The Budget Question Came Back Reversed: A Fifty-Asset District Cooling Fleet Under Ten Years of Asset Investment Planning

A fifty-asset district cooling portfolio was run through Maximo Asset Investment Planning with a ten-year horizon and a fixed capital target. Five of eight budget levels came back feasible, the requested target was not one of them, and the cheapest plan was not the one with the lowest capital spend.

Share

The Budget Question Came Back Reversed: A Fifty-Asset District Cooling Fleet Under Ten Years of Asset Investment Planning

Every asset owner in district energy has the same budget conversation, and it rarely changes shape. The fleet is getting older, failures are getting more expensive, and next year's capital budget looks like this year's. Which assets do we refurbish, which do we replace, in which year, and what happens to risk if finance cuts the number by ten percent? Most organisations answer that with a spreadsheet, a few engineering opinions and a prioritisation meeting.

On 17 October 2026, Mohamed Chihaoui published a hands-on account of answering it inside Maximo instead. The lab runs Maximo Asset Investment Planning 9.2, on top of Maximo Manage 9.2 with Maximo Optimizer, against a scenario that will be immediately recognisable to anyone operating cooling plant: fifty assets across three district cooling plants, thirty chilled-water pumps of 250 kW and twenty air handling units rated at 40,000 cubic metres per hour. Every screen was captured on a live system, and the full set of scenario outputs is published.

The result is worth reading twice, because the interesting part is not that the optimiser produced a plan. It is that the plan came back with the question reversed. The strategy asked for roughly 350,000 a year of capital with a plus ten and minus twenty percent band. The tool returned eight budget levels, of which five were feasible, and the requested target was not among them. Higher spending was infeasible, not because the money was short, but because every budget level has a floor as well as a ceiling, and 315,000 a year for ten years is more renewal work than this fleet can usefully absorb with the two interventions on offer.

That is a capital planning answer rather than a software answer, and it is the sort of thing a spreadsheet will never tell you, because a spreadsheet does not know what the fleet can physically absorb.

The Fleet, and the Gap That Justifies the Exercise

The scenario is deliberately ordinary. Three plants, named North, South and East, with two asset classes doing the work of keeping them running. Pumps were installed between 1996 and 2019, air handling units between 2000 and 2018. Average ages are 17.1 years for pumps and 16.8 years for the units. Eleven of the thirty pumps and nine of the twenty units are already twenty years old or older.

The money is equally ordinary, which is what makes the case useful. The pump register carries a replacement value of 1.72 million and the air handling units 2.72 million, so replacing everything would cost about 4.4 million. The plants spend about 350,000 a year on capital renewal. That gap, between what a full replacement would cost and what the organisation actually funds, is the problem the tool exists for, and it is a gap that exists in every mature estate regardless of industry.

Data loading follows the standard route. Asset types are added to the ASSETTYPE synonym domain, here CHWPUMP with an internal value of PUMP and AHU with an internal value of FACILITIES, and the fifty assets are loaded through the MXAPIASSET REST API, which is how a real register would arrive. Application import works as well. Each asset carries four fields that the model reads: the asset type, which decides the failure curve and the intervention catalogue; the install date, which produces chronological age; the replacement cost, which the reactive replacement measure multiplies; and the priority, which the effective age formula uses.

Two structural facts about the tool are worth stating plainly, because they contradict the intuition most planners bring. First, AIP is a strategic planning tool, not a work management tool. Manage already answers what is broken today and what preventive maintenance is due next month. AIP looks five to thirty years ahead and asks which intervention, in which year, buys the most value for the money available. Second, all the optimisation data lives at the asset type level, not on individual assets. Assets of one type share a failure curve and an intervention catalogue. What makes two pumps of the same type different is their own data in Manage.

That second point is the one that decides whether a first project succeeds. A register with five hundred asset types and no installation dates is not a planning input, it is a data remediation job.

The Value Framework Is the Actual Deliverable

If there is one section of this lab that deserves to be read by finance as well as engineering, it is the value framework, because it is where qualitative judgements about failure become numbers the optimiser can trade against money.

The framework has four goals, and the template ships with two measures each: reliability, financial performance, employee and public safety, and environmental respect. Every measure is a Maximo formula, and every measure is either a failure cost, meaning what one end-of-life failure costs, or a key performance indicator, meaning how a failure moves an indicator that is scored against ideal, acceptable and unacceptable ranges.

