Coverage bounds on floating photovoltaics at Colorado River reservoirs

An hourly analysis of seven reservoirs, with uncertainty propagated and outputs validated against independent measurements.

Steps Ventures · independent analysis, not affiliated with any agency, district, utility or investor · correspondence: mike@stepsventures.com · code and data: github.com/steps-re/colorado-river

1. Motivation

Reservoir evaporation on the Colorado River is a large and poorly instrumented loss. Lake Mead alone loses on the order of half a million acre-feet a year, comparable to a mid-sized state's municipal demand. Covering water with photovoltaic modules is an appealing response because it appears to address supply and generation simultaneously, and coverage targets of 15–20% have entered the policy conversation.

Such a target is only meaningful if both benefits scale with coverage. This paper tests whether they do, and asks what bounds the answer.

2. Study sites and data

Seven Reclamation reservoirs with hydroelectric generation. Asterisk marks an evaporation rate that is a screening estimate rather than a measurement.

Surface area. Daily surface area from Reclamation's own operating record, 2024–2026. Reclamation publishes daily pool elevation and daily storage for every reservoir here, and surface area is the derivative of one against the other: a reservoir gaining dV acre-feet while rising dh feet has a surface of dV/dh acres. Rather than differentiate noisy daily data point by point, we fit the hypsometry a valley filling with water actually follows, V = a(h − h₀)^b, to the whole record and differentiate it in closed form. Three parameters against roughly four thousand daily observations, smooth and increasing by construction. Changing the estimator's window moves the result by 0.1% at Mead and under 7.5% everywhere it is used.

What this is and is not, because an operator would object to the framing. Reclamation's published daily storage is itself computed from measured elevation through their own area-capacity relation. Differentiating storage against elevation therefore RECOVERS that relation rather than measuring the reservoir independently. One reviewer put it as doing calculus on a lookup table, and that is fair. It matters for what the result can be claimed to be. It is not an independent measurement, and its agreement with Reclamation's own figures is not corroboration, it is arithmetic. What it IS is the operating area-capacity relation, recovered per reservoir and evaluated at every day's actual elevation, which is exactly the quantity the model needs and is strictly better than our satellite composite for the reason the canal test demonstrated. Where a published table is available it should simply be used. We derive it because the tables are not published in machine-readable form for all seven reservoirs.

Two further caveats an operator would add. The daily series mixes provisional and approved records, and we do not distinguish them. And a three-parameter hypsometry is a smooth fit to a reservoir whose bathymetry is terraced in places, so it will be least reliable exactly where the pool geometry changes character.

This replaced our own Sentinel-2 measurement, which we had used through six rounds of review. The satellite reads low at every reservoir: 4.6% at Mohave, 6.9% at Mead, 13% at Flaming Gorge and Navajo, 17% at Blue Mesa and 25.9% at Powell. The ordering is the tell, since the gap tracks how convoluted the shoreline is, and Powell's drowned side canyons are the extreme case. A 30 m water mask loses narrow arms, shadowed banks and mixed shoreline pixels. Reviewers flagged that risk repeatedly and none of them measured it. Lake Havasu keeps a published static area, because it is held within four feet and its hypsometry cannot be recovered from an elevation record that does not move.

Using daily rather than annual areas also lets the model see reservoirs that move: Powell's surface varied 21.5% within 2024 alone, Blue Mesa 18.2%, Navajo 12.6%. Mead varied 7.1% and Mohave 0.1%. Evaporation here still multiplies an annual depth by an annual mean area, so it captures the level change but not the covariance between area and evaporation rate within the year. That covariance is now measured rather than assumed away. Lake Mead's USGS record is monthly, so it supplies a real open-water seasonality: evaporation peaks in month 8, two months after peak net radiation, because the lake spends spring and early summer storing heat and gives it back in the autumn. Running the daily areas against that shape, and separately against each reservoir's own Penman shape, moves basin evaporation by +0.3%. The per-reservoir detail is in the limitations section.

Evaporation. Lake Mead (6.22 ft/yr) and Lake Mohave (5.64 ft/yr) use USGS eddy-covariance and energy-balance flux measurements. These are measured depths and are independent of surface area, so they may validly be applied to a separately measured area. The Mead record now runs through 2023 rather than 2019, and aggregated the way USGS aggregate theirs (the most probable column, an unweighted mean of complete calendar years) the 8 qualifying years in it average 6.25 ft/yr corrected, 0.5% over the depth used here and well inside the 8% flux uncertainty the model already carries, so the extension corroborates the published rate rather than moving it. 2019 is held out of that mean because its own latent-heat flux implies 33% more evaporation than it reports (7.17 against 5.39 ft measured, the gap taken relative to the reported figure), and 6 of its 12 months are flagged estimated. The methods page shows the test and both means. Lake Mohave has not been remeasured since 2019, and its depth is far less certain than Mead's in a way the model used to hide: Closure uncertainty at Lake Mohave averages 10.8% of its annual evaporation against 2.1% at Lake Mead, 5 times as much, because a larger share of Mohave's measured turbulent flux comes off the surrounding desert (period energy balance ratio 0.82 against 0.98). The model now carries 10.8% at Mohave rather than the 5% it used to share with Mead. The sampled sigmas differ by less than that 5x, because Mead's is held at 5%: the report's own stated measurement uncertainty of 5-7% is wider than the 2.1% closure spread measured there, so it binds instead. At Mohave closure is the wider of the two and binds. Lake Powell uses the Reclamation/DRI measured programme. Lake Havasu has no flux station and USGS have confirmed that one is only in planning, so its depth is derived rather than measured. Reclamation's LCRAS accounting covers Mead, Mohave and Havasu on one convention and two of those three have measured flux, so the convention is calibrated where it can be checked and applied where it cannot. Lake Havasu's evaporation depth is 7.52 ft/yr, bracketed 7.05-7.99. The accounting volume over Reclamation's own area gives 7.00 ft/yr net of rain; adding the 0.43 ft/yr that falls on the lake gives 7.43 gross, and the calibration's own bias against measured flux at Mead and Mohave is -0.09 ft/yr with a per-year spread of 0.23. That replaces an asserted 5.2-7.4 ft/yr bracket 2.4 times wider, whose ends were two different readings of the same quotient rather than a measured range. A higher depth means more water saved per acre covered, so this raises Lake Havasu's standing rather than lowering it. Havasu is already the reservoir this work rates highest, so the correction cuts against our own conclusion and is reported for that reason. Upper Basin rates are screening estimates; those reservoirs are only now being instrumented.

Solar resource. PVGIS-NSRDB hourly irradiance and temperature at each reservoir's coordinates, converted to AC output per installed MW with a floating cooling adjustment (cell temperature rise scaled to 0.82 of a ground mount, power coefficient −0.35%/°C), which averages a 1.01 uplift and is immaterial to any conclusion.

Prices. Day-ahead locational marginal prices at each reservoir's own balancing authority. At five of the seven that is a measured full year 2024, 8,784 hours, no gap-filling; Flaming Gorge and Blue Mesa are shape-transferred from a partial 2026 record and are described in the next paragraph. Lake Mead prices off the Nevada Power balancing authority, the Arizona dams off the Arizona Public Service balancing authority, and Navajo off PNM. These are balancing-authority aggregations, not a node at each dam: no dam has its own published node. Nodes were identified from CAISO's APNode atlas; the EDAM load aggregation points carry prices for external balancing authorities, while the imbalance-market intertie names do not.

