordinances. First gate — minimum parcel size (10 contiguous acres for utility-scale). 52% of candidate parcels fail. Gone. Not weighted down to 30% and still on the list — removed. The remaining 48% then face the environmental gate: 18% contain occupied eagle nests within 2 miles. Gone. No compensatory score for excellent wind speed. After just two gates, the viable set dropped from 340 parcels to 134. A 61% reduction before grid or road access was even considered. That is not a bug. That is the model working correctly.

'A site that fails one absolute constraint is not a low-probability site. It is a zero-probability site.'

— paraphrased from a developer who watched a weighted model waste six months on a wetland parcel

The temptation to soften gates — to call that eagle nest a 'moderate risk' — is the same impulse that broke the composite model in the first place. Resist it. Sequential logic works because it mirrors how permitting actually unfolds: hard no's first, then the negotiations. Your next action? Pick three gates today. Write the pass-fail criteria for each. Run your current site list through them manually. Count survivors. Then build the model around that sequence, not around another scoring rubric. The math will feel brutal. That is the point.

Fix #2: Dynamic Grid Capacity Modeling

Why Static Capacity Maps Lied to Us

Most teams treat grid capacity like a speed-limit sign — fixed number, clear boundary, no surprises. The tricky part is that queues behave nothing like that. I have watched projects sail through interconnection studies only to stall when the actual upgrade cost came back three times the estimate. That happens because your model used nameplate capacity: the substation transformer rating printed on a one-page PDF from 2019. Meanwhile, three other developers stacked their own 200 MW queues ahead of you — same bus, same transformer, but your model still showed available headroom. The real constraint isn't the nameplate; it's the queue depth.

Queue-Based vs Nameplate Capacity

Swap the static field for a time-weighted queue model. Public data sources — OASIS filings, EIA-860 generator lists, even regional ISO interconnection reports — let you reconstruct the effective capacity left after committed projects. One analyst I know scripted a scraper for PJM's queue dashboard; he found that four of his top-ten sites had less than 30 MW truly available, despite the transformer nameplate showing 150 MW. The gap was all queue. His revised rankings dropped those sites from Tier 1 to Tier 3 overnight. That hurts. But it is cheaper than spending six months on an interconnection deposit for a site that was already dead.

The catch: queue data is messy. Filing dates shift, projects withdraw quietly, and upgrade timelines stretch. A simple queue-depth lookup is better than nothing — but it still assumes all queued projects actually get built. Spoiler: many don't. So you need a decay factor: weight queue position by historical attrition rates for each ISO. Typical withdrawal runs 30% to 60% depending on the region. Ignore that, and you overcorrect, filtering out sites that would have cleared.

'Queue data without attrition is like a weather forecast that assumes it never rains — technically predictive, practically useless.'

— paraphrased from a developer I met at a grid workshop, 2023

Time-Dependent Cost Adders for Upgrades

Even after queue-adjusted capacity, you still face upgrade cost uncertainty. A site that needs a new breaker and a relay panel might cost $200k today — but if six other projects queue up in the same feeder corridor next year, the shared upgrade allocation could double. Dynamic cost adders solve this by modeling when you interconnect relative to others. Build a simple Monte Carlo overlay: sample queue arrival rates from historical patterns, compute shared upgrade costs per scenario, then rank sites by the expected total — not the cheapest single-path estimate. One client of ours ran this and discovered their highest-ranked site (cheapest upgrade, best irradiance) had a 70% probability of triggering a $4M substation expansion within 18 months. They swapped to a site with slightly higher base cost but far lower queue collision risk. That site is now operational. The first one is still in queue hell.

Trade-off worth noting: dynamic modeling eats compute time and data cleaning hours. Small teams with three analysts cannot maintain a live OASIS scraper plus a queue-attrition model plus upgrade cost sims. My advice — start with the decay-weighted queue depth only. That single fix usually flips enough rankings to justify the effort. Add the full Monte Carlo upgrade layer only after you have burned through the low-hanging fruit of bad queue assumptions. Wrong order? Yes. But it works.

