Peak load costs money — plant managers have known this since demand charges exist. What few know: classic load management and AI control are two different things. One cuts, the other avoids. In mid-sized manufacturing, the difference now decides over six-figure amounts per year.
We see in projects that many plants already have load shedding installed — and still regularly run through the expensive peak. The reason is simple: by the time the load shedding reacts, the peak is already in the metering interval. The utility sees the maximum value, the demand charge rises for a year.
What classic load management delivers — and where it ends
Classic load shedding works threshold-based. You define a value — say 800 kW — and the system switches a consumer off as soon as the current load approaches the threshold. This works mechanically reliably and has been standard for decades.
Three weaknesses become visible in practice. First, the intervention is reactive: when the threshold approaches, something is switched off — often in the middle of a process. Second, the intervention is binary: either the machine runs fully or not at all. Third, the system never learns: the tenth peak of a week is treated exactly like the first, although the pattern has long been known.
In plants with uniform load, these weaknesses are uncritical. In plants with volatile load — typical in food production, print shops, metalworking — they cost measurable production. Anyone who hears the load shedding more than twice per shift knows what we mean.
What learning models do differently
At its core, AI control is a forecast plus a control instruction. The model receives historical data — consumption, shift schedule, outdoor temperature, order mix, plant states — and learns patterns in it. From this it forecasts the next 15 to 60 minutes of load in operation and gently decides whether and which consumer should be throttled.
Three effects stand out in the data we see from projects:
- Earlier reaction. Where load shedding reacts at 95 % of the threshold, AI control already starts to quietly counteract at 70 %.
- Gentler interventions. Instead of switching a machine off completely, it runs at reduced speed for 8 minutes. The peak is flattened, not chopped off.
- Learning effect over time. After three to six months, the model knows the recurring peak triggers per shift and can predict them before they even arise.
On average, in the plants we support, we see 15 to 30 % less peak load at the same production. For a mid-sized plant with 1.5 million kWh per year and a demand charge of €100 per kW, this reduces the annual grid costs by mid five-figure to low six-figure amounts.
What stays with the plant manager — and why that is good
An honest message first: the AI supports the plant manager instead of replacing him. It takes no responsibility for critical machines. It is the additional layer that coordinates the non-critical consumers and warns the plant manager in time when an intervention on a critical machine becomes necessary.
We build the control in three levels:
- Non-critical + automatable: air conditioning, ventilation, charging processes of the wallboxes, heat pump buffer storage. The AI controls independently — the plant manager sees in the dashboard what happens.
- Critical + plannable: welding systems, dryers, cold stores with a setpoint temperature corridor. The AI gives recommendations — the plant manager decides per shift what he releases.
- Critical + unplannable: machines tied to orders, safety-relevant units. The AI leaves them alone — it only warns and suggests alternatives.
This separation is the decisive point. An AI that controls everything fails at the first audit. An AI that knows clear limits and sticks to them becomes a tool that shift supervisors trust.
What you need to get started
The most important question in the first conversations is how clean the data connection is, long before the choice of algorithm. Without a clean data foundation, no model runs.
Minimum requirement in practice:
- Load profile in 15-minute intervals for at least 6 to 12 months
- Shift schedule and order mix from the ERP or MES system (at least for the main production lines)
- Outdoor temperature and weather forecast — most peak drivers are temperature-correlated
- Plant states from the PLC (programmable logic controller — the “computer” of a plant) — for the non-critical consumers, switching and power data are sufficient
If these data sources lie continuously on one platform, AI control is the next logical step rather than the hard one — the energy twin from our digital twin examples is exactly this platform. We implement this with eco2lot so that the platform first measures, then optimizes, then automates. So AI comes at the end of a clean data chain, not at the beginning.
Anyone who starts with AI without a clean data chain builds a glasshouse on sand. Anyone who cleans up the chain first often already has 60 % of the optimization through visibility alone.
Realistic effects — and their limits
AI control is no miracle weapon. There are plants where it brings little: very constant loads, hardly any shiftable consumers, already highly optimized classic load management. Here we often see only 5 to 8 % reduction — that still pays off, but is far from spectacular.