The lab's edited measures show the arithmetic of consequence in a form anyone can argue with. Service interruption risk is set at 40,000 per pump failure and 15,000 per air handling unit failure. Unplanned downtime is 36 hours and 24 hours respectively. Reactive replacement cost uses each asset's own replacement cost with an emergency premium, 35 percent for a pump because of expediting, overtime and temporary chillers, and 25 percent for a unit. Lost time injury risk applies a probability to a cost: one pump failure in twenty causes a lost-time injury costing 250,000, one unit failure in forty does the same. Death or serious injury risk uses a much smaller probability against a much larger number, one in two thousand for a pump and one in five thousand for a unit, at 18 million. Emissions risk multiplies tonnes of carbon, a price per tonne of 150, and the proportion of a failure that releases them, one in ten.

The author is explicit that these figures are illustrative rather than benchmarks, and that they belong in the framework only after somebody has written down where each number comes from: incident history, insurance claims, energy prices, a carbon price. That instruction is the most important sentence in the lab, because the optimiser is only as defensible as the consequences it is given. A plan built on unexamined defaults will be argued down in the first finance meeting and will not be rebuilt.

The template trap is worth its own paragraph, because it is a live risk on any first attempt. The shipped formulas are written for a power utility and reference asset types DTRF345, a transformer, and CB345, a circuit breaker. A typical measure reads as a nested conditional that prices a transformer at a quarter of its replacement cost and everything else at an eighth. Publish that unchanged and every pump and air handling unit in a cooling plant falls into the else branch, priced with consequences chosen for a different industry, and the optimiser will produce a confident, well-presented, wrong plan. Every measure has to be edited before publication, and only a published framework is used by the optimisation.

There is one more discipline hidden in the framework tab: only one framework can be published at a time, and re-publishing replaces it. That has a governance consequence. A framework is a statement of organisational values about failure, so it should be versioned and approved, not edited by whoever happens to be running the tool that week.

Curves, Effective Age and Two Ways to Spend Money

The asset data tab is where engineering judgement enters the model, and it reduces to five numbers per asset type: a failure curve, an effective age formula, an annual maintenance cost near end of life, and a small catalogue of interventions. An asset type becomes ready for optimisation when it has all of them plus at least one replacement.

The curve is a Weibull distribution defined by two parameters, and the lab explains both in a sentence each. Shape describes how the failure rate changes with age: above one it rises, which is wear-out; at one it is constant, which is random failure that age cannot predict; below one it falls, which is infant mortality. Scale is the age by which roughly 63 percent of a population has failed, which the interface rounds to 65 percent.

For the pumps the lab uses a shape of 3.2 and a scale of 25 years, justified by clear wear-out in bearings, seals and impeller erosion against a 25-year design life. For the air handling units it uses 2.6 and 22, justified by wear-out with more spread because coils corrode at different rates. Those are defensible engineering positions rather than measured outputs, which is exactly the right status for them at a first pass, and the lab says where the numbers should come from instead: failure history, manufacturer data, or judgement that can be defended in a room.

Effective age is the more interesting mechanism, because it is where condition data enters the plan. Used raw, the formula is just chronological age from the install date. The lab multiplies by 1.15 for assets at priority one or two, on the reasoning that those pumps run as duty machines while the others stand by. That single multiplier is the pattern that turns AIP from a register report into an engineering model: any attribute can go into the formula, including run hours from a meter, an inspection grade or a Maximo Health score, and it is what makes two pumps of the same age receive different plans.

The annual maintenance cost near end of life, 6,500 for a pump and 9,000 for a unit, is the input that most first attempts omit, because it looks like an accounting detail. It is not. Without it the optimiser cannot see that keeping an old asset is itself a cost that grows with age, so it under-replaces, and the resulting plan looks cheap because half of its cost is sitting in the maintenance budget.

The interventions complete the model. Each has a type, capital and operating cost, an optional linked job plan, a lead time and a downtime, and an effect on effective age: replacement resets to new, refurbishment extends remaining life by a percentage. Pumps get replacement at 58,000 capital and 4,500 operating, four months lead time and 16 hours downtime, against an overhaul at 14,000 and 2,500, one month and 24 hours, extending life by 40 percent. Units get replacement at 135,000 and 8,000, six months and 72 hours, against refurbishment at 32,000 and 4,000, two months and 48 hours, extending life by 35 percent.

