On paper, scheduling is simple: decide when, where and in what order production runs. In a refinery or chemical plant, that decision collides with physics - intermediates that cannot wait, units that cannot pause, cleaning rules that dominate the calendar - and becomes one of the hardest optimisation problems in industry. Here is what scheduling in the process industry actually does, how the puzzle is solved, and what decides in practice whether it works.
In Brief
- Scheduling decides when, where and in what order production runs - in a process plant it collides with physics that discrete manufacturing rarely faces.
- The number of possible schedules is astronomical (a 19-digit figure for just 20 runs on one unit), so it is solved with a mix of exact methods, heuristics, metaheuristics and increasingly AI.
- Planning, scheduling and the mass balance each see the same physical stream differently - the same word rarely means the same thing.
- Most scheduling projects fail not on the solver but on the data it trusts: no shared definition of stream, tank and unit.
Part of the series , The refinery optimisation triad
- Scheduling in the process industry, explained - you are here
- Why refineries count in tonnes: the mass balance, explained
- One data backbone, 25 years (series centrepiece)
- When the crude tanker doesn’t arrive: how fast can a refinery replan?
The discipline
What is scheduling in the process industry?
Scheduling is the short-term, detailed planning of plant operations: it determines when, where and in what order products, batches and movements actually run - on which units, using which tanks and lines, in which sequence. The typical horizon is the next few hours to days; typical outputs are transfer schedules, unit programmes and tank assignments.
It sits between two sibling disciplines. Planning decides what to produce, how much and at what economics, over weeks to a month. The mass balance closes the loop by reconciling process measurements and laboratory data into what actually flowed. In short: planning is the target, scheduling the execution, the mass balance the actual result. So far the textbook. Now the plant floor.
A plan is a promise. The schedule is how you keep it. The mass balance tells you whether you did.
The daily questions
The four questions a schedule answers every day
- Sequence. In what order do the different products run on a given unit - and what does each transition cost?
- Resources. Which reactors, vessels, filling lines or tanks are assigned to which production lot or campaign?
- Changeovers. How are set-up and cleaning times minimised? A unit often needs elaborate cleaning before the next product may start.
- Storage and availability. When do intermediate tanks run full, and when must raw materials, additives or lab releases be ready?
The industry connection
Why is scheduling harder in the process industry than in discrete manufacturing?
Because in a plant, unlike an assembly hall, you cannot simply put the work-in-progress on a shelf. Process production adds constraints that discrete manufacturing rarely sees:
| Discrete manufacturing | Process industry | |
|---|---|---|
| Work in progress | Countable parts, storable on a shelf | Flows and batches; intermediates often perishable or tank-limited |
| Interrupting a job | Usually possible | Often impossible - reactions and continuous units cannot simply pause |
| Changeover | Retooling a machine | Cleaning, flushing, contamination rules, quality release |
| Process types | Mostly one paradigm | Continuous and batch processes must be synchronised with each other |
| Grouping logic | Lot sizes | Campaigns; similar products bundled to cut cleaning effort and cross-contamination risk |
Shelf life alone changes everything: an intermediate that must be processed within hours removes whole families of otherwise attractive schedules. Add the coexistence of continuous and batch operations, plus campaign planning, and the space of feasible schedules becomes as narrow as it is immense.
The mathematics
Why is plant scheduling so hard to compute?
Because the number of possible schedules explodes. Sequencing just 10 product runs on a single unit already allows more than 3.6 million orderings; with 20 it is over 2.4 quintillion – a 19-digit number. A real plant adds dozens of units, shared tanks and pipelines, cleaning and contamination rules, shelf-life limits, personnel availability and shipping dates, all at once. Scheduling belongs to the class of NP-hard problems: as the plant grows, finding the provably best schedule quickly demands more computing time than anyone has.
Under hard constraints that may never be violated, a plant schedule tries to improve several objectives at once - and they routinely pull in different directions: total duration, changeover and cleaning time, utilisation of the expensive units, adherence to customer and shipping deadlines, and how briefly intermediates sit in tanks.
The methods
How are plant scheduling problems solved in practice?
Since computing every option exactly takes too long, practice works with a spectrum of methods - increasingly in combination:
| Approach | Strength | Limit |
|---|---|---|
| Exact methods (MILP) | Provably optimal solutions | Hit time limits as plants and horizons grow |
| Heuristics and rules of thumb | Fast, robust, easy to explain | Good results, rarely the best ones |
| Metaheuristics (genetic algorithms, simulated annealing) | Near-optimal schedules in short time, at plant scale | No optimality guarantee; tuning required |
| AI and machine learning | Predicts durations and disturbances; supports fast rescheduling | Learns from plant data - and inherits every flaw in it |
Which method is better depends on plant size, planning horizon and how much time a decision may take. Many operators use both: an exact method for the structural questions, fast heuristics for the daily rescheduling that every unit trip, delayed vessel or off-spec batch forces. What AI adds in a plant context is primarily the ability to explore a much broader range of scenarios instead of the two or three a manual review allows. But every one of these capabilities consumes plant data, and inherits every inconsistency in it.
The real bottleneck
Why do mathematically perfect schedules fail in the plant?
Because a solver is only as good as the data it optimises over. A schedule built on a wrong tank level, a stale laboratory analysis, or a unit whose real availability lives in someone’s spreadsheet is optimal on paper and infeasible in the plant. When the same material carries three different definitions in planning, scheduling and the ERP, the solver optimises three slightly different definitions at once. The outcome is always the same: the schedulers stop trusting the tool, correct it by hand - and drift back to Excel.
The root of that definition problem is structural. In a process plant, the same physical stream lives in all three optimisation disciplines at once - but each looks at it for its own purpose:
One physical stream - three legitimate views
| Planning sees | a target: how much of the stream should flow this month, and at what economics |
| Scheduling sees | a sequence: which movements, between which units and tanks, in what order today |
| Mass balance sees | a result: what actually flowed, reconciled against measurements and inventories |
None of these views is wrong. Each describes the same physical stream for its own purpose, at its own level of detail. The trouble begins when nothing connects them - and the same stream quietly becomes three unrelated objects.
The limiting factor of most scheduling projects is not the solver. It is the data the solver believes.
This is the quiet sequence behind every optimisation success in the process industry: data maturity first, then optimisation, then AI. The foundation beneath the solver keeps values that share one source rather than three - aggregated definitions of materials, tanks and units across planning, scheduling, mass balance, laboratory and the ERP, plus a time dimension that keeps plan-versus-actual comparisons valid over the years. What such a foundation looks like after 25 years in operation: One data backbone, 25 years.
FAQ
Frequently asked questions
Can process plants use scheduling tools built for discrete manufacturing? Only to a point. Generic tools tend to assume work-in-progress can be stored and jobs can be paused - assumptions a plant breaks daily. Process scheduling needs first-class support for shelf-life limits, tank capacities, cleaning and contamination rules, campaigns, and the mix of continuous and batch operation.
Is scheduling the same as an APS or MES system? No. Scheduling is the task; APS (advanced planning and scheduling) and MES (manufacturing execution) are categories of software that support it. Many plants schedule with specialised refinery tools, some with APS modules, and more than a few - for better or worse - with spreadsheets. The task stays the same regardless of the tool.
How often does a schedule change? Daily as a rule, hourly when needed. Every unit trip, delayed vessel, off-spec batch or late order can invalidate the current sequence - which is why fast, trustworthy rescheduling matters more in practice than finding the one perfect schedule. Providing the shared foundation that makes rescheduling trustworthy is exactly what the boTec DataBridge is built to be.