SolarEdge firmware is known to silently drop the Modbus TCP listener
(port refuses connections while still shown as enabled). Write a
connection_state=1 point on first failed poll and 0 on recovery so the
outage is visible in Grafana instead of just a data gap.
ConnectionException is re-raised from read() - it is fatal for the
whole cycle, not a flaky single block. Marker points carry no tags
(registers unreadable during outage), so the tile query skips the
serial filter.
Also includes threshold tweaks on the power stat tiles.
The battery register reads negative on discharge (raw since day
one), but the tile label promised '+ Entladen / - Laden'. Flip the
three raw battery1_power/current queries with * -1 so positive
means power flowing into the house - the same mental model as the
grid tile (+ Bezug). Full history renders consistently; surplus
panels already use the correctly-signed derived fields.
Verified against simultaneous SolarEdge app readings (PV 40 W,
battery discharge 520 W, house 560 W): battery1_power negative
means discharge, and power_ac on this DC-coupled Energy Hub is
the inverter's TOTAL AC output - it already contains the battery
contribution. The old formula double-counted the battery, so
house_load collapsed (44 W shown vs 560 W real) and the PV tile
showed battery-inflated totals, which looked like PV and house
consumption were swapped.
New derived fields: house_load = power_ac + grid, and
pv_power = power_ac - battery (true PV production). Dashboard PV
tiles and both yield panels now use pv_power / integral(pv_power)
- the energy_total lifetime counter also counts battery
throughput and is not a PV yield.
Short or garbled gateway responses are rejected on register count
instead of half-decoding (previously: 'unpack requires a buffer',
OverflowError on garbage scale factors). Scale factors outside
+/-10 and non-finite floats are dropped before they reach line
protocol.
offgrid is written as integer offgrid_state: the existing DB
cannot be migrated (mixed inmem/tsi1 shards block all deletes)
and the old boolean field conflicts with integer writes, which
poisoned the spool - HTTP 400 now drops the batch and logs the
InfluxDB error body instead of re-spooling it forever.
grafana_dashboard.json: fully German dashboard (InfluxQL,
InfluxDB 1.x) with off-grid state, SoC (SoE*SoH/100) and
consumption-source panels; status tiles use integers plus value
mappings because Grafana drops string/boolean fields from
InfluxQL time series.
dj7nt