Even-Order Harmonics: Why APFs Miss Them

Ask a plant engineer what their harmonic problem looks like, and you will usually get the same three numbers: 5th, 7th, 11th. Those are the characteristic orders of a balanced three-phase load, the ones every filter is tuned for, and the ones most harmonic studies report first.

Then the filter goes in, the 5th and 7th drop as promised — and the transformer still runs hot. The neutral still carries current it should not. Something in the spectrum is still there.

In medium-frequency melting shops and rectifier plants, that "something" is usually the even orders: 2nd, 4th, 6th — and further up.

Where even harmonics actually come from

Even-order harmonics appear whenever three-phase symmetry breaks. In practice that means:

• Half-wave or asymmetrical rectification — one leg conducting differently from the others
• Asymmetric thyristor firing angles, common in older rectifier and furnace supplies
• DC bias or residual DC component in the load current
• Single-phase loads sharing a bus with three-phase equipment
• Unbalanced furnace operation, especially during melt-down and electrode movement

None of these are exotic. They are ordinary conditions in exactly the industries where harmonic problems are most expensive.

And they are easy to under-estimate, because the limits themselves are modest: IEEE 519 caps even-order harmonics at 25% of the corresponding odd-order limit. A 2nd harmonic at 1% looks negligible next to a 5th harmonic at 4%. But even-order content behaves differently — it drives DC flux in transformer and reactor cores, produces asymmetric heating, and can push a core toward saturation long before the total THD figure looks alarming.

Why conventional APFs don't solve this

Most active filters are engineered — correctly — around the odd-order spectrum. That shows up in three ways:

1. Control bandwidth and detection logic are optimised for characteristic orders, where the energy usually is.
2. Even-order content is often treated as measurement error, because in a healthy balanced system it should be close to zero. Some units reject it to keep the phase-locked loop stable.
3. Filtering a spectrum you don't measure is guesswork. Without explicit even-order detection, a unit can run for years and never touch that part of the problem.

The result is a filter that does exactly what its specification says, on a site that needed something the specification never asked for.

Designing for the whole spectrum

Our units are built to compensate even and odd orders in the same device — 2nd, 4th, 6th alongside 3rd, 5th and 7th — across the bands that actually carry pollution in industrial grids.

The second point in that figure is the one that matters commercially. Detecting and removing an even-order component demands a different detection strategy from the one used for characteristic harmonics — the unit has to separate a genuinely small, symmetry-breaking signal from the noise that surrounds it, then act on it without destabilising the control loop. That is engineering work, not a checkbox in a datasheet.

The number that decides whether it survives the site

Any filter can be specified for a clean bus. The question is what happens when the bus is already badly distorted before the filter is connected.

On a recent field installation, phase voltage THD measured 30.7% / 27.4% / 28.6% — roughly double the level at which many standard units begin to current-limit, derate, or drop out of regulation. The unit kept compensating normally, holding a stable 151–156 A RMS per phase at 50 Hz.

This is the difference between a product validated in a laboratory and one validated on a grid. At 30% voltage distortion, the grid voltage is no longer a clean reference — zero crossings drift, the PLL is under stress, and the current loop is being asked to track a reference derived from a badly deformed signal. Units that quietly stop working rarely announce it; they simply stop contributing, and the site blames the load.

Output quality is part of the specification

Compensation current is only useful if it is clean. A filter that injects a ragged, spike-laden current trades one distortion problem for another — extra high-frequency content, additional reactor heating, audible noise, and a waveform that fails its own commissioning test.

Spike-free output is not a cosmetic property. It is the evidence that the switching strategy, the output filter and the current loop are working together rather than fighting each other.

Built for the grid the plant actually has three  design decisions carry most of the field reliability:

• Voltage distortion tolerance — designed for furnace and rectifier duty, with stable operation specified at voltage THDu ≥ 15%
• Immunity — no false operation, no oscillation, no tripping out in complex electromagnetic environments, so production continuity is not traded for power quality
• Wide input voltage range — the core resists saturation and is not clamped by harmonic voltage, which is what keeps the unit alive on a weak or distorted feeder

What to ask before you specify

If you are buying an active filter for a furnace or rectifier site, five questions separate a real answer from a brochure:

1. Does the vendor state even-order capability explicitly — with the orders listed, not just "2nd to 50th"?
2. What is the field data at THDu above 10%, and was it measured on site or in a lab?
3. What does the unit do at 25–30% THDu — derate, current-limit, or trip? Ask for the curve, not the claim.
4. Can they show an oscilloscope capture of the output current, rather than a screenshot of a THD figure?
5. How is the even-order component detected without destabilising the control loop?

A vendor who can answer all five has done the work. A vendor who answers three has a product that will work on somebody else's site.

If you are dealing with medium-frequency furnaces, rectifier banks, or any non-linear load where conventional filters cannot hold regulation, send us your harmonic spectrum and voltage THD data. We will tell you honestly whether our equipment is the right fit — and if it is not, what would be.