Fuel tank

Fuel that leaves the tank without the engine burning it is the most expensive litre on the site.

The twin knows how much the generator should burn from its load. Set that against what left the tank and the gap is fuel you can’t account for. The catch is the sender: cheap level senders read low and jump in steps, so we calibrate before we raise a flag.

500 L horizontal cylinder, resistive level sender, refuel by truck

Start with one cluster of sites See how we prove results

Capacity: 500 L
Litres per step, half full: 9.1 L
Smallest loss we can show: 24 L a month
Days to empty when full: 80 days

Anatomy and nameplate

1Level sender. 63 steps of 10 mm. Read through a tested strapping table.
2Generator load. Gives the burn the tank should show.
3Refuel events. Matched against the delivery record.
4Tank temperature. Optional. Diesel grows about 1% for every 12 °C.

Fuel tank nameplate
Tank
Volume	500 L
Shape	Horizontal cylinder, 700 mm diameter × 1300 mm long
Pickup	30 mm above the bottom
Refuel	Order at 20% left, fill to 95%
Sender
Type	Resistive reed chain with float
Resolution	10 mm steps, 63 steps over the travel
Dead bands	25 mm at the bottom, 40 mm at the top
Litres per step	5.1 L near the bottom, 9.1 L at half full

Telecom site power chain

The grid or the generator feeds the site through the transfer switch. The rectifier makes −48 V DC for the radios and keeps the battery charged. The air conditioner keeps the battery cool.

Air
AC power
−48 V DC
Diesel
Has its own twin page

The physics

The equations behind every curve on this page.

Volume from level

V = L · [r²·acos((r − h)/r) − (r − h)·√(2rh − h²)]

A round tank holds few litres per centimetre near the bottom and top, many in the middle. A tested strapping table replaces it when the sender is not straight.

Expected burn

Σ F(P) · Δt

From the generator’s own fuel curve and its load, every 15 minutes.

Unexplained loss

ΔV_measured − burn − refuel

What is left after the engine and the truck are accounted for.

Days to empty

V_usable / average daily burn

5.9 L a day at the reference site.

Level against load

The expected level only falls while the generator runs, by what its load burns. Anything else that moves the level shows up as a gap.

Tank level against the level the load explains

Reference site, July

Expected from generator load
Sender, tested strapping table

Left the tank without being burned: 45 L
Generator at the time: Stopped
Flagged: At the next 15-minute reading

The expected line falls only while the generator runs, by what its load burns. A drop with the engine stopped is fuel that left the tank without being burned. Your team confirms why.

loss = ΔV_measured − Σ F(P)·Δt − refuel

The sender, and refuel planning

Before we put a number on a gap we check the sender against a metered fill. After that, the same level and burn tell the truck which sites to visit.

Sender reading against the tested strapping table

63 steps of 10 mm, dead bands at the bottom and top

Tested strapping table
Read as a straight percent of 500 L
Cylinder formula, sender taken as full height

Litres per step, half full: 9.1 L
Worst error, straight percent: 17 L

A round tank holds few litres per centimetre near the bottom and top and many in the middle. The sender also moves in steps. Read it as a straight percent and the tank looks like it is losing fuel when it is not.

V = L · [r²·acos((r − h)/r) − (r − h)·√(2rh − h²)]

Days to empty across a cluster of 40 sites

Today’s level and each site’s own burn rate

Days to empty

Sites to put on the next truck: 3 of 40

Each site burns at its own rate. Days to empty, not the calendar, decides which sites go on the next truck.

days = V_usable / average daily burn

Failure modes

What breaks, what it looks like, what the twin does about it.

Failure mode	The physics	Signals that show it	How the twin catches it
Unexplained drop	Fuel leaves the tank with no load to explain it: a leak, a loose fitting or fuel drawn off.	Level falls with the generator stopped, or faster than its load explains.	Shows fuel that left the tank without being burned, how much and when. Your team confirms why.
Sender fault or bad calibration	A stuck float, a broken reed or a straight-line setting on a round tank.	Level steps that do not match the strapping table. Readings frozen through a long run.	Checks every step against the tested strapping table before any gap is reported.
Water in fuel	Condensation and bad deliveries settle water at the bottom, where the pickup draws.	Water sensor, where fitted. Rough running after a refuel.	Raises a drain-and-check work order when a water reading appears.
Refuel recorded but not received	The delivery note says more than the tank gained.	Level rise at the refuel smaller than the recorded volume.	Sets the measured rise against the delivery record for every refuel.

What the twin does not do

The twin shows fuel that left the tank without being burned, and when. It does not say why. That is for your team to confirm.

How we catch it early

A physics twin and a PINN, working on one fuel tank.

The twin is a set of equations for this machine, tuned to its nameplate. Give it the flow, speed and conditions and it returns what every reading should be. It always gives the same answer for the same inputs.

A physics-informed neural network, or PINN, sits on top. It is trained on the plant’s own data and it is penalized whenever it breaks a physics rule. So it can only learn what the equations leave out: wear, fouling, friction. It cannot invent an answer that the machine could never produce.

Step 1
Measure
Flow, pressure, current, temperature, vibration. Whatever the asset already has.
Step 2
Physics twin
The equations of this machine, set to its nameplate. It says what the readings should be right now.
Step 3
Residual
Measured minus expected. Demand, weather and speed are already taken out.
Step 4
PINN
A small neural network that learns the cause of the gap, under the physics rules.
Step 5
Finding
A named cause, a date, and the kWh or hours at stake. An engineer reviews it.

loss = (measured − predicted)² + λ · (physics residual)²

The second term is the rulebook. It is zero when the network obeys the equations.

Why this beats black-box AI

A black-box model learns patterns from history. It needs examples of each failure, and it goes quiet when the pump runs at a point it has not seen. The twin starts from physics, so it works on day one. The PINN adds what physics cannot know about this particular machine.

That is also what makes the savings countable. The baseline is a calculation, not a guess, so the gap between baseline and meter is a number a customer can check.

What the twin checks on this asset

Unexplained loss above 2 sender steps plus 3% of burn
Level steps consistent with the strapping table
Refuel volume against the delivery record
Days to empty against the 5-day truck lead time

	Fixed alarm	Black-box AI	Twin + PINN
Needs failure history to learn	No	Yes, lots	No
Says what is wrong	No	Rarely	Yes, by name
Works at a flow it has never seen	No	Often fails	Yes, the equations still hold
Puts a number on the saving	No	No	Yes, in kWh and dollars
Lead time on slow wear	Late	Varies	Weeks to months
Every number can be traced	Yes	No	Yes, to an equation

What the twin puts a number on

Energy and cost, from the equations.

Smallest loss we can show

24 L a month

Two sender steps (9.1 L each at half full) plus 3% of the month’s burn

Refuel trips avoided

4 a year per site

10 on a calendar sized for the worst month, 6 when trucks go on days to empty

Straight-percent error

up to 17 L

Reading the sender as a straight share of 500 L, against the strapping table

What we do for you

We watch it. We catch it. We prove it.

Our reliability engineers run the twin on your fuel tank, review every finding, and help your crew fix it. After the repair we compare the meter to the baseline and write the result in the ledger.

Start with one cluster of sites See the 90-day pilot

More assets

More from telecom towers, then the lead asset of each other industry.