Then I realised couple of things, an humbling experience:
1) given any position on earth, you can compute exactly what's the optimal inclination at any given point in time for a PV to maximize the energy production. Sure, there are reflection and secondary irradiation conditions (eg.: there is a lake close to it), but again, assuming the environment is static, it's way faster to just compute it statically rather than dynamically. Also, in most scenarios Beam irradiance from diffusion (the beam hitting the object) is order of magnitude higher than from reflective one (the same beam bouncing on a 3rd object first).
2) In mechanics movable part are the things to avoid. They have lower MTBF (mean time before failure) and as such they introduce complexity and increase cost
3) Economics is a key component of engineering. There is a cost to everything, the computational power, the energy needed by the servo, etc, etc. Given 1 and 2, a dynamic solution simply has a lower ROI than a static one.
I really appreciate the OP exploration here: there is a good overview of basic control theory and a good foundation of ML (although don't be deceived, this is a very simple modelling task that OP is overkilling with a way more complex model). That said, for everyone reading, this is not something you want to do in a real world situation.
Commodity fixed angle ground mount photovoltaics arrays are low cost.
If you do the dollar and kWh produced in year calculation for spending $40,000 on fixed mount ground pv, and $40k on a combination of pv panels on trackers, and compare the kWh proxied by both... The fixed ground mount comes out far ahead.
A tracking mount can make sense only if you have a VERY small amount of space to work with and want the absolute most kWh per month per square meter of area occupied on the ground. And don't care about money much.
But I do wonder if heliostats might see quite a revival in agrovoltaics: there, you want a certain distance between panels anyways, and perhaps the plants won't mind if you steal a little more light off-noon in exchange for less shadowing at noon. Electricity supply/demand would certainly applaud this bias, in a market with lots of photovoltaics a Wh at noon is certainly worth less than those closer the the periphery of the daily sun cycle.
And if you do agrovoltaics right, the structure will be expensive anyways (making the markup for heliostat insignificant) because imho it's still just an unfinished prototype if the structures for holding the panels aren't designed to double as an overhead rail system for farming powertools that could become a considerable efficiency gain over the century-old game of tractor vs mud.
(I spent quite a lot of time on an idea for rooftop solar thermal power and was trying to build a prototype when the solar panel prices started crashing. It pretty soon became inescapable fact that small scale solar thermal with all its moving parts just wasn't viable any more. I'd be surprised if even the large mirror-farm CSP is competitive these days.)
I'm no engineer, so I can't determine whether it would work, but on the surface it looks like it should?
A 'joke' we had in engineering school: Anyone can design something to do X for $5, but it's an engineer's job to get the same results for $3.
In this solar project, the metric should be a comparison to the yield from pointing at the sun based on lat,long, and (earth) time.
Really, the analysis would have to include anticipated costs of installation and maintenance in comparison to a dumb array.
Perhaps the ingenious author could consider xy or xyz movement in an intermittently shadowed environment instead of 2-axis rotation. This might be be a better job for machine learning, or just a well-known control system problem.
Here is how they do it. https://gml.noaa.gov/grad/solcalc/calcdetails.html
Isn't this exactly the problem that can be better solved using real power data instead of values expected from theory ?
The trickiest part might be getting an accurate time on whatever cheap controller your using.
Just need your location and the time (quarter/season, month, day, hour, minute) and you'll know where the sun is in relation to the location given.
Maybe I'm missing something, but i would use a simpler algorithm which doesn't need ML. On day 0, plug in the latitude and allow the system to traverse the range of angles, finding the optimal one at the time - ie: yielding maximum power. Let it run 3-5 times during the day, then fit those points to the theoretical path of the sun across the sky. Now your system is calibrated, without needing any other input. As the seasons change, the system will always know which angle to face for optimal power.
[1]Micro hydro:
https://en.m.wikipedia.org/wiki/Micro_hydro
[2]Micro hydro power with turgo generator:
I wonder how much more accurate your system is and whether the tradeoff is worth the added expense of a motor + the additional maintenance cost of moving parts.
