What is the target for the MPC algorithm?

[T1983M]

Hej all,

i am running the MPC alorithm for some days now together with my heat pump. The Setup is with Sonos TRVs with Direct Valve control and it is generally working like a charm.

But to what goal does the MPC optimize?
To keep the temperature stable by heating in bursts or by setting the valve % so that it just stays steady state? With my heatpump i prefer if all the radiators and the floor heating are moving between 40% and 60% valve opening. if they are higher consistently i will change my heating curve to adjust it back, the same if they are lower. I try to aim for a steady stae in the Heatpump and the valves

Now i am observing that some rooms are in (or close to) this steady state, while others open the valve for 10ish minutes every hour.

Basically all rooms carry a Type 22 radiator that behave simmilar, yet in the Workaway one i get these bursts....
Is that local Maximums the MPC optimizes towards. Is there a way to motivate it to go for more open valves. Or is this more a PID behaviour i should look for? I also rather have it a bit to warm / cold and have the valves open/close more gradual....

[cl445]

Great question. MPC ("Model Predictive Control") in BT works on 5-minute steps with a 30-minute prediction horizon (MPC_STEP_SECONDS=300, MPC_HORIZON_STEPS=6, see utils/calibration/mpc.py).

The cost function it minimizes is roughly:

cost = Σ_t [ overshoot² · 8.0  +  setpoint_deviation²  +  control² · 0.05  +  Δcontrol² · 0.2 ]

With the defaults:

• mpc_overshoot_penalty = 8.0 — overshooting the target is penalized 8× more than a normal deviation
• mpc_control_penalty = 0.05 — small penalty against valve effort (energy)
• mpc_change_penalty = 0.2 — moderately penalizes valve-position changes (avoid flapping)

Practical consequence: MPC strongly prefers continuous moderate heating over bursts, because bursts create overshoot risk (8× penalty) and require larger valve changes. For a heat pump this matches the most efficient operating regime — heat pumps prefer steady moderate output over on/off bursts.

The predictor uses the learned heating_power and heat_loss_rate (from utils/thermal_learning.py) plus a solar-gain estimator to predict the temperature trajectory over the next 30 min, then picks the valve sequence that minimizes the cost.