Flaming Gorge and Blue Mesa sit in markets that began publishing day-ahead prices only in 2026: PacifiCorp East when CAISO's Extended Day-Ahead Market went live, and the SPP West footprint when SPP RTO West replaced the Western Energy Imbalance Service, which was real-time imbalance only and had no day-ahead product. For these two, the hour-of-day ratio of measured local prices to a reference node over the overlapping window is applied to the reference node's full year. The result is shape-transferred, not measured, and is treated as such throughout. The overlap window is June and July only, because those markets launched in 2026, so the ratio is a summer ratio. Summer and winter diurnal price shapes differ materially in both regions, driven by cooling load and solar penetration, so winter revenue at these two reservoirs is the least reliable number in this study. Neither reservoir's result should be read as a measured annual price. Blue Mesa's ratio is computed from Colorado River Storage Project settlement points, the delivery locations for the power that dam actually generates.

The measured differences matter. Against the Desert Southwest nodes, the Palo Verde hub runs $7.91/MWh lower at midday with a third more negative-price hours; against Blue Mesa's own market it runs 26% lower at midday. Pricing the basin off a single hub understated revenue at every reservoir.

Dam generation. Glen Canyon uses measured sub-daily USGS discharge below the dam. The remaining plants have no public sub-daily tailrace gauge and use a load-following diurnal and seasonal shape normalised to published annual energy.

3. Methods

For each reservoir and each coverage fraction c, at hourly resolution over a full year:

where A is measured surface area, ρ areal density, T plant nameplate, G(t) dam generation, λ the share of idle line available to solar, L on-shore load, κ its coincidence with solar, S(t) per-MW solar output, p(t) price, σ evaporation suppression and e the evaporation rate. Energy not delivered is partitioned into inverter clipping, line spill, negative-price withholding, storage round-trip loss and stranded charge; the partition is verified to close to <10⁻¹¹ of gross generation across 448 parameter combinations.

Where storage is modelled, dispatch is a day-ahead greedy: charge from energy the line cannot carry or that would sell at a negative price, discharge into the day's best-priced hours with remaining headroom, subject to a 10% state-of-charge reserve, 0.88 average lifetime capacity fade and 365 full-equivalent cycles per year. Dispatch has perfect price foresight within each day, which flatters storage.

Costs are annualised with a capital recovery factor. Capital expenditure is split into a DC term scaling with array size and an AC term scaling with the inverter rating, so oversizing an array past its export limit incurs only DC cost. Suppressed evaporation is not monetised: no mechanism exists under the 1922 Compact or the Arizona v. California decree for a private party to own or sell water salvaged from evaporation on a mainstream federal reservoir. Cost per acre-foot is therefore reported as a public cost-effectiveness measure net of power sales.

3a. A 54-year accounting record, and what it caught

Reclamation's Lower Colorado accounting publishes reservoir evaporation as a volume, annually, back to 1971, and reconciles to Decree Accounting within about a hundred acre-feet. That is independent of both the satellite work and the USGS flux campaigns, so dividing those volumes by our measured areas is a genuine external test of the evaporation depth this model uses. It is now one of the validation tests.

Two of the three Lower Basin reservoirs passed comfortably, Mead within 7% and Mohave within 4%. Lake Havasu came in 35% low, and Havasu is the reservoir this paper singles out as the strongest candidate.

The cause was a bad input rather than a bad method. Havasu's rate had been set at 5.21 ft/yr, derived from an LCRAS volume of 98,246 acre-feet. That figure appears nowhere in Reclamation's 1971–2024 series, which runs between 126,000 and 135,000 acre-feet with a mean of 130,664 and is unusually stable across five decades. The 2019–2024 mean of 131,992 acre-feet over 18,864 acres gives 7.00 ft/yr, which is what the model now uses and which the paper had already bracketed as plausible when it flagged this rate as the weakest in the set.

Worth being explicit about the direction, because almost every correction in this project has run the other way. A higher evaporation depth means more water saved per acre covered, so this one makes reservoir solar at Havasu look better, not worse. That is exactly why it deserves more scepticism than the ones that cut against the technology, and two reviewers duly supplied it.

Their objection is correct and it limits what this correction establishes. A volume divided by an area is only a depth if the volume and the area describe the same water. LCRAS does not publish the surface it divides by, so 7.00 ft/yr is LCRAS's volume over LCRAS's own nominal area of 18,864 acres, while the model applies that depth to our measured 15,330 acres. If those two extents differ, and they plainly do, the quotient carries the ratio between them. The rate this replaced carried the same defect in the other direction, which is why the code had flagged it as the weakest in the set all along.

So the honest statement is narrower than a correction. Havasu's evaporation depth is not measured, it is inferred from an accounting volume whose denominator is unknown, and 5.2 to 7.4 ft/yr remains the defensible bracket rather than any point inside it. We have asked Reclamation's water accounting group directly what area their reservoir evaporation divides by. Until that comes back this is the least secure input in the model, and it belongs to the reservoir the paper otherwise rates highest.

The validation test built alongside it needed the same correction. At Mead and Mohave the comparison is genuinely external, because those rates come from USGS eddy-covariance flux and owe nothing to LCRAS. At Havasu it compares an LCRAS quotient against that same quotient and returns 0.1%, which is arithmetic rather than agreement. It is now labelled circular, which drops the external validation count from 17 to 16 and is the more useful number.

3b. What drives the evaporation, and what a canal does not fix

The argument that prompted this section was not made against reservoir floats. It was made for them. Reviewing an earlier draft, Upmanu Lall argued that canal solar "does not compute in terms of economics or potential water savings", because the usual truss-and-shade design "leaves the aerodynamic pathways for evaporation fully open, and the evaporation is driven as much if not more by vapor pressure deficit as it is by incoming radiation", while floating systems on reservoirs "reduce the edge effect and can dramatically reduce the evaporation". We could not evaluate either half of that with one annual depth and one area, so we built the daily physics. Eleven years of daily weather at each reservoir's own coordinates, from NASA POWER, run through the Penman open-water combination equation rather than FAO-56 reference ET. The distinction matters: reference ET is defined for short grass with a fixed canopy resistance and an albedo of 0.23, and a lake has no stomata and an albedo near 0.06.

Splitting each day into its radiative and aerodynamic parts settles the question quantitatively. Across the three Lower Basin reservoirs the aerodynamic term, the one carrying vapour pressure deficit and wind, is 44 to 47% of total evaporative demand, falling to 25% at Blue Mesa. That range is not sampling noise, it is the choice of reanalysis: 47% on NASA POWER, built on MERRA-2, and 44% on the ERA5-based archive we used first. We report both, because the gap between two mainstream reanalyses is wider than anything this analysis can resolve, and a single figure would imply a precision the underlying meteorology does not have. The conclusion holds either way. Radiation remains the larger single term, and the objection is substantially right, because close to half of what leaves these reservoirs leaves through a pathway that shading alone does not touch.

