314 karma · joined November 20, 2021
See section 2: https://www.ti.com/lit/ml/slup239a/slup239a.pdf
And if you don't mind coughing up when it counts... TPS7A02 is 25nA, $0.45/ku. Triple the price, but more than four orders of magnitude better leakage.
Most of the inverter designs offload all the complexity to the software controller in an attempt to keep the power component choice and placement simple. The cool control stuff is mostly available in published IEEE papers, particularly from 2015-2020. It's not open-access, but it's definitely easier to get ahold of IEEE papers than a PV inverter. Once you know what you need to implement, the rest is just software engineering.
I've written my fair share of datasheets now, and while most of us are trying to do a good job and be clear and helpful, sometimes the stuff below the spec tables in the datasheet is, uh... less good than we'd like, for any number of reasons (inexperience, no time, someone left halfway through writing the datasheet, someone forgot to clean up copy-paste from the other datasheet with the slightly different thing, etc). I guess my point is: trust, but verify.
There's a handful of projects where the size of the solar field is large enough to make it economic to step up from 400V or 800V bus. I've seen many 1000V buses, a few 1200V and 1500V buses. Honestly it's exactly the same circuits, just with higher voltage ratings; all your switching elements are still giant hockey pucks, you're still doing a three phase hex bridge, etc. The half-assed flyback is sometimes replaced with something a little less braindead.
It is true that in most cases the inductor current isn't changing direction though.
But spin-up time... Yeah that still sucks :)
There's actually a bunch of ways to make buck-boost style converters that can do both functions, like the cuk and sepic topologies. There's also bidirectional bridge converters that can change input and output direction - you see this a lot on hybrid vehicles (12V to 48V or vice versa) and electric vehicles (48V to 400V and vice versa).
If you're careful about the converter design (keep the high-current loops extremely short, use counter-rotating loops to tightly confine magnetic field within those short loops, use a soft-switching topology to reduce EMI and sharp edges at the switch node, switch quickly or use multiple parallel converters at different phase offsets to reduce magnitude of current ripple), you can get decently low noise. There's a good Jim Williams app note about this.[0]
But it's almost never worth it to do this, since there's LDOs with two or three orders of magnitude better noise voltage. There's a time and a place for a really low noise converter; usually EMI constrained galvanically isolated converters like medical supplies or scientific instruments need them and aren't too sensitive to the cost or development effort. But even then, you'll often find LDOs cascaded on the outputs just afterward, since a good LDO can add another two orders of magnitude of ripple rejection in the switching frequency band.
[0] https://www.analog.com/media/en/technical-documentation/appl...
Step 2: Stop shorting the inductor. Inductors cannot instantaneously change current, so the now-built-up magnetic field continues pushing current into the switching node. The magnetic field and the inductor current linearly ramp down over time.
The forced current will push charge onto the parasitic capacitance of the switch (from switch to ground), the inductor itself (from inductor output to inductor input), and the reverse diode capacitance (anode to cathode).
Since capacitor voltage is charge over capacitance, once enough charge is forced onto the capacitance at the switch node, eventually the voltage from the switch node to the output capacitance is high enough to turn the diode on in forward conduction. The rest of the inductor current is forced into the output capacitance until the remaining magnetic field in the inductor is depleted.
Step 3) Repeat very fast to reduce inductor size and ripple current required (100s of kHz or MHz speed). Vary the duration for which the inductor is shorted in step 1 according to how much charge you need to put on the output capacitor. You could figure this out open-loop by noting that output current at the high voltage side is in charge per second, output voltage is equal to charge over output capacitance, calculating the time taken for the ramp to grow and decay, etc. Or you could design a closed loop control scheme that looks at the output voltage and converts it to shorted duration for you (this is what most integrated circuit boost converters do).
In summary, you dump current into a inductor to build up a magnetic field, then you use the inductor's magnetic field to yeet current up over a large voltage difference.
1) LM317 and LM7805 are not LDOs. LM317 is a series regulator, which is way more flexible and has some infrequent but difficult-to-emulate use cases. LM7805 is a linear regulator but has substantial voltage drop relative to a modern LDO. Enthusiast/hobbyist space would do well to consider alternatives to the LM317/LM7805 if they're designing PCBs with surface mount components and very simple power rail needs, but for quick and dirty through hole designs there's few well-known alternatives that are truly LDOs - maybe MCP1700?
2) Enthusiast/hobbyist market isn't too sensitive to bulk pricing. Case in point: LM317 is about a dollar in hobbyist/enthusiast quantities, and LM7805 is comparable. TI lists 1ku TO-220 pricing at $0.65. You're definitely right for real products with mass manufacturing, but hobbyists don't generally haggle over pennies.
