| """Leak-free fuel features for the (zone, target hour, horizon h) rows of the EU model, from the daily cache of fetch_fuel.py. |
| |
| Leakage rule. A forecast of target day T at horizon h is issued on day T - h (h = 2: the day before the day-after-tomorrow's |
| price auction results are known; the last known PRICE day is T - h + 1). The fuel value used is the one of the CLOSE of the |
| day BEFORE issue, f = T - h - 1: a full day more cautious than necessary (the MIBGAS index of the issue day and the 11:00 |
| EUA auction of the issue day are public before the day-ahead auction results at 13:00), so no feature can contain a number |
| that was not settled when the forecast was made. Weekends and holidays carry the last published value forward (the EUA is |
| auctioned on weekdays only). |
| |
| Features (per row; the same value for every zone - the fuel is a EUROPE-wide level, the zone decides what it does with it): |
| gas, gas_d7, gas_d28 gas day-ahead level (EUR/MWh) and its 7-/28-day change |
| gas_m1_spread month-ahead minus day-ahead gas: where the market expects the level to go |
| eua, eua_d7, eua_d28 carbon level (EUR/tCO2) and its changes |
| ccgt_mc, ccgt_mc_d7, _d28 marginal cost of a combined-cycle gas plant, gas / 0.55 + 0.202 / 0.55 * eua (EUR/MWh_e): |
| the number a gas-priced hour clears near, with the two inputs already combined |
| full, full_d28 storage fill (per cent) and its 28-day change, only when the cache has the AGSI+ column |
| fuel_age days between f and the newest value actually published at or before f (a stale feed is visible) |
| """ |
| import numpy as np |
| import pandas as pd |
|
|
| CCGT_EFF = 0.55 |
| EF_T_PER_MWH_TH = 0.202 |
| BASE = ["gas", "gas_d7", "gas_d28", "gas_m1_spread", "eua", "eua_d7", "eua_d28", "ccgt_mc", "ccgt_mc_d7", "ccgt_mc_d28", "fuel_age"] |
| STORAGE = ["full", "full_d28"] |
|
|
|
|
| def calendar_frame(fuel): |
| """The cache on a gapless daily index, filled forward, plus `age` = days since the value was actually published.""" |
| idx = pd.date_range(fuel.index.min().normalize(), fuel.index.max().normalize(), freq="D") |
| f = fuel.reindex(idx) |
| published = pd.Series(np.where(f["gas_da"].notna(), idx.to_numpy(), np.datetime64("NaT")), index=idx) |
| age = (pd.Series(idx, index=idx) - published.ffill()).dt.days |
| out = f.ffill() |
| out["age"] = age |
| return out |
|
|
|
|
| def daily_table(fuel): |
| """Per fuel day f: the level and change columns, as a frame indexed by f.""" |
| c = calendar_frame(fuel) |
| t = pd.DataFrame(index=c.index) |
| t["gas"] = c["gas_da"] |
| t["gas_d7"] = c["gas_da"] - c["gas_da"].shift(7) |
| t["gas_d28"] = c["gas_da"] - c["gas_da"].shift(28) |
| t["gas_m1_spread"] = c["gas_m1"] - c["gas_da"] |
| t["eua"] = c["eua"] |
| t["eua_d7"] = c["eua"] - c["eua"].shift(7) |
| t["eua_d28"] = c["eua"] - c["eua"].shift(28) |
| mc = c["gas_da"] / CCGT_EFF + EF_T_PER_MWH_TH / CCGT_EFF * c["eua"] |
| t["ccgt_mc"] = mc |
| t["ccgt_mc_d7"] = mc - mc.shift(7) |
| t["ccgt_mc_d28"] = mc - mc.shift(28) |
| if "full" in c and c["full"].notna().any(): |
| t["full"] = c["full"] |
| t["full_d28"] = c["full"] - c["full"].shift(28) |
| t["fuel_age"] = c["age"] |
| return t |
|
|
|
|
| def for_rows(table, target_dates, h): |
| """Fuel features for target DAYS (any date-like array) at horizon h: the table row of f = T - h - 1. |
| Days before the table starts (or after it ends: the newest value is carried) give NaN / the last row.""" |
| f = pd.DatetimeIndex(target_dates).normalize() - pd.Timedelta(days=int(h) + 1) |
| rows = table.reindex(f) |
| last = table.index.max() |
| late = f > last |
| if late.any(): |
| rows = rows.copy() |
| rows.loc[late, :] = table.iloc[-1].to_numpy() |
| rows.loc[late, "fuel_age"] = table.iloc[-1]["fuel_age"] + (f[late] - last).days |
| return rows.reset_index(drop=True) |
|
|
|
|
| def add_to_frame(df, table, date_col="date", h_col="h"): |
| """Append the fuel columns to a rows frame that has a date and an h column, one call per distinct h.""" |
| parts = [] |
| for h, g in df.groupby(h_col, sort=False): |
| feats = for_rows(table, g[date_col].to_numpy(), h) |
| feats.index = g.index |
| parts.append(feats) |
| out = pd.concat(parts).reindex(df.index) |
| return pd.concat([df, out], axis=1) |
|
|