That distinction decides the canal comparison. We mapped the basin's conveyance network from OpenStreetMap: 19,787 km of canals, ditches and drains inside the main-stem corridor, carrying about 37,057 acres of open water against 251,258 acres across the seven reservoirs. Canals are therefore roughly 13% of the combined surface, and around 211,783 AF/yr is at stake on them. The usual canal-solar design is a truss-mounted array spanning the channel, which removes the radiative pathway and leaves the aerodynamic one substantially open. On these numbers that caps what shading a canal can recover at roughly the radiative share, whereas a cover floating on still water suppresses both paths. Published modelling of California canals reports 20 to 82% evaporation savings, and Project Nexus is measuring it in the field. Our contribution is only the decomposition: shading and suppressing are not interchangeable, and the gap between them is the aerodynamic share. On this specific point the physics supports Lall's position rather than ours.

We then checked whether transferring the reservoir decomposition to canals was legitimate, by running the same physics at the canals' own coordinates, resolved from the mapped centrelines rather than assumed. It is: All-American 44%, Central Arizona Project 40%, Coachella 41%, Gila Gravity 43%, against 42% across the Lower Basin reservoirs. The atmospheric demand over the Imperial and Yuma valleys is not meaningfully different from the demand over Mead and Mohave, which is unsurprising given they share a climate.

We also tried to confirm the canal widths from satellite, since OpenStreetMap tags width on only 1% of segments and everything else rests on a conservative 12 m default. It does not work, and the way it fails is informative. Measuring water inside a corridor around each mapped centreline and dividing by length returns:

The All-American Canal is 60 to 70 m wide in its main reaches. A measurement returning 5.9 m is not an imprecise answer, it is a failed one, recovering roughly a tenth of the water actually there. A positive control settles why: the identical recipe run over Lake Havasu returns 18,994 acres against a published figure near 19,000. The method is excellent on wide open water and collapses on narrow linear water, because a canal is one to six pixels across at 10 m and nearly every pixel is a mixture of water and bank. Confirming canal widths needs 1 m imagery, not Sentinel-2.

A third source bounds the canal figure from the other side. The National Hydrography Dataset carries surveyed polygons for water wide enough to map as area, with an area attribute attached, so for those reaches canal surface is measured rather than inferred. Across the basin it returns 260 canal polygons totalling 10,858 acres, against the 37,057 acres our centreline-times-width estimate gives for the whole network. These are not competing measurements of the same thing. NHD polygonises only the wide reaches and leaves everything narrower as a line, so its total is a measured lower bound; ours covers the full network at an assumed width. Read together they say the canal surface is uncertain by roughly a factor of three, and that the assumed 12 m default is almost certainly too narrow for the large conveyances and too wide for the laterals. Nothing in this paper's conclusion turns on which is right, but a canal-solar case would.

The same dataset is a useful check in a way that is easy to misread. Its reservoir polygons correspond to full pool rather than current level: it puts Navajo at 15,363 acres against a published full-pool figure of 15,610 and against the 11,269 acres Reclamation's elevation record gives as the recent operating mean. All three are correct about different questions. It is a reminder that a mapped water body and an operating water surface are not the same object, and that our own areas had to come from the elevation record for exactly that reason.

That failure also explains the reservoir result earlier in this paper, and turns an assertion into evidence. Our satellite areas came in low by 4.6% at Mohave and 25.9% at Powell, and we attributed the gap to mixed pixels in narrow canyon arms. The canal test is the controlled version of exactly that: water narrower than a few pixels goes undetected. The ordering of the reservoir errors follows shoreline complexity, and Powell, whose surface is largely drowned side canyons, is the extreme case.

One limit survives the location check. This is atmospheric demand at canal locations, not canal geometry. A canal has a fetch of metres rather than kilometres, the water moves, and many reaches are lined and sunk below grade, so the wind field over a narrow channel is not the wind field over open water. The share is therefore location-appropriate and geometry-naive, and a real canal number needs the field measurement Project Nexus is taking. And Lall's stronger claim, that float systems cost the same as or less than ground-mount because US labour dominates and float manufacture scales, runs against the benchmark this paper uses, which puts floating capital above ground-mount. That disagreement is unresolved here and it moves the answer more than the evaporation physics does.

Two checks came out of the same daily series. First, our Penman totals run 30% to 93% above measured evaporation. That is not a defect to be hidden but the documented behaviour of any combination equation without a heat-storage term, which overestimates in the warm season and underestimates in the cold, with published overestimates of 24 to 36%. It is why the levels in this paper come from measured flux, and why the physics above is used only for the ratio, which largely survives the bias. Second, we tested whether multiplying an annual mean area by an annual mean depth misstates evaporation, since both vary within the year and their covariance is discarded. The error is under 2.3% at every reservoir, so the annual product stands.

We also tried to recover evaporation as the residual of Reclamation's own daily water balance, from gauged inflow, release and the change in storage. It would be an evaporation measurement entirely independent of the meteorology. The first attempt returned 1.25 ft/yr at Lake Powell against an expected 5.83, which is a failure rather than a measurement.

The reason turned out to be a term Reclamation publishes and we had ignored. Powell holds roughly 4.9 million acre-feet in bank storage, water held in the sandstone around the reservoir rather than in it, against annual evaporation nearer 450 thousand acre-feet. Daily movement in and out of the banks is therefore an order of magnitude larger than the quantity being solved for. Adding the published bank-storage series takes Powell's residual from 1.25 to 3.07 ft/yr, or from 21% of the expected value to 53%. Flaming Gorge barely moves, 1.68 to 1.74, which is consistent: its banks hold 0.13 maf against Powell's 4.9.

So the balance is now half-closed rather than broken, and what remains bounds ungauged inflow and gauging error. It is worth naming what cannot substitute here: satellite gravimetry measures total water storage at a resolution of roughly 300 km, and Lake Powell sits inside a single footprint, so it cannot separate a reservoir's banks from its basin. Bank storage at this scale is available because Reclamation computes and publishes it, and for no other reason.

4. Validation

Tests are labelled by what they can establish. Only external tests validate anything: they compare a model output against a value produced independently, by a different method. Circular tests compare against figures the model was normalised to and prove only that the scaling arithmetic is correct. Internal tests are round-trips, and range tests check a constant against published bounds. 16 of 16 external validations pass (22 of 22 overall).

The Lake Mead evaporation test reads 16.6% low by construction: our surface is a 2024–2026 measurement and Reclamation's volume is a 2017–2021 average over a larger lake. Lake Mohave, held on a seasonal guide curve so that its area is stable between the two periods, agrees to 4.6% and is the tighter test. Outline tests are directional: full-pool shorelines should exceed a drawn-down measured surface.

5. Uncertainty

Monte Carlo propagation, 2,000 draws, seed 20260807. Parameters are not drawn independently: a Gaussian copula links capital cost to O&M (ρ 0.75, since a deep, storm-exposed, mussel-infested site raises both), capital cost to areal density (ρ −0.35, since wide mooring corridors lower density and raise cost per watt together), capital cost to discount rate (ρ 0.25), and suppression to evaporation rate (ρ 0.20). Achieved rank correlations match the targets to within 0.015. Independent draws would understate the tails.