I wonder why solar farms don't use active tracking, is that added maintenance + equipment cost just not worth it?
This is incorrect, especially for solar farms
Panels are indeed only about 1/3 of the cost. Additional components, labour, inverter, mounting - another 1/3.
The last third is indeed soft(ish) cost, but this includes profit (duh..), certification, survey, tax, fees, etc. This can be reduced, but it is not going to magically disappear ...
Another fun fact : newer mega farms in low-altitude deserts are considering no-axis no-mounting zero-tilt - just laying the panels on the ground ... It all comes down to cost vs yield
A nice project by the way. Did you ever compare the results to pre-calculated angles based on time/location/season?
In my personal case I have 12+9 classic chains modules, I need more than 2x physical space to transform them with a dual-axis tracking setup. That means it's cheaper just add some fixed panels eastward and westward to catch extra power earlier and later.
Also in those terms: lithium storage is very expensive BUT for self-consumption is still the cheaper option to have electricity for more time, just arriving to a meaningful production 1/1.5h earlier and later in the day does not help much given it's added cost.
In costs terms: these days it's even cheaper (in TCO terms) having hot water heated by p.v. than the more efficient thermal because that cost more, have more moving parts and regular maintenance that just making an a bit bigger p.v.
The real issue in all cases is that to have enough power to really pay back the investment "quickly" we need much non-shadowed southward space witch can be found somewhere but far from everywhere. A similar issue is for EVs: I like the idea of charging them "for free" from solar, BUT since I normally use a vehicle during the day or I use it only sometimes or I have two or more in a round-robin scheme. Also lithium storage lifetime is an issue, on scale the production capacity and recycling are issues. Until we solve them just produce some more Wh it's meaningless...
We use intelligent power management over the fiber network to tell certain loads when to turn on or off or to change their operating parameters based on power conditions.
I’ve been daydreaming about building a ml based forecaster that just gives the next few hours weather outlook based on pressure, temperature, humidity, and a wide Nigel image of the sky.
I know it is doable because I can do it myself, and probably without any intuition about the pressure. It would automatically calibrate the model wights by feedback from the actual events vs the forecast. This would be really useful for me at least, in managing battery usage and otherwise managing the various systems that store energy like air compressors and large mass refrigeration.
https://www.youtube.com/watch?v=qguTFa9tj3c
Dyson Sphere Program · Covering Half a Planet with Solar Panels
https://www.youtube.com/watch?v=MKxkWgknkco
Dyson Sphere Program - Solar Panels
https://www.youtube.com/watch?v=yO78pXYnjFA
Full day night cycle of solar panels | Dyson Sphere Program
Will changing the orientation of the panel really have some effect on these (besides the drift in axis position) ?
Control theory is better if you know what you’re doing. ML is technical debt for sure.
Better to add more panels.
A better way to ensure more power output is to have a set of panels with a small battery back to automate cooling of the panels and cleaning of the panels.
https://www.youtube.com/watch?v=MiADday0mDA
https://en.wikipedia.org/wiki/Wax_motor
tl;dr - sun rises, temperature rises, wax material expands, motor actuates, solar panel tracks the sky as if it was a sun flower. Maybe the gloop is sentient or something.
And the next day your panels are staring in the wrong direction.
The latter exist but don't seem more profitable (in the case of PV panels because of lifetime problems due to more heat), and while we have gotten a bit better with motors I don't think there even is a lot of headroom to gain much efficiency.
But even then, I got panels on my roof last year. A nearby tree that was trimmed when we installed them and now the shade partially covers panels in the morning. If we had gimbles perhaps they'd be more efficient. Maybe not, and maybe we should trim the tree, but the point is that perhaps there is a use case that hasn't been thought of yet. Say, dropping cells on mars in a relatively unknown environment.
Of course, a $15 receiver might not have the best sensitivity, so reception might actually be a practical concern. On the other hand, you only need 4 satellites for a coarse fix, and you could seemingly tolerate a many minute cold fix time.
