What predictive control means and how it works

Predictive control (MPC) is a predictive control method that optimizes heating automatically based on electricity price, weather forecast and your home's thermal properties. This page explains how it works and when to use it.

Predictive control in brief

Predictive control looks ahead, predicts price and temperature, and computes a heating plan that keeps comfort with minimum cost.

You can also use the Predictive control AI assistant to review settings and suggest improvements.

Predicts: electricity price, outdoor temperature and indoor temperature evolution.
Computes a plan: when to heat more and when less while staying within comfort limits.
Controls devices: pushes the computed plan to compatible devices.

Benefits of predictive control

Lower electricity bill without constant tweaking.
Stable comfort temperature.
Fewer manual rules — mostly target and allowed range.
Learns how your home heats and cools over time.

Learning modes — what are they?

Learning modes help you start by collecting data and validating predictions before full predictive control.

Learning (beta)

Learns from history and controls automatically.
Good quick start.
Beta: first weeks may require monitoring.

Learning (shadow)

Learns and predicts but does not control.
Best risk-free start.
Switching to predictive control later is easy.

Troubleshooting (common issues)

Problem: prediction is clearly wrong

check sensor placement
check sampling and gaps
start with Learning (shadow)

Problem: predictive control cannot create a plan (limits too tight)

widen comfort limits temporarily
check heating power
use 2R2C for high inertia targets

Problem: oscillation or unstable control

increase min ON / min OFF
widen comfort range slightly
verify power is realistic

Recommendation: how to start

For most users, the safest path is to start with Learning, confirm the sensor and data quality, and only then move to MPC.

Practical path

Days 1–3: Learning (shadow) Ã¢â‚¬â€ collect data without control.
Days 3–14: Learning (beta) Ã¢â‚¬â€ optional controlled learning.
After week 1–2: Predictive control Ã¢â‚¬â€ more predictable control that respects your limits.

If you already have good data and correct sensor placement, you can start directly with predictive control.

Decision tree: Learning or predictive control?

Start with Learning (shadow) if:

little or no history
uncertain sensors or sampling
you want predictions without control

Start with Learning (beta) if:

you want easy start and quick benefit
you can monitor early weeks
comfort range and sensor are reliable

Move to predictive control if:

you want predictable control with clear limits
comfort is critical
you have enough data

Checklist before switching to predictive control

Sensor: placed sensibly (no drafts/sun/near heater).
Sampling: regular updates with no long gaps.
Heating power: matches reality.
Comfort range: realistic min/max.

Control modes: predictive control vs Learning — detailed comparison

Feature	Learning (shadow)	Learning (beta)	Predictive control
Controls device?	No (prediction only)	Yes	Yes
Best use	Safe start and data validation	Quick start while data accumulates	Precise, predictable control with clear limits
Needs good data	Yes	Yes	Yes
Explains ?why?	Partly	Partly	Yes
Start recommendation	First step	Second step	Once data and basics are correct

Most important practical difference

Learning is a soft start to validate measurements and gather data.

Predictive control is usually the best final state for stable comfort and clear limits.

Setup in practice (very detailed)

1) Devices and requirements

Controlled load: e.g. radiator, floor heating, water heater.
Controller device: Shelly switch/socket controlling the load.
Temperature sensor: recommended per zone.
Measurements: ensure temperature and (if needed) power are visible.

2) Sensor placement

Place the sensor to measure real room comfort.

Height: about 1.1–1.5 m.
Distance: not on/above heater.
Avoid drafts: not near doors/windows/vents.
Avoid sun: no direct sunlight.
Living area: where you want comfort.

Common mistake: Typical mistake: sensor too close to heater → system thinks room is warmer than it is.

3) Sampling and data quality

Learning and predictive control need regular data.

Temperature: 5–15 min recommended; 30 min may work for slow targets.
Gaps: long gaps reduce quality.
Power: if used, should be reliable during ON periods.

4) Heating power

Set power as close to reality as possible.

Radiators: nominal power from label.
Multiple heaters: sum powers.
If unsure: start slightly low and validate.

5) Comfort range (min/max)

Too tight limits can make it impossible to create a plan.

Start wider: e.g. 20–23°C.
Tighten later: once model matches.
Night/day: lower night minimum to save.

How predictive control works step by step

1. Set target and range. For example 22–23°C.
2. System learns behavior. How fast it heats/cools and power.
2b. Choose model (RC or 2R2C). 2R2C separates air and thermal mass.
3. Fetch price and weather prediction. Typically 24–48h ahead.
4. Compute the plan. Minimize cost while staying within limits.
5. Push schedule to devices. Often also as offline schedule.
6. Update continuously. Recompute when conditions change.

2R2C thermal model — when and why

2R2C models room air and thermal mass separately, useful for high inertia heating.

When to choose 2R2C

floor heating or high thermal mass
clear delay between air and mass
you want preheating in cheap hours

What parameters mean (simplified)

R air↔mass: how quickly room air exchanges heat with structures. Slower exchange usually means more delay.
R mass↔outdoor: how easily heat escapes outdoors through the building. In practice: insulation level.
C air / C mass: how much heat air and structures can store. More mass reacts slower but can be steadier.
Heating power: zone heating power. If it is far from reality, predictions and control may be off.

Tip: start with presets if unsure and validate using preview.

RC vs 2R2C quick comparison

RC: simpler and often enough for radiators.
2R2C: more accurate for high inertia and preheating.

Example: floor heating

Predictive control preheats during cheap hours and reduces during expensive hours while staying above your minimum.

Example: water heater

Predictive control fills the tank during cheap hours but can still heat if needed to maintain comfort.

When should you use predictive control?

Comfort-critical heating
High price volatility
You don't want complex manual rules

Identification test (excitation)

Run a controlled ON/OFF test to produce clear data and improve thermal model identification.

Read the excitation guide

Comfort schedule constraints

Define time-based min/max limits and let MPC use them directly in optimization.

Read the comfort schedule guide

Predictive control FAQ

Answers to common questions are available in the FAQ under "Predictive control and temperature".

Go to FAQ