Flood-water capture and aquifer storage on the Somerset Levels
Working note v3. 2026-08-07.
A working note on the water-accountancy model. The current statement of this work is Banking winter water for a dry summer.
A water-accountancy assessment on open data
Working note v3. 2026-08-07.
1. The question, and what this model is
Somerset’s rivers carry far more water in winter than anyone can use, and too little in the summers when irrigation needs it. The question this work addresses is narrow and quantitative:
If flood-season water were diverted from the lower Parrett system into an off-line storage cell, treated, and injected into an aquifer, how much could be recovered for summer irrigation - and what would it do to river flows?
What the model is. A daily water-volume accountancy model. It tracks every cubic metre through a fixed chain - river, diversion pump, surface reservoir, treatment, injection, aquifer store, recovery pump, irrigation
- over 6,025 days from 2010 to mid-2026, on open data throughout (Environment Agency gauged flows and rain gauges, British Geological Survey hydrogeology and borehole index, Environment Agency LiDAR, Natural England designations). Mass balance closes to less than one cubic metre across the entire record; that is checked on every run and is the model’s only guarantee.
What it can say. How much water is available above a licence threshold; how much a scheme of a given size could capture, store and return; how reliability responds to the size of each component; and which component limits the outcome.
What it cannot say. It is not a hydraulic model - no channel routing, no flood levels, no travel times. It is not a groundwater flow model - the aquifer is a bucket with a leakage rate, not a simulated head field. It contains no water quality, no ecology, no costs, and no ground investigation. Wherever a number below rests on an assumption rather than a measurement, the sentence that uses it says so.
2. What we know about the water
The contributing area was delineated from a 50 m digital elevation model rather than assumed. The Parrett and Brue/Huntspill systems together drain 2,174 km² above their coastal outlets.
| km² | |
|---|---|
| Contributing area (Parrett + Brue/Huntspill) | 2,174 |
| Gauged by nine Environment Agency stations | 905 |
| Ungauged remainder, modelled | 1,329 |
The nine gauged records are measurements. The ungauged 1,329 km² is a modelled estimate from a soil-moisture and quick-flow runoff model driven by Thiessen-weighted rain gauges, with regional parameters transferred from the eight gauged catchments it was calibrated against. It reproduces those catchments with a median Nash-Sutcliffe efficiency of 0.75, and - importantly, because the diversion trigger sits high on the flow-duration curve - reproduces the 12.5% exceedance flow to within 4%. It is about 5.5% light on total volume, which is the conservative direction.
One gauged catchment, Somerton, was excluded from those regional parameters: its delineated catchment is 73% smaller than its published area, so its fitted parameters absorb a geometry error rather than describing catchment behaviour.
Divertible resource. With a diversion trigger at the 12.5% exceedance flow (31.6 m³/s on the nine-gauge sum) and a hands-off flow at the 19% exceedance (22.2 m³/s), 2,318 Mm³ passed the licence test over the 16.5-year record - about 141 Mm³ a year. That is the resource. It is roughly forty times what the scheme below actually takes.
[figure not carried over: map_system.png]
Figure 1. The system: terrain, river network, flow gauges, delineated catchments and the hydrogeology. Shows where the water is measured and where the candidate aquifers outcrop. Does not show designations, land use, or anything about the scheme itself.
3. What we know about the ground
Two candidate storage settings emerged from the geological screen. They are not two versions of the same thing; they behave differently and deliver different products.
Aquifer A - local sandstone, 4-12 km. Highly-productive Triassic sandstone with significant intergranular flow, outcropping over 240.7 km² within the study area. The British Geological Survey describes it as a principal sandstone aquifer up to 600 m thick, yielding up to 125 L/s, “hard but of good quality, becoming saline beneath the confining Mercia Mudstone”.
The borehole evidence is unusually good. Of 3,324 BGS borehole records in the search area, 948 lie within 2 km of the highly-productive outcrop - 245 directly on it - and every one has a scanned log available. The deepest on the outcrop reaches 1,153 m.