Panels that work with light from either side are preferred for this use.
NREL’s model puts soft costs at 44% for a 100MW fixed-tilt utility-scale plant in 2021:
https://www.nrel.gov/docs/fy22osti/80694.pdf
The percentage goes up as the installation size goes down, but it is not 70%
Low latitude?
For heating, photovoltaics supplying a heat pump is starting to give direct thermal a run for the money (well, not actually for the money yet, direct thermal is still cheaper, but at least in terms of how much you could harvest from a given roof area)
If money isn't an objection at all, e.g. if you strive for that sense of achievement of a good setup, there are hybrid modules that pick up the 20% or so photovoltaics achieves and still funnel the remaining energy into heating a liquid medium.
I had all the thermodynamics worked out and it would have been something like 5x as cost effective as photovoltaic. Then the cost of photovoltaic panels dropped 10x in a year. C'est la vie, at least my roof is covered in PV now. I've thought of running some tubing under the panels to pre-heat water for our solar hot water system but these days it's scarcely worth the bother (at least where I live which is pretty much perfect for solar power.)
Some example: https://www.convertenergy.co.uk https://dualsun.com/
It is a thing already: http://www.sulasindustries.com/technology/
The concentrator mirrors had to be aimed within about 2 degrees to aim the light onto the active area of the cells. I tried using a pair of photodiodes as the sensor, and it worked well in early prototypes. We flew out to Australia before the system was fully integrated into the car, though. In the real setup it just wasn't precise enough. On top of that, the aerodynamically shaped acrylic window surfaces created weird refractions and reflections that made the accuracy poor as well.
I had a couple weeks to figure something out, with a bunch of PIC18F4680 MCUs on hand, and access to parts from Dick Smith's. I ended up buying a small PAL security camera to experiment with. I went to a nearby photo center and got some overexposed film negative to use as an IR-passing filter. With enough layers of film, the camera image was completely black other than a white dot when pointed at the sun. Looking at the analog video signal on a scope, I was able to rig up a couple fine-tuneable voltage dividers and then use the MCU's dual comparator peripheral to generate interrupts on sync pulses and on white pixels. I could then count scan lines, and detect scan lines with the sun in them, giving me a Y coordinate precise to a fraction of a degree. I also got an X coordinate by timing between syncs and white pixels, but I didn't need it for control. I then mounted the camera on the concentrator mechanism, and wrote a basic PID controller.
It worked pretty well, and we were able to happily concentrate sunlight while driving at highway speeds.
It turned out the linear servo mechanism had a design flaw, where it would lose too much mechanical advantage at the extremes, and in the presence of road vibrations it would jam. We discovered this fairly early on in the race. Luckily I had wireless control of the motor over our telemetry system, allowing me to keep the motor from burning out. We drove a couple hours with the system jammed, taking the power hit over losing race time. Someone realized we would be driving over a "cattle grid" soon, and we had the idea to try running the motor at full torque to see if the shock would be enough to unjam the mirrors. It worked, and we suddenly started getting several hundred watts of additional power! After that day, we tied a string to the mechanism and routed it up to the driver cockpit. Whenever the mirrors got stuck after that, we simply radioed the driver and they gave the string a yank.
Just skimming their web site, range is qualified as "WLTP drive cycle". I assume that means software enforced torque and speed limits, and no one is going to want to drive with that enabled.
Using 4 independent motors implies they're using in-hub motors. They don't seem to mention a top speed anywhere. Unless their motors have dynamically adjustable air gaps (which I am guessing they don't), the motors will have a fixed KV constant. That means the top speed of the vehicle will be limited by battery voltage at some point. Maybe that limit is high enough not to matter, but it's funny they don't specify it.
Using in-hub motors also means there's a lot of mass on the wheels themselves. I'm not a mechanical engineer, but my understanding is that adding wheel mass makes suspension design a lot more difficult. On top of that, the rest of the car is presumably a couple hundred lb of batteries, and then a bunch of lightweight composites. The batteries will likely be in the floor, which is good for vehicle dynamics. The overall car will be significantly lighter than what people are used to. I think it's possible that the car will have a relatively rough ride, and could possibly have more noise than usual in the passenger compartment.
