Smith Predictor Simulator — Dead-Time Compensation vs PI
Compare in real time the step responses of a standard PI controller and a Smith predictor on a first-order process with dead time. See intuitively how dead-time compensation reduces overshoot and speeds up settling.
Parameters
Process gain K_p
—
Time constant T_p
s
Dead time L
s
PI proportional gain K_c
—
Model error (internal-model mismatch)
%
Shifts the internal model's T_p and L above the real process so you can see the Smith compensation degrade.
The integral time T_i is fixed to T_p by the simplified IMC rule (T_i = 5 s at the defaults). The simulation uses dt = 0.05 s over 0–60 s.
While paused, moving a slider updates the results immediately.
Live Readouts
—
Dead time L [s]
—
Model error
—
Standard PI overshoot
—
Smith predictor overshoot
—
Standard PI settling time (5%)
—
Smith predictor settling time (5%)
Step Response y(t) (real-time plot)
Red = standard PI (with dead time, oscillatory) / blue = Smith predictor (well settled) / yellow dashed = internal prediction y_m (no dead time) / grey dashed = setpoint r=1 / red band = dead-time interval L
Theory & Key Formulas
First-order plus dead-time process. K_p is the process gain, T_p the time constant and L the dead time:
$$G(s) = \frac{K_p}{T_p s + 1}\,e^{-L s}$$
Standard PI control law. K_c is the proportional gain and T_i the integral time:
Smith predictor internal model G_m and the modified feedback signal returned to the PI controller:
$$G_m(s) = \frac{K_p}{T_p s + 1}\,e^{-L s}, \qquad e(t) = r - \bigl(y(t) - y_{m,\mathrm{delayed}}(t) + y_m(t)\bigr)$$
If the internal model matches the real process, the control loop behaves as if it were seeing a process without dead time and the stability margin is recovered. This tool fixes the integral time T_i = T_p (simplified IMC rule).
What Is the Smith Predictor Simulator?
🙋
I often hear the term "dead time" on temperature-control jobs. Why does it make PI control so hard?
🎓
Roughly speaking, when there is a delay of L seconds between acting and seeing the result, the PI controller can no longer judge in real time whether it is "doing too much or too little". If you raise the gain K_c to track faster, it over-corrects during that delay and the loop oscillates. Set K_c to 1.5 and dead time L=3 s in the simulator above: you should see the red standard-PI response wobble.
🙋
You're right, the red line overshoots and oscillates. The blue Smith predictor settles cleanly by comparison. Why such a difference?
🎓
The Smith predictor keeps a "copy of the process" inside. It combines the current output y(t) from the real process with two predictions from the internal model, "with dead time" y_m(t−L) and "without dead time" y_m(t), to rebuild the feedback signal handed to the PI controller. Written out: y(t) − y_m(t−L) + y_m(t). When the model matches the real process, this signal turns into the output of a "process without dead time".
🙋
The dead time disappears?! But in the real world the dead time doesn't go away, does it?
🎓
Right: only the dead time "as seen by the control loop" disappears. The real process still has its L-second delay, and the response cannot start before L seconds. In the simulator, too, the blue line only starts moving at t=3 s. But once it starts, it tracks quickly and without oscillation. The dead time itself cannot be removed, but because the PI "no longer has to see it", a more aggressive gain design becomes possible.
🙋
It sounds like all upside. Are there no drawbacks?
🎓
The biggest weakness is that "the effect degrades when the internal model drifts from the real process". Set the "Model error" slider in the simulator to 30%. The real process stays at T_p=5 s, L=3 s, but the internal model believes T_p=6.5 s, L=3.9 s and compensates accordingly, so the settling time of the blue line stretches from 12.7 s to about 24 s (the PI integral time also follows the internal model, so the red line changes a little too). On site the common case is that operating conditions change the physical properties and shift the dead time L or time constant T_p. That is why the standard practice is to keep the PI gain a little conservative and leave margin against model error.
Frequently Asked Questions
When the dead time L is comparable to the time constant T_p (defaults L=3 s, T_p=5 s), standard PI reaches 30–50% overshoot around K_c=1.5 and settling becomes difficult above K_c=2. The Smith predictor does not oscillate at the same K_c and can in principle run stably up to a larger K_c. Move the sliders in the simulator and compare the two.
With L=0 the Smith feedback signal y − y_m_delayed + y_m becomes y_m + (y − y_m) = y, exactly the same behaviour as standard PI. As L is increased from zero the standard-PI response deteriorates, whereas the Smith predictor keeps its performance as long as the internal model is accurate.
The simplified IMC (internal model control) tuning rule is known to balance overshoot and settling time when T_i is set equal to the process time constant T_p. This tool fixes T_i = T_p to keep the comparison clear. On real plants T_i is tuned independently as well.
Yes, in principle a PID is fine. The Smith predictor is a framework independent of the controller structure, and combining the internal model with a PID controller gives a PID-Smith predictor. However, the main purpose of dead-time compensation is to suppress the destabilising effect of the integral term, so PI usually gives sufficient performance and designs without D are common.
