# Smith Predictor > Use when you must compensate the dead time of a feedback loop with the smith-predictor structure: given a first-order-plus-dead-time plant model and given PI controller gains, run the delay-free plant model on the controller output, hold the model output in the delay line by the dead time, and subtract the delayed model output from the measured plant output to form the predictor feedback signal, so the primary controller closes on a delay-free compensated error. Produces the compensated error signal of the primary loop, the closed-loop step responses of the predictor-compensated loop, which re-emits the delay-free design response after the dead time, and of the same loop without compensation, and the step-response metrics of both loops: overshoot, settling time to the 2 percent band and final deviation. Trigger: smith predictor, dead time compensation, time delay compensation, transport delay, delay free model prediction, predictor feedback signal. - Skill: `ashfordeou/smith-predictor` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add ashfordeou/smith-predictor` - Raw SKILL.md: https://api.skillmd.com/api/skills/ashfordeou/smith-predictor/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: Apache-2.0 - Author: ashfordeOU (https://skillmd.com/u/ashfordeou) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/ashfordeou/smith-predictor --- # Smith Predictor (gnc-autonomy/control/smith-predictor) Use when the task is dead-time compensation for a single feedback loop on a first-order-plus-dead-time (FOPDT) plant: the primary PI controller is designed for the delay-free plant model, and a Smith predictor removes the dead time from the loop's characteristic equation so the closed loop tracks as if the delay were not there, re-emitting the delay-free response after the dead time. This leaf implements the O. J. M. Smith dead-time compensator ("Closer Control of Loops with Dead Time", Chemical Engineering Progress 53(5):217-219, 1957) in pure Python, stdlib only, deterministic and offline. It pairs with pid-control-design for the PI gains this leaf takes as given inputs, and with digital-control-design for the sampled-data toolbox that would discretize a continuous plant before it reaches this leaf. ## Domain quick reference - Plant and internal model: P(s) = P0(s) exp(-theta s) with the delay-free plant model P0(s) = K / (tau s + 1); the internal model is P0(s) driven by the controller output, never a state observer. - Primary controller: C(s) = kp + ki / s; kp and ki are given inputs, never tuned or placed by this leaf. - Predictor feedback signal: the current model output minus the model output delayed by theta, p(s) = P0(s) (1 - exp(-theta s)) U(s), the 1 - exp(-s theta) dead-time block of the Smith anchor. - Compensated error: e_comp = r - y_measured - (y_model - y_model_delayed); the measured plant output, delayed by the dead time, is corrected by the difference between the current and delayed model output so the primary loop closes on a delay-free error. - Setpoint closed-loop identities: the conventional loop keeps the dead time in its characteristic equation, Y/R = C P0 exp(-theta s) / (1 + C P0 exp(-theta s)); the predictor loop with a perfect model removes it, Y/R = C P0 exp(-theta s) / (1 + C P0), so the characteristic equation 1 + C P0 is the delay-free design denominator and the dead time survives only as an output shift. - Worked-case cancellation: at K = 1.0, tau = 1.0, kp = 0.5, ki = 0.5, C(s) P0(s) = 0.5/s exactly (the controller zero cancels the plant pole at s = -1), so the delay-free closed loop is CL = 0.5 / (s + 0.5), step response g(t) = 1 - exp(-0.5 t), and the predictor loop settles at theta + 2 ln(50) seconds. - Discrete simulation: exact sampled first-order recurrence y[k+1] = a y[k] + b u[k-D] with a = exp(-dt/tau), b = K(1-a), an integer pure-delay line D = round(theta/dt), and a delay-free internal model ym[k+1] = a ym[k] + b u[k]; the PI controller runs on whichever error the configuration selects, e_comp under the predictor or e otherwise. - Perfect-model assumption: the internal model uses the true plant gain, time constant and dead time exactly; model mismatch and time-varying delay are out of scope. ## Workflow 1. Fix the plant and controller: the FOPDT plant P(s) = P0(s) exp(-theta s) with given K, tau and theta, and the given PI gains kp and ki (never tuned or placed here). 