tune korekce z clear sky

db/routines/R__078_fn_pv_forecast_delta_profile.sql

@@ -64,19 +64,101 @@ as $fn$
         (extract(hour from (b.interval_start at time zone tz.tz_name))::int * 60)
         + extract(minute from (b.interval_start at time zone tz.tz_name))::int
       ) / 15 as slot_of_day,
+      (b.interval_start at time zone tz.tz_name)::date as day_local,
       extract(epoch from (now() - b.interval_start)) / 86400.0 as age_days
     from best b
     cross join tz
     where b.rn = 1
-    group by b.interval_start, slot_of_day, tz.tz_name
+    group by b.interval_start, slot_of_day, day_local, tz.tz_name
+  ),
+  -- Denní „clear-ish“ skóre: preferujeme dny s hladkou křivkou výroby a vysokou denní energií
+  -- relativně k ostatním dnům v okně (mraky dělají vysokofrekvenční šum na 15min, který není dobrý anchor pro slot bias).
+  day_energy as (
+    select
+      s.day_local,
+      sum(s.actual_total_w)::numeric / 4000.0 as energy_kwh
+    from slots s
+    group by s.day_local
+  ),
+  ref as (
+    select percentile_cont(0.5) within group (order by de.energy_kwh) as med_kwh
+    from day_energy de
+  ),
+  slot_steps as (
+    select
+      s.*,
+      lag(s.actual_total_w) over (partition by s.day_local order by s.interval_start) as prev_actual_w
+    from slots s
+    where s.slot_of_day between 20 and 80
+      and s.actual_total_w > (select threshold_w from bounds)
+  ),
+  day_jump as (
+    select
+      ss.day_local,
+      percentile_cont(0.5) within group (order by abs(ss.actual_total_w - ss.prev_actual_w)) as med_jump_w
+    from slot_steps ss
+    where ss.prev_actual_w is not null
+    group by ss.day_local
+  ),
+  day_med as (
+    select
+      s.day_local,
+      percentile_cont(0.5) within group (order by s.actual_total_w) as p50_actual_w
+    from slots s
+    where s.actual_total_w > (select threshold_w from bounds)
+    group by s.day_local
+  ),
+  day_stats as (
+    select
+      de.day_local,
+      de.energy_kwh,
+      dj.med_jump_w,
+      dm.p50_actual_w,
+      case
+        when (select med_kwh from ref) is null or (select med_kwh from ref) <= 0 then 0.5
+        else greatest(
+          0.0,
+          least(
+            1.0,
+            (de.energy_kwh - (select med_kwh from ref) * 0.55)
+            / nullif((select med_kwh from ref) * 0.35, 0)
+          )
+        )
+      end as w_energy,
+      case
+        when dj.med_jump_w is null or dm.p50_actual_w is null then 0.35
+        else greatest(
+          0.0,
+          least(
+            1.0,
+            1.0
+            - (
+              dj.med_jump_w
+              / nullif(greatest(300.0, dm.p50_actual_w * 0.25), 0)
+            )
+          )
+        )
+      end as w_smooth
+    from day_energy de
+    left join day_jump dj on dj.day_local = de.day_local
+    left join day_med dm on dm.day_local = de.day_local
   ),
   filtered as (
     select
       s.slot_of_day,
       (s.forecast_total_w - s.actual_total_w) as error_w,
-      exp(-s.age_days / nullif((select half_life_days from bounds), 0)) as w
+      exp(-s.age_days / nullif((select half_life_days from bounds), 0))
+        * (
+          0.05
+          + 0.95
+            * greatest(
+              0.0,
+              least(1.0, coalesce(ds.w_energy, 0.35) * coalesce(ds.w_smooth, 0.35))
+            )
+        ) as w
     from slots s
     cross join bounds b
+    left join day_stats ds on ds.day_local = s.day_local
     where s.slot_of_day between 0 and 95
       and (s.actual_total_w > b.threshold_w or s.forecast_total_w > b.threshold_w)
   ),
@@ -119,4 +201,4 @@ as $fn$
 $fn$;
 
 comment on function ems.fn_pv_forecast_delta_profile(int, timestamptz, timestamptz, numeric, int) is
-  'Aditivní delta profil chyby PV forecastu po 15min slotu dne (96 slotů). Zdroj: forecast_accuracy, vážení exp(-age/half_life_days). Vrací JSON {deltas:[{slot_of_day, delta_w, sample_count}], ...}. Interní minimální cutoff dat (2026-04-06 Europe/Prague) brání učení z nekonzistentní historie před kompletním plněním actual.';
+  'Aditivní delta profil chyby PV forecastu po 15min slotu dne (96 slotů). Zdroj: forecast_accuracy, vážení exp(-age/half_life_days) * day_weight (preferuje „clear-ish“ dny: vyšší denní energie vs median v okně + nižší median skoků výkonu mezi 15min v daylight bandu). Vrací JSON {deltas:[{slot_of_day, delta_w, sample_count}], ...}. Interní minimální cutoff dat (2026-04-11 Europe/Prague) brání učení z nekonzistentní historie před kompletním plněním actual.';