Real-World Applications
Temperature control of long pipelines: In chemical plants and steelworks fluid often travels through long pipes from the heater to the measurement point, producing dead times of seconds to minutes. Standard PI oscillates or responds slowly, so Smith or PID-Smith predictors are widely used. Steam-temperature control, feed preheating and reactor-inlet temperature control are typical examples.
Quality control of paper, film and sheet thickness: In continuous sheet production there is a delay of L seconds between the sheet leaving the rolls and reaching the thickness gauge. Standard PI responds slowly and product variation grows, so a Smith predictor is used to achieve both speed and quality. Continuous casting and rolling are similar.
Composition control and analyser delay: When the top composition of a distillation column or the outlet concentration of a reactor is measured by an online analyser, sampling and analysis add a delay of several minutes. Standard PID cannot avoid oscillation, so Smith predictors or model predictive control such as DMC (dynamic matrix control) are introduced.
Remote control with network delay: With remote control over IoT and communication links, the communication delay L can no longer be ignored and the Smith-predictor idea is attracting renewed attention. It is being studied as a way to compensate communication delay in robot control over the internet, remote supervisory control (SCADA) and swarm control.
Common Misconceptions and Points to Note
The most common misconception is to think that "adding a Smith predictor makes the dead time itself disappear". In reality the physical delay of the process remains, and you still have to wait L seconds after acting before the output starts to move. The blue Smith line in the simulator does not move at all before t=3 s. What the Smith predictor removes is "the dead time inside the loop transfer function", not the start time of the step response. The accurate understanding is "a more aggressive gain design becomes permissible".
The next most common is to underrate the accuracy of the internal model. The Smith predictor compensates dead time fully on the assumption that the internal model G_m matches the real process, but on real plants there is always model error. Using the "Model error" slider to shift the internal model's T_p and L above the real process, the Smith settling time stretches from 12.7 s (0%) to 16.4 s (10%), 23.7 s (30%) and 30 s (50%). When model parameters drift on site the compensation becomes imperfect and can even be worse than standard PI. The standard practice is to set the PI gain slightly below the theoretical optimum to leave margin against model error.
Finally, "adding a Smith predictor to a process with little dead time" is over-engineering. When L is small compared with the time constant T_p (roughly L/T_p < 0.3), properly tuned standard PI already performs well and the added benefit of a Smith predictor is small (with L=1 s and T_p=5 s in this tool, standard PI settles in 7.4 s while Smith takes 10.8 s, i.e. slower). Conversely, the maintenance cost of the internal model and the identification effort rise, so the cost-performance worsens. The Smith predictor shows its true worth on "dead-time-dominant" processes with roughly L/T_p > 0.5. Vary the ratio of L to T_p in the simulator and watch how the performance gap changes.
Set the process parameters with the sliders: process gain K_p (0.5–5), time constant T_p (0.5–20 s) and dead time L (0–10 s). The defaults K_p=1.0, T_p=5 s, L=3 s describe a dead-time-dominant process
Adjust the PI proportional gain K_c (0.1–5, default 1.5) and watch the responses of the standard PI controller and the Smith predictor side by side. The integral time is fixed at T_i=T_p. The "Model error" slider (0–50%) shifts the internal model's T_p and L above the real process so you can see the compensation degrade
The responses to a setpoint step r=1 (red = standard PI, blue = Smith, yellow dashed = internal prediction) are animated, and the overshoot and settling time (±5%) of both are shown on the cards
Worked Example
Consider a temperature loop with the default parameters Kp=1.0, Tp=5.0 s, dead time L=3.0 s and Kc=1.5. Standard PI control is oscillatory with an overshoot of about 41% and a 5% settling time of about 25.8 s, whereas the Smith predictor with the same Kc has almost 0% overshoot and settles in about 12.7 s. Dead-time compensation suppresses the destabilising effect of the integral term and shortens settling considerably.
Practical Notes
The Smith predictor is sensitive to model uncertainty. In this tool, shifting the internal model's T_p and L by 10% lengthens the settling time from 12.7 s to 16.4 s, and 30% to 23.7 s. On real plants periodic parameter identification is essential
For a dead-time-dominant process with L larger than T_p (e.g. L=5 s, T_p=1 s), standard PI does not settle even at K_c=0.3 (overshoot above 100%), while the Smith predictor at K_c=1.5 gives 3% overshoot and settles in 7 s. With a 10% model error, however, this worsens to 11% and 21 s, so model accuracy decides the performance
This tool is a discrete simulation with dt=0.05 s, so L is rounded to a multiple of it. In an implementation the control period and computation delay become dead-time errors directly. For short dead times (L=0.5–1 s with T_p=5 s) standard PI settles faster (8.9–7.4 s versus 10.3–10.8 s for Smith), so the Smith predictor is unnecessary