Results. Right-sized array is the largest whose spill stays under 5% for that draw.

5.1 Variance attribution

Partial rank correlation coefficients, computed on the same sample. These identify which inputs drive the spread in each output.

Two patterns matter. Areal density is the strongest driver of cost per acre-foot at every reservoir, which means layout efficiency, not module price, is where the cost uncertainty lives. And in the Upper Basin the evaporation rate dominates (ρ −0.96 at all three sites) because those rates have never been measured. The Reclamation and Upper Colorado River Commission instrumentation programme now underway would therefore resolve most of the Upper Basin uncertainty on its own.

5.2 Is the Lake Havasu result an artifact of the export proxy?

Havasu's advantage could plausibly be an artifact of the assumed line share rather than a property of the site. Sweeping that parameter across its full range and ranking reservoirs by capacity retained relative to their own unconstrained case:

Havasu ranks first at 16/19 constrained line-share values (line share < 1.0). At a line share of 1.0 no reservoir is constrained, every site retains 100%, and the ranking is a tie carrying no information; that case is excluded. So the finding is conditional on that parameter rather than robust to it. The objection partly lands: Havasu leads across most of the constrained range but not all of it, and its rank is not independent of the export proxy.

It is not robust to load coincidence, which is the assumption that actually decides it:

Below roughly 25% coincidence the advantage disappears entirely. How much Central Arizona Project pumping is genuinely simultaneous with solar output, and genuinely dispatchable to it, is therefore the measurement that would settle this site — not the transmission question.

The spread on array size is dominated by the share of idle line assumed available, which is the parameter standing in for the interconnection study this analysis lacks. The spread on cost per acre-foot is dominated by capital cost and O&M, reflecting genuine disagreement about mooring in deep, strongly fluctuating water and about biofouling by quagga mussels in the Lower Basin.

6. Results

6.1 The benefits do not scale together. Suppressed evaporation is linear in coverage by construction. That is an assumption, not a result, and it is the optimistic direction: heat a cover prevents from leaving as vapour warms the water instead and raises evaporation on the surrounding open surface, so real suppression falls below linear as coverage rises (section 8). Because the finding is that energy saturates while water does not, granting water the more generous treatment only strengthens it. Deliverable energy saturates: once array capacity exceeds available headroom, additional modules generate energy that cannot leave. A single coverage target therefore cannot be optimal for both objectives.

6.2 The coverage bound is reservoir-specific and spans an order of magnitude. Taking the coverage at which line spill first exceeds 10%: Havasu 8.5%, Mead 7.25%, Powell 4.25%, Blue Mesa 2.5%, Mohave 1%, Flaming Gorge 1%, Navajo 0.5%. A 15–20% target exceeds the bound at every reservoir studied under the export assumption used here; since that assumption is a scalar proxy rather than an interconnection study, the ranking between reservoirs is more robust than the absolute bound. Reservoirs with large surfaces behind small powerplants are the worst candidates, which is the opposite of what a surface-area-driven screen selects.

6.3 On-shore load, not transmission, distinguishes the best site. Lake Havasu retains 333 MW at median (P10–P90 192–471) against 1,218 MW at Lake Mead (P10–P90 390–2,112). Read those spreads before reading the medians: Mead's spans more than five to one, so its point estimate carries little weight on its own. The comparison that survives the uncertainty is relative to each line rather than absolute. Havasu's array is several times its own interconnection while Mead's is a fraction of Hoover's, despite Havasu's line being roughly one seventeenth the size, because roughly 300 MW of Central Arizona Project pumping load sits on its shoreline. Its surface is also administered by BLM and Arizona State Parks rather than the National Park Service. Load coincidence is consequently the parameter that decides that site.

6.4 Cost of suppressed water. Median $19,139 per acre-foot at Lake Mead and higher elsewhere, against $325–700 to purchase conserved agricultural water and $2,500–3,500 for seawater desalination. Non-generating covers are cheaper than FPV but do not close the gap:

The chemical monolayer is the only entry cheaper per acre-foot than purchased conservation, and it is not a counter-example: wind disperses the film continuously, field suppression falls far below laboratory values, and it has been trialled and abandoned repeatedly on open water. Among options demonstrated at scale, none approaches the conservation benchmark.

The raw-water cover row is new, and it is the cheapest credible entry in the table. An earlier version of this work reported that no published cost for a raw-water evaporation cover at reservoir scale could be found, and priced the category off a sealed, food-grade potable-water quote instead. That was wrong by roughly an order of magnitude. A Queensland Government-funded assessment prices a floating cover at the industrial specification, the relevant duty for water this size, at A$75 per square metre installed over a thirty-five year life, which annualises to about $3,854 per acre-foot. Its other readings, an agricultural specification and an executive-summary range, bracket it below. Those are May 2020 Australian prices and this figure escalates them by 1.33 to 2026, weighting polymer and installation equally: Australian construction input costs and global HDPE resin are both up around 40% since mid-2020, US construction cost indices 25–27%. On cost alone that beats every other cover here and lands just above the desalination band.

The other supply chain moves it much further, and in the direction that matters least. Australia is a small, high-labour market. China makes most of the world's geomembrane, and a Chinese manufacturer lists floating-cover-grade sheet at US$1.40–3.90 per square metre as material. Applying the Australian report's own material-to-installed ratio of about 3.1 puts a China-supplied installed cover near US$4–12 per square metre, roughly a fifth of the escalated Australian figure. So the true cost of a cover is bracketed by a high-labour market above and a Chinese supply chain below, and that spread is about fivefold. It is still not a resolved number.

How far that source should be trusted, since a good deal now rests on it. It is a July 2020 final report by named academics at the University of Southern Queensland, funded by the Queensland Government, and one author published a peer-reviewed review of the same subject in 2023. It is not itself peer reviewed. Its prices are May 2020 Australian and are not inflated here, so the figure today is higher. Its scope is farm dams: 96% of the 243,000 storages it surveys are under two hectares. And it contradicts itself, giving service lives of five to ten, fifteen and thirty-five years in different sections, and a size limit of two hectares in one place and five in another. It also states that the Los Angeles Reservoir is 175 hectares when it is 175 acres, an error this paper can check because it prices that same project elsewhere. The direction of its cost finding is consistent across every reading. Its precision is not, which is why the range above is carried rather than a single number.

What rules it out here is neither cost nor size. It is level change. The same report states that floating covers “must be tethered to avoid beaching and obstructing spillways” and are “not suitable for storages experiencing large water level fluctuations”. Mead sits 189 ft below full pool and Powell 179 ft, which is the most extreme fluctuation case in the basin. That is a direct quotation rather than a modelling assumption, and it disqualifies the technology on these reservoirs on its own. Size compounds it: covers are deployed as tethered rafts of about a hectare each and the report calls them generally limited to storages under two hectares, against Mead's roughly 30,900. A separate search for Chinese reservoir-cover deployments corroborates this independently and from the country that manufactures most of the world's geomembrane: it found no large-scale commercial installation at all, only pond and evaporator field tests. Those tests also measured 51–75% suppression for solid and modular covers and 14–60% for spheres, against the 90% assumed in the table, so the water yield above is probably optimistic too. So the honest reading of this table is not that covers are expensive. It is that the cheap ones cannot be built on water that moves like this, which is the same class of objection this paper makes to shade balls, and it applies with equal force to floating solar.

