Production Scheduling: Plan, Reality, and the Gap

What is production scheduling?

Production scheduling is the process of deciding, for every resource in a plant, what will run on it, in what sequence, for how long, starting when, and in what order. It is the translation layer between a customer order arriving in the ERP and a machine actually cutting, moulding, filling or stamping a part. When it works, the plant produces the right products in the right sequence at the right time with minimum changeover loss. When it doesn't, the plant produces the wrong things in the wrong order and the shift leader spends half the day firefighting.

I have been commissioning the systems that feed production schedules since 1989 — Simatic S5 warehouse management and material-flow control in the early years, OPC UA and IoT gateways feeding cloud-MES today. What has not changed in 37 years of this work is the single most important observation about production scheduling: the schedule leaves the planner's office, and within an hour on the floor, reality has diverged from it. The plants that handle this well and the plants that struggle are separated by exactly one capability: how cleanly reality flows back up to the scheduler.

The four planning horizons — stop conflating them

The first source of confusion in nearly every scheduling conversation I have with a new customer is that "production scheduling" is being used as one word for four different things. They live at different horizons, run on different tools, have different owners, and fail in different ways. Getting them apart is the prerequisite for any sensible scheduling decision.

Each of these layers can be done well or badly independently of the others. The common failure is that a plant buys an APS tool to solve what is actually a dispatching problem, or tries to schedule at MPS granularity what requires finite capacity APS, or — most commonly — runs all four layers in a single Excel spreadsheet that has no idea what is happening on the floor. The rest of this article is about the fourth problem and how to fix it without buying anything expensive.

The reality gap: the single most important observation about scheduling

After 37 years of walking into plants with their APS printouts in hand and then walking onto the actual floor, I can tell you what will be true 19 times out of 20. The schedule says machine 4 is running order 114293 at 11:00. The machine is running order 114301. Order 114293 is sitting in the buffer because material for 114301 arrived first. Order 114304 has been moved up because the customer called. The scheduler does not know any of this yet.

Within the first 45 to 90 minutes of any shift, the gap between the planned schedule and the actual floor behaviour is already 15 to 30 percent of sequence positions. This is not a failure of the algorithm. It is the normal physics of a real plant. Stops happen. Quality issues happen. Material shortages happen. Operators make sensible local decisions based on what they see in front of them. The floor is not ignoring the schedule out of malice; the floor is adapting faster than the schedule can.

The one rule I've learned in 30+ years of this work: the quality of a production schedule depends far more on the quality of the feedback from the floor than on the sophistication of the algorithm. An average scheduling engine with real-time order status from every machine beats a state-of-the-art AI scheduler running on yesterday's data. Every single time. This is why plants that spend 80 percent of their scheduling budget on the APS tool and 20 percent on the feedback loop get worse results than plants that do the reverse.

The three data qualities finite scheduling actually requires

An APS tool that promises finite-capacity scheduling can only deliver on that promise if three specific data qualities are in place upstream. Missing any one of them turns finite-capacity scheduling back into infinite-capacity fiction with extra steps. This is the honest checklist every plant evaluating an APS or scheduling module should run before signing anything.

1. Real capacity, not nameplate. The scheduler needs to know that machine 4 runs at 86 percent of its nameplate on steady state with 71 percent availability and 4 percent reject, giving an effective throughput number that is 57 percent of the OEM datasheet. Nobody gets this from the OEM. It only exists in historical OEE data. Without real capacity data, the APS allocates orders to the number on the datasheet — which is also known as over-booking.

2. Real changeover matrix. The full n×n matrix of changeover times between every product pair, captured from real production events, not from the estimated numbers that have been in the ERP master data since 2008. Changeover sequencing is the single biggest lever in most scheduling engagements, and the quality of the sequence depends entirely on whether the matrix is honest. Automatic setup-event capture from MDE/BDE is the only path to a real matrix.

3. Real-time order status, bidirectional. When order 114293 starts on machine 4, the ERP knows within seconds. When it stops, the ERP knows. When it finishes, the ERP knows the actual quantity and the actual time. Without this loop, the APS is re-planning every night based on where orders were 14 hours ago, which in a fast-moving plant is nowhere useful. This requires proper ERP-MES integration, and it is where most "scheduling improvement projects" quietly die.

