Hi,
my solarforecast is completely off - ist shows almost 10time more energy per hour then it should.
I have 20kWp solar but mostly north so the max peak I have ever seen was 10kW.
If I take then the irradience forecast of 550W/m2 I might be able to produce ~ 6kWh/h at noon.
somehow the forecast graphs show 60kWh so around ~10times more.
If I compare it with the DESS forecast which shows ~17kWh/h in peak it does not get better.
Any ideas how this can happen? I already reset the forecast useage but seems not to improve. The consumption forecast looks OK btw.
so I am asking myself how is the algorithm working here?
The production forecast is based primarily on past experience. In other words, given the weather conditions and the position of the sun, the system will produce x kWh. That’s why Victron recommends a runtime of approximately four weeks for a reasonably reliable forecast. By resetting the previous forecast data, you could achieve the complete opposite. Give the system a few days to recognize your parameters.
“based on experience” - this is a common/standard statement I know - I had the system running for several month already. It was never working properly for the solar engery forecast. That’s why I ask how is the forecase being calculated - I would like to understand the mechanism behind.
If I take e.g. yesterday which looks like comparable with todays sun forecast it should look like:
yesterdays real data
Forecast for today
to be factor ~10 of would be the worst forecast alg. I have ever seen. The consumption forecast looks OK with 10% difference.
A simple math calculation shows that 60kWh production in one hour is completely impossible with a 20kWp (north) and irradiance forecast of Germany.
Of course, that’s nearly impossible for a north-facing solar system. But how does the system know it’s a north-facing solar system? Or how much power is installed in which orientation? The system learns. With 5 hours of sun from the east, I get x kWh. Tomorrow, there will be 3 hours from the east, so x/5 × 3 =… This is a very simplified representation, but it captures the principle. My forecast has been running for 2.5 years now, and it’s pretty accurate. I work outdoors and rely on good weather forecasts. These days, I prefer to look at the solar forecast in VRM instead of the weather report.
Your problem is that you have just recalibrated the forecast to the beginning and the system is now starting again from the beginning
Nope please read my statement that I was running it for several month before without any improvement. Resetting was the last hope to fix it.
- It does not even need to know the orientation since a 20kWp system cannot deliver more than the peak. So 60kWh/h is three times above theoretical max already.
- PLUS the systems is even after reset running several days now. Simply extrapolate the last days radiation + solar energy production then you would get somehow accurate figures.
So there must be other influencing factors that I would like to know.
Another observation: when parts of the day are gone the forecast in many cases will be corrected to a realistic figure
Here the same today graph as above just few hours later shows now <4kWh/h for the rest of the day.
This cannot be explained by learning - it seems to be a bug. Forcasted total of the day drops from 362kWh to 24.5kWh.
With a little thought, you’ll figure it out yourself. You write that you ran the forecast from November onwards. How much radiation did the north-facing system receive at that time? The output should tend towards 0. Right? At some point, the time came when a little radiation occasionally reached the system in the morning or evening. Right? That led to a sudden increase in PV production. Right? So much for the past few events. Now the system says that if the 2-3 degrees higher sun position causes such an increase, the increase will continue for the next few days and reach your factor of 10 in production. That’s not correct and won’t be the case, but the system would have noticed this error after a few days. I don’t know exactly how far back the DESS looks at the data and whether different behavior will occur next year based on current experience. But I think you’ll observe a similar problem in the fall during the spontaneous decline, when there is suddenly no more radiation on the north-facing roof. I don’t know how you could handle it differently. This year is also likely to be extreme for your north-facing system. Hazy all winter and then for a certain time the direct radiation on the panels
I still believe it is a very bad forecasting algo but if you theory is right the forecast should dramatically improve the next days. I will post regular updates then lets see.
BTW: where did you read november? System is running since April last year.
I don’t think such a deviation is normal.
Could you post your site id (the numeric one in the VRM url) then somebody from the forecasting team could have a look, why the forecasts are so high for your system.
Did you set up the geolocation of your system properly?
2 Likes
423190 installation ID
Geolocation is correct
What maybe is not standard: I set the ESS target SOC on a higher level if the electricity price is below a certain threshold. Unfortunately DESS does not support this (only difference between min max costs). But this should not influence the solar forecast.
The weather forecast is just really bad.
For me, the forecasted energy looks every day exactly the same, when there is full sun, the predicted energy is on the low side but sort of okay, but when it is a completely cloudy day, it is off by miles. It just looks like it doesn’t ever get a real weather input.
Also, at 16.00 every day, the sun is fully away from the solar panels, but it keeps giving the same, way too high amount in the forecast.
The system doesn’t learn anything for some reason.
Just to have a status update:
Forecast
reality
So factor 3 today (mixed sun + clouds). Not as bad as yesterday (factor 9.x)
Lets wait a few days and also wait for a sunny comparable day.
For solar forecast there is no learning needed. It is just take last x days calculate the efficiency of the system (kwh/h and actual radiation/h) and then calculate using the forecasted radiation the next days. No rocket sience can be done by anyone.
BTW: Victron’s PV forecast is done by Solcast and it’s one of the best forecast services you can get!