2. Compute the sampled first-order lag recurrence with first_order_lag_recurrence and the integer delay line length with delay_steps. 3. Run the delay-free companion simulation (theta = 0, no predictor) with simulate_closed_loop to get the design response the predictor loop must reproduce after the dead time. 4. Run the predictor-compensated loop with simulate_closed_loop (predictor=True): the delay-free plant model runs on the controller output, its output is held in the delay line by the dead time, and the predictor feedback signal is formed by subtracting the delayed model output from the measured plant output to close the primary loop on the compensated error. 5. Run the conventional loop without compensation at the SAME given gains with simulate_closed_loop(predictor=False) for the comparison. 6. Read the step-response metrics of both loops with step_metrics: overshoot percentage, peak time, settling time to the 2 percent band and final deviation. 7. Confirm the predictor delay-shift identity with delay_shift_max_error: the predictor loop output equals the delay-free companion output shifted by the dead time, the discrete witness that the dead time left the characteristic equation. 8. Confirm the deterministic checks with the contract test scripts/test_smith_predictor.py. ## Worked example Plant: FOPDT K = 1.0, tau = 1.0 s (P0 = 1/(s+1)), worked dead time theta = 2.0 s, robustness case theta = 5.0 s. Primary controller: PI with the given gains kp = 0.5, ki = 0.5. Simulation: dt = 0.01 s, 60 s horizon, unit step reference. - Recurrence and delay line: first_order_lag_recurrence(1.0, 1.0, 0.01) gives a = 0.990049834, b = 0.009950166, a + b = 1.0; delay_steps gives D = 200 at theta = 2.0 s and D = 500 at theta = 5.0 s. - Delay-free companion: matches the continuous closed form 1 - exp(-0.5 t) with max abs error 0.001655 over the horizon (the forward-integral discretization residual), overshoot 0.000 and settling 7.83 s against the continuous 2 ln(50) = 7.824 s. - Worked case theta = 2.0 s, predictor loop: overshoot_pct 0.000, settling_time 9.83 s, final_deviation about 0. Monotone response: y(4.0) = 0.633527, y(8.0) = 0.950313, y(12.0) = 0.993245, y(20.0) = 0.999875, the delay-free response re-emitted 2.0 s later. Predictor delay-shift identity: max |y_sp[k] - y_df[k-200]| = 2.220e-16. - Worked case theta = 2.0 s, conventional loop, same gains: overshoot_pct 50.212 (peak 1.502 at 6.00 s), settling_time 25.82 s, final_deviation about 0.000022: y(4.0) = 1.002158, y(8.0) = 1.165740, y(12.0) = 0.841161, y(20.0) = 0.946013, oscillating through the delayed characteristic equation. - Robustness case theta = 5.0 s, predictor loop: overshoot_pct 0.000, settling_time 12.83 s; y(6.0) = 0.395100, y(10.0) = 0.918155, y(20.0) = 0.999442, y(40.0) = 1.000000; delay-shift identity max |y_sp[k] - y_df[k-500]| = 0.000e+00. - Robustness case theta = 5.0 s, conventional loop, same gains: overshoot_pct 2182.315, settling_time None (the oscillation never holds the band; still growing at the final sample), final_deviation 11.909619: y(6.0) = 0.501578, y(10.0) = 2.502479, y(20.0) = -2.403046, y(40.0) = -10.009273. The 5.0 s delay pushes the loop past stability at the same gains that were stable with the predictor. - ValueErrors: first_order_lag_recurrence(0.0, 1.0, 0.01) raises "plant gain K must be positive, got 0.0"; first_order_lag_recurrence (1.0, 1.0, 0.0) raises "sample time dt must be positive, got 0.0"; delay_steps(-0.1, 0.01) raises "dead time theta must be non-negative, got -0.1"; delay_steps(2.5, 1.0) raises "dead time theta must be an integer multiple of dt, got theta 2.5 at dt 1.0"; delay_steps(0.0, 0.0) raises "sample time dt must be positive, got 0.0"; pi_output(0.1, 0.0, 0.01, 0.0, 0.5) raises "proportional gain kp must be positive, got 0.0"; pi_output(0.1, 0.0, 0.01, 0.5, -0.1) raises "integral gain ki must be non-negative, got -0.1"; simulate_closed_loop(sim_time=0.0) raises "simulation time sim_time must be positive, got 0.0". ## Verification - Confirm first_order_lag_recurrence and delay_steps match the recurrence and delay-line anchors within tight tolerance, with the a + b = K and b/(1-a) = K identities holding. - Confirm the delay-free companion simulation matches the continuous closed form 1 - exp(-0.5 t) within the discretization residual. - Confirm the predictor loop at theta = 2.0 s and theta = 5.0 s matches the worked overshoot, settling time, final deviation and sample values, and that delay_shift_max_error confirms the delay-shift identity against the delay-free companion. - Confirm the conventional loop at the SAME given gains degrades as theta grows: theta = 2.0 s overshoots and settles late; theta = 5.0 s diverges (settling_time None), the isolated effect of removing the predictor structure. - Confirm ValueError rejection of a non-positive plant gain, time constant or sample time, a negative or non-multiple dead time, a non-positive proportional gain, a negative integral gain, a non-positive simulation time, and mismatched or out-of-window inputs to delay_shift_max_error. - Run the contract test offline: python3 scripts/test_smith_predictor.py (deterministic, no imports beyond math). ## Related leaves - gnc-autonomy/control/pid-control-design: tunes and places the PI gains kp and ki this leaf takes as given inputs; this leaf never tunes, places or margin-checks a controller. - gnc-autonomy/control/digital-control-design: the sampled-data toolbox that discretizes a continuous plant with a zero-order hold; this leaf simulates at a fixed sample time constant and never designs in the z-domain. - gnc-autonomy/control/observer-design: a full-order state observer for unmeasured states; this leaf's internal model is a delay-free copy of the plant driven by the controller output, not a state estimate. ## Pitfalls - Treating the internal model as a state estimator: the delay-free plant model P0(s) run on the controller output is not a Luenberger or Kalman state estimate; it carries no estimator gain and no error dynamics. - Tuning kp and ki here: this leaf never derives, places or margin-checks the PI gains; a change in the predictor loop's behavior at fixed gains comes only from the dead time theta, never from a gain change. - Assuming the predictor helps disturbance rejection: the delay-free characteristic equation 1 + C P0 governs the setpoint path only; the worked comparison is a setpoint step and makes no disturbance-path claim. - Feeding a non-multiple dead time: delay_steps requires theta to be an integer multiple of dt within STEPS_EPS; a fractional number of samples has no exact pure-delay line and raises ValueError instead of rounding silently. - Expecting the predictor to fix an unstable model mismatch: the perfect-model assumption is the model itself; a mismatch between the internal model and the true plant is out of scope and not simulated. ## Behavior contract (gate 3) Run the deterministic contract test (stdlib unittest, offline): python3 scripts/test_smith_predictor.py The test covers the recurrence and delay-line identities (workflow step 2), the delay-free companion collapse (step 3), the predictor loop worked values at theta = 2.0 s and theta = 5.0 s (step 4), the conventional-loop comparison at the same given gains (step 5), the step-response metrics of both loops (step 6), the predictor delay-shift identity (step 7), the zero-delay collapse, determinism across repeated runs, and ValueError rejection of every non-physical input enumerated in the Worked example. ## Compliance - Standards referenced, not reproduced: ARP4754A is a proprietary SAE standard (name plus paraphrase only, per standards-map.yaml); the Smith predictor relations above are standard control engineering methodology from O. J. M. Smith (1957), summary-only. - compliance: STANDARDS-REF, gated: false.