Estimate and validate linear models from multipleinput/singleoutput (MISO) data to find the one that best describes the system dynamics.
After completing this tutorial, you will be able to accomplish the following tasks using the command line:
Create data objects to represent data.
Plot the data.
Process data by removing offsets from the input and output signals.
Estimate and validate linear models from the data.
Simulate and predict model output.
Note: This tutorial uses timedomain data to demonstrate how you can estimate linear models. The same workflow applies to fitting frequencydomain data. 
This tutorial uses the data file co2data.mat
,
which contains two experiments of twoinput and singleoutput (MISO)
timedomain data from a steadystate that the operator perturbed from
equilibrium values.
In the first experiment, the operator introduced a pulse wave to both inputs. In the second experiment, the operator introduced a pulse wave to the first input and a step signal to the second input.
Load the data.
load co2data;
This command loads the following five variables into the MATLAB Workspace:
Input_exp1
and Output_exp1
are the input and output data from the first experiment, respectively.
Input_exp2
and Output_exp2
are the input and output data from the second experiment, respectively.
Time
is the time vector from 0 to 1000 minutes, increasing in equal increments of 0.5
min.
For both experiments, the input data consists of two columns of values. The first column of values is the rate of chemical consumption (in kilograms per minute), and the second column of values is the current (in amperes). The output data is a single column of the rate of carbondioxide production (in milligrams per minute).
Plot the input and output data from both experiments.
plot(Time,Input_exp1,Time,Output_exp1) legend('Input 1','Input 2','Output 1') figure plot(Time,Input_exp2,Time,Output_exp2) legend('Input 1','Input 2','Output 1')
The first plot shows the first experiment, where the operator applies a pulse wave to each input to perturb it from its steadystate equilibrium.
The second plot shows the second experiment, where the operator applies a pulse wave to the first input and a step signal to the second input.
Plotting the data, as described in Plotting the Input/Output Data, shows that the inputs and the outputs have nonzero equilibrium values. In this portion of the tutorial, you subtract equilibrium values from the data.
The reason you subtract the mean values from each signal is because, typically, you build linear models that describe the responses for deviations from a physical equilibrium. With steadystate data, it is reasonable to assume that the mean levels of the signals correspond to such an equilibrium. Thus, you can seek models around zero without modeling the absolute equilibrium levels in physical units.
Zoom in on the plots to see that the earliest moment when the operator applies a disturbance to the inputs occurs after 25 minutes of steadystate conditions (or after the first 50 samples). Thus, the average value of the first 50 samples represents the equilibrium conditions.
Use the following commands to remove the equilibrium values from inputs and outputs in both experiments:
Input_exp1 = Input_exp1... ones(size(Input_exp1,1),1)*mean(Input_exp1(1:50,:)); Output_exp1 = Output_exp1... mean(Output_exp1(1:50,:)); Input_exp2 = Input_exp2... ones(size(Input_exp2,1),1)*mean(Input_exp2(1:50,:)); Output_exp2 = Output_exp2... mean(Output_exp2(1:50,:));
The System Identification Toolbox™ data objects, iddata
and idfrd
,
encapsulate data values and data properties into a single entity.
You can use the System Identification Toolbox commands to conveniently
manipulate these data objects as single entities.
In this portion of the tutorial, you create two iddata
objects,
one for each of the two experiments. You use the data from Experiment
1 for model estimation, and the data from Experiment 2 for model validation.
You work with two independent data sets because you use one data set
for model estimation and the other for model validation.
Note: When two independent data sets are not available, you can split one data set into two parts, assuming that each part contains enough information to adequately represent the system dynamics. 
You must have already loaded the sample data into the MATLAB^{®} workspace, as described in Loading Data into the MATLAB Workspace.
Use these commands to create two data objects, ze
and zv
:
Ts = 0.5; % Sampling interval is 0.5 min
ze = iddata(Output_exp1,Input_exp1,Ts);
zv = iddata(Output_exp2,Input_exp2,Ts);
ze
contains data from Experiment 1 and zv
contains data from Experiment 2. Ts
is the sampling interval.
The iddata
constructor requires three arguments
for timedomain data and has the following syntax:
data_obj
= iddata(output,input,sampling_interval);
To view the properties of an iddata
object,
use the get
command. For example,
type this command to get the properties of the estimation data:
get(ze)
ans = Domain: 'Time' Name: '' OutputData: [2001x1 double] y: 'Same as OutputData' OutputName: {'y1'} OutputUnit: {''} InputData: [2001x2 double] u: 'Same as InputData' InputName: {2x1 cell} InputUnit: {2x1 cell} Period: [2x1 double] InterSample: {2x1 cell} Ts: 0.5000 Tstart: [] SamplingInstants: [2001x0 double] TimeUnit: 'seconds' ExperimentName: 'Exp1' Notes: {} UserData: []
To learn more about the properties
of this data object, see the iddata
reference
page.
