OK,( let's install this system!

[SafetyThird]

That's great thanks, something for me to aim towards

[Swwils]

Someone wrote a home assistant integration for HA that can do the forward planning, there is an official firmware for it too but it's locked to a single uk seller as far as I know.

There is also a LXP bridge that will let you do it.

[Oldgreybeard]

Someone wrote a home assistant integration for HA that can do the forward planning, there is an official firmware for it too but it's locked to a single uk seller as far as I know.

There is also a LXP bridge that will let you do it.

I've been trying out the forecast entity in HA, which has a go at estimating generation for the day ahead, but it isn't brilliant, TBH. The problem is that the available weather data just isn't good enough. Our nearest local forecast is for a weather station about 5 miles away, yet even that gets the sunshine forecast badly wrong most days. The problem seem to be timing. The forecast is mostly right in terms of what is going to arrive here, but it very often gets the timing out by a few hours. This means that we can get a clear spell earlier or later than forecast, making a very big difference to the actual generated energy.

Seems a really hard nut to crack, and probably needs a fair degree of AI to get it to try and replicate the way someone might interpret the data and decide what settings to use, knowing the issues with timing. I've been running an experimental forecast generation dashboard for a few weeks now, but I can't say that I use it at all for planning. I've just taken a screenshot for today, and that illustrates exactly why this isn't yet really useful. The dashed line on the solar production plot is the forecast of generation for today. It's clear that this is time-shifted a fair bit relative to the actual generation:

Forecast generation.jpg (63.87 KiB) Viewed 1725 times

[Swwils]

Solcast is spot on for me almost all the time.

[SafetyThird]

Looks like I’ll have to get on to Home Assistant and start playing. Now to see if I can find an RPi at a reasonable price

[Oldgreybeard]

HA is pretty flexible and easy enough to set up. The hardest part I found was getting data out of the inverter. The forecasting problem doesn't relate to any shortcomings in the way the forecast entity in HA works at all, it comes down to the unreliability of the local Met Office data that it relies on. No app is ever going to give a decent prediction when the core weather data is subject to so much change and error.

Yesterday, for example, the morning was far less bright than forecast and the afternoon was a lot brighter. Just down to a band of cloud coming over us that the forecast predicted would be a few miles away at that time. I can understand this, and can't see an easy way around it knowing how very different the actual weather is when compared to the forecast. For example, right now the forecast is for cloud, yet we have blue skies and sunshine. I suspect the cloud either passed over earlier or will arrive here a bit later, as that's the normal pattern for the forecast errors, one of timing.

[Joeboy]

Could the data from a light sensitive sensor working in real-time be incorporated?

I found that by growing the battery stack the vagaries of the weather were largely nulled out. Used to be a bit of a scramble when I only had 4.8kWh storage. Thought it worth mentioning.

[Oldgreybeard]

Could the data from a light sensitive sensor working in real-time be incorporated?

Could use all those big light sensors on the roof?

The problem here always seems to be timing, and that's critical when deciding whether to switch the system to auto mode or time-of-use-mode. Before switching to auto mode the day before I need to be 100% confident that the forecast generation the following day is going to be enough to both meet the baseload and to charge the batteries up enough to last the following night. If the forecast happens to be two or three hours wrong in terms of event timing then there is a high risk that there won't be enough generation.

All the generation forecast feature does take the local weather forecast data (in our case via the Met Office API for a weather station about 5 miles away) and combine that with the maximum generation for our exact location, using a model that is near-identical to that used by PVGIS. This model includes terrain data, so can predict the local horizon. The effect of weather is to reduce the output at any instant depending on the forecast cloud and rain cover. Unfortunately clouds move faster or slower than the Met Office predict a fair bit of the time, and sometimes bands of cloud just miss us completely.

I now play safe and just keep the system in time-of-use mode most of the time, with a 55% SoC charge threshold. This means we keep the house going using off-peak electricity overnight, but that the battery won't usually charge from it unless its pretty low from the day before. If it looks set to be clear for a day or two I'll switch the system to auto and run off the batteries pretty much alone 24/7. Until there's a better weather forecast methodology then I think this is the best I can do.

[Joeboy]

Could the data from a light sensitive sensor working in real-time be incorporated?

Could use all those big light sensors on the roof?

The problem here always seems to be timing, and that's critical when deciding whether to switch the system to auto mode or time-of-use-mode. Before switching to auto mode the day before I need to be 100% confident that the forecast generation the following day is going to be enough to both meet the baseload and to charge the batteries up enough to last the following night. If the forecast happens to be two or three hours wrong in terms of event timing then there is a high risk that there won't be enough generation.

All the generation forecast feature does take the local weather forecast data (in our case via the Met Office API for a weather station about 5 miles away) and combine that with the maximum generation for our exact location, using a model that is near-identical to that used by PVGIS. This model includes terrain data, so can predict the local horizon. The effect of weather is to reduce the output at any instant depending on the forecast cloud and rain cover. Unfortunately clouds move faster or slower than the Met Office predict a fair bit of the time, and sometimes bands of cloud just miss us completely.

I now play safe and just keep the system in time-of-use mode most of the time, with a 55% SoC charge threshold. This means we keep the house going using off-peak electricity overnight, but that the battery won't usually charge from it unless its pretty low from the day before. If it looks set to be clear for a day or two I'll switch the system to auto and run off the batteries pretty much alone 24/7. Until there's a better weather forecast methodology then I think this is the best I can do.

Could do! It becomes complicated quickly and then throw in the chaos factor of weather and
I read a few bits and bobs about the Tesla batteries and their AI back in the day. Even at that level of financial power for development there were issues. Not sure how it is now? Anyone got a powerwall?

[SafetyThird]

Yesterday I added two more US3000C batteries to the pack, taking me to 21Kwh. I think that should allow me to shift something close to 80-90% of our usage to off-peak E7 tariff electricity. To get any further, I'd need to add a second inverter and up the amount I can draw from 3.5kw to 7kw, as when the heat pump is running and we're cooking dinner in the evening we can easily be pulling 6-7kw for a while.