Fix #3: Monte Carlo for Land Access Uncertainty

Parameterizing landowner willingness from county tax records

Most teams skip this: they assign a flat 70% or 80% 'landowner acceptance rate' and move on. That single number is a lie—it hides the fact that willingness varies block by block, sometimes by an order of magnitude. I have seen a model flag a 50-acre parcel as 'high suitability' because of solar irradiance and road proximity, while the actual landowner had already rejected two developers in three years. The fix is ugly but concrete. Pull county tax roll data—look for parcels held in LLCs versus individual ownership, look for recent transfers (flippers rarely lease), and look for agricultural exemptions that signal active farming income. Parameterize each site's probability as a beta distribution, not a point estimate. A beta(2,5) versus beta(5,2) changes everything when you run 10,000 draws. The catch is data hygiene—tax records are filthy, full of missing fields and stale addresses. But even a rough distribution beats a deterministic yes/no that feels clean and is wrong.

Modeling lease negotiation duration

Access isn't a switch—it's a timeline. What usually breaks first in a real portfolio is not the rejection but the wait. A landowner says 'maybe' in January, then goes quiet until April, then demands a higher option fee in June. By then your interconnection queue slot is gone. The Monte Carlo approach here is brutal but honest: sample a negotiation duration from a lognormal fit to your own project history, or if you have none, use public PPA filing dates as a proxy. I've seen durations range from 47 days to 314 days for similar-sized parcels in the same county. That spread matters. One run gives you a portfolio where all five sites close in month eight; the next run gives you two stragglers that push the whole program past the RFP deadline. The trade-off? You need at least 15–20 historical data points to fit a distribution that isn't junk. Without them, use a triangular distribution with expert-elicited min/mode/max—it's crude but beats pretending every deal closes in 90 days flat.

'We modeled land access as a binary gate. Then we ran 5,000 paths and realized the gate was actually a revolving door with a broken hinge.'

— developer quoted during a post-mortem after losing a 200 MW allocation round

Portfolio-level probability of securing N sites

Here is where the stochastic approach earns its keep. A deterministic model tells you: 'Site A works, site B works, site C works—three sites, proceed.' Monte Carlo tells you: 'There is a 64% chance you secure at least three sites, a 22% chance you secure exactly two, and a 14% chance you get one or zero.' That changes capital allocation decisions overnight. You might hedge by optioning six sites instead of three, knowing the correlation between landowner decisions in the same school district is 0.3, not zero. Quick reality check—correlation coefficients are the parameter nobody estimates, and they dominate the portfolio tail. I fixed a client's model once by switching from independent Bernoulli draws to a Gaussian copula with weak positive correlation; the probability of total failure tripled. Not because the data was worse—because the deterministic model had been silently assuming each landowner's decision was coin-flip independent. Wrong. Farmers talk to each other. One rejection can cascade.

The hard part is stopping yourself from over-engineering. You do not need a full Bayesian hierarchical model for a 50-site screen. A simple Monte Carlo with 2,000 iterations, three input distributions, and a correlation matrix of 0.2–0.4 is enough to reveal the fragility that deterministic models hide. Run the simulation, inspect the histogram of 'sites secured,' and ask: would I sign a lease based on the mean or the 10th percentile? If you answer 'the mean,' you are repeating the original mistake—treating all sites as equally accessible just with extra steps.

When Not to Model Accessibility at All

When the Model Costs More Than the Decision

Sometimes the smartest accessibility model is a blank sheet of paper. Not because data is scarce—but because the choice itself doesn't need a model. I have watched teams spend three weeks building a weighted suitability raster for a two-turbine project where the developer already owned the land. That hurts. The model told them nothing they didn't already know from a single Saturday drive.

The tricky part is recognizing which decisions are actually decisions. If your mandate says 'site within 5 km of Substation A—period'—you are running a compliance check, not a siting trade-off. The list of candidate parcels is the deliverable. Adding distance decay curves or permitting timelines just hides the real constraint, which is policy, not geography. Quick reality check—if your boss can point to a map and say 'here, and only here' before you open your GIS, close the model. Make a spreadsheet instead.

Early-Stage Prospecting with Thin Data

When you're screening a state or a region for the first time, your land-use attributes are probably wrong anyway. Parcel boundaries might be six months old. Wetland buffers shift with every rain. In that fog, a complex accessibility model creates a false sense of precision—it looks rigorous but it's really just propagating noise in a pretty color ramp. I have seen analysts spend a month calibrating a Monte Carlo land-access simulation when a simple Google Maps overlay and three phone calls would have eliminated 80% of the uncertainty.

