Predictive Controllers are a group of model-based predictive controllers. Because they are model-based, as the name suggests, a model of the system is required in order to design an explicit controller for the system. Typically, we would identify the model of the system through one of parameter estimation methods in system identification or by explicitly modeling the system dynamics and then design a controller to fulfill the given design specifications. Note that predictive control is not the only model-based control design method available. Among other model-based controllers are pole-placement and linear quaratic (LQ) control methods. If the underlying process is linear, if there are no constraints (hardly the case in real plants and actuators), and if the set-point or if the reference trajectory is fairly simple, then pole-placement, LQ methods and model-based predictive control are almost equivalent. When this is the case, they result in the same structure after some manipulation (the way the controller parameters are generated are typically different between these three methods) and they have a sufficient number of degrees of freedom.

Predictive Control consists of broad methods of designing controllers; therefore, various methods with special cases exist in how predictive controllers are designed. Among the popular ones are the Smith Predictor (perhaps the earliest predictive controller, 1959 by O.J. Smith), Generalized Predictive Control (GPC, Clarke et. al, Automatica, 1987), Dynamic Matrix Control (DMC), Extended Horizon Adaptive Control (EHAC), Extended Predictive Self-Adaptive Control(EPSAC), and Unified Predictive Control (UPC). Besides the downside of requiring an explicit modeling of the process, predictive controllers are very useful in control design. Consider these. you are worried laymen/technicians might have trouble adjusting the gains of the controller that you have painstakingly designed because they might have trouble tuning the controller? Not a worry, predictive controllers are a less hassle to tune (bite that robust adaptive control);

you are worried about handling multivariable processes? Predictive controllers can be derived for MIMO processes. Hakuna matata!

your process is nonlinear and you are worried about explicitly handling nonlinearities? No pressure. Predictive controllers are extensible to nonlinear processes (see, for example, this) without the need for local or global linearizing transformations;

so you’ve designed your controller and you have checked off instability among your worries but you have to contest with physical hardware saturation? Easy-peasy, lemon-squeezy; one of the most reliable control design methodologies that can well-handle process constraints in a systematic fashion during controller design are predictive controllers (this is their richest property, IMO).

you have a non-minimum phase system or your system is open-loop unstable and you are confused as to how you would handle these absurd zeros? Check!

your process dynamics is very complicated and you do not know how you can explicitly factor in stochastic disturbances and other parametric uncertainties into your controller? Check! Prior setpoints or reference trajectories can be robustly tracked by formulating feed-forward action explicitly into your design.

So predictive controllers are broad and very rich in the class of problems they can solve. Particularly, they are widely used in the chemical processing industry.

Model Predictive Control and LRPCs: Take the way we cross a road intersection for example. It is not sufficient that a road is not busy with passing cars on it. As you cross, you constantly look ahead (your prediction horizon) to anticipate oncoming vehicles and update your movement (control parameters) using the information that is available to you now ( your past observations). So MPCs employ the long-range prediction horizon concept i.e., they anticipate future disturbances based on past observations and generate a control sequence that is a proposal among likely alternatives that can well-handle possible disturbances at the next time step.

In classical PID control, it is difficult to guarantee closed-loop stability when there are (non-)parametric uncertainties to deal with. Model Predictive Controllers employ a receding (which is a finite) horizon concept in establishing a control sequence in order to assure that closed-loop stability is generally guaranteed.

In principle, MPCs are long-range predictive controllers(LRPC). Contrary to classical control laws, they are potent if there is a process dead-time or if the setpoint/reference trajectory is well-known ahead of time. Predictions constitute the bulwark of model predictive controllers. Matter-of-factly, the prediction that this family of controllers have in their structure explains their robust performance when correctly implemented in process control. MPCs incorporate an internal model (e.g. a CARIMA (Controlled Auto Regressive Integral Moving Average) model) of the plant in their dynamics to avoid model mismatch between the plant and predictor. This allows them to anticipate the future "behavior" of a plant's output and mitigate such errors before the plant reaches such a "future time".

Accurate prediction over a selected horizon is critical to a successful MPC implementation.

Enter GPCs: Like the other answer mentions, GPCs are a class of MPCs and LRPCs . The basic algorithm involves:

At a current time, kk, we predict an output, ykyk, over a certain output horizon, nyny, based on a mathematical model of the plant dynamics (see figure above). The predicted output is a function of future possible control scenarios.

the index of performance is quadratic assuring that we can exactly compute the control law based on a computation of the gradient of the performance index, J,J,with respect to the control input uu

the control law is computed by minimizing a performance index with respect to the first of the future control sequence, nunu

From the proposed control scenarios, the strategy that delivers the best control action to bring the current process output to the setpoint/reference trajectory is chosen.

The chosen control law is applied to the real process input only at the present time kk.

the control law is linear in measured values and the prediction.

The above procedure is repeated at the next sampling instant leading to an updated control action with correctness based on latest measurements. In literature, this is referred to as the receding horizon concept.

Hi,

From a general point of view, if you create a predictive model of a process , you can predict the future behaviour of the process. Otherwise, if this predictive model is able to control the process, you have a model able to predict and control the behaviour of the process. That means your controller is able to send a feedback to the process (to your devices) in order to rectify the process if there are a deviation. Furthermore, if your physical process is not stable itself, with a predictive and control model (with a feedback) you can create a stable close-loop.

I wait that my answer help you ever if it is a very gerenic answer.

Dear Hima,

Historically, the predictive control with internal model, which is a special case of optimal control, was born at the end of the 1970s, more exactly this method was invented by a Frenchman, J. Richalet, in 1978 and generalized by DW Clarke in 1987 in agreement with major industrial groups in the United States and Europe (Shell and Adersa). The first name was at the beginning and after what is generalized the scientific community of the field calls it MPC.