Suppression yield is in any case bounded by physics: an acre of covered water saves only the depth that would have evaporated, about 5.6 acre-feet a year in the Lower Basin, whereas a desalination plant's output per unit footprint has no such ceiling. Covers distribute industrial capital cost across large areas to harvest a thin layer.

6.4a What if the reservoir gets covered anyway? Every row above assumes floating solar is competing against bare water. It may not be. Covers are under consideration for reservoir temperature and biodiversity management rather than for evaporation, and this river already has that problem: as Powell fell, Glen Canyon released warmer water, bass-suitable days rose from about two a year to about forty-one, and smallmouth bass reached the last humpback-chub stronghold. Cool-mix flows work and are contested by hydropower. If a cover goes in for thermal reasons, an array is not an additional cost on a bare surface. It substitutes for a cover that was going to be bought regardless, and the avoided cost belongs on its side of the ledger.

A correction, because the first version of this section was wrong. It priced the avoided cover off LADWP's potable-water quote, the only figure then available, and concluded that floating solar's water became free once it displaced 46% of that cost. Verified raw-water cover prices are fifteen to thirty-six times lower. On those, the credit is worth tens of per cent rather than an order of magnitude, and floating solar does not approach the reuse or desalination band at any verified cover price. The earlier conclusion is withdrawn.

So the substitution argument is real and it is second-order: Between 3% and 34% off FPV's cost per acre-foot. It does not reach the reuse or desalination band at any verified cover price. The bottom row is retained only to show what the earlier version was leaning on. It is a sealed, food-grade product for treated drinking water, and it is not what anyone would buy to manage the temperature of a 30,000-hectare reservoir.

The premise is the bigger problem, and it is not about price at all. The same assessment that supplies these costs puts the effective upper size limit of floating covers at 2 hectares. Lake Mead is about 30,900. Nothing in the record shows a floating cover of any kind working at reservoir scale, so “the reservoir gets covered anyway” is not an established counterfactual here. The credit above should be read as an upper bound on a mechanism whose premise is itself unproven on water this size.

Three further caveats stand. Substituting an array for a solid cover loses suppression, 75% against 90%. The credit applies only to surface that would have been covered anyway, and covering part of a reservoir for power need not be the same placement as covering it for temperature control. And no cover programme on these reservoirs is funded or announced. The argument came from a critic of this analysis and was worth pricing properly. Priced properly, it does not change the answer.

6.4b Has anyone, anywhere, actually built this? Every number above is a cost model, and a cost model can be argued with. The cheaper test is whether any basin facing the same physics has done the thing. This paper spent fourteen review rounds on one river without asking, which was a gap, so here is the global record of surface intervention on reservoirs.

Nothing purpose-built for evaporation control has ever operated above 1,000 hectares. The largest operating cover of any kind on Earth is the Los Angeles Reservoir at 71 ha, which is 0.23% of Lake Mead. The two surface projects that exceed 1,000 ha are floating solar built for power, and the larger sits on a flooded coal-mining subsidence lake with no drawdown duty. The one large evaporation trial that did reach this scale, 1,011 ha at Lake Hefner in 1958, measured 9% and was abandoned because wind above 13 mph drove the film to the lee shore.

China is the decisive case, and it cuts against surface intervention. It operates the world's largest floating-solar fleet, manufactures most of the world's geomembrane, and has seen deeper drawdowns than the United States: Poyang Lake fell from more than 350,000 hectares to 81,400 in 2022. Facing that, it did not cover the water. It built the following instead.

The detail worth sitting with is Longyangxia. It is the most-cited hydropower-plus-solar hybrid in the world, 850 MW paired with a 38,300-hectare reservoir, and the solar is on land. The surface was there and was not used.

This is evidence about feasibility and priority rather than proof of impossibility, and it should be read that way: several of these technologies are young, and floating solar is scaling fast on water bodies that do not fluctuate. But seventy years of proposals, in every arid basin with the money to try, have not produced a single operating evaporation cover at even a tenth of the scale this question requires. That applies to floating solar on Mead and Powell exactly as it applies to the covers above it in the table.

6.5 Backfilling lost hydropower capacity. An array sized to replace generation a dam can no longer produce is not competing for export capacity, because the derate releases it. Modelled against a 58% Hoover derate (2.32 TWh/yr), arrays fill 40–57% of the hourly deficit across the array and storage combinations examined (up to 3.4 GW with 2 GW/8 hr). We did not find a configuration that exceeds it, and the physical reason is clear — a capacity derate is a continuous loss while solar is diurnal and seasonal — but this is a bound observed over the configurations modelled, not a proof. Seasonal storage, demand response or a hybrid firm resource would raise it. Combined break-even falls from $94/MWh with no storage to $65/MWh at 1,000 MW of four-hour storage, but most of that is not attributable to the array. At the capacity values used, the battery earns more in capacity payments than its own annualised capital costs and is profitable standing alone, so bundling offsets the array's losses with the battery's profit. Held apart, the array's own break-even improves only from $94 to $84/MWh. Merchant capture is $24 to $32/MWh depending on the market. Storage capital is net of the 30% credit it still qualifies for; the array itself receives none. Eight-hour storage over-invests. The residual gap is a question of contracted value for avoided replacement power, not of dispatch. Two caveats are material: hydrology is cyclical, so a derate may reverse and return the capacity to the dam; and wheeling and basis costs from each dam to a trading hub are not modelled.

7. Limitations

1. Export capacity is a scalar, not an interconnection study. Plant nameplate minus dam output, scaled by an assumed share, omits the generator step-up transformer, switchyard and breaker ratings, N-1 contingency, reactive capability and existing firm reservations. This is the dominant limitation and the reason no result here establishes feasibility at a specific site.

7a. How much water is actually at stake

Everything above bounds what floating solar can do at each reservoir. Summing across all seven at their own computed bounds gives the size of the whole opportunity: 49,865 acre-feet a year, for 5,316 MW of floating panel.

That number only means something against an alternative of known size. Reuse was surveyed across the basin for the first time in 2025: 26% of treated municipal wastewater is currently reused, ranging from 85% in Nevada to under 1% in Utah, and raising the basin to 50% would free about 1.3 million acre-feet a year. That is roughly 26 times the water every reservoir in this study could suppress if all of them were covered to the limit their transmission allows, and it costs on the order of a sixth as much per acre-foot.

One correction that cuts against the comparison, stated here rather than elsewhere. That 1.3 million acre-feet is a gross volume of treated effluent, not net-new river water. Reuse only creates new supply where the effluent is currently leaving the system. Inland wastewater discharged to a river is somebody's return flow and is already counted downstream, so recycling it upstream extinguishes that flow rather than adding to it. Las Vegas already credits its return flow to Lake Mead and Phoenix already recycles almost all of its effluent, which means a meaningful share of the headroom is not available twice. Southern California's ocean outfalls are where the net-new water actually is. We have not seen a published decomposition of the 1.3 MAF into net-new and already-credited, and until someone produces one the ratio above should be read as an upper bound on reuse's advantage rather than a settled figure.

