ON-OFF controlled offline half-bridge smps is unstable?

[grizedale]

Hello,

We are designing an offline , isolated half-bridge SMPS.

Vin = 90-265VAC.
Vout = +/-50V
Power (out) = 270W


Our load is highly transient in nature.

It is either fully on or fully off......and goes suddenly from one to the other.

I am just thinking that given this situation, normal , feedback compensation methods are going to be less useful.

I am wondering, to get better transient response, should we not just implement this as an ON-OFF ("BANG-BANG" ) controller?

We could get rid of the slow, opto-isolated feedback scheme and just use a comparator on the secondary side......-every time the output voltage "hits" the regulation value.........the comparator switches....and this is transferred through a fast digital isolator to the primary side...where switching would be simply and suddenly stopped.

Digital isolator
https://www.silabs.com/Support Documents/TechnicalDocs/Si8410.pdf

We have a current sense transformer on the primary side, so we can use that to stop switching whenever the primary current gets too high.

The problem is........its half-bridge topology, so obviously theres an LC filter at the output.........now.......this on-off type of control scheme will tend to have a kind of "burst mode" effect.........there will be bursts of switching followed by no switching..........so what happens i wonder, when the "burst-frequency" ever gets the same as the resonant frequency of the output LC filter?

....would this cause vast ringing and instability?...and make this "on-off" scheme impractical?

 

[WimRFP]

When dealing with LC filters, you have always some over- or undershoot.

If the load is removed, the energy in the L will increase the capacitor voltage. To reduce this effect, you need a large C and a small L.

Regarding the control loop, you may us a two loop approach. The first (fast loop) is a current controlled one (so you have to load current or inductor current). Using current results in a first order behavior (despite the LC circuit). The set value for the current loop is determined by a slower voltage loop.

The control loop can be made faster by using the capacitor current to control fully on or fully off. This design is somewhat more elaborate. Te get some feeling for it, you may simulate an on/off buck topology based on current measurement (start with system level HSPICE blocks instead of a full HW implementation).

 

[FvM]

If you expect bad control system behaviour with a continuous PI or PID controller, you should expect nothing but worse results with on-off control, I think.

In my opinion, the best method to achieve good dynamic load regulation is a feedforward structure, that directly translates output load changes into primary control action, without waiting for an ouput voltage change. The voltage control loop has to perform only minor corrections.

I already mentioned the method related to PFC converter control, where you basically need to restrict the voltage control loop bandwidth.

Unfortunately, I can't give you a ready-to-use design drawing, because I never implemented it in an analog design. But it works well with digitally controlled high power systems and I don't see a reason, why it shouldn't be feasible with classical analog controllers.

 

[WimRFP]

FVM: measuring the capacitor current of the output capacitor parallel to the load is a feedforward scheme, so I fully agree with you. Just a voltage control loop gives bad response (from analog experience with hysteresis controllers).

Unfortunately I don't have something at hand that will fit your needs.

 

[FvM]

measuring the capacitor current of the output capacitor parallel to the load is a feedforward scheme

Right, I overlooked that you referred to capacitor current sensing in your previous post. Although feedforward is the ultimate solution, I'm not sure if it's really needed for the present application. I have it in designs with PWM frequency below 10 kHz

 

[grizedale]

Hi,

I think the control schemes that you mention are too expensive for us.

I am just basically talking about a very simple hysteretic type scheme......the power stage will give the hysteresis....and as mentioned..........soon as vout hits regulation value.......shut switch off..........this will basically result in burst mode.....there will be some vout ripple........but with the fast digital isolator signalling to the primary as soon as vout nudges regulation level, i really just cannot see what can go wrong here?


i just can't beleive for example, that we would get unstable oscillation on vout of several volts pk-to-pk, like many articles suggest.

if vout hits regulation level, ...this is immediatley signalled to the primary and it immediately stops switching...so theres no way the output can keep rising at all far.

.. the secondary voltage then falls below regulation level, and primary resumes switching........i would be happy if there was say 4% ripple on vout or less.

 

[WimRFP]

Hello,

If you have built your circuit allready, try it, and see how it behaves. If not, run some simulation. I also designed several hysteresis control loops and 2 loops that turned on and and off a power inverter.

If the current through the LC filter is only limited by the inductance of the LC filter itself, then you will store lots energy in it and that will result in voltage rise, after the controller shuts down the inverter. In my case the inverter was power, and current limited cycle by cycle, so the current through the inductor of the LC was limited, even when under shorted circuited load.

From a system modelling point of view, the current, voltage and power limitation results in a virtual resistor in series with the LC circuit, changing the transfer function of the LC network in a very positive way. Because of this virtual resistor, the ripple was about +/-4%. I could reduce the ripple by changing the hysteresis and adding some D action, but this would result in more switching loss in the circuit that turns on and off the inverter.

Sometimes simple LC hysteresis loops function well because of the ESR of the filter capacitor.

Regarding price, If your inverter (that will be used in burst mode) isn't current limited by itself, implementing a mixed current/voltage feedback scheme doesn't increase the price much (especially when you have some secondary current measurement allready). There are schemes where some part of the input before the LC filter is taken into the loop. It can have benefits, but they don't have my preference.

 

[mtwieg]

FVM: measuring the capacitor current of the output capacitor parallel to the load is a feedforward scheme, so I fully agree with you. Just a voltage control loop gives bad response (from analog experience with hysteresis controllers).

Unfortunately I don't have something at hand that will fit your needs.

Isn't measuring output capacitor current equivalent to measuring the derivative of output voltage, though? Looking at it that way, it still seems like feedback, not feedforward. Measuring load current, on the other hand, sounds like it could work.