The rule of thumb I use: if your dataset has more than 30% missing values for parcel ownership or road access, a list beats a model. Write down the top 20 sites by name, check if they're actually for sale, and move on. That sounds crude—and it is. But crude and right beats polished and wrong every time. What usually breaks first in early-stage work is not the accessibility logic; it's the assumption that a public tax parcel database reflects private willingness to sell. No model can fix that silence.

'The most expensive mistake in distributed generation siting is modeling a question that nobody asked you to answer.'

— overheard at a renewable energy conference, 2023

Small Portfolios Where Manual Vetting Is Cheaper

For portfolios under ten sites, manual vetting often wins on speed and depth. One phone call to a county planning office can reveal a transmission moratorium that your dynamic grid capacity model would miss entirely. The catch is that analysts hate admitting a model is overkill—it feels like wasted effort. But the math is simple: if your model takes forty hours to build and runs for two hours, and manually checking ten sites takes twenty hours of phone calls and windshield surveys, the manual route saves you twenty-two hours. Plus you get the unwritten rules—the planning commissioner who hates solar, the subdivision that was approved last week but isn't on any map. Those don't show up in a CSV.

The pitfall here is pride. We have all inherited a model from a predecessor that was 'almost ready' and spent a week debugging it instead of just calling the utility. That is not diligence; it's sunk-cost logic wearing a lab coat. For small portfolios, the best accessibility tool is a car key and a cell phone. Not every problem is a modeling problem—some are just coordination problems wearing a GIS interface.

Open Questions & FAQ

How do you actually validate accessibility assumptions?

Most teams skip this entirely — they run the model, like the output, and ship the map. That hurts. I have seen a siting committee greenlight a portfolio where every 'accessible' parcel sat behind a single-lane bridge with a 12-ton weight limit. The tricky part is that validation doesn't mean comparing your model to itself. You need a small, ugly field audit: pick 10–15 sites your model scores as high-accessibility, then call the county permitting office or check Google Street View for gate presence. Quick reality check — if three of those fifteen have no legal road frontage, your accessibility layer is lying to you. One practitioner I know built a simple log: for each candidate site they recorded 'gate found? Y/N' and 'road width less than 20 ft? Y/N.' After 40 sites the pattern was obvious — his model overrated parcels near state highways because it ignored the last mile of private easement. That is cheap validation, not a research project.

'A model that never fails a field test has either been tested wrong or calibrated to the wrong question.'

— paraphrased from a transmission planner after his third false-positive site visit

What software supports sequential permitting gates?

Off-the-shelf GIS tools rarely do. ArcGIS Pro with the ModelBuilder add-in can approximate a sequential gate using conditional loops, but you are essentially duct-taping a state machine onto a raster calculator. The catch is that most renewable siting platforms treat permitting as a single binary column — 'permit possible? yes/no' — which is exactly the flat-accessibility trap we are trying to kill. We fixed this by writing a lightweight Python script that reads a queue of permit steps (zoning variance → environmental review → interconnection deposit → road access agreement) and only advances a site if the previous gate cleared. That script is maybe 200 lines. Not sexy. But it caught a project that would have died at step two because the county required a 50-foot setback that the parcel couldn't accommodate. For commercial tools: DNV GL's WindFarmer and UL's HOMER have sequential gating in their advanced modules, but you pay for the overhead. Open-source alternative? Use NetworkX to model permit steps as a directed graph — each node has a cost and a probability of failure. That takes an afternoon to prototype.

Can machine learning replace these fixes?

No, and anyone who says yes is selling consulting hours. ML is great at pattern recognition in high-dimensional data — road density, slope, land cover — but accessibility is a sequential, constraint-satisfaction problem, not a classification task. A random forest can tell you 'this parcel looks accessible,' but it cannot tell you that the access agreement requires a 90-day public comment period that starts only after the environmental review clears. That is a logic chain, not a correlation. Where ML helps is in softening the Monte Carlo inputs: instead of guessing a uniform distribution for 'probability of landowner refusal,' you train a model on historical lease-signing rates by county. That reduces variance in your Monte Carlo runs. But the gate structure itself? Hard-coded state machine, every time. I have seen teams try to feed permit sequences into an LSTM — the output was beautiful and wrong. Wrong order. Not yet.