The two are not legally equivalent, and the difference is the one this paper has returned to throughout: suppressed evaporation stays in the reservoir and has no owner, while reused effluent is a supply a utility can contract for. The reuse survey also carries a severe caveat from its own authors, who describe a data desert in which most states do not track reuse systematically and plants had to be contacted individually. Neither point changes the ranking. An option twenty-six times larger and several times cheaper does not lose to a smaller, dearer one because its accounting is imperfect.

Does that ranking survive a different sizing convention?

It should be tested, because the bound above is ours. A reader is entitled to suspect that a paper which bounds coverage at the tie and then reports a small number has assumed its own conclusion.

Groups working on floating solar here do arrive at areas several times larger, and two conventions account for it: sizing against full pool rather than the drawn-down surface the reservoirs actually hold, which is 2.1x at both Mead and Powell, and a 10 to 20% cover target with no transmission bound, roughly another 2x. But it is worth being precise about what that disagreement is, because it is not mainly about water. Practitioners sizing arrays this way are generally optimising the joint array-and-hydro system for energy revenue against intraday prices, and treating suppressed evaporation as an ancillary benefit to be tallied afterwards rather than as the objective. On that framing the larger areas are not a competing answer to the question below. They are the answer to a different one, and the comparison that follows should be read as “what happens if you size for water”, not as a rebuttal of anyone sizing for power.

So the table below removes our bound entirely. It expresses the full-pool convention as coverage of the real lake and runs it back through the same hourly simulation, at Mead and Powell combined.

Two assumptions in that test are contestable, and both favour our conclusion, so they belong here rather than in a footnote. The first is that "10 to 20% cover" means of full pool rather than of the surface the reservoir currently holds. That reading roughly doubles the implied array, and we have not confirmed it with the groups who use the figure. If it means the current surface instead, the gap between the two conventions is about half what is shown. The second is that the transmission tie stays fixed while the array grows. That is precisely the assumption a proponent of the larger convention disputes, because their design pairs the array with pumped storage, behind-the-meter load, or hydro rescheduling specifically to relax it. This model contains none of those. So the rows below do not show what a 20%-cover design would cost as its proponents would build it. They show what it costs on the interconnection that exists today, which is a bound on the free option and not a refutation of the design.

Sizing convention	Array (GW)	Water suppressed (AF/yr)	Transmission spill	Cost per AF (p50)	Against reuse
This paper: bounded by the existing tie	4.2	39,835	10-10%	$21,058–$22,646	6.0–6.5x
10% of full pool, no transmission bound	15.8	146,566	62-75%	$22,596–$24,705	6.5–7.1x
15% of full pool, no transmission bound	23.6	219,471	74-83%	$22,974–$24,980	6.6–7.1x
20% of full pool, no transmission bound	31.4	292,309	80-87%	$23,180–$25,101	6.6–7.2x

Mead and Powell only, the two reservoirs where a full-pool convention is the likely source of the difference. Spill and cost ranges span the two. Cost is the Monte Carlo p50 at a fixed coverage, net of power sales, against reuse at $3,500/AF, the expensive end of the range in reuse_resource.py. Computed by fpv_coverage_scenarios.py.

The larger convention does suppress substantially more water, and that is the honest half of the disagreement. It is also not free, and the direction of the cost is the finding: an acre-foot gets dearer as coverage grows, because gross cost scales with the array while revenue saturates against a fixed tie. Transmission spill climbs from about a tenth to over four-fifths. You buy six times the panel to get six times the water and throw away most of the electricity that was supposed to pay for it.

Within those assumptions the ranking does not depend on where we set the coverage bound. Reuse is several times cheaper per acre-foot at the tie-limited size and further ahead at the larger one, because the marginal panel earns less as spill grows. That is a narrower claim than “floating solar loses”, and it is the one the model supports: relax the tie and the cost side of this comparison has to be rebuilt, not merely rescaled.

One thing does not depend on any of it. A build on this scale is not a modelling detail: 15% of full pool across both lakes is over 20 GW of floating solar, against roughly 10 GW installed worldwide to date.

7b. The same question at other reservoir elevations

Every area in this paper is historical, and a twenty-five year asset is not. Reclamation publishes projected elevations in its 24-Month Study, but only as PDF, and lifting numbers out of PDF tables is the kind of step that fails quietly. It is also unnecessary. The hypsometry fitted to recover surface area, V = a(h − h₀)^b, gives area at any elevation in closed form, so the model can simply be evaluated at whatever pool level the question calls for. That answers more than one projection would: minimum power pool, dead pool, the 2007 shortage tiers, or any elevation a reader wants to supply.

Lake Mead	Elevation	Surface	MW at today's ceiling
2024 operating mean	1,076 ft	76,444	2,691 MW
1,075 ft shortage tier 1	1,075 ft	81,105	2,856 MW
1,025 ft shortage tier 3	1,025 ft	63,518	2,236 MW
950 ft minimum power pool	950 ft	37,774	1,330 MW
895 ft dead pool	895 ft	19,608	690 MW
Lake Powell	Elevation	Surface	MW at today's ceiling
2024 operating mean	3,582 ft	75,566	1,560 MW
3,525 ft target	3,525 ft	58,844	1,214 MW
3,490 ft minimum power pool	3,490 ft	43,640	901 MW
3,370 ft dead pool	3,370 ft	12,330	254 MW

Read the last column as what the same coverage percentage buys as the lake shrinks, not as a revised bound. It moves a long way: at minimum power pool Lake Mead's surface is half what it is now, and Powell's is well under half. Two effects run in opposite directions and only one is in this table. A falling reservoir offers less surface to cover, which shrinks the array. It also produces less hydro, which frees room on the shared line and would let a larger share of whatever is built actually sell. Anyone using these numbers for a real decision needs both, and that is a reservoir-operations question rather than an interconnection screen.

The dead-pool rows are extrapolation. Those elevations sit 130 ft and 170 ft below anything in the record the hypsometry was fitted to, so they indicate a direction rather than a quantity.

7c. How much of the inflow is actually gauged

Adding bank storage took the Powell water balance from broken to half-closed, and what remained was ungauged inflow and gauging error. USGS daily discharge lets that be stated rather than gestured at. Against Reclamation's published inflow, the main tributary gauges account for 100% at Powell, 76% at Flaming Gorge, 71% at Blue Mesa and 46% at Navajo. The remainder is side canyons, small tributaries, groundwater and gauge error.

One caveat limits how hard this can be pushed, and it is the reason the Powell figure is so close to unity: Reclamation's published inflow is itself partly computed from these same gauges, so agreement is closer to a consistency check than to an independent measurement of what is missing. The reservoirs where the fraction is low, Navajo especially, are the ones where the statement means something.