To modify
data properties, you can use dot notation or the set
command. For example, to assign channel
names and units that label plot axes, type the following syntax in
the MATLAB Command Window:
% Set time units to minutes ze.TimeUnit = 'min'; % Set names of input channels ze.InputName = {'ConsumptionRate','Current'}; % Set units for input variables ze.InputUnit = {'kg/min','A'}; % Set name of output channel ze.OutputName = 'Production'; % Set unit of output channel ze.OutputUnit = 'mg/min'; % Set validation data properties zv.TimeUnit = 'min'; zv.InputName = {'ConsumptionRate','Current'}; zv.InputUnit = {'kg/min','A'}; zv.OutputName = 'Production'; zv.OutputUnit = 'mg/min';
You can verify that the InputName
property of ze
is changed, or "index" into this property:
ze.inputname;
Tip
Property names, such as 
You can plot iddata
objects using the plot
command.
Plot the estimation data.
plot(ze)
The bottom axes show inputs ConsumptionRate
and Current
, and the top axes show the output ProductionRate
.
Plot the validation data in a new figure window.
figure % Open a new MATLAB Figure window plot(zv) % Plot the validation data
Zoom in on the plots to see that the experiment process amplifies the first input ( ConsumptionRate
) by a factor of 2, and amplifies the second input (Current
) by a factor of 10.
Before you begin, create a new data set that contains only the first 1000 samples of the original estimation and validation data sets to speed up the calculations.
Ze1 = ze(1:1000); Zv1 = zv(1:1000);
For more information about indexing into iddata
objects,
see the corresponding reference page.
Frequencyresponse and stepresponse are nonparametric models that can help you understand the dynamic characteristics of your system. These models are not represented by a compact mathematical formula with adjustable parameters. Instead, they consist of data tables.
In this portion of the tutorial, you estimate these models using
the data set ze
. You must have already created ze
,
as described in Creating iddata Objects.
The response plots from these models show the following characteristics of the system:
The response from the first input to the output might be a secondorder function.
The response from the second input to the output might be a firstorder or an overdamped function.
The System Identification Toolbox product provides three functions for estimating the frequency response:
etfe
computes
the empirical transfer function using Fourier analysis.
spa
estimates
the transfer function using spectral analysis for a fixed frequency
resolution.
spafdr
lets
you specify a variable frequency resolution for estimating the frequency
response.
Use spa
to estimate the frequency response.
Ge=spa(ze);
Plot the frequency response as a Bode plot.
bode(Ge)
The amplitude peaks at the frequency of about 0.7 rad/s, which suggests a possible resonant behavior (complex poles) for the first inputtooutput combination  ConsumptionRate
to Production
.
In both plots, the phase rolls off rapidly, which suggests a time delay for both input/output combinations.
To estimate the step response from the data, first estimate a nonparametric impulse response model (FIR filter) from data and then plot its step response.
% model estimation Mimp = impulseest(Ze1,60); % step response step(Mimp)
The step response for the first input/output combination suggests an overshoot, which indicates the presence of an underdamped mode (complex poles) in the physical system.
The step response from the second input to the output shows no overshoot, which indicates either a firstorder response or a higherorder response with real poles (overdamped response).
The stepresponse plot indicates a nonzero delay in the system, which is consistent with the rapid phase rolloff you got in the Bode plot you created in Estimating the Empirical Step Response.
To identify parametric blackbox models, you must specify the input/output delay as part of the model order.
If you do not know the input/output delays for your system from the experiment, you can use the System Identification Toolbox software to estimate the delay.
In the case of singleinput systems, you can read the delay
on the impulseresponse plot. However, in the case of multipleinput
systems, such as the one in this tutorial, you might be unable to
tell which input caused the initial change in the output and you can
use the delayest
command instead.
The command estimates the time delay in a dynamic system by estimating a loworder, discretetime ARX model with a range of delays, and then choosing the delay that corresponding to the best fit.
The ARX model structure is one of the simplest blackbox parametric structures. In discretetime, the ARX structure is a difference equation with the following form:
$$\begin{array}{l}y(t)+{a}_{1}y(t1)+\dots +{a}_{na}y(t{n}_{a})=\\ {\text{b}}_{\text{1}}u(t{n}_{k})+\dots +{b}_{nb}u(t{n}_{k}{n}_{b}+1)+e(t)\end{array}$$
y(t) represents the output at time t, u(t) represents the input at time t, n_{a} is the number of poles, n_{b} is the number of b parameters (equal to the number of zeros plus 1), n_{k} is the number of samples before the input affects output of the system (called the delay or dead time of the model), and e(t) is the whitenoise disturbance.
delayest
assumes that n_{a}=n_{b}=2
and
that the noise e is white or insignificant, and
estimates n_{k}.
