Repository navigation
RoofCast: an open-source PV forecast for EMHASS, 8 Open-Meteo models calibrated per roof face #1341
Replies: 3 comments 6 replies
|
Thanks for publishing this. Measuring the level factor on the forecasts as issued the day before, and training on the Previous Runs API, are the two things I'd have checked first, so it's good to see both. I haven't run RoofCast, but I read the EMHASS side of
I rendered the new template on Home Assistant 2026.11, on the Solcast fallback with real data and on rows shaped like yours, and ran the payload through EMHASS 0.18.8's runtime parsing offline at 15 and 30 min steps. I haven't sent it to a live EMHASS or run it against your sensor, so a report from your setup would be welcome. On a native input, that's David's call, but I think a recipe would cover most of it. The cookbook index already lists custom forecast injection as a wanted recipe, and On the multi-inverter work, thanks for the offer. Two things first. It's on master only, not in 0.18.8 yet, so testing means running master. And do your two Deyes each have their own battery and strings, and can you command them separately in master/slave mode? EMHASS ties each battery to one inverter, so a shared bank behind a master/slave pair is best left as one inverter with the combined limits. If they're separate, EMHASS wants one PV series per inverter, so you'd add up the faces on each Deye. Your |
|
Hi, thank you both for the discussion. |
|
Follow-up on P10: I backtested Setup: EMHASS 0.18.8 offline, Energy cost over the 37 days:
The follow figures for 0.3 and 0.5 are corrected for the end SOC, which was 15 % and 18 % against 17 % at bias 0. The difference is valued at the mean buy price. Uncorrected they are 188.39 and 188.75 €, so the correction is as large as the effect. The rigid runs all end at the 15 % floor and need none. The differences are within noise. In follow mode P10 saves 0.42 € (0.3) and 0.82 € (0.5) over 37 days. With each day's change in stored energy valued the same way, the daily differences have a standard deviation of 0.13 and 0.20 €, which puts the uncertainty on the total at roughly ±0.8 and ±1.2 €. In rigid mode P10 came out 0.64 and 2.89 € dearer, against ±1.4 and ±2.1 €. The saving is also not tied to bad forecasts: on the 6 days where actual PV came in under 70 % of the last P50 issued the day before, P10 saved 0.05 and 0.18 € in follow mode. Part of the reason is probably that this P10 says little that the P50 doesn't. It sat near 0.65 × P50 for the first month and 0.53 × P50 in the last week, so bias 0.3 and 0.5 come down to cutting the P50 by roughly 10 % and 18 % every day. Limits: one installation, one autumn month, perfect load and price forecasts, and an hourly step instead of the 15-min step I run live. Each hour is priced at its last quarter-hour price. The simulation leaves out PV curtailment at negative sell prices (25 hours), the EV and other deferrable loads, nowcast corrections, and the import cap I use against the capacity tariff, so it says nothing useful about peak imports. I'm keeping the bias at 0. The P10 key stays in the recipe because at bias 0 it changes nothing, and it lets people test their own setup. |
Uh oh!
There was an error while loading. Please reload this page.
Hi all,
After a chat on Reddit with one of the EMHASS contributors, I've published the PV forecast engine I use with EMHASS. It's called RoofCast: https://github.com/JustAnotherPU/roofcast
Why I built it
My roof has three faces (SSW, west, ESE), and a tree and the roof itself shade parts of it. Solcast and the stock Open-Meteo integration were both consistently off for it, and the EMHASS plan suffered for it.
What it does
detailedForecastattribute (30-min slots, today plus two days, P10/P50/P90) and a per-array splitResults on my roof
Day-ahead error over the last 30 days: 21 % for the calibrated ensemble, 32 % for the stock Open-Meteo integration, 44 % for Solcast. That's one roof in one location, so treat it as an indication.
How it plugs into EMHASS
At each MPC run I pass the curve as the
pv_power_forecastruntime parameter (W, one value per optimization step). The repo has an example HA package inha/pv_forecast.yaml: an input_boolean to switch between this forecast and another source, a freshness check, and a rest_command for EMHASS. The rest_command part is a generalised version of my own scripts and is marked untested, so I'd like to hear how it behaves on other setups.Where I'd like help
I'm following the multi-inverter work too: I run two Deye hybrids in parallel (master/slave) and I'm happy to test it.
Thanks for EMHASS.
All reactions