Best regards

In power electronics, Model Predictive Control (MPC) is a form of Predictive control. Furthermore, Finite Control Set-MPC (FCS-MPC) and Continuous Control Set-MPC (CCS-MPC) are the two forms of MPC in general.

MPCs employ the long-range prediction horizon concept i.e., they anticipate future disturbances based on past observations and generate a control sequence that is a proposal among likely alternatives that can well-handle possible disturbances at the next time step.

Predictive Controllers are a group of model-based predictive controllers. Because they are model-based, as the name suggests, a model of the system is required in order to design an explicit controller for the system. Typically, we would identify the model of the system through one of parameter estimation methods in system identification or by explicitly modeling the system dynamics and then design a controller to fulfill the given design specifications. Note that predictive control is not the only model-based control design method available. Among other model-based controllers are pole-placement and linear quaratic (LQ) control methods. If the underlying process is linear, if there are no constraints (hardly the case in real plants and actuators), and if the set-point or if the reference trajectory is fairly simple, then pole-placement, LQ methods and model-based predictive control are almost equivalent. When this is the case, they result in the same structure after some manipulation (the way the controller parameters are generated are typically different between these three methods) and they have a sufficient number of degrees of freedom.

Predictive Control consists of broad methods of designing controllers; therefore, various methods with special cases exist in how predictive controllers are designed. Among the popular ones are the Smith Predictor (perhaps the earliest predictive controller, 1959 by O.J. Smith), Generalized Predictive Control (GPC, Clarke et. al, Automatica, 1987), Dynamic Matrix Control (DMC), Extended Horizon Adaptive Control (EHAC), Extended Predictive Self-Adaptive Control(EPSAC), and Unified Predictive Control (UPC). Besides the downside of requiring an explicit modeling of the process, predictive controllers are very useful in control design. Consider these. you are worried laymen/technicians might have trouble adjusting the gains of the controller that you have painstakingly designed because they might have trouble tuning the controller? Not a worry, predictive controllers are a less hassle to tune (bite that robust adaptive control);

you are worried about handling multivariable processes? Predictive controllers can be derived for MIMO processes. Hakuna matata!

your process is nonlinear and you are worried about explicitly handling nonlinearities? No pressure. Predictive controllers are extensible to nonlinear processes (see, for example, this) without the need for local or global linearizing transformations;

so you’ve designed your controller and you have checked off instability among your worries but you have to contest with physical hardware saturation? Easy-peasy, lemon-squeezy; one of the most reliable control design methodologies that can well-handle process constraints in a systematic fashion during controller design are predictive controllers (this is their richest property, IMO).

you have a non-minimum phase system or your system is open-loop unstable and you are confused as to how you would handle these absurd zeros? Check!

your process dynamics is very complicated and you do not know how you can explicitly factor in stochastic disturbances and other parametric uncertainties into your controller? Check! Prior setpoints or reference trajectories can be robustly tracked by formulating feed-forward action explicitly into your design.

So predictive controllers are broad and very rich in the class of problems they can solve. Particularly, they are widely used in the chemical processing industry.

Model Predictive Control and LRPCs: Take the way we cross a road intersection for example. It is not sufficient that a road is not busy with passing cars on it. As you cross, you constantly look ahead (your prediction horizon) to anticipate oncoming vehicles and update your movement (control parameters) using the information that is available to you now ( your past observations). So MPCs employ the long-range prediction horizon concept i.e., they anticipate future disturbances based on past observations and generate a control sequence that is a proposal among likely alternatives that can well-handle possible disturbances at the next time step.

In classical PID control, it is difficult to guarantee closed-loop stability when there are (non-)parametric uncertainties to deal with. Model Predictive Controllers employ a receding (which is a finite) horizon concept in establishing a control sequence in order to assure that closed-loop stability is generally guaranteed.

In principle, MPCs are long-range predictive controllers(LRPC). Contrary to classical control laws, they are potent if there is a process dead-time or if the setpoint/reference trajectory is well-known ahead of time. Predictions constitute the bulwark of model predictive controllers. Matter-of-factly, the prediction that this family of controllers have in their structure explains their robust performance when correctly implemented in process control. MPCs incorporate an internal model (e.g. a CARIMA (Controlled Auto Regressive Integral Moving Average) model) of the plant in their dynamics to avoid model mismatch between the plant and predictor. This allows them to anticipate the future "behavior" of a plant's output and mitigate such errors before the plant reaches such a "future time".

Accurate prediction over a selected horizon is critical to a successful MPC implementation.

Enter GPCs: Like the other answer mentions, GPCs are a class of MPCs and LRPCs . The basic algorithm involves:

At a current time, kk, we predict an output, ykyk, over a certain output horizon, nyny, based on a mathematical model of the plant dynamics (see figure above). The predicted output is a function of future possible control scenarios.

the index of performance is quadratic assuring that we can exactly compute the control law based on a computation of the gradient of the performance index, J,J,with respect to the control input uu

the control law is computed by minimizing a performance index with respect to the first of the future control sequence, nunu

From the proposed control scenarios, the strategy that delivers the best control action to bring the current process output to the setpoint/reference trajectory is chosen.

The chosen control law is applied to the real process input only at the present time kk.

the control law is linear in measured values and the prediction.

The above procedure is repeated at the next sampling instant leading to an updated control action with correctness based on latest measurements. In literature, this is referred to as the receding horizon concept.

I will recommend you to watch videos of Prof. J. A. Rossiter on youtube in order to understand clearly the concept of Predictive control hence model predictive control

https://www.youtube.com/watch?v=4kCcXGDvjU8