Extract metrics from a Kostal PV inverter and write them directly to VictoriaMetrics or InfluxDB v2.
Repositories: GitHub · Codeberg mirror
Set VM_HOST to write directly to VictoriaMetrics. VM_PORT defaults to 8428, and VM_TOKEN optionally sets a bearer token. InfluxDB writes are disabled by default; set INFLUX_ENABLED=true to use InfluxDB alone or alongside VictoriaMetrics.
When VictoriaMetrics is enabled, a background goroutine fetches /yields.json?total=0
and backfills whatever the inverter still remembers, best resolution first. It runs
beside the scrape loop, never inside it: its own paced VictoriaMetrics client
(-historyPace, default 250ms between requests) and its own long timeouts, so a
multi-megabyte import can never delay a 5-second reading. It repeats every
-historyInterval (default 6h, 0 runs once, negative disables) because
DayCurves is a rolling 31-day window: a gap left by an outage is only
recoverable at 10-minute fidelity until it ages out. Every tier
writes the same four metrics — kostal_AC_Power_W, kostal_grid_consumed_watts,
kostal_grid_injected_watts, kostal_own_consumed_watts — in watts, and is
identified by a source label:
source |
from | covers | resolution |
|---|---|---|---|
| (absent) | live measurements.xml polling |
now | 5s |
history_10m_avg |
DayCurves |
rolling 31 days | 10-minute average power |
history_daily_avg |
MonthCurves |
rolling 13 months | one day's energy as flat average power |
history_monthly_avg |
YearCurves |
up to 20 years | one month's energy as flat average power |
Every tier only fills gaps: the low-resolution ones skip calendar days that already
hold any sample, and history_10m_avg skips 10-minute buckets that already hold
one. Energy is therefore never counted twice, and restarting the daemon is cheap
and idempotent. Re-importing a tier means deleting its series first:
curl -X POST --data-urlencode 'match[]={source=~"history_.*"}' \
http://victoriametrics:8428/api/v1/admin/tsdb/delete_seriesDays with zero yield are skipped, so a genuine outage stays a visible gap.
A day's energy is written as 24 hourly samples of constant average power (Wh / 24).
That is deliberately lossy in shape but exact in total: the dashboards' hourly
integration sum_over_time(avg_over_time(x[1h])[1d:1h]) reproduces the inverter's
own daily figure to the watt-hour.
These inverters have no NTP: the clock sits at an arbitrary offset from UTC and
free-runs from there (mine is 65 minutes slow). History timestamps are in that
clock, so they are useless until corrected. HISTORY_OFFSET defaults to auto,
which measures now - measurements.xml DateTime at every import and rounds it to
the curve's 10-minute bucket. Pin a Go duration such as 1h5m to override.
make test-e2e runs the import three times — fresh, restart, restart with a
changed clock correction — against a throwaway VictoriaMetrics seeded with
live-like data: 5.5 years of samples with the gaps the real database has, and
-search.maxSamplesPerSeries scaled down to keep the production ratio between a
coverage probe and the limit. An empty TSDB only exercises the first-import path
and hides every interesting failure: result alignment, per-series sample limits
and re-import behaviour all need data that is already there.
The dashboards live in dashboards/. The source label makes each
backfill tier its own Prometheus series, so aggregate it away to render one line:
sum without (source) (kostal_own_consumed_watts)
sum without (source) (
sum_over_time(avg_over_time(kostal_grid_consumed_watts[1h])[$__interval:1h])
)
The "Sample provenance" panel on each dashboard keeps the tiers visible:
clamp_max(count_over_time(kostal_AC_Power_W{source="history_daily_avg"}[$__interval]), 1)
The latest published container image is ghcr.io/msf/kostal2influx:v0.7; ghcr.io/msf/kostal2influx:latest currently points to the same image.
docker pull ghcr.io/msf/kostal2influx:v0.7My inverter has an old firmware, so it doesn't have the http://hostname/api/dxs.json
So I couldn't use work like: kostal-dataexporter
So, looking at the source code of the page: http://hostname/pages/livechart.html
there's a XML endpoint at http://hostname/measurements.xml which is used to read the data.
I've created a 
to monitor my system.
It uses raw data, but it also uses synthetic metrics to at a glance read total power generation and consumption (see kostal_inverter_msf). Here's the json config of the dashboard