To estimate the delay in this system, type:
delayest(ze)
ans = 5 10
This result includes two numbers because there are two inputs: the estimated delay for the first input is 5
data samples, and the estimated delay for the second input is 10
data samples. Because the sampling interval for the experiments is 0.5
min, this corresponds to a 2.5
min delay before the first input affects the output, and a 5.0
min delay before the second input affects the output.
There are two alternative methods for estimating the time delay in the system:
Plot the time plot of the input and output data and read the time difference between the first change in the input and the first change in the output. This method is practical only for singleinput/singleoutput system; in the case of multipleinput systems, you might be unable to tell which input caused the initial change in the output.
Plot the impulse response of the data with a 1standarddeviation confidence region. You can estimate the time delay using the time when the impulse response is first outside the confidence region.
Model order is one or more integers that define the complexity of the model. In general, model order is related to the number of poles, the number of zeros, and the response delay (time in terms of the number of samples before the output responds to the input). The specific meaning of model order depends on the model structure.
To compute parametric blackbox models, you must provide the model order as an input. If you do not know the order of your system, you can estimate it.
After completing the steps in this section, you get the following results:
For the first input/output combination: n_{a}=2, n_{b}=2, and the delay n_{k}=5.
For the second input/output combination: n_{a}=1, n_{b}=1, and the delay n_{k}=10.
Later, you explore different model structures by specifying modelorder values that are slight variations around these initial estimate.
In this portion of the tutorial, you use struc
, arxstruc
, and selstruc
to
estimate and compare simple polynomial (ARX) models for a range of
model orders and delays, and select the best orders based on the quality
of the model.
The following list describes the results of using each command:
struc
creates a matrix of modelorder
combinations for a specified range of n_{a}, n_{b},
and n_{k} values.
arxstruc
takes the output from struc
,
systematically estimates an ARX model for each model order, and compares
the model output to the measured output. arxstruc
returns
the loss function for each model, which is the
normalized sum of squared prediction errors.
selstruc
takes the output from arxstruc
and
opens the ARX Model Structure Selection window, which resembles the
following figure, to help you choose the model order.
You use this plot to select the bestfit model.
The horizontal axis is the total number of parameters — n_{a} + n_{b}.
The vertical axis, called Unexplained output variance (in %), is the portion of the output not explained by the model—the ARX model prediction error for the number of parameters shown on the horizontal axis.
The prediction error is the sum of the squares of the differences between the validation data output and the model onestepahead predicted output.
n_{k} is the delay.
Three rectangles are highlighted on the plot in green, blue, and red. Each color indicates a type of bestfit criterion, as follows:
Red — Best fit minimizes the sum of the squares of the difference between the validation data output and the model output. This rectangle indicates the overall best fit.
Green — Best fit minimizes Rissanen MDL criterion.
Blue — Best fit minimizes Akaike AIC criterion.
In this tutorial, the Unexplained output variance (in %) value remains approximately constant for the combined number of parameters from 4 to 20. Such constancy indicates that model performance does not improve at higher orders. Thus, loworder models might fit the data equally well.
Note:
When you use the same data set for estimation and validation,
use the MDL and AIC criteria to select model orders. These criteria
compensate for overfitting that results from using too many parameters.
For more information about these criteria, see the 
In this tutorial, there are two inputs to the system and one output and you estimate model orders for each input/output combination independently. You can either estimate the delays from the two inputs simultaneously or one input at a time.
It makes sense to try the following order combinations for the first input/output combination:
n_{a}=2:5
n_{b}=1:5
n_{k}=5
This is because the nonparametric models you estimated in Estimating Impulse Response Models show that the response for the first input/output combination might have a secondorder response. Similarly, in Estimating Delays in the MultipleInput System, the delay for this input/output combination was estimated to be 5.
To estimate model order for the first input/output combination:
Use struc
to
create a matrix of possible model orders.
NN1 = struc(2:5,1:5,5);
Use selstruc
to
compute the loss functions for the ARX models with the orders in NN1
.
selstruc(arxstruc(ze(:,:,1),zv(:,:,1),NN1))
Note:

This command opens the interactive ARX Model Structure Selection window.
Note: The Rissanen MDL and Akaike AIC criteria produces equivalent results and are both indicated by a blue rectangle on the plot. 
The red rectangle represents the best overall fit, which occurs for n_{a}=2, n_{b}=3, and n_{k}=5. The height difference between the red and blue rectangles is insignificant. Therefore, you can choose the parameter combination that corresponds to the lowest model order and the simplest model.
Click the blue rectangle, and then click Select to choose that combination of orders:
n_{a}=2
n_{b}=2
n_{k}=5
To continue, press any key while in the MATLAB Command Window.
It makes sense to try the following order combinations for the second input/output combination:
n_{a}=1:3
n_{b}=1:3
n_{k}=10
This is because the nonparametric models you estimated in Estimating Impulse Response Models show that the response for the second input/output combination might have a firstorder response. Similarly, in Estimating Delays in the MultipleInput System, the delay for this input/output combination was estimated to be 10.