Usually you can also provide elevation, azimuth, and peak power to the algorithm so that is has a base line until it has gathered enough data to learn about your PV plant and how it reacts to the weather and possible shadows but they skipped that in favor of an easier setup I guess (otherwise you would need to specify a lot of details in case of a plant with multiple orientations)… So it’s really just a matter of time for it to learn and a north-facing PV plant in the winter is the worst scenario for that kind of machine learning I would say! It can only get better now in spring.
And please note: And it’s not for free. Victron must be paying for it and I’m curious how long they can afford to not charge us for DESS…
PS: and you’re wrong: I’ve been working in that business and PV forecast, especially for setups like yours with multiple orientations is not easy-peasy! Especially not if you just want to learn the behavior from production data without any plant-specific details!! Just have a look at the shadows from surroundings (buildings, trees, hills) in winter vs. summer! They make a huge difference and your model needs to be able to capture that on a seasonal basis while differentiating it from the effect of clouds!!!
1 Like
I disagree as we are not looking for a longterm forecast. If you just take average resulting values of the last few days + radiation forecast you will have a pretty accurate prediction for one day ahead. Yes shadows and sun position changes but what will be the impact within ONE day? If it is some percent off track - no problem but victron is off for factors.
I would agree if you would need to create a prediction for a longer period, but as (at least here) I get also only the price prediction at 2pm for only the next day - long term solar forecast makes no sense anyway.
I 100% disagree because you seem to not incorporate the weather and it’s forecast into your equation, which is pretty much crucial to get it right, especially in climates where you don’t get 100% clear sky every day!
Of course - and that’s what I already tried to express - if they had details of your PV plant like orientations and peak power they could use clear sky forecast as a good base line / upper limit. But apparently they don’t.
1 Like
Of course I do - radiation (lets use irradiance as more accurate term) forecast is depending on the weather. Bad weather low irradiance forecast - sunny weather high irradiance forecasted.
today (monday) mixed weather tomorrow slightly worse and in two days we have bad weather forecast with only 0.3h sunshine and a lot of rain. On friday the weather forecast states 9.2h sunshine no rain. Which fits typically to the irradiance forecast (as indicator).
Lets have a look at the last 7 days:
Irradiance the last days
Real kWh produced per day
The integral of the irradiance fits okish to the real delivered energy. Not perfect but the trend is visible.
Obviously this is a simplified view but it works more accurate as what I see currently in VRM. E.g. if you have multiple directions foggy weather increases the portion from north PV and sunny weather increases the south one if your system is not equally distributed.
Data of today:
Forecast
Reality
111kwh vs 29.4kwh → factor 3.7 of track.
Ok, now I see - you’re talking about an irradiation forecast that’s already based on the weather and most likely provided by Solcast too…
So yes: I agree that the final PV forecast in VRM performs very bad in view of that and it would - as I said previously - must likely do much better if it would take into account the real parameters of the PV plant (that is orientation and peak power). Instead it tries to guess these numbers and something in the historical data must have flawed it, a combination of hard shadows and sun peaks maybe… I’m very curious how it will look like when 1 year of production data was gathered…
PS: I myself have 2 pv plants - one to the East and one to the West where in the winter there is almost no overlap due to shadows from the house in-between and in the summer there is more than 1 hour of overlap. The data of both plants is feed to VRM seperately (via dbus-opendtu plugin) and the PV forecast in VRM is pretty accurate so far (10..20% off). But the big difference here is that it can learn about both PV plants independently. If I would provide only aggregated data, i.e. the totals of everything, it certainly would be much worse.
PPS: it is possible to use Solcast directly and provide PV plant data seperately with proper parameters (orientation, peak power), e.g. via HomeAssistant: GitHub - BJReplay/ha-solcast-solar: Solcast Integration for Home Assistant There you can also inspect and tweak the dampening values (which are meant to model shadows from the surroundings). Most likely this will be much more accurate than what you get from VRM (but surely doesn’t improve DESS performance 
).
The interesting thing is that Victron does this for all of Europe? Worldwide? For all orientations, angles of attack, shading, and so on. Of course, this compromises accuracy somewhat, but without specifying hundreds of parameters. When I look at the problems with configuring the few parameters in DESS and some of the opinions expressed on the matter, I can’t imagine the chaos it would cause if every user had to specify 25 parameters for orientation, kWp, and shading. What would then be going on here in the forum? I fear Victron would have already given up on the DESS project in frustration.
Does not need to be complex, as it does not need to be super accurate - I have shown already the most simpliest way to calculate (rule of three). Of cause knowing all parameters (each single PV size, angles etc…) would improve - but you are right most users would be overwhelmed.
Comming back to my orginal question:
How is the algo working and which parameters are influencing the result. (no guessing)
The forecast is so unbelievable bad (in my and many other cases) that there must be something wrong. If we would know how it works it would most likely clarify many issues. Making a secret out of it does not help.
forecast of today
result
39.2 vs 116kWh factor 2.9.
If I would be optimistic then it is maybe improving - factor is now <3 (starting from >10) - we will see.
Today could be the first day that the forecast is correct:
irradience forecast is <50% of yesterday and
Energy Forecast is similar
15.2 compared to yesterdays 38.2kWh.
Lets see if this works for tomorrow as well as tomorrow should be sunny again.