We see large effects, on the other hand, in plants with three characteristics:
- Volatile load due to order mix, shift jumps, air conditioning
- Several flexibly controllable consumers (at least three to five units above 50 kW)
- An energy officer who takes the data seriously and implements the recommendations
In this combination the effects are regularly in the double digits. And they grow over time, because the model gets better and the plant manager learns to trust the system.
How BS-Systeme implements it
We have built the three-stage principle firmly into eco2lot: measuring with our industrial IoT platform (sensor layer), optimizing with our OEE and production management software (evaluation layer), automating with AI control as the top layer. Each stage can run on its own — you can buy them one at a time instead of all three at once.
For plants that are still at the very beginning, we start with measurement. For plants that already have data, we jump directly into optimization. AI control is usually armed after 4 to 8 months of platform operation — when the data foundation is in place and the model has learned enough. This structure saves costly lessons and delivers measurable interim results against which you can check the investment.
What peak load really costs — and where the effect becomes visible
So that the managing director can decide, a second look at the cost mechanics is worthwhile. Peak load hits the electricity bill in three places.
Demand charge. In industrial contracts, the highest 15-minute average power of the year is multiplied by the demand charge. At €100 per kW, a peak of 1,200 kW therefore costs €120,000 — per year, with effect for the next twelve months. A single badly run shift in February can tip the annual result.
Grid fee reductions for predictable load shifting. In Germany, plants that can reliably shift loads out of the grid’s high-load time windows receive discounts on the demand-based grid fee — depending on the grid area and their own consumption structure. AI control can document this shift cleanly and thus secure the recognition.
Balancing effects. With directly marketed electricity (power purchase agreement, PPA for short), peaks additionally affect balancing. Here every deviation from the schedule costs — AI control helps to keep the schedule clean.
We recommend measuring the effect from the start in three figures: the maximum 15-minute peak per month, the hours with loads above 80 % of the previous year’s peak, and the number of avoidable interventions per week. These three KPIs usually reflect reality better than a single percentage.
Where the limits of automation really lie
There are three situations in which AI control acts much more cautiously than marketing sometimes promises.
Short-term order changes. If the shift spontaneously starts up an additional line, the model has no historical reference for the new situation. In such cases it falls back on the stored safety rules — that is intended, but in the specific case it sometimes costs optimization potential.
Safety-relevant machines. Cold stores with a temperature corridor, compressed air storage with minimum pressure, UPS feeds — wherever safety comes first, the AI never controls on its own. It can give recommendations, the final decision is made by the plant manager or a stored safety circuit.
Data gaps. If a meter fails for 14 days or a sensor is miscalibrated, the recommendations are correspondingly uncertain. A good platform detects these data gaps automatically and reduces the depth of intervention during this time — it switches back to classic load shedding mode until the data is reliable again.
In projects we call this the “honest corridor”: the AI is only as good as the data and only as bold as the safety logic allows. Anyone who defines this cleanly in the first workshop ends up with a system that the workforce trusts — and that is the real prerequisite for the double-digit effects.
Two short examples from practice (anonymized)
Food production in Bavaria, 80 employees. Starting point: regular peaks above 1,100 kW, load shedding on three consumers, plant manager with high skepticism towards “AI”. Approach: 6 weeks of measuring without intervention, then 4 weeks of shadow mode, armed from month 4. Result after 9 months: 22 % lower peak load, load shedding permanently deactivated on two of the three consumers, plant manager an active sponsor of the system. The decisive point: shadow mode. It showed the plant manager that the model chooses the same interventions he would have chosen himself — only earlier and more precisely.
Corrugated board plant in Hesse, three lines, 200 employees. Starting point: hard-to-plan order mix, frequent shift jumps, heat pump with a large buffer storage. Approach: connection of the three lines, the heat pump and the two largest dryers to the data platform; model trained on a 12-month load profile plus ERP order data. Result after 7 months: 17 % lower peak load, dryer interventions reduced by 60 %, EnPI “kWh per tonne of board” improved by 9 %. The plant manager describes the result laconically: “The shift notices nothing. The electricity bill does.”