To estimate model order for the second input/output combination:
Use struc
to
create a matrix of possible model orders.
NN2 = struc(1:3,1:3,10);
Use selstruc
to
compute the loss functions for the ARX models with the orders in NN2
.
selstruc(arxstruc(ze(:,:,2),zv(:,:,2),NN2))
This command opens the interactive ARX Model Structure Selection window.
Note: The Akaike AIC and the overall best fit criteria produces equivalent results. Both are indicated by the same red rectangle on the plot. 
The height difference between all of the rectangles is insignificant and all of these model orders result in similar model performance. Therefore, you can choose the parameter combination that corresponds to the lowest model order and the simplest model.
Click the yellow rectangle on the far left, and then click Select to choose the lowest order: n_{a}=1, n_{b}=1, and n_{k}=10.
To continue, press any key while in the MATLAB Command Window.
In this portion of the tutorial, you estimate a continuoustime transfer function. You must have already prepared your data, as described in Preparing Data. You can use the following results of estimated model orders to specify the orders of the model:
For the first input/output combination, use:
Two poles, corresponding to na=2
in
the ARX model.
Delay of 5, corresponding to nk=5
samples
(or 2.5 minutes) in the ARX model.
For the second input/output combination, use:
One pole, corresponding to na=1
in
the ARX model
Delay of 10, corresponding to nk=10
samples
(or 5 minutes) in the ARX model.
You can estimate a transfer function of these orders using the tfest
command. For the estimation, you
can also choose to view a progress report by setting the Display
option
to on
in the option set created by the tfestOptions
command.
Opt = tfestOptions('Display', 'on');
Collect the model orders and delays into variables to pass to tfest
.
np = [2 1]; ioDelay = [2.5 5];
Estimate the transfer function.
mtf = tfest(Ze1, np, [], ioDelay, Opt);
View the model's coefficients.
mtf
mtf = From input "ConsumptionRate" to output "Production": 91.75 s  1.272 exp(2.5*s) *  s^2 + 37.76 s + 1.009 From input "Current" to output "Production": 5.186 exp(5*s) *  s + 0.5066 Continuoustime identified transfer function. Parameterization: Number of poles: [2 1] Number of zeros: [1 0] Number of free coefficients: 6 Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties. Status: Estimated using TFEST on time domain data "Ze1". Fit to estimation data: 85.89% (simulation focus) FPE: 6.424, MSE: 6.259
The model's display shows more than 85% fit to estimation data.
In this portion of the tutorial, you create a plot that compares the actual output and the model output using the compare
command.
compare(Zv1, mtf)
The comparison shows about 60% fit.
Use the resid
command
to evaluate the characteristics of the residuals.
resid(Zv1, mtf)
The residuals show high degree of autocorrelation. This is
not unexpected since the model mtf
does not have
any components to describe the nature of the noise separately. To
model both the measured inputoutput dynamics and the disturbance
dynamics you will need to use a model structure that contains elements
to describe the noise component. You can use bj
, ssest
and procest
commands,
which create models of polynomial, statespace and process structures
respectively. These structures, among others, contain elements to
capture the noise behavior.
In this portion of the tutorial, you estimate a loworder, continuoustime transfer function (process model). the System Identification Toolbox product supports continuoustime models with at most three poles (which might contain underdamped poles), one zero, a delay element, and an integrator.
You must have already prepared your data, as described in Preparing Data.
You can use the following results of estimated model orders to specify the orders of the model:
For the first input/output combination, use:
Two poles, corresponding to n_{a}=2 in the ARX model.
Delay of 5, corresponding to n_{k}=5 samples (or 2.5 minutes) in the ARX model.
For the second input/output combination, use:
One pole, corresponding to n_{a}=1 in the ARX model.
Delay of 10, corresponding to n_{k}=10 samples (or 5 minutes) in the ARX model.
Note: Because there is no relationship between the number of zeros estimated by the discretetime ARX model and its continuoustime counterpart, you do not have an estimate for the number of zeros. In this tutorial, you can specify one zero for the first input/output combination, and no zero for the secondoutput combination. 
Use the idproc
command
to create two model structures, one for each input/output combination:
midproc0 = idproc({'P2ZUD','P1D'}, 'TimeUnit', 'minutes');
idproc
accepts a cell array that contains two strings that specify the model structure for each input/output combination:
'P2ZUD'
represents a transfer function with two poles ( P2
), one zero ( Z
), underdamped (complexconjugate) poles ( U
) and a delay ( D
).
'P1D'
represents a transfer function with one pole ( P1
) and a delay ( D
).
The example also uses the TimeUnit
parameter to specify the unit of time used.
