2vzorky pro korekci predikce

db/routines/R__078_fn_pv_forecast_delta_profile.sql

@@ -11,7 +11,7 @@ create or replace function ems.fn_pv_forecast_delta_profile(
   p_data_to timestamptz default now(),
   p_half_life_days numeric default 14,
   p_threshold_w int default 150,
-  p_top_n_days int default null,
+  p_top_n_days int default 2,
   p_non_top_day_factor numeric default 0.02,
   p_day_weight_gamma numeric default 1.0
 )
@@ -227,4 +227,4 @@ as $fn$
 $fn$;
 
 comment on function ems.fn_pv_forecast_delta_profile is
-  'Aditivní delta profil chyby PV forecastu po 15min slotu dne (96 slotů). Zdroj: forecast_accuracy, vážení exp(-age/half_life_days) * day_weight (clear-ish dny) * volitelně top_n_days (jen N nejlepších kalendářních dní podle w_energy*w_smooth, ostatní ztlumené) * power(day_weight, day_weight_gamma). Vrací JSON {deltas:[{slot_of_day, delta_w, sample_count}], ...}. Cutoff dat od 2026-04-12 Europe/Prague.';
+  'Aditivní delta profil chyby PV forecastu po 15min slotu dne (96 slotů). Zdroj: forecast_accuracy, vážení exp(-age/half_life_days) * day_weight (clear-ish dny) * top_n_days (default 3 = jen 3 nejlepší kalendářní dny podle w_energy*w_smooth, ostatní ztlumené non_top_day_factor; explicitní NULL = tier vypnut, váží se všechny dny) * power(day_weight, day_weight_gamma). Vrací JSON {deltas:[{slot_of_day, delta_w, sample_count}], ...}. Cutoff dat od 2026-04-12 Europe/Prague.';