Its leakage rate is the one aquifer parameter here with an observational basis: 0.0167 per day, a 41-day half-life, fitted from Environment Agency borehole recession records. Water put into this formation does not stay long.
Aquifer B - distant chalk, 42 km. A confined chalk body, assumed to behave as a classic aquifer-storage-and-recovery target: leakage 0.00005 per day, three hundred times tighter than A. Water put here stays.
The gap that matters more than either. The BGS “Aquifer Properties” layer holds 64 sites in the study area, 26 of them on the highly-productive Triassic - but it is a site index only. It publishes an identifier, a site name and a grid reference. It contains no transmissivity and no storativity values, and no related table, attachment or open API supplies them.
So of the four aquifer numbers the model depends on - storage capacity, leakage rate, recovery efficiency and injection rate - three are assumptions with no measurement behind them, and the fourth (A’s leakage) rests on recession fitting rather than a pumping test. Section 6 returns to what that costs.
4. What the scheme can do
First, a correction to the reservoir
Earlier versions of this model carried a surface reservoir of 5,000,000 m³ over 350,000 m² - a cell 14.3 m deep. On the flat ground near sea level where such a cell would sit, that geometry is not achievable, and it was not a harmless simplification: the model computes evaporation from water-surface area, so an impossibly small area understated evaporation throughout.
A LiDAR terrain screen of the buildable area - after excluding SSSI, SAC, SPA, Ramsar and settlement - produced real candidates. The best, S1, holds 5 Mm³ as a 1.690 km² cell 2.96 m deep, retained by 1.27 km of embankment. A second, S2, holds the same volume over 2.871 km² but needs 4.71 km of embankment because it sits on flat moor with no natural relief to help retain it.
| earlier assumption | S1 | S2 | |
|---|---|---|---|
| surface area at capacity | 0.350 km² | 1.690 km² | 2.871 km² |
| mean depth | 14.3 m | 2.96 m | 2.00 m |
| evaporation over the record | 2.83 Mm³ | 8.55 Mm³ | 17.92 Mm³ |
| as a share of water captured | 2.2% | 6.5% | 13.4% |
The correction costs the scheme three to six times more evaporation than was previously assumed, and for S2 evaporation becomes the second largest loss in the whole system. Reservoir area is now derived from the measured terrain curve and cannot be specified independently of capacity.
[figure not carried over: stage_curve_S1.png]
Figure 2. Stage-volume and stage-area for candidate S1, with the 5 Mm³ design point marked. Shows how much water the terrain holds at each water level and the surface area that comes with it. Does not show foundation conditions, peat depth, seepage, land availability, or anything about whether the site could be consented.
Reliability, measured as shortfall rather than as a percentage
“Percentage of demand met” is not a decision variable. Nine per cent shaved evenly off every year is absorbed; the same nine per cent concentrated into one drought is a lost crop. The results below are therefore reported as the distribution of shortfall.
All three options use the same sited reservoir (S1), the same inflow, the same period and the same total recovery pumping.
Full record, 2010-2026:
| A alone | B alone | A + B | |
|---|---|---|---|
| worst single year, deficit as % of that year’s demand | 84.1% | 81.8% | 59.1% |
| Jun-Aug deficit | 45.6% | 32.4% | 9.2% |
| longest continuous deficit | 158 days | 178 days | 121 days |
| baseflow returned to river | 62.7 Mm³ | 0.0 Mm³ | 29.0 Mm³ |
| (demand met, for continuity) | 62.4% | 71.1% | 91.4% |
The driest three consecutive years in the record - 2015-2017, identified from the flow data rather than chosen:
| A alone | B alone | A + B | |
|---|---|---|---|
| worst single year deficit | 84.1% | 0.0% | 18.6% |
| Jun-Aug deficit | 71.4% | 0.0% | 3.7% |
| longest continuous deficit | 158 days | 0 days | 72 days |
| baseflow returned | 8.5 Mm³ | 0.0 Mm³ | 4.0 Mm³ |
| (demand met) | 40.4% | 100.0% | 93.8% |
Two things in that pair deserve emphasis. The portfolio is much the best over the record as a whole - but in the drought, the confined chalk alone is perfect and the portfolio is not. In a dry sequence you want banked water, and giving half the recharge to a store with a 41-day half-life wastes it. B alone, however, returns nothing to the river in any year.