View the two resulting models.
midproc0
midproc0 = Process model with 2 inputs: y = G11(s)u1 + G12(s)u2 From input 1 to output 1: 1+Tz*s G11(s) = Kp *  * exp(Td*s) 1+2*Zeta*Tw*s+(Tw*s)^2 Kp = NaN Tw = NaN Zeta = NaN Td = NaN Tz = NaN From input 2 to output 1: Kp G12(s) =  * exp(Td*s) 1+Tp1*s Kp = NaN Tp1 = NaN Td = NaN Parameterization: 'P2DUZ' 'P1D' Number of free coefficients: 8 Use "getpvec", "getcov" for parameters and their uncertainties. Status: Created by direct construction or transformation. Not estimated.
The parameter values are set to NaN
because they are not yet estimated.
Set the time delay property of the model object to 2.5
min and 5
min for each input/output pair as initial guesses. Also, set an upper limit on the delay because good initial guesses are available.
midproc0.Structure(1,1).Td.Value=2.5; midproc0.Structure(1,2).Td.Value=5; midproc0.Structure(1,1).Td.Maximum=3; midproc0.Structure(1,2).Td.Maximum=7;
Note: When setting the delay constraints, you must specify the delays in terms of actual time units (minutes, in this case) and not the number of samples. 
procest
is an iterative estimator
of process models, which means that it uses an iterative nonlinear
leastsquares algorithm to minimize a cost function. The cost
function is the weighted sum of the squares of the errors.
Depending on its arguments, procest
estimates
different blackbox polynomial models. You can use procest
,
for example, to estimate parameters for linear continuoustime transferfunction,
statespace, or polynomial model structures. To use procest
,
you must provide a model structure with unknown parameters and the
estimation data as input arguments.
For this portion of the tutorial, you must have already defined
the model structure, as described in Specifying the Structure of the Process Model. Use midproc0
as
the model structure and Ze1
as the estimation
data:
midproc = procest(Ze1, midproc0); present(midproc)
midproc = Process model with 2 inputs: y = G11(s)u1 + G12(s)u2 From input "ConsumptionRate" to output "Production": 1+Tz*s G11(s) = Kp *  * exp(Td*s) 1+2*Zeta*Tw*s+(Tw*s)^2 Kp = 1.1804 +/ 0.29988 Tw = 1.6566 +/ 629.69 Zeta = 15.917 +/ 6039.2 Td = 2.425 +/ 56.972 Tz = 109.2 +/ 57.37 From input "Current" to output "Production": Kp G12(s) =  * exp(Td*s) 1+Tp1*s Kp = 10.264 +/ 0.048403 Tp1 = 2.0488 +/ 0.054899 Td = 4.9175 +/ 0.034373 Parameterization: 'P2DUZ' 'P1D' Number of free coefficients: 8 Use "getpvec", "getcov" for parameters and their uncertainties. Status: Termination condition: Maximum number of iterations reached. Number of iterations: 20, Number of function evaluations: 276 Estimated using PROCEST on time domain data "Ze1". Fit to estimation data: 86.2% FPE: 6.08, MSE: 5.984 More information in model's "Report" property.
Unlike discretetime polynomial models, continuoustime models let you estimate the delays. In this case, the estimated delay values are different from the initial guesses you specified of 2.5
and 5
, respectively. The large uncertainties in the estimated values of the parameters of G_1(s)
indicate that the dynamics from input 1
to the output are not captured well.
To learn more about estimating models, see Process Models.
In this portion of the tutorial, you create a plot that compares the actual output and the model output.
compare(Zv1,midproc)
The plot shows that the model output reasonably agrees with the measured output: there is an agreement of 60% between the model and the validation data.
Use resid
to perform residual analysis:
resid(Zv1,midproc)
Because the sample system has two inputs, there are two crosscorrelation plots of the residuals with each input, as shown in the following figure.
Autocorrelation and CrossCorrelations of Residuals with the First Input
The crosscorrelation between the second input and residual errors is significant.
After MATLAB displays the first plot, press Enter to view the crosscorrelation with the second input, as shown in the following figure.
CrossCorrelations of Residuals with the Second Input
In the preceding figure, the autocorrelation plot shows values outside the confidence region and indicates that the residuals are correlated.
Change the algorithm for iterative parameter estimation to LevenbergMarquardt.
Opt = procestOptions;
Opt.SearchMethod = 'lm';
midproc1 = procest(Ze1, midproc, Opt);
Tweaking the algorithm properties or specifying initial parameter guesses and rerunning the estimation may improve the simulation results. Adding a noise model may improve prediction results but not necessarily the simulation results.
This portion of the tutorial shows how to estimate a process model and include a noise model in the estimation. Including a noise model typically improves model prediction results but not necessarily the simulation results.
Use the following commands to specify a firstorder ARMA noise:
Opt = procestOptions;
Opt.DisturbanceModel = 'ARMA1';
midproc2 = procest(Ze1, midproc0, Opt);
You can type 'dist'
instead of 'DisturbanceModel'
. Property names are not case sensitive, and you only need to include the portion of the name that uniquely identifies the property.
