Forecasting Consumption From the Wrong Variable: A Cryogenic Storage Example

There is a forecasting pattern worth stealing from an unusual domain: laboratory liquid nitrogen supply.

The naive model is demand driven. Samples frozen per month, multiplied by nitrogen per operation, equals consumption. It fails badly.

The real model is dominated by an idle term. A vacuum insulated vessel loses nitrogen continuously whether anyone touches it or not. That baseline, the static evaporation rate, is measured with the vessel closed and undisturbed under defined conditions. Actual consumption is always higher, because every lid lift and every warm object introduced adds heat the measurement deliberately excludes.

consumption = static_boil_off(vessel, ambient) + handling_overhead(access_events)

In a busy store the first term dominates. Halving throughput barely moves the bill, which is exactly the counterintuitive result you get whenever a system's cost is driven by a standing rate rather than a transactional one.

Two operational consequences

1. Size the delivery interval from measured litres per week in situ, not from the throughput figure. Vendor evaporation rates are laboratory conditions, not your room.
2. Log per vessel, not per site. A rising trend at unchanged workload is the earliest practical indicator of vacuum degradation, and it shows up months before any visible symptom. Aggregate site level logging hides it.

Where else this shape appears

Idle cloud spend, standing energy load, warehouse refrigeration. If your cost is mostly standing, optimising the transactional path is theatre.

Full write up: https://cryolab.co.uk/liquid-nitrogen-supply/