The changeover matrix — where the real money is

Ask any shift leader which order of products produces the least setup time on a given line and you will get an answer. Check it against the actual matrix captured from 12 months of MES event data and you will usually find the answer is 15 to 25 percent off optimum. Not because the shift leader is wrong — because human memory cannot hold an honest n×n matrix for a line running 40 products, and the intuition that says "we always run A before B" was true in 2019 and has not been true since they changed the tool design.

The single most reliable win in my scheduling engagements is this: capture the real changeover matrix from MES data, feed it into whatever scheduler the plant uses, and let the sequence optimise against it. The output gain is typically 3 to 8 percent on a changeover-heavy line, without touching anything else, without buying a more expensive scheduler, without retraining operators. It is the definition of the cheap win, and almost nobody has the matrix because nobody has captured the setup events cleanly enough.

Five scheduling anti-patterns

Over 30+ years of commissioning these systems across the food, beverage, automotive, wood and chemical industries, five failure patterns account for most of the scheduling projects that disappoint. Naming them usually prevents them.

1. The daily-reschedule trap. The APS reruns every morning with the latest data and spits out a fresh schedule. By 10:00 the schedule has already been superseded by reality. By day 4, nobody on the floor trusts the schedule because they know it will change tomorrow. The schedule becomes advisory at best, decorative at worst. Fix: longer planning freeze for the next 24-48 hours, continuous rescheduling only for the horizon beyond.

2. The frozen-schedule trap. The opposite failure. The schedule is computed weekly and treated as gospel for the full seven days. By Wednesday it is fiction, but nobody will admit it because the process says the schedule is frozen. The shift leader runs a parallel unofficial schedule in their head. Fix: explicit rescheduling triggers (major stop, big order change, material delay) rather than a fixed cadence.

3. The Excel shadow schedule. The official schedule lives in the APS. The real schedule lives in the shift leader's Excel. The real-real schedule lives on the whiteboard in the production meeting. Three versions of truth, guaranteed to disagree. Fix: one schedule of record, visible on the floor terminals, editable by those with authority, flowing back into the ERP automatically.

4. The infinite-capacity ghost plan. The MPS was computed by MRP assuming infinite capacity. The result is technically feasible on paper and physically impossible on the floor. Every week the plant misses the schedule and every week everyone blames execution. Fix: finite-capacity APS in the short horizon, even a simple one, so the schedule at least respects real machine capacity.

5. The AI scheduling pilot on top of paper. The plant buys a machine-learning scheduling engine because a conference speaker recommended it. The engine is genuinely clever. It is also fed by manual data entry updated once per shift. The algorithm's best-case decision is based on data that was already six hours old when it arrived. Fix: fix the feedback loop first, add the AI after — in that order, never the reverse.

A real case: Klocke Pharma packaging

Klocke Group is an international contract manufacturer in pharma, cosmetics and nutritional supplements, with a packaging operation in Weingarten that runs a heavy mix of blister, sachet and ampoule formats. Scheduling in this kind of environment is textbook difficult: many short-run orders, heavy format changeovers, strict GMP compliance, and — the variable that makes it interesting — a plant floor with essentially no LAN infrastructure to the machines, because the packaging lines predate modern industrial networking.

The engagement started at a single line. Within three weeks it had scaled to every line at the Weingarten site. The connectivity pattern is the one I use on almost every retrofit: digital-I/O gateways tapping existing machine signals for part counts and stop events, no PLC intervention, no production interruption, no cabling project. The ERP side was a Navision integration through a file interface — not the most elegant protocol in the world, but entirely sufficient for what was needed: the order status and master data come down from Navision, the machine cycle events and stop reasons get mapped back to the production orders in the MES, and the schedule-to-reality loop closes cleanly within minutes instead of shifts.

The measured outcomes after the first weeks of operation:

• 7 hours of additional production time per week, recovered from stops and changeovers that were now visible and actionable
• 12 % improvement in output, driven by sequencing decisions that now had honest changeover data to work from
• 8 % improvement in availability, through structured action on the top stop categories visible in real time

None of this required a new APS tool. It required the one thing every scheduling system depends on and most plants never actually build: a clean, cheap, real-time feedback loop from every line back to the order system. Once that was in place, the existing scheduling process immediately produced better sequences — because it was finally working with real data instead of estimates.