Compare the resulting model to the measured data.
compare(Zv1,midproc2)
The plot shows that the model output maintains reasonable agreement with the validationdata output.
Perform residual analysis.
figure resid(Zv1,midproc2)
Press Enter to view the crosscorrelation of the residuals with the second input.
The next plot shows that adding a noise model produces uncorrelated residuals: the top set of axes show that the autocorrelation values are inside the confidence bounds. This indicates a more accurate model.
In this portion of the tutorial, you estimate several different types of blackbox, inputoutput polynomial models.
You must have already prepared your data, as described in Preparing Data.
You can use the following previous results of estimated model orders to specify the orders of the polynomial model:
For the first input/output combination, use:
Two poles, corresponding to n_{a}=2 in the ARX model.
One zero, corresponding to n_{b}=2 in the ARX model.
Delay of 5, corresponding to n_{k}=5 samples (or 2.5 minutes) in the ARX model.
For the second input/output combination, use:
One pole, corresponding to n_{a}=1 in the ARX model.
No zeros, corresponding to n_{b}=1 in the ARX model.
Delay of 10, corresponding to n_{k}=10 samples (or 5 minutes) in the ARX model.
About ARX Models. For a singleinput/singleoutput system (SISO), the ARX model structure is:
$$\begin{array}{l}y(t)+{a}_{1}y(t1)+\dots +{a}_{na}y(t{n}_{a})=\\ {\text{b}}_{\text{1}}u(t{n}_{k})+\dots +{b}_{nb}u(t{n}_{k}{n}_{b}+1)+e(t)\end{array}$$
y(t) represents the output at time t, u(t) represents the input at time t, n_{a} is the number of poles, n_{b} is the number of zeros plus 1, n_{k} is the number of samples before the input affects output of the system (called the delay or dead time of the model), and e(t) is the whitenoise disturbance.
The ARX model structure does
not distinguish between the poles for individual input/output paths:
dividing the ARX equation by A, which contains
the poles, shows that A appears in the denominator
for both inputs. Therefore, you can set n_{a} to
the sum of the poles for each input/output pair, which is equal to 3
in
this case.
The System Identification Toolbox product estimates the parameters $${a}_{1}\dots {a}_{n}$$ and $${b}_{1}\dots {b}_{n}$$ using the data and the model orders you specify.
Estimating ARX Models Using arx. Use arx
to compute the
polynomial coefficients using the fast, noniterative method arx
:
marx = arx(Ze1,'na',3,'nb',[2 1],'nk',[5 10]); present(marx) % Displays model parameters % with uncertainty information
marx = Discretetime ARX model: A(z)y(t) = B(z)u(t) + e(t) A(z) = 1  1.027 (+/ 0.02917) z^1 + 0.1675 (+/ 0.04214) z^2 + 0.01307 (+/ 0.02591) z^3 B1(z) = 1.86 (+/ 0.1896) z^5  1.608 (+/ 0.1894) z^6 B2(z) = 1.612 (+/ 0.07417) z^10 Sample time: 0.5 minutes Parameterization: Polynomial orders: na=3 nb=[2 1] nk=[5 10] Number of free coefficients: 6 Use "polydata", "getpvec", "getcov" for parameters and their uncertainties. Status: Estimated using ARX on time domain data "Ze1". Fit to estimation data: 90.7% (prediction focus) FPE: 2.75, MSE: 2.719 More information in model's "Report" property.
MATLAB estimates the polynomials A
, B1
, and B2
. The uncertainty for each of the model parameters is computed to 1 standard deviation and appears in parentheses next to each parameter value.
Alternatively, you can use the following shorthand syntax and specify model orders as a single vector:
marx = arx(Ze1,[3 2 1 5 10]);
Accessing Model Data. The model you estimated, marx
, is a discretetime idpoly
object. To get the properties of
this model object, you can use the get
function:
get(marx)
a: [1 1.0266 0.1675 0.0131] b: {[0 0 0 0 0 1.8600 1.6085] [0 0 0 0 0 0 0 0 0 0 1.6117]} c: 1 d: 1 f: {[1] [1]} IntegrateNoise: 0 Variable: 'z^1' ioDelay: [0 0] Structure: [1x1 pmodel.polynomial] NoiseVariance: 2.7611 Report: [1x1 idresults.arx] InputDelay: [2x1 double] OutputDelay: 0 Ts: 0.5000 TimeUnit: 'minutes' InputName: {2x1 cell} InputUnit: {2x1 cell} InputGroup: [1x1 struct] OutputName: {'Production'} OutputUnit: {'mg/min'} OutputGroup: [1x1 struct] Name: '' Notes: {} UserData: [] SamplingGrid: [1x1 struct]
You can access the information stored by these properties using dot notation. For example, you can compute the discrete poles of the model by finding the roots of the A
polynomial.
marx_poles=roots(marx.a)
marx_poles = 0.7953 0.2883 0.0570
In this case, you access the A
polynomial using marx.a
.
