Predict Far Ahead, Keep Integer Decisions Near Term

Separate the prediction horizon from the integer horizon: retain a long look ahead while relaxing distant on/off decisions to reduce computation.

Operating chillers or electrolyzers requires looking far enough ahead to account for future demand and storage levels. Yet choosing every on/off decision as an integer over that entire horizon can make optimization expensive. Can we retain the long view while shortening the interval over which we search for on/off combinations?

Nurkanović’s study of horizon-tail relaxations offers an intuitive approach: predict far ahead, but concentrate integer decisions on the near future. Allow continuous relaxations of distant on/off decisions, then recompute the plan when the next measurement arrives.

1. Separate the prediction horizon from the integer horizon

Call the full look-ahead length the prediction horizon N, and the initial interval that retains integrality the integer horizon M. Within M, each on/off input must be either 0 or 1.

Consider three chillers controlled at three-minute intervals with a 90-minute prediction horizon. Here N is 30. Keeping every on/off decision binary requires 30 decisions per chiller, giving 90 binary variables in total.

With M set to 5, each chiller must be either on or off during the first 15 minutes. Over the remaining 75 minutes, its input may take any value between 0 and 1.

Formulation First 15 minutes Remaining 75 minutes
Integrality over the full horizon On/off decisions are 0 or 1 On/off decisions are 0 or 1
Relaxed tail integrality On/off decisions are 0 or 1 Continuous values between 0 and 1 are allowed

The controller still looks 90 minutes ahead, but the number of freely optimized binary variables falls from 90 to 15. Continuous variables and constraints remain over the full horizon, so runtime need not fall in the same proportion. The reduction targets the combinatorial search.

2. Keep the distant future in the calculation

Shortening the prediction horizon itself to 15 minutes would make it harder to account for demand or temperature changes beyond that point. Tail relaxation keeps the distant costs, dynamics, and constraints in the approximate problem while relaxing integrality there. The distant future still influences the current decision.

A future chiller input might therefore be 0.4. This is a relaxed value allowed for computational purposes; it is not automatically equivalent to running the equipment for 40% of that interval. The first input actually applied comes from the initial integer interval, so it is on or off.

MPC does not execute the entire plan at once. It applies the first input and optimizes again using the next measurement. As time advances, a previously relaxed interval enters the near-term integer horizon. This repeated planning structure reduces the need to search over every distant on/off combination at every update.

The appealing idea is to treat N and M as separate design choices. How far ahead should the controller look? How far ahead should it retain binary on/off decisions? Those lengths can be chosen independently. The paper also studies fixing the tail to a previously obtained binary plan; the focus here is on relaxing integrality while leaving future operating choices open.

3. Reduce computation after the measurement arrives

The paper combines this idea with advance computation. Before a measurement arrives, the controller solves a continuous relaxation at a predicted state and prepares a mixed-integer quadratic program (MIQP). After the measurement, it updates the initial state and solves the MIQP to choose near-term on/off decisions together with continuous inputs.

In a three-chiller simulation with N = 30 and M = 5, the continuous-tail variant computed feedback 15.56 times faster on average than full AS-MINMPC, which retains integrality over the entire horizon. Closed-loop cost increased by 0.27%. This comparison illustrates the effect of combining long predictions with a short integer horizon within the same experiment. The reported time is the feedback computation after the measurement arrives. Paper, §V

Reference

Nurkanović, A. (2026). Real-time nonlinear mixed-integer model predictive control with horizon-tail relaxations. arXiv, 2611.09187v1. Paper · DOI · Code