2. Dam output is shaped by measured daily releases, not an assumed seasonal pattern. Six of the seven reservoirs previously used a load-following shape with a summer-heavy seasonal weight, on the reasoning that power demand peaks then. Rule-curve reservoirs do not work that way: releases answer to water obligations, equalisation tiers and downstream orders rather than to power prices, so the months when the shared line is busy are an operational fact and not something to assume. Reclamation publishes daily release for six of them, so that pattern is now measured and only the within-day shape remains modelled, which is the right division given there is no public sub-daily tailrace gage except below Glen Canyon. It moved three ceilings down and none up: Mohave 1.25 to 1.0%, Blue Mesa 2.75 to 2.5%, Navajo 0.75 to 0.5%. Lake Havasu keeps a modelled shape, having no published sub-daily release series. Reclamation does publish Parker Dam releases daily, and this analysis uses them elsewhere; what is missing is the within-day shape.
3. This is still an interconnection screen, not a reservoir operations model. A hydroclimatologist reviewing it made that distinction and it is the right one. The dam's generation enters only as a claim on the shared line, hour by hour. Nothing here represents rule curves, storage targets, the 2007 Interim Guidelines, or the fact that a dam's releases are decided by water obligations rather than by power prices. Any question about how FPV would change reservoir operation, or how operation would change FPV, is outside what this can answer.
4. A flat annual evaporation depth carries no seasonal information, and the seasonality runs against the co-benefit story. We apply a measured annual depth times coverage. That is defensible for an annual total and says nothing about timing. It matters because deep reservoirs like Mead and Powell store heat through the summer and release it as latent flux into the autumn and winter, so their evaporation does not peak when the sun does. Solar output peaks in June. The water benefit peaks months later. Neither this model nor the coverage argument depends on the two coinciding, but anyone reasoning that shading and generating are naturally complementary in time should know that on these reservoirs they are not.
5. Surface areas move within the year and evaporation uses an annual mean. Areas now come from a stage-area curve driven by Reclamation's daily elevations, so they follow the reservoir rather than freezing it at a summer composite. Within 2024 the surface varied by 21.5% at Powell, 18.2% at Blue Mesa and 7.1% at Mead. Evaporation nonetheless multiplies an annual mean area by an annual depth. Weighted by each reservoir's own evaporation volume, ignoring within-year seasonality moves basin evaporation by +0.3%. The largest per-reservoir mean is +1.4% at Blue Mesa, and the worst single year in the record is 5.9% at Blue Mesa. Both seasonal shapes agree on the direction at every reservoir, so the result does not rest on which shape is right. The term is positive at every reservoir except Lake Mead: the surface is largest in summer when the rate is highest, so an annual product UNDERSTATES evaporation there. Understating evaporation understates the water floating solar could save, which raises its cost per acre-foot. That is the direction that flatters our own conclusion, which is why it is stated rather than buried. At +0.3% it is far inside the flux and area uncertainties the model already carries. Whether it reorders the reservoirs is not something the size of the term settles on its own: the closest two are only 1.8% apart in cost per acre-foot against a largest term of 1.4%. Applying every term and re-sorting leaves the order unchanged, so the ranking stands. A fixed array against a moving surface is the larger unmodelled issue, since an array is built once and the fraction it covers changes as the lake does.
6. Salvaged evaporation is not delivered water. An acre-foot not evaporated at Flaming Gorge or Navajo sits hundreds of river miles above where Lower Basin demand is, and reaches it only after channel losses, further reservoir evaporation downstream, and whatever the operating rules do with it in between. We report a change in one reservoir's mass balance and call it suppressed evaporation, which is what it is. Treating it as usable yield at a demand node would require a routing study we have not done, and would reduce the Upper Basin numbers rather than raise them.
7. The export bound is the assumption that most constrains the answer, and a developer could buy their way out of it. Every array here is sized against whatever the dam's existing line is not already using. That is the right frame for a project that wants to avoid capital it cannot recover, and it is the wrong frame for a developer willing to pay for network upgrades, a different point of interconnection, or a dedicated line. Nothing in this analysis says FPV fails under a paid-for interconnection. It says FPV does not fit inside the interconnection that already exists, which is a narrower and cheaper claim. The coverage bounds should be read as bounds on the free option, not on the technology.
8. Several value streams are counted at zero. The array is credited only with energy sold at the day-ahead price and, where storage is present, capacity. It receives nothing for resilience, local reliability, deferred distribution investment, reduced algal growth, or the capacity value of the solar itself. Each is real and none is large enough to close a tenfold gap on cost per acre-foot, but a reader should know the tally is deliberately narrow rather than complete.
9. The design space searched is narrow. One module density, one mooring concept, full-shade geometry. Designs that trade energy density for water value, wider float spacing, partial shading, or hybrid shade-and-generate structures, are not represented and could shift cost per acre-foot in a direction this model cannot see.
10. Two thousand draws is enough for the cost figure and marginal for the capacity one. Re-running the whole Monte Carlo under three additional seeds moves Lake Mead's median cost per acre-foot by about 1% and its P90 by about 4%, but moves the median right-sized capacity by about 6%. The cost conclusion is not sampling noise. Any capacity figure quoted to three significant digits is, which is a further reason to read the interval rather than the median.
11. The weather year and the price year are not the same year. Hourly solar output comes from a 2015 radiation year and dam discharge from 2015 gauge records, while prices are 2024. The diurnal shape of both survives that, but the pairing of any particular cloudy day to any particular price day is arbitrary, so the real correlation between weather and price is absent. We can bound what it is worth: replacing the 2024 prices with their hour-of-day means, which destroys day-to-day variation entirely, moves Lake Mead's capture price from $18.88 to $20.80 per MWh, about 10%. A correctly paired year would probably sit below the shape-only figure rather than above it, because low-output days are high-price days and the anticorrelation is what a same-year pairing would capture. Treat capture prices as carrying roughly a tenth of slack from this alone.
12. Evaporation suppression is treated as linear in coverage. Heat not lost as latent flux mixes into the bulk water and raises evaporation on the surrounding open surface, making suppression sub-linear at high coverage. The magnitude is unmeasured at reservoir scale.
13. The Upper Basin price transfer rests on a summer window. Flaming Gorge and Blue Mesa prices come from a June-July 2026 hour-of-day ratio applied to a full year. Price shapes are seasonal, so this misprices winter by an unknown amount. Both external reviewers identified it as the weakest step in the analysis. It cannot be resolved until those markets have published a full year.
14. Upper Basin evaporation rates are unmeasured. Flaming Gorge and Blue Mesa prices are shape-transferred from a partial local record rather than measured over a full year, because their markets did not produce day-ahead prices before 2026.
15. Storage dispatch has perfect intraday price foresight, which flatters storage.
16. Solar receives no firm capacity credit, which is conservative rather than correct.
17. No investment tax credit is applied, and this is correct rather than conservative. The One Big Beautiful Bill Act, enacted 4 July 2025, terminates the section 48E credit for solar placed in service after 31 December 2027, preserving it only for projects that began construction on or before 4 July 2026. No reservoir floating-solar project met that deadline: none has a federal surface lease, a completed environmental review or an interconnection study, and none could reach commercial operation inside 2027 against a development timeline of five to ten years. Applying a 30–50% credit would assume support the law no longer offers this class of project. Revenue is therefore merchant energy plus a capacity credit on storage, with no power purchase agreement premium.
18. Interconnection cost is modelled; wheeling is not applicable. Transmission access charges in these markets are billed to load rather than to generators, and per-reservoir nodal pricing already carries congestion and losses, so no separate wheeling charge applies. Generator interconnection facilities and local network upgrades do apply and are included at Berkeley Lab's median of $30/kW for completed projects, with $167/kW and $194/kW available as the skewed upper cases.
19. The conservation benchmark is heterogeneous. Published conservation prices cover fallowing, efficiency and delivered transfers, which differ in reliability, shepherding losses and permanence. The comparison supports an order-of-magnitude ranking, not a like-for-like price.
20. Institutional feasibility is assumed, not assessed. No leasing path currently exists for private solar on federal reservoir surface, and most of these surfaces are National Recreation Areas.