The model marx
describes system dynamics using three discrete poles.
Tip
You can also use 
Learn More. To learn more about estimating polynomial models, see InputOutput Polynomial Models.
For more information about accessing model data, see Data Extraction.
About StateSpace Models. The general statespace model structure is:
$$\begin{array}{l}\frac{dx}{dt}=Ax(t)+Bu(t)+Ke(t)\\ y(t)=Cx(t)+Du(t)+e(t)\end{array}$$
y(t) represents the output at time t, u(t) represents the input at time t, x(t) is the state vector at time t, and e(t) is the whitenoise disturbance.
You must specify a single integer as the model order (dimension
of the state vector) to estimate a statespace model. By default,
the delay equals 1
.
The System Identification Toolbox product estimates the statespace matrices A, B, C, D, and K using the model order and the data you specify.
The statespace model structure is a good choice for quick estimation
because it contains only two parameters: n
is the
number of poles (the size of the A matrix) and nk
is
the delay.
Estimating StateSpace Models Using n4sid. Use
the n4sid
command to specify
a range of model orders and evaluate the performance of several statespace
models (orders 2 to 8):
mn4sid = n4sid(Ze1,2:8,'InputDelay',[4 9]);
This command uses the fast, noniterative (subspace) method and opens the following plot. You use this plot to decide which states provide a significant relative contribution to the input/output behavior, and which states provide the smallest contribution.
The vertical axis is a relative
measure of how much each state contributes to the input/output behavior
of the model (log of singular values of the covariance matrix).
The horizontal axis corresponds to the model order n
.
This plot recommends n=4
, indicated by a red rectangle.
To select this model order, select 4
from
the Model Order dropdown list and click Apply.
By default, n4sid
uses a free parameterization of the statespace form. To estimate a canonical form instead, set the value of the SSParameterization
property to 'Canonical'
. You can also specify the inputtooutput delay (in samples) using the 'InputDelay'
property.
mCanonical = n4sid(Ze1, 3,'SSParameterization', 'canonical','InputDelay', [4 9]); present(mCanonical); % Display model properties
mCanonical = Discretetime identified statespace model: x(t+Ts) = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) + e(t) A = x1 x2 x3 x1 0 1 0 x2 0 0 1 x3 0.0738 +/ 0.05922 0.6083 +/ 0.1624 1.445 +/ 0.1284 B = ConsumptionR Current x1 1.844 +/ 0.1756 0.5631 +/ 0.1225 x2 1.064 +/ 0.1679 2.309 +/ 0.1226 x3 0.2769 +/ 0.0966 1.878 +/ 0.1061 C = x1 x2 x3 Production 1 0 0 D = ConsumptionR Current Production 0 0 K = Production x1 0.8676 +/ 0.03183 x2 0.6864 +/ 0.04166 x3 0.5113 +/ 0.04379 Input delays (sampling periods): 4 9 Sample time: 0.5 minutes Parameterization: CANONICAL form with indices: 3. Feedthrough: none Disturbance component: estimate Number of free coefficients: 12 Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties. Status: Estimated using N4SID on time domain data "Ze1". Fit to estimation data: 91.37% (prediction focus) FPE: 2.456, MSE: 2.339 More information in model's "Report" property.
Note:
mCT1 = n4sid(Ze1, 3, 'Ts', 0, 'InputDelay', [2.5 5]) mCT2 = ssest(Ze1, 3,'InputDelay', [2.5 5]) 
Learn More. To learn more about estimating statespace models, see StateSpace Models.
About BoxJenkins Models. The general BoxJenkins (BJ) structure is:
$$y(t)={\displaystyle \sum _{i=1}^{nu}\frac{{B}_{i}(q)}{{F}_{i}(q)}{u}_{i}\left(tn{k}_{i}\right)}+\frac{C(q)}{D(q)}e(t)$$
To estimate a BJ model, you need to specify the parameters n_{b}, n_{f}, n_{c}, n_{d}, and n_{k}.
Whereas the ARX model structure does not distinguish between the poles for individual input/output paths, the BJ model provides more flexibility in modeling the poles and zeros of the disturbance separately from the poles and zeros of the system dynamics.
Estimating a BJ Model Using pem. You
can use pem
to estimate the
BJ model. pem
is an iterative method and has the
following general syntax:
pem(data,'na',na,'nb',nb,'nc',nc,'nd',nd,'nf',nf,'nk',nk)
To estimate the BJ model, type:
na = 0; nb = [ 2 1 ]; nc = 1; nd = 1; nf = [ 2 1 ]; nk = [ 5 10]; mbj = polyest(Ze1,[na nb nc nd nf nk]);
This command specifies nf=2
, nb=2
, nk=5
for the first input, and nf=nb=1
and nk=10
for the second input.