Crash-safety wise, I wouldn't want to drive that car on a busy highway. In the US, it would likely struggle to pass crash safety testing. I wonder if they are certifying it as a low speed vehicle, which would put it in the same regulatory class as a golf cart. It would have to have a 40kph speed limit on level ground.
It will probably cost more than a Tesla. At the end of the day, slapping a solar roof on a Tesla would probably get a lot of folk to the grocery store and back every few days. Mounting the same panels on the ground and plugging in the Tesla would be even more effective.
1954 - first practical silicon solar cell, Timeline of Solar Cells, https://en.wikipedia.org/wiki/Timeline_of_solar_cells
whereas 54, 60, 72-cell pv panels with aluminum frames made from 156mm cells are manufactured in metric dimensions.
The "if you know what you’re doing" here does not refer to the ability to understand control theory. It means that if you know the underlying dynamics, there is mathematically nothing better than controlling those dynamics. Flying a plane, oscillating a circuit, etc. are all things we can do very well without ML because we have exact models of the physical phenomena. Playing chess has no dynamics, control theory is useless. Anything where the dynamics are not "nice" differential equations, ML is probably easier at learning the dynamics than coming up with an ansatz.
There is a surprising amount of structure imposed by the assumption that the dynamics are differential equations, even if you don't know what the differential equations look like. As a consequence, adaptive control laws generally converge a lot faster (like, orders of magnitude faster) than MDP-based RL approaches on the same system being controlled.
The other advantage is that you can prove stability and in some cases have an idea of your performance margin with control theory. THis is important if you eg want your system to receive any sort of accreditation or if you want to fit it into the systems engineering of a more complex system. There's a reason autopilots don't use RL, and it isn't that RL can't be made to work. It's that you can't rigorously prove how robust the RL policy is to changes in the airplane dynamics.
https://oregon.public.law/rules/oar_860-039-0010
From the cursory look it seems like most states have limits like those. A lot of them are generous enough so you don't have to worry about it. But there are some states where covering an entire roof of a moderately sized building would put your over the limit.
https://www.ncsl.org/research/energy/net-metering-policy-ove...
Wait, panels are that cheap? Could you point me to a good place to buy them?
I don't know what, where, or how many you want, but these are by the pallet (~30ea) and about $6500. You can get much cheaper with used ones.
You get conservatively ~3hr of effective peak power per day (so a 400W panel will give you 1200Wh). You also need a DC converter for these to charge a battery or go to AC. Of course you'll want to mount them at the correct angle with a good view of the sun.
But used panels, at 80% rated output, are even cheaper, as low as $150. A few of those will fail, so you keep spares. They are often repairable in a few minutes if you are not afraid of a soldering iron, e.g. replace a diode or MOSFET.
Any roofer will put in mounting brackets, and almost any electrician is happy to put in the panel.
https://www.remodelingcosts.org/solar-panel-costs-increase-s...
That includes labor, mounting hardware, inverters and grid tie in. It also assumes high efficiency panels.
Watch out before buying older technology (lower efficiency) panels. Some have significant efficiency losses per degree Celsius increase in temperature.
Also, installation of the panels currently costs more than panels. They don't say (or I didn't find) how efficient the optimal fixed mount is, but the agent starts at 80%, so assume some fixed position is 80% of optimal. They increase that to 96%, so they reduce the number of solar panels by about 13%. If the installation labor cost increases more than ~26% because of the servo mount, then the servo mount hardware and frame would need to be cheaper than fixed mounts for it to break even. Similarly, the amount of aluminum being consumed by solar panel installions is non-trivial, and the movable frame is likely to increase that.
However, this is still a cool hobby project with a nice writeup.
That's partly the racquet people set up to snag government grant dollars, but still true.
