Key Takeaways
- Averaged models settle control questions quickly, while switched models are the only place ripple, harmonics and switching loss actually appear.
- Solver family and time step size set the accuracy and run time balance, and a step refinement study proves the setting is good enough.
- Topology and study goal decide the method, so reaching for the most detailed model available adds cost and hides errors rather than reducing risk.
The modelling method you pick for a converter decides how much you can trust the waveform on your screen.
Averaged models, switched models, solvers and time step sizes each answer a different question, and each charges you something in accuracy or run time. Grid-tied conversion is where that choice bites hardest, since global solar PV additions passed 600 GW for the first time in 2025, with wind additions near 160 GW. Each one sits behind a converter somebody modelled first.
No single method carries a converter from concept to production. You’ll change methods as the question changes, moving from an averaged model that settles a control loop to a switched model that shows ripple and loss. Getting good at this means knowing which detail you’re buying each time.
What converter simulation predicts before any hardware gets built
Converter simulation predicts the electrical behaviour of a switching power stage before you commit to a board. It gives you duty cycle response, inductor current ripple, device stress and stability margins from equations you can inspect. The value is in testing conditions that would be slow or destructive on a bench.
A 400 V to 48 V DC/DC stage for a telecom rack is a fair case. You can sweep the input from 320 V to 460 V, drop the load to 10%, and watch the output recover inside a few minutes. The same sweep on hardware needs a supply, an electronic load, a scope and a day of setup.
What you’re really buying is the ability to be wrong cheaply. An unstable loop found at light load costs an afternoon of compensator redesign, and the same discovery on a populated board costs a respin. That trade only holds if the model is faithful enough for the question, which is where method selection starts.
Averaged models and switched models answer different design questions
The main difference between averaged and switched converter models is what each does with the switching action. An averaged model replaces the switch with its duty cycle equivalent and runs orders of magnitude faster. A switched model keeps every transition and shows you ripple, harmonics and switching loss.
Loop compensation work is where averaging earns its place. A boost converter at 100 kHz needs roughly 30 ms of settling before you can judge a load step, and an averaged model gives you that in under a second. The same window in a switched model means resolving 3,000 cycles and every edge inside them.
| Modelling approach | Questions it answers well | What you give up |
| Averaged state-space model | Loop stability, bandwidth and load step response | Ripple, harmonics and loss tied to switching |
| Ideal switched model | Ripple, duty cycle behaviour and discontinuous conduction | Device loss and heating, since the switch is ideal |
| Switched model with device loss data | Efficiency across the load range and junction temperature | Longer runs and reliance on datasheet parameters |
| Small-signal frequency response | Gain and phase margin read off a Bode plot | Large-signal behaviour such as saturation and startup |
| System model holding source and load | Interaction between the converter, its source and later stages | Fine detail at each switching edge |
Averaging hides the exact things a switched model exists to show. You won’t see inductor current going discontinuous at light load, and you can’t estimate switching loss from a model with no switching in it.
Solver choice decides how faithfully switching edges get resolved
Solver choice controls how a simulation advances through time and how it copes with the discontinuity at each switching instant. Variable-step solvers shrink the step near an edge and stretch it through smooth intervals, which suits detailed converter work. Fixed-step solvers hold a constant interval and keep run time predictable.
Stiffness usually forces the decision. A snubber capacitor of a few nanofarads beside an output capacitor of several hundred microfarads spans time constants five orders apart. A non-stiff solver such as ode45 crawls through that, cutting its step to satisfy the fast state while the slow state barely moves. A stiff solver such as ode23t handles it far faster.
Zero-crossing detection matters just as much. It lets the solver land on the instant a diode stops conducting instead of stepping past it, and turning it off to chase speed gives you a diode current that dips negative. When a model carries dozens of switches, the search itself becomes the slow part, and that’s where a fixed-step formulation earns its keep.
“Averaging hides the exact things a switched model exists to show.”
Time step selection sets your accuracy and run time balance
Time step selection is the single biggest lever on how long a converter simulation runs and how much switching detail survives. A workable rule keeps the step at 1% or less of the switching period, then tightens it until the numbers you care about stop moving.
Consider a 100 kHz converter with a 10 microsecond period. A 100 nanosecond step gives 100 points per cycle, which resolves ripple cleanly and places each edge close to where it belongs. Stretch to 1 microsecond and ripple amplitude reads low, duty cycle quantizes to 10% increments, and the efficiency figure comes out wrong.
The test that settles the argument is a step refinement study. Run the model at your chosen step, halve it, and compare the results that drive your design. Once halving moves peak current and average loss by less than 1%, more resolution buys nothing.
Before you trust a converter waveform, run through these checks.
- Confirm the switching frequency in the waveform matches the source setting.
- Check that inductor current ripple scales the way inductance and duty cycle predict.
- Compare input and output power, since a gap wider than the modelled loss signals trouble.
- Watch the light-load case for discontinuous conduction an averaged model would hide.
- Repeat the run at half the time step and confirm the headline numbers hold.
How switching losses and thermal behaviour enter the model
Switching losses reach a converter model in one of three ways, and each buys a different level of fidelity. You can ignore the switch drop entirely, apply conduction and switching energy from datasheet curves, or model device physics with parasitic capacitance and gate behaviour. Most production work sits in the middle option.
Efficiency targets in high-power computing justify that effort. Data centres used 415 TWh of electricity in 2024, about 1.5% of global electricity consumption, and nearly all of it passes through at least two conversion stages first. A converter reading 94% efficient in an ideal model and 91% on the bench sends you hunting a fault that was never there.
Loss data belongs inside the model rather than in a spreadsheet beside it. When conduction loss, switching energy and a thermal network share one model, junction temperature feeds back into on-resistance and the loss climbs with heat the way it does on hardware. SPS SOFTWARE keeps those component models open and editable, so you can see which loss equation is applied and set the parameters your datasheet provides.
Converter topology and study goal both shape method choice
Topology changes which simplifications stay honest. Hard-switched DC/DC stages average well, because the switching action is fast compared with the loop you’re studying. Resonant and line-frequency converters resist averaging, since the behaviour you want lives in the switching pattern itself.
Take a three-phase inverter feeding an induction motor. The outer speed loop settles over hundreds of milliseconds, so an averaged inverter block feeding a machine model answers that question in seconds. The inner current loop responds inside a few switching periods, and judging it needs a switched model with the modulation carrier represented.
An AC/DC front end pushes you the other way. Line current harmonics are the whole point of that study, and they only exist in a model that keeps the rectifier switching. Writing the question down before opening the model prevents the common error of reaching for the most detailed method available. Detail you don’t need still costs run time and gives a modelling error more places to hide.
Method discipline turns simulation results into engineering judgment
Good converter simulation is a sequence of deliberate method choices, each made because a specific question needed answering. The averaged model, the switched model, the solver and the time step are instruments rather than a ranking of quality. Picking the wrong instrument wastes hours or hides a fault you needed to see.
Engineers who work this way build a habit of asking what a model is allowed to tell them before they read the result. That habit compounds across projects. It becomes a quiet confidence about which results to act on and which ones deserve a second run at a finer step.
Converter templates in SPS SOFTWARE make that habit cheap to keep, since the DC/DC, AC/DC and inverter starting points carry solver and step settings you can inspect and change. What you get out of a converter model has always tracked the care you put into choosing how to build it.
“Picking the wrong instrument wastes hours or hides a fault you needed to see.”