Display the model information.
present(mbj)
mbj = Discretetime BJ model: y(t) = [B(z)/F(z)]u(t) + [C(z)/D(z)]e(t) B1(z) = 1.823 (+/ 0.1792) z^5  1.315 (+/ 0.2368) z^6 B2(z) = 1.791 (+/ 0.06435) z^10 C(z) = 1 + 0.1066 (+/ 0.04015) z^1 D(z) = 1  0.7453 (+/ 0.027) z^1 F1(z) = 1  1.321 (+/ 0.06939) z^1 + 0.5911 (+/ 0.05516) z^2 F2(z) = 1  0.8314 (+/ 0.006445) z^1 Sample time: 0.5 minutes Parameterization: Polynomial orders: nb=[2 1] nc=1 nd=1 nf=[2 1] nk=[5 10] Number of free coefficients: 8 Use "polydata", "getpvec", "getcov" for parameters and their uncertainties. Status: Termination condition: Near (local) minimum, (norm(g) < tol). Number of iterations: 6, Number of function evaluations: 13 Estimated using POLYEST on time domain data "Ze1". Fit to estimation data: 90.75% (prediction focus) FPE: 2.757, MSE: 2.689 More information in model's "Report" property.
The uncertainty for each of the model parameters is computed to 1 standard deviation and appears in parentheses next to each parameter value.
The polynomials C
and D
give the numerator and the denominator of the noise model, respectively.
Tip Alternatively, you can use the following shorthand syntax that specifies the orders as a single vector: mbj = bj(Ze1,[2 1 1 1 2 1 5 10]);

Learn More. To learn more about identifying inputoutput polynomial models, such as BJ, see InputOutput Polynomial Models.
Compare the output of the ARX, statespace, and BoxJenkins models to the measured output.
compare(Zv1,marx,mn4sid,mbj)
compare
plots the measured output in the validation data set against the simulated output from the models. The input data from the validation data set serves as input to the models.
To perform residual analysis on the ARX model, type the following command:
resid(Zv1,marx)
Because the sample system has two inputs, there are two plots showing the crosscorrelation of the residuals with each input. Press Enter to view the crosscorrelation with the second input.
To perform residual analysis on the statespace model, type the following command:
resid(Zv1,mn4sid)
Finally, to perform residual analysis on the BJ model, type the following command:
resid(Zv1,mbj)
All three models simulate the output equally well and have uncorrelated residuals. Therefore, choose the ARX model because it is the simplest of the three inputoutput polynomial models and adequately captures the process dynamics.
In this portion of the tutorial, you simulate the model output.
You must have already created the continuoustime model midproc2
,
as described in Estimating Process Models.
Simulating the model output requires the following information:
Input values to the model
Initial conditions for the simulation (also called initial states)
For example, the following commands use the iddata
and idinput
commands
to construct an input data set, and use sim
to
simulate the model output:
% Create input for simulation U = iddata([],idinput([200 2]),'Ts',0.5,'TimeUnit','min'); % Simulate the response setting initial conditions equal to zero ysim_1 = sim(midproc2,U);
To maximize the fit between the simulated response of a model
to the measured output for the same input, you can compute the initial
conditions corresponding to the measured data. The best fitting initial
conditions can be obtained by using findstates
on
the statespace version of the estimated model. The following commands
estimate the initial states X0est
from the data
set Zv1
:
% Statespace version of the model needed for computing initial states
midproc2_ss = idss(midproc2);
X0est = findstates(midproc2_ss, Zv1);
Next, simulate the model using the initial states estimated from the data.
% Simulation input Usim = Zv1(:,[],:); Opt = simOptions('InitialCondition',X0est); ysim_2 = sim(midproc2_ss, Usim, Opt);
Compare the simulated and the measured output on a plot.
figure plot([ysim_2.y, Zv1.y]) legend({'model output','measured'}) xlabel('time'), ylabel('Output')
Many controldesign applications require you to predict the future outputs of a dynamic system using the past input/output data.
For example, use predict
to
predict the model response five steps ahead:
predict(midproc2,Ze1,5)
Compare the predicted output values with the measured output values. The third argument of compare
specifies a fivestepahead prediction. When you do not specify a third argument, compare
assumes an infinite prediction horizon and simulates the model output instead.
compare(Ze1,midproc2,5)
Use pe
to compute the prediction error Err
between
the predicted output of midproc2
and the measured
output. Then, plot the error spectrum using the spectrum
command.
[Err] = pe(midproc2,Zv1); spectrum(spa(Err,[],logspace(2,2,200)))
The prediction errors are plotted with a 1standarddeviation confidence interval. The errors are greater at high frequencies because of the highfrequency nature of the disturbance.