8. Conclusion

Coverage targets in the 15–20% range exceed the export-limited bound at every reservoir studied, and the bound varies by more than tenfold between sites, so a basin-wide coverage figure is not a useful specification. Floating photovoltaics cannot be justified at these reservoirs as a water measure: the cost per acre-foot of suppressed evaporation exceeds purchased agricultural conservation by roughly an order of magnitude under every parameter combination sampled, and no alternative cover technology closes that gap. What remains is a hypothesis rather than a finding, and we state it as one because two independent reviewers pointed out that we had never tested it. Declining reservoir elevation is removing hydropower capability from these dams, and that leaves room on lines that were built for full plants. Solar can occupy some of that room. Whether it should is a different question from the one this paper answers, and answering it needs things absent here: a representation of what the plant can still produce as a function of head, a capacity and ancillary-services market rather than energy alone, and storage sized against what hydro actually provides. Until that work exists, the honest statement is that reservoir solar is not a water project, that its economics are least bad where load sits on the shoreline, and that whether it is a good energy project at these sites is untested. Establishing whether any particular site is buildable additionally requires interconnection studies that are not publicly available.

8b. Every headline number, and how far to trust it

Two reviewers working independently on measurement and verification said the same thing about this work: it is strongest on transparency and weakest on auditability. Every number is disclosed somewhere, but a reader has to read everything to audit anything. This table is the answer. It is generated from the model outputs rather than typed, so it cannot drift from what the model actually produces.

The status column is the important one. Measured means an instrument recorded it. Derived means arithmetic on measured inputs. Modelled means this work produced it and it inherits every assumption behind it. Inferred means we reasoned to it from something adjacent and could be wrong. Only two rows here are measurements, and one of the model's own inputs is openly unresolved.

9. Reproducibility

All inputs are public. Monte Carlo results are seeded. Code:

Two things about rerunning this that are easy to get wrong. The full-year 2024 price archive is an input rather than an output: it is regenerated with fetch_upper_basin_overlap.py --rebuild-base, deliberately outside the normal run, because a re-fetch moves every published price without anyone deciding it should. And the weather fetch will refuse to write if any reservoir fails or if a result is non-physical, after an episode where rate limiting produced a valid-looking file containing no reservoirs at all.

References

1. Bureau of Reclamation, Upper Colorado Region. Reservoir daily data. usbr.gov/uc/water/hydrodata. Daily pool elevation, storage, total release, inflow and bank storage, per site. The source of every surface area, the dam-output shape at six reservoirs, and the bank-storage term that half-closed the water balance.
2. Bureau of Reclamation. RISE. data.usbr.gov. Daily elevation and storage for Lake Havasu, which the Upper Colorado service does not carry.
3. NASA Langley Research Center. POWER daily agroclimatology (MERRA-2 based), power.larc.nasa.gov. Temperature, humidity, wind, shortwave radiation and precipitation at each reservoir's own coordinates, 2015–2026. Replaced an ERA5-based archive that reached its quota; the two disagree on the aerodynamic share by three points and both figures are reported.
4. Penman, H. L. Natural evaporation from open water, bare soil and grass. Proceedings of the Royal Society A 193 (1948) 120–145. The open-water combination equation used here, as distinct from FAO-56 reference ET, which assumes a grass canopy.
5. McJannet, D. et al. Comparison of techniques for estimating evaporation from an irrigation water storage. Water Resources Research 49 (2013). Open-water evaporation method comparison.
6. Global lake evaporation by the Penman method with an equilibrium temperature approach. Journal of Hydrometeorology (2025). Reports combination equations without a heat-storage term overestimating by 24–36%, which is the behaviour seen here and the reason evaporation levels come from measured flux rather than from the physics.
7. US Geological Survey and US Environmental Protection Agency. National Hydrography Dataset, hydro.nationalmap.gov. Surveyed water polygons with area attributes; the measured lower bound on canal surface, and the check that showed mapped water bodies are full pool rather than operating level.
8. US Geological Survey National Water Information System. Daily discharge at the main tributary gauges above Powell, Flaming Gorge, Blue Mesa and Navajo. Used to state the gauged fraction of reservoir inflow.
9. McKuin, B. et al. Energy and water co-benefits from covering canals with solar panels. Nature Sustainability 4 (2021) 609–617. The canal-solar evaporation-savings estimate this work compares against.
10. Bureau of Reclamation. Lower Colorado River Mainstream Evaporation and Riparian Evapotranspiration Losses. 2023. LCRAS and HDB datasets; reservoir areas and evaporation volumes (Tables 7–10).
11. Earp, K. and Moreo, M. Evaporation from Lake Mead and Lake Mohave, Nevada and Arizona, 2010–2019. US Geological Survey Open-File Report, 2021. Eddy-covariance and energy-balance flux measurements.
12. Moreo, M. and Swancar, A. Evaporation from Lake Mead, Nevada and Arizona, March 2010 through February 2012. USGS Scientific Investigations Report, 2013.
13. Western Area Power Administration. Colorado River Storage Project regional fact sheet. 2026. Plant capacities and average annual energy.
14. Bureau of Reclamation. Hoover Dam frequently asked questions: generation and capacity.
15. Upper Colorado River Commission and Bureau of Reclamation. Reservoir Evaporation Project (Flaming Gorge, Blue Mesa, Navajo), in progress.
16. European Commission Joint Research Centre. Photovoltaic Geographical Information System (PVGIS), PVGIS-NSRDB radiation database.
17. Copernicus Sentinel-2 surface reflectance, accessed via Google Earth Engine.
18. OpenStreetMap contributors. Reservoir water polygons, retrieved via the Overpass API.
19. California Independent System Operator. OASIS day-ahead locational marginal prices and APNode atlas (queryname ATL_APNODE), 2024 and 2026.
20. Southwest Power Pool. RTO West day-ahead locational marginal prices by settlement location, 2026. Colorado River Storage Project settlement points.
21. US Geological Survey National Water Information System. Discharge below Glen Canyon Dam, gauge 09380000.
22. Los Angeles Department of Water and Power. Los Angeles Reservoir shade-ball deployment (2015) and floating-cover cost comparison.
23. Ramasamy, V. and Margolis, R. Floating Photovoltaic System Cost Benchmark. National Renewable Energy Laboratory, 2021.
24. Gadzanku, S., Lee, N. and Dyreson, A. Enabling Floating Solar Photovoltaic Deployment: Exploring Hydropower-Floating PV Hybrids. NREL/TP, 2022.
25. National Renewable Energy Laboratory. Cost Projections for Utility-Scale Battery Storage, 2025 update.