More importantly, the motor drive electronics would have to be designed specifically to support that. A motor's top speed is limited by the supply voltage because a spinning motor generates its own voltage called back-EMF. When back-EMF matches the supply voltage, the motor drive can't push any current through the motor because there's not enough of a voltage difference. When the motor spins even faster than that, its back-EMF exceeds the supply voltage. This does something funny to the motor drive. The voltage across the FETs in its H-Bridge circuit swaps polarity and the FETs become "reverse biased". A reverse biased FET acts like a diode, and allows current to flow through it. As a result, the motor generates a braking torque. The only way for a motor drive to avoid that would be to have an additional semiconductor in series with the H-Bridge circuit. That would make the drive more expensive, and lower efficiency. It could be done, but I doubt it.
It's also hard to optimize a motor to work well at low speeds, but then have it be efficient at high speeds, even if not using it. One of the ways motors lose efficiency are in what are called "eddy current losses". Just the fact that the motor is spinning means that metal is moving through alternating magnetic fields, and that induces eddy currents in the metal. Those Eddy currents generate a braking torque, and heat up the metal.
This all assumes the motors are permanent magnet brushless motors. I think that's a safe assumption since they are in-hub. There are other types of motors, such as induction, that work differently and don't have the same "speed limit" or Eddy current losses when freewheeling. Most electric vehicles use one or two induction motors. They aren't in-hub, though, because they don't scale down in size well.
Besides, the panel can't move, only rotate and tilt. Have you actually read the article? (With rotate and tilt only, pointing directly at the sun should maximize the power output from a single panel even with shade from other objects. You can get a very small amount of movement from the fact that the point of rotation is outside of the plane of the panel. But that's not mentioned at all, is it?)
In any case, if you solve a problem with machine learning that already has a non-machine learning solution, you will get this kind of comment. If on top of that you don't compare the existing solution and yours and show that yours significantly improves performance, it just looks like doing ML for the kicks.
Be sure to specify if you need liftgate service at the destination or not, because that will affect cost, otherwise by default a pallet by LTL freight will need a loading dock to receive.
Which is why I was so fascinated by someone saying that you could get panels at 250USD/400W=0.625USD/W; I suppose it's possible that all the other stuff (electrician, mounting, inverters) is the difference, but a factor of 5? That feels like an chance to do something hacky and come out way ahead (like, say, DIYing a panel to run your computers, thus cutting out rewiring the house and needing an inverter).
One of the factors in cheaper solar is that the panels have gotten bigger. Panels grew from 250w to 300, 350, and now to 400w and 450w. The 450w panels are 82 inches by 42 inches, so taller than the average person. Larger panels require less mounting and less labor, so even if they cost more they might be slightly cheaper to install.
I think that utility scale solar will eventually beat residential solar on price because of less labor per watt to install. I sometimes think about a solar system that could be set on top of a house in a few hours and would contain the inverters and interlocks and be wired into a single breaker in the house electrical panel. A truck operator/installer and electrician could do two installs a day and the labor price would be significantly less. We are so far from this currently, with site surveys, permitting processes, individual panels in custom configurations and so on that result in several days of work spread out over months. I don't know if it could ever happen, but it is fun to think about at times.
Utility scale solar beats residential solar by a huge margin already. New utility scale solar projects on the California grid are priced around $20/MWh [1] compared to feed in tariff rates of $0.08923/kWh = $89.23/MWh [2].
As a renter who can't have solar installed on my home I find it pretty objectionable that my electricity rates subsidise expensive toy systems of homeowners.
[1] https://www.pv-magazine.com/2021/09/30/us-utility-scale-sola...
[2] https://www.pv-magazine.com/features/archive/solar-incentive...
Prices in the UK & Europe are lot lower per watt.
For example, https://www.leroymerlin.es/fp/88110787/panel-solar-risen-445...
In Germany, around 2019/2020 depending on the size of the installation it was around 1€ to 1,30€ fully and professionally installed on your roof. Right now with the increasing demand and very bad availability we are back at about 1,50€ to 1,80€ per Watt.