Wow, three different questions.
Tom,
The transfer function for a simple temperature system is:
G*exp(-s*d*T)/(tau1*s+1)*(tau2*s+1)
G is the gain in degrees per percent output.
tau1 is the time constant of the thermal mass
tau2 is the time constant of the temperature sensor.
T is the update time.
d is the number of update periods of dead time.
There could be a third time constant if the heater takes time to heat up to radiate heat.
The equation for this is:
Temp1
= exp(-T/tau1)* Temp1(n-1) + G*(1-exp(-T/tau1))*Control%(n-d) + Ambient temp.
Temp2
= exp(-T/tau2)*Temp2(n-1) + ( 1-exp(-T/tau2))*Temp1(n-1)
Where Temp2 is the resulting temperature as a function of the control output percentage. This implements a simple temperature with a gain ( G), deadtime ( d*T ), and two low pass filters using tau1 and tau2. Run some simulations using excel. I did this off the top of my head since I don't have access to my Mathcad right now.
I did this on the Hotrod.zip in the download area and on my ftp site.
ftp://ftp.deltamotion.com/public/PDF/Mathcad - TempPID.pdf
10BaseT, I can model systems. If one uses the PID to control the model without the dead time then the PID will not wind up and over react because it will see the model responds. See the hotrod.zip. The trick is that the model is not perfect and must be corrected. The estimated temperatures predicted in the model must be put in a delay queue these delayed values are thencompared with the current meassured temperature. A fraction of the error is then added to the estimated temperature to get a new corrected estimated temperature. It is a little more difficult than that because I would also compare the delayed estimated derivatives ( rates ) with the current temperature derivatives. This is not easy to do in a PLC. That is why there are DCS systems and temperature controllers. Attempting to PID control a system with deadtime with just the PID alone is extremely difficult and often a waste of time and money. If one doesn't have a model then feed forwards should be used so the PID gains can be much lower. Using feed forwards or bias as a function of the SP and its derivitives is common practice in motion control.
Alaric's non linear case is a good reason to limit power. Does limiting the output limit power? An analog output limits the voltage only. A PWM just limits the time the current is on. Where are we controlling I*V? A more sophisticate system is required. What happens when the system is warmer and resistance is higher? We have current source feedback on our V to I product to compensate for the change in resistance in a servo spool as temperature changes.