One confound, stated: A and B carry different assumed capacities (10 and 20 Mm³), so A+B also has more total storage than either alone. All three capacities are assumptions.
[figure not carried over: portfolio_timeseries.png]
Figure 3. The portfolio in operation. Shows the surface cell cycling annually, the fast sandstone store never accumulating, the tight chalk store filling over fourteen years and then saturating, and irrigation shortfall concentrated in 2010-12 and 2017. Does not show water quality, head distribution within either aquifer, or what happens after 2026.
The trade-off between the two products
Splitting recharge between A and B trades one product against the other, but not linearly.
| recharge share to A | demand met | baseflow to river |
|---|---|---|
| 0.00 (all to chalk) | 71.6% | 0.0 Mm³ |
| 0.25 | 93.7% | 14.2 Mm³ |
| 0.50 | 91.5% | 29.0 Mm³ |
| 0.75 | 81.7% | 45.3 Mm³ |
| 1.00 (all to sandstone) | 62.4% | 62.7 Mm³ |
Baseflow rises steadily with the share sent to the sandstone. Reliability does not fall steadily: it peaks at a 25% share and is worse at 0% than at 25%. An all-chalk scheme is not the reliability optimum, because a store that fills and stops cycling stops being useful.
[figure not carried over: share_sweep.png]
Figure 4. Irrigation reliability and baseflow against the nominal recharge share, with and without the ability to redirect unused share between stores. Shows the shape of the trade-off and the gap that operational flexibility is worth. Does not show cost, and the delivered share differs from the nominal share wherever a store is saturated.
5. What was learned
Capture is set by the scheme, not by the river. Across configurations in which the divertible resource ranged from 1,111 to 5,605 Mm³ - a fivefold range - the water actually captured stayed within 91 to 131 Mm³. Doubling the measured river while holding licence stringency constant reduced capture, because the trigger rises with the flow and diversion then occurs on fewer days. Pump utilisation sits at 7-8% throughout. Finding more water in the catchment does not help a scheme that is already limited by its own plant.
Recovery capacity is the irrigation lever; injection capacity is a baseflow purchase. Increasing recovery pumping from 15,000 to 240,000 m³/day moved reliability from 66% to 93%, with a sharp knee at 60,000 m³/day beyond which nothing changes at all. Quadrupling injection capacity over the same model moved reliability by less than one point while raising baseflow 2.5-fold. (Both sweeps predate the reservoir geometry correction, so the levels shift slightly; the directions and the location of the knee do not.)
Confined and unconfined settings deliver different products, not better and worse versions of one. On identical inputs, a confined store returns 0 Mm³ to the river and meets ~98% of demand; an unconfined store returns 64.5 Mm³ and meets ~64%. Neither is the right answer without first deciding which product is wanted.
Flood attenuation is negligible, and the bound is arithmetic. Across the twenty largest flood events in the record, the median reduction in peak flow is 1.37% and the largest is 2.19%. The ceiling is fixed: diverting at the licensed 260,000 m³/day removes 3.01 m³/s from a river peaking at 120-240 m³/s. On top of that, six of the twenty events began with the cell already more than 90% full, because the flows that fill it are exceeded long before the peak arrives. This is a water-resource scheme with incidental flood benefit, not a flood scheme, and it should not be presented as one.
A saturated, under-cycled store degrades. The chalk fills over fourteen years and then sits full, refusing further recharge. Directing all recharge to it recovers only 60.4 Mm³ of the 109.7 Mm³ the scheme could otherwise inject - the rest is simply not accepted. Storage that does not cycle stops earning.