FAQ

What is the difference between MPS, APS and production scheduling?
MPS (Master Production Schedule) decides which end items in which weeks, lives in the ERP, and assumes infinite capacity. APS (Advanced Planning & Scheduling) decides which orders run on which machines in which sequence over 1-4 weeks, respecting real capacity and constraints. "Production scheduling" colloquially covers both plus real-time dispatching on the floor. They are not interchangeable. Treating them as one thing is the most common source of scheduling project failure.

Do we need an APS tool, or is MES-based scheduling enough?
For most mid-market discrete manufacturers with up to 200-300 active orders per week, MES-based scheduling with finite capacity awareness and a clean changeover matrix is enough. Dedicated APS tools pay off in high-complexity environments — highly mixed product portfolios, deep BOM explosions, strong multi-stage dependencies, or when the scheduling problem genuinely becomes combinatorial. The honest test: is your current scheduling poor because the algorithm is weak, or because the data going into it is weak? In 80 percent of cases, it is the data.

How often should production schedules be recomputed?
There is no universally correct answer, but the working rule I use is: freeze the next 24-48 hours (so the floor can trust it), reschedule continuously beyond that horizon based on incoming reality, and explicitly trigger a reschedule on material-level disruptions rather than on a fixed daily cadence. Continuous rescheduling with a rolling freeze window is the pattern that combines planning stability with adaptability.

What is schedule adherence and why does nobody measure it honestly?
Schedule adherence is the percentage of orders that ran on the planned machine, in the planned sequence, within the planned time window. It is the single most revealing scheduling KPI and almost nobody measures it because it requires comparing the planned schedule (from the APS) against the actual event stream (from the MES) order by order. Typical honest schedule adherence in a mid-market plant is 40-65 percent. Plants that are above 80 percent have almost invariably invested in the feedback loop, not in the scheduling algorithm.

How do we handle scheduling when orders change constantly?
Two things help. First, a rolling freeze window (24-48 hours) where orders do not move, and an open horizon beyond where changes are allowed. Second, automated priority rules (customer tier, due date slack, changeover cost) so that urgent incoming changes can be slotted without manual replanning of the whole sequence. The goal is not a perfect plan; it is a plan that degrades gracefully under change.

What about AI-powered or autonomous scheduling?
Genuinely useful in 2026 — but almost always in plants that already have clean data. Machine-learning schedulers add real value when real-time machine state, accurate changeover matrix and honest order status are already flowing. Without those three inputs, an AI scheduler just produces slightly more confident fiction. In my engagements over the past three years, every successful AI-scheduling pilot was preceded by 12-24 months of data-layer work. Every unsuccessful one started with the AI.

Can old machines without modern interfaces be included in a real-time scheduling loop?
Yes, and the belief that they can't is the most common reason scheduling feedback loops never get built. A machine from 1995 without any digital interface can almost always be connected through digital-I/O gateways tapping existing cycle signals and stop-reason contacts, without PLC intervention, without production interruption, within a few hours per machine. I have never yet encountered a packaging line, press or CNC that genuinely could not report cycle counts and stop events. The Klocke case in this article is an entire plant connected this way.

How does SYMESTIC support production scheduling?
Real-time bidirectional integration with SAP, Infor, proAlpha, Navision and Microsoft Dynamics so order status flows back from the floor to the ERP continuously; automatic capture of changeover events and durations to build an honest n×n matrix; finite-capacity awareness derived from live OEE data; shopfloor-terminal dispatching so operators see the current sequence and the impact of deviations. Plus the gateway library (OPC UA, MQTT, digital-I/O) that connects any generation of machine — including the ones that predate industrial networking. See SYMESTIC Production Planning.

Related: MES · APS System · MRP · Production Planning · Machine Data Acquisition · BDE · OEE · SYMESTIC Production Planning

About the author

Martin Brandel

MES Consultant & Project Lead at SYMESTIC. 30+ years in industrial automation and MES integration — Simatic S5 warehouse and material-flow control from 1989, paint-line commissioning in Eastern Europe and China with Hermos AG, and since 2000 at SYMESTIC building the machine connectivity that feeds the cloud-MES platform today. Since 2019 leading end-to-end MES projects from first enquiry to go-live. Dipl.-Ing. Nachrichtentechnik. · LinkedIn