scripts/analysis/pv_delta_profile_diagnostics.sql

@@ -0,0 +1,222 @@
+-- Diagnostika: z kterých kalendářních dní (Europe/Prague) se skládá váha pro delta profil
+-- (stejná logika jako ems.fn_pv_forecast_delta_profile: best → slots → day_stats → day_rank → váhy w).
+--
+-- Uprav params (site_id, okno, half_life, threshold, top_n_days / non_top / gamma) a spusť v psql.
+-- Jedna řádka = jeden kalendářní den v okně; p_top_n_days mění tier u vah (ne počet řádků).
+
+WITH params AS (
+  SELECT
+    2::int AS site_id,
+    (now() - interval '60 days')::timestamptz AS p_data_from,
+    now()::timestamptz AS p_data_to,
+    14::numeric AS half_life_days,
+    150::int AS threshold_w,
+    NULL::int AS p_top_n_days,
+    0.02::numeric AS p_non_top_day_factor,
+    1.0::numeric AS p_day_weight_gamma
+),
+tz AS (
+  SELECT coalesce(nullif(trim(s.timezone), ''), 'Europe/Prague') AS tz_name
+  FROM ems.site s
+  JOIN params p ON s.id = p.site_id
+),
+cutoff AS (
+  SELECT timestamptz '2026-04-11T22:00:00Z' AS min_ts
+),
+bounds AS (
+  SELECT
+    greatest(p.p_data_from, p.p_data_to - interval '120 days', (SELECT min_ts FROM cutoff)) AS ts_from,
+    p.p_data_to AS ts_to,
+    greatest(p.half_life_days, 1) AS half_life_days,
+    greatest(p.threshold_w, 0) AS threshold_w
+  FROM params p
+),
+best AS (
+  SELECT
+    fa.interval_start,
+    fa.pv_array_id,
+    fa.forecast_power_w,
+    fa.actual_power_w,
+    fa.forecast_created_at,
+    row_number() OVER (
+      PARTITION BY fa.interval_start, fa.pv_array_id
+      ORDER BY fa.forecast_created_at DESC
+    ) AS rn
+  FROM ems.forecast_accuracy fa
+  CROSS JOIN bounds b
+  JOIN params p ON fa.site_id = p.site_id
+  WHERE fa.interval_start >= b.ts_from
+    AND fa.interval_start < b.ts_to
+    AND fa.actual_power_w IS NOT NULL
+    AND fa.forecast_created_at <= fa.interval_start
+),
+slots AS (
+  SELECT
+    b.interval_start,
+    sum(b.forecast_power_w)::numeric AS forecast_total_w,
+    sum(b.actual_power_w)::numeric AS actual_total_w,
+    (
+      (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, day_local, tz.tz_name
+),
+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
+),
+day_rank AS (
+  SELECT
+    ds.day_local,
+    row_number() OVER (
+      ORDER BY
+        (coalesce(ds.w_energy, 0.35) * coalesce(ds.w_smooth, 0.35)) DESC,
+        ds.day_local DESC
+    ) AS rn,
+    (coalesce(ds.w_energy, 0.35) * coalesce(ds.w_smooth, 0.35)) AS day_score
+  FROM day_stats ds
+),
+filtered AS (
+  SELECT
+    s.day_local,
+    s.slot_of_day,
+    exp(-s.age_days / nullif((SELECT half_life_days FROM bounds), 0))
+      * (
+        CASE
+          WHEN (SELECT p_top_n_days FROM params) IS NULL THEN 1::numeric
+          WHEN (SELECT p_top_n_days FROM params) < 1 THEN 1::numeric
+          WHEN dr.rn <= (SELECT p_top_n_days FROM params) THEN 1::numeric
+          ELSE greatest(
+            0::numeric,
+            least(1::numeric, coalesce((SELECT p_non_top_day_factor FROM params), 0.02))
+          )
+        END
+      )
+      * (
+        0.05
+        + 0.95
+          * power(
+            greatest(
+              0.0,
+              least(1.0, coalesce(ds.w_energy, 0.35) * coalesce(ds.w_smooth, 0.35))
+            ),
+            greatest(
+              0.25,
+              least(coalesce((SELECT p_day_weight_gamma FROM params), 1.0), 8.0)
+            )
+          )
+      ) AS w
+  FROM slots s
+  CROSS JOIN bounds b
+  LEFT JOIN day_stats ds ON ds.day_local = s.day_local
+  LEFT JOIN day_rank dr ON dr.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)
+),
+by_day AS (
+  SELECT day_local, sum(w) AS w_sum, count(*)::bigint AS slot_rows
+  FROM filtered
+  GROUP BY day_local
+),
+tot AS (
+  SELECT sum(w_sum) AS w_tot FROM by_day
+)
+SELECT
+  dr.rn AS day_rank,
+  ds.day_local,
+  round(ds.energy_kwh::numeric, 2) AS energy_kwh,
+  round(ds.w_energy::numeric, 3) AS w_energy,
+  round(ds.w_smooth::numeric, 3) AS w_smooth,
+  round(dr.day_score::numeric, 4) AS day_score,
+  round(coalesce(bd.w_sum, 0)::numeric, 2) AS sum_w_in_filtered,
+  coalesce(bd.slot_rows, 0::bigint) AS slot_rows,
+  round(
+    (100.0 * coalesce(bd.w_sum, 0) / nullif((SELECT w_tot FROM tot), 0))::numeric,
+    2
+  ) AS pct_of_total_weight
+FROM day_stats ds
+JOIN day_rank dr ON dr.day_local = ds.day_local
+LEFT JOIN by_day bd ON bd.day_local = ds.day_local
+ORDER BY dr.rn;
+
+-- Shrnutí okna + výstup funkce (stejné parametry jako v params výše):
+-- SELECT (ems.fn_pv_forecast_delta_profile(
+--   (SELECT site_id FROM params),
+--   (SELECT p_data_from FROM params),
+--   (SELECT p_data_to FROM params),
+--   (SELECT half_life_days FROM params),
+--   (SELECT threshold_w FROM params),
+--   (SELECT p_top_n_days FROM params),
+--   (SELECT p_non_top_day_factor FROM params),
+--   (SELECT p_day_weight_gamma FROM params)
+-- ))->'data_from' AS data_from,
+--      (ems.fn_pv_forecast_delta_profile(...))->'data_to' AS data_to;