6. Risks, and what would falsify this
The dominant risk, by a wide margin, is the confined/unconfined assumption. The same scheme, on identical inputs, meets 98% of demand if aquifer leakage is 0.00005/day and 64% if it is 0.0167/day. That single unmeasured parameter moves the answer further than every sizing decision in this note combined - and as Section 3 sets out, there is no open measurement of transmissivity or storativity for either candidate. The BGS aquifer-properties dataset that would settle it publishes locations only.
Everything below is secondary to that.
- The reservoir sites are terrain screening only. No ground investigation, no peat depth, no seepage or permeability testing, no land ownership, no agricultural land classification, no archaeology, no flood-risk sequential test. A high rank means the ground is the right shape.
- S2 sits on peat, with founding, seepage and buoyancy questions that are not assessed. S1 sits on mineral ground above the moor and largely avoids them - one reason to prefer it beyond its embankment efficiency.
- A third of the inflow is modelled, not measured. The 1,329 km² ungauged contribution comes from a calibrated model that is ~5.5% light on volume.
- The evaluation contains one severe drought (2011), and it falls in the period when the stores are still filling from empty. Deficits in 2010-12 are therefore part drought and part start-up transient, and the record is not long enough to separate them. Sixteen years is thin for an asset with a sixty-year life.
- Recovery efficiency is assumed at 0.6 with no site basis; it scales deliverable yield roughly linearly.
- LiDAR covers 58% of the study extent. The unscreened remainder removes about 15 km² of otherwise-eligible ground, all of it rising ground in the south-east - the same setting that produced the best candidate.
- Roads were not screened. No open main-road dataset was obtainable without a commercial key. A road can be diverted at a cost; this belongs in costing, not siting, but it has not been done.
What would falsify the case. A pumping test on the local sandstone returning a leakage rate at the observed end of the range, together with a recovery efficiency materially below 0.6, would leave the scheme unable to bank water to August and reduce it to a baseflow-support proposition. Conversely, a confined response would make the irrigation case straightforward. Both outcomes are consistent with what is currently known, which is the problem.
7. Recommendations
The value of information exceeds the value of further optimisation. Sizing anything more finely is premature. The spread between the confined and unconfined cases is larger than the spread across every capacity choice examined here, and it can be closed by measurement rather than by modelling.
In order:
- Commission a pumping test on the local Triassic sandstone, sited on the highly-productive outcrop, reporting transmissivity, storativity and a recovery test. This single measurement resolves the dominant uncertainty and determines whether the scheme is an irrigation proposition, a baseflow proposition, or both.
- Read the borehole logs already identified. 948 records lie within 2 km of the target outcrop, all with scanned logs, ranked by depth and by whether BGS also holds them as water wells. This is desk work, it costs nothing but time, and it will establish sandstone thickness and the depth of the confining Mercia Mudstone before any drilling is commissioned.
- Ground-investigate the two reservoir candidates. S1 and S2 are the two ends of a build-cost against operating-cost trade-off - 1.27 km of embankment and 13 m of lift, against 4.71 km and none - and S2 additionally carries 110% more evaporation. Neither can be costed without knowing what is under it.
- Extend the flow record backwards as far as the archive allows. The design case for storage is the drought, and this record contains one, awkwardly placed at the start.
- Only then size the scheme. When sizing does happen, the evidence here says to concentrate on recovery capacity, which has a clear knee, and to treat additional injection capacity as a purchase of river support rather than of reliability.
The scheme is not obviously unviable, and the portfolio result - reliability close to the confined case while returning substantial water to the river - is the most promising finding in this work. But it rests on an aquifer characterisation that does not yet exist, and no amount of further modelling will supply it.
Provenance
All data open: Environment Agency (gauged flow, rainfall, groundwater levels, LiDAR composite DTM, source protection zones), British Geological Survey (hydrogeology 625k, onshore borehole index, aquifer properties index), Natural England (SSSI, SAC, SPA, Ramsar), Office for National Statistics / Ordnance Survey (built-up areas). Open Government Licence v3 except where stated. Everything is EPSG:27700, volumes in m³, fluxes in m³/day. Mass balance closes below 1 m³ on every run reported.