Those two options per type are the whole strategic choice. A fleet with only replacements has one lever and a plan that is entirely a question of timing. A fleet with overhauls has a trade-off, and the optimiser's job becomes deciding where a 14,000 overhaul buys more remaining life than a 58,000 replacement.

Eight Budget Levels, Five Feasible

The strategy is the input that defines the question. Ten-year horizon starting in 2027, a capital target of about 350,000 a year with an upper buffer of 10 percent and a lower buffer of 20 percent, and an operating target of 60,000 with buffers of 15 and 30 percent. All fifty assets are added, either individually, through a saved Manage query, which is the route for thousands of assets, or with an add-all action. A strategy holds up to 10,000 assets, and adding them runs in the background.

One design decision inside AIP explains the whole result, and it is easy to miss. A single run does not produce a single answer. The optimiser solves the same problem at several budget levels: unconstrained, the target itself, and steps of five percent up to the upper buffer and down to the lower buffer. With those buffers the run produced eight levels. Each level caps annual capital, and it also sets a floor 10 percent below that cap, so a plan has to spend inside a band every year rather than saving up and spending in one heroic year. Levels that cannot be met are marked infeasible and left out of the scenario list.

For this fleet, five of the eight came back feasible, and the infeasible ones were the target and both higher levels. Running the optimisation for fifty assets over ten years took about ten minutes.

The comparison is the case study's payload. Unconstrained, the model plans 21 replacements and 29 overhauls, 2.19 million of capital over the decade, and a total cost of ownership of 7.79 million. At five percent below target, in a band from 300,000 to 333,000, it plans 28 replacements and 29 overhauls, 2.98 million of capital, and a total cost of ownership of 7.78 million. At ten percent below target, in a band from 286,000 to 318,000, the intervention counts and capital are identical at 28 and 29 and 2.98 million, and total cost of ownership falls to 7.72 million, the lowest of the eight. At fifteen percent below, the same plan again, with total cost of ownership at 7.75 million. At twenty percent below, two replacements drop out, leaving 26, capital falls to 2.79 million, and total cost of ownership rises to 7.89 million.

Then the instruction that separates a practitioner from a tool operator: re-ask the question at a realistic target. A strategy is locked once it has run, so the second question gets a second strategy over the same assets, this time 300,000 a year with a tighter plus five and minus ten percent band. All five of its budget levels came back feasible, and its target plus five percent level, a band of 284,000 to 315,000, reproduced the same answer: 28 replacements, 29 overhauls, 2.98 million of capital, 7.72 million of total cost of ownership.

Plot every feasible level from both runs on one chart of annual capital cap against total cost of ownership and the finance conversation collapses into a single curve. Below about 300,000 a year, every thousand saved comes back as risk cost. Around 315,000 to 318,000 sits the minimum. At 333,000 the cost rises slightly. At 350,000 and above, the yearly spending floor cannot be met at all. That is the picture to take into a budget meeting, because it says something a spreadsheet cannot: there is a floor as well as a ceiling, and the floor is a property of the fleet, not of the finance department.

Unconstrained Is Not a Plan, and Capital Is Not Cost

Three readings of that output matter more than the numbers themselves, and the lab states all three explicitly.

The first is that an infeasible budget level is an answer, not an error. Teams new to optimisation treat a failed level as a data problem to be fixed, and spend a week looking for the bug. Here the failure is the finding: at 315,000 a year for ten years there is more renewal work available than the fleet can absorb, given two intervention types, four and six month lead times, and the outage windows each intervention requires. A capital programme that size would have to be spent on work the plants cannot host, or on assets that do not need it yet. That is a conversation about absorption capacity, and it belongs in the room with engineering rather than in a ticket queue.

The second is that the unconstrained scenario is not a plan. With no budget, the optimiser spends 420,000 in 2027, holds roughly 300,000 to 360,000 a year until 2030, and then almost nothing, ending at 14,000 in 2036. It has the lowest capital spend at 2.19 million and only 21 replacements. No capital programme is funded that way, and the reason becomes visible on the key performance indicator chart: the improvement peaks near 68 percent in 2032 and then falls back as the un-replaced fleet ages again. The levelled scenarios spend more capital and end with a slightly lower total cost of ownership, 7.72 million against 7.79 million, because the additional replacements remove risk cost. Spreading and raising the spend costs nothing over the life of the plan.