 

[Orson Cart]

The fastest loop response for a half bridge is most easily obtained with average current mode control inside a voltage loop.

 

[grizedale]

Our output inductor is 47u and its peak current is 6.7A.

So it can store up to 1.5mJ.

The output capacitor is 680u and 40V...so it is already storing 0.544 Joules.

The extra 1.05mJ from the inductor wont make the output caps voltage rise far at all....so when the switch switches off...we wont see any rise in volts on the output cap...so no overshoot.

 

[FvM]

Assuming that your hysteretic scheme won't cause unsuitable output overshoot, there are other unpleasant properties involved, I fear. You'll need to set the primary switcher to maximum power when on, resulting in high peak currents and lower efficiency. And you get strong audible transformer noise. Saying a hysteretic controller can work for the application, implies that a regular linear control loop can do to (and most likely better).

 

[WimRFP]

Isn't measuring output capacitor current equivalent to measuring the derivative of output voltage, though? Looking at it that way, it still seems like feedback, not feedforward. Measuring load current, on the other hand, sounds like it could work.

The main (fast) loop uses the inductor current. This is a first order behavior so can be easily handled by a hysteresis control scheme.

The voltage loop is the slow one and that controls the set point of the inductor current loop.

When the load current increases, this causes a voltage drop across the capacitor and this will be detected by the voltage feedback. Gradually the set point of the current loop will be increased to get a new equilibrium ("inductor current" = "output current").

If you know that the output current increases with 1 A, you may directly add 1A to the set point of the inductor current control loop. This bypasses the voltage control loop. In my opinion this is a feed forward scheme.

Instead of measuring inductor current and output current separately (and sum/subtract them in the end), you can measure the capacitor current directly as capacitor current is "inductor current" – "output current".

---------- Post added at 13:06 ---------- Previous post was at 13:01 ----------

Our output inductor is 47u and its peak current is 6.7A.

If this is the absolute maximum output current that can be provided by the inverter stage, you are right, your scheme might work when there are no other significant delays.

 

[mtwieg]

The main (fast) loop uses the inductor current. This is a first order behavior so can be easily handled by a hysteresis control scheme.

The voltage loop is the slow one and that controls the set point of the inductor current loop.

When the load current increases, this causes a voltage drop across the capacitor and this will be detected by the voltage feedback. Gradually the set point of the current loop will be increased to get a new equilibrium ("inductor current" = "output current").

If you know that the output current increases with 1 A, you may directly add 1A to the set point of the inductor current control loop. This bypasses the voltage control loop. In my opinion this is a feed forward scheme.

Yeah I'm with you so far.

Instead of measuring inductor current and output current separately (and sum/subtract them in the end), you can measure the capacitor current directly as capacitor current is "inductor current" – "output current".

I think I get what you're saying... so the goal is to null capacitor current to zero, which should help keep output voltage constant. Sounds good in theory, but the current measurement would have to be filtered carefully in order to get rid of ripple and be usable. And that imparts phase shift on the feedforward, which could hurt stability. I'll have to try and simulate it sometime.

---------- Post added at 08:50 ---------- Previous post was at 08:19 ----------

Saying a hysteretic controller can work for the application, implies that a regular linear control loop can do to (and most likely better).

I kind of feel the same. A hysteretic controller may give the fastest rise time on transients, but will suffer far more ripple and overshoot, and thus a much longer settling time than a well compensated linear feedback scheme. Personally I would try to simulate all options to get a rough idea of which is better for the application.

 

[WimRFP]

Hello,

You shouldn't filter the measured current to get rid of the ripple due to the switching of the hysteresis block, you will lose the first order system behavior (for the current feedback loop).

You should only filter components due to higher order behavior (think of parasitic capacitance of the inductor and ringing effects). Of course grizedale has to remove the ripple due to switching of the half bridge inverter itself.

 

[mtwieg]

Hello,

You shouldn't filter the measured current to get rid of the ripple due to the switching of the hysteresis block, you will lose the first order system behavior (for the current feedback loop).

You should only filter components due to higher order behavior (think of parasitic capacitance of the inductor and ringing effects). Of course grizedale has to remove the ripple due to switching of the half bridge inverter itself.

Yes, I was referring to filtering ripple caused by the switching of the inverter, not the on/off switching of the entire converter.

---------- Post added at 09:02 ---------- Previous post was at 08:54 ----------

The fastest loop response for a half bridge is most easily obtained with average current mode control inside a voltage loop.

By the way Orson, where do you actually measure current in a half bridge ACMC scheme? Do you measure primary current and rectify it before taking the averaging, or measure the secondary choke current, or measure the switch currents? I can't find an answer in my reference books.

 

[WimRFP]

If you want to do some simulation, I would advice you to start with a buck converter. You can "measure" current in spice very easily without fully implementing current sense circuits in "hardware". You can even use a voltage controlled switch as on/off device (you can specify hysteresis in most cases). Using less actual hardware results in short runtime.

You will be amazed about the dynamic response of a current-voltage controlled hysteresis system.

If you have a good feeling about how to tweak the control parameters, you can start with an on/off controlled inverter.

Because of the inverter's high frequency, in combination with relative large response time of the LC filter, runtimes will be long.

I once did a feasibility study for a low power DC to DC converter system that should have very high efficiency over a very large range of loads (from uA to A). An on/off-controlled inverter had high potential as the inverter always operated in its region of highest efficiency, or was off (with only some component leakage current). The simulations were al based on a voltage feedback loop as the inverter itself behaved more like a current source. It was never built because the client stopped the project where this converter was required.

 

[Orson Cart]

1.05 mJ, correct the overshoot will be small at switch off, careful to watch the transient behaviour when you re-enable it...!