The kind of DC-DC converters that work well for megabattery to 10kVDC conversion will look very different from the kind of DC-DC converters that step up your 3.3V rail to 5V for some low-power peripheral, and may actually have individual components that completely reverse current direction for more efficient current transfer. You pretty much need a transformer to handle the high power transfer and voltage ratio mismatch. Depending on the pack voltage you might use multiple stages cascaded, but typical 400V-800V packs can step up to 10kV in a reasonable number of turns (12-25). The battery pack side probably has IGBTs or SiC FETs driving some kind of large bridge switcher (or several parallel bridges); the 10kV side probably has some big chonker diodes in a rectifier bridge, though they conceptually could be replaced with synchronous switches if you could find thyristors with fast enough switching speed and better efficiency (usually it's not worth it). Technically this topology runs the transformer primary current in both directions (hence the rectifier at the output) so I guess this is the "AC" stage in the middle... But it's worth pointing out that the AC portion is incidental to operation, unlike something like a Tesla power wall using an AC inverter to feed power into the AC grid, and a grid-connected charger converting that back to DC.
[0] https://support.google.com/product-documentation/answer/9682...
Understanding the technology and understanding its attendant risks aren't completely orthogonal, but neither does one imply the other - or else the kids who "instinctively know about technology" would also instinctively know the risks. On the other hand, it's probably faster and easier for someone who understands the technology to learn about the attendant risks, and understanding the technology may be required to personally validate the risks or to invent the best remediation strategies. But a considerable amount of parenting is, and always has been, received wisdom, for better or worse. I can see an argument to describe "lazy" as neglectful, but I take issue with "stupid": the scope of a parent's best reasonable effort to understand and mitigate technological risks for their children should be conditioned on what level of understanding they can realistically achieve, or else "neglect" is reduced to an IQ test.
When an airplane crashes due to a design failure, thorough investigation often yields a specific design element or combination of elements acting as a root cause; and we have a much broader range of corrective actions available, including modifications to the design for improved redundancy, fault tolerance, higher likelihood of correct manufacture, lower defect rate, easier cockpit control, etc. We can also make process changes, like modifying the maintenance schedule, updating the maintenance checklist, pre-screening critical components for correlated early signs of defect, updating pilot training manuals, etc.
There are also plenty of cases where aircraft crashes occurred due to malfeasance. In rare cases, there might be a design mechanism that can prevent malfeasance, but aircraft are overwhelmingly designed to be built, maintained, and operated by people who aren't trying to misuse them for mass violence (sort of like malls, or public schools). But there's legions of process controls implemented on the hiring and training of the people directly involved with aircraft manufacture, maintenance, and operation.
And yet, Boeing still made a plane that dropped out of the sky twice. And that wasn't even necessarily malfeasance - a good chunk of it can likely be chalked up to second or third order effects of cost saving measures and siloed design teams. I guess you could make an argument for malfeasance, but something something incompetence indistinguishable from malice... in any case, the technical analysis of the design problem is totally solved, but what about the process control by which the design came to be flawed in the first place? I genuinely don't know whether they've modified or added to their process controls to prevent the same design flaw, let alone others not directly related.
Scores of new process controls were put into place in the wake of 9/11 (which incidentally resulted in about the order of magnitude of casualties observed in all US mass shootings since (I think? There's a few different numbers floating around, depending on the threshold for which a mass shooting is recorded)). Many additional "process controls", if you'll allow the tortured reading of the concept, were implemented outside of aviation - a multi-trillion dollar double decade war campaign, passenger luggage and body scanning, and mass surveillance systems, to name a few. 20 years later, and none of this prevented a pilot from doing some unauthorized loops in the Seattle sky before crashing, and very likely nothing except for a lack of killing intent prevented him from flying his plane into a building downtown. Were the process controls implemented after 9/11 effective? I genuinely don't know the answer. Maybe aviation terrorism just has a low base rate to begin with. Maybe I'm uninformed, and our process controls have foiled numerous attempts at terrorism. It's hard to tell with rare events.
Unlike with engineering design work, where you can deduce real responsibility for failure and synthesize a test case to prove the soundness of a design, debugging complex and rare psychological or sociological phenomena is substantially more inexact. Mass shooters are a statistical anomaly against the population as a whole. People are aware of process controls and can deliberately plan subversion tactics. The snarled mess of genetic predispositions and environmental insults that drive someone to terrorism are not guaranteed to be the same across multiple people, especially at the long tail of the distribution. Preventative policy proposals can be misguided or undermined for numerous unrelated reasons, never achieving the desired effect.
None of this is to say we shouldn't try to find and implement effective process controls on mass violence... I just hope to offer an explanation why mass violence can't be approached with the same rigor as human-designed engineering systems with any predictable return on investment.
All this to say, this is a hard problem, composed of thousands of overlapping adjacent elements, we have very little insight (or reliable insight-gathering mechanisms) applicable to the system under investigation, and so throwing money or effort at the investigation is a gamble that a feasible solution exists at all, rather than a process of concrete deduction with a deterministic endpoint. Maybe the optimal amount of scrutiny should be different.