The third is that below a point, saving capital costs more. From five to fifteen percent below target the plan barely changes, which is a useful finding in itself: it means the fleet has a wide plateau where the same work can be funded at different levels. At twenty percent below, two pump replacements fall out of the horizon, capital falls by about 190,000, and total cost of ownership rises by about 170,000, because the risk cost of the assets left running is higher. The 190,000 saved now comes back as risk, slightly larger and slightly later.

One last detail in the scenario pages is worth understanding before anyone reads a plan output as a work list. The plan details table lists 81 rows for 57 interventions in the horizon, because the model also plans beyond the horizon in order to price the lifecycle correctly. A typical line shows an air handling unit refurbished in 2028 and replaced in 2040, outside the ten years. Filter by target year to get next year's programme, or the list will look like it disagrees with the counts.

The Access Traps That Will Break a First Attempt

Nothing in this lab is conceptually difficult, and almost all of the friction is administrative, which is why the access section deserves to be read before a workshop is booked rather than during one.

AIP is a Manage add-on with the component key aip. It is activated from the MAS catalog or added to the Manage workspace components alongside Optimizer, and the operator then builds new images and runs an update database. On this system that took about an hour, and Manage was down for part of it. When it finishes, a job deploys the optimisation model to Maximo Optimizer and sets the system properties that connect the two.

The first trap is specific to essential RBAC mode. Ticking the add-on in the MAS admin interface fails with a permission error naming the ManageWorkspace, and that is by design rather than a defect: a cluster administrator adds the component to the workspace specification directly. The second trap is a log message that never appears, so teams wait for a line that will not come. Look instead for the script update process ending and an update database result of zero.

The third trap is authorisation, and it has three layers. No security group can open the application until read access is granted on the application itself. That is not sufficient, because the app is a React front end that reads and writes through eight REST object structures, each secured separately, so all eight have to be granted explicitly. Skip that step and the application opens and then fails with a specific error saying the read action is not allowed on the plan schedule object. And even after everything is granted, a user who is already signed in keeps the old security profile until every session for that user has ended, which on a test system is most quickly resolved by restarting the interface pod.

Two more small traps will consume an afternoon. Creating the first intervention of a type with a double click adds a blank row in the background, and saving then fails complaining that the intervention template number is required. And the application has no classic presentation, so the familiar load-application URL returns an empty page; the correct entry point is the Graphite path.

The practical consequence is that a first AIP workshop needs an administrator with cluster rights, a maintenance window for the add-on install, and a dedicated planner security group rather than administrator rights, booked before anyone starts. None of that is exotic. All of it is invisible in a feature overview.

Practical Implications

Start with one asset type, not fifty assets across five classes. Get a single type to ready, run it, and argue about the output with engineering and finance before adding the next. The first run always produces a consequence value that nobody agrees with, and finding it with one type is cheap.

Treat the value framework as the deliverable and the optimiser as arithmetic. Write down the provenance of every measure, version the framework, and approve it before publication, because only the published framework drives the plan and re-publishing silently replaces it.

Use the effective age formula to carry condition. A formula that is just age wastes the mechanism. Duty cycle, run hours, inspection grades or a Health score in that formula is what separates two identical pumps in the plan, and it is the difference between a register report and an engineering model.

Read the infeasible levels. A budget level that cannot be met is telling you something about absorption capacity, lead times and outage windows, which is a capital planning conversation, not a defect to be fixed.

Publish work orders when the work is approved, not to see what happens. A generated work order commits the intervention, and later optimisation runs plan the flexible work around it, so a speculative work order quietly freezes a decision.

Bottom Line

The most valuable output of this exercise is not a plan. It is the shape of the curve, and the discovery that the fleet has a spending floor as well as a ceiling. A fifty-asset cooling portfolio with an average age around seventeen years can usefully absorb roughly 300,000 to 320,000 a year of renewal, no more; asking for 350,000 does not buy more renewal, it just becomes infeasible. Below about 300,000 the savings come straight back as risk cost.

That is a defensible position to take into a budget negotiation, and it is defensible precisely because the numbers are illustrative and traceable rather than authoritative. Replace the consequence values with your own incident history, put real run hours into the effective age formula, and the same method produces a plan your engineers and your finance team have both already argued over. What the optimiser contributes is consistency: it applies the same logic to all fifty assets and all ten years, which no spreadsheet and no prioritisation meeting manages for long. What it cannot contribute is judgement about what a failure costs, and that remains the asset owner's job, permanently.