"""Sun*+ ライブラリ: 全国版 特徴量生成 (Sun* V2 の特徴量定義を完全に含み, 全国向け特徴を追加)。 出力の特徴列は先頭が V2 と同名同順 (V2 モデルを同じ特徴行列で評価可能)、その後ろに Sun*+ 追加特徴が続く。""" from pathlib import Path import numpy as np, pandas as pd, xarray as xr from scipy.ndimage import uniform_filter HOLD_V2 = ["0658", "0664", "0670"] REGIME_B_START = "2024-10-01" VARS6 = ["temp_c", "wind_u_ms", "wind_v_ms", "humidity_pct", "wind_speed_ms", "precip_mm"] TERRAIN_VARS = ["elevation_m", "slope_deg", "aspect_sin", "aspect_cos", "tpi_500m", "tpi_2km", "tpi_10km", "relief_1km", "relief_5km", "coast_dist_km", "land_mask", "lc_tree_frac", "lc_crop_frac", "lc_built_frac", "lc_water_frac", "lc_grass_frac", "lc_shrub_frac"] ERA_VARS = ["2m_temperature_c", "2m_dewpoint_temperature_c", "10m_u_component_of_wind", "10m_v_component_of_wind", "100m_u_component_of_wind", "100m_v_component_of_wind", "surface_pressure_hpa", "total_precipitation", "surface_solar_radiation_downwards", "total_cloud_cover", "low_cloud_cover", "medium_cloud_cover", "high_cloud_cover", "total_column_water_vapour", "convective_available_potential_energy", "mean_sea_level_pressure_hpa", "skin_temperature_c", "instantaneous_10m_wind_gust", "total_sky_direct_solar_radiation_at_surface"] REGION = {**{p: 0 for p in range(11, 25)}, **{p: 1 for p in range(31, 37)}, **{p: 2 for p in range(40, 47)}, 48: 3, 49: 3, 54: 3, 55: 4, 56: 4, 57: 4, 50: 5, 51: 5, 52: 5, 53: 5, **{p: 6 for p in range(60, 66)}, **{p: 7 for p in (66, 67, 68, 69, 81)}, **{p: 8 for p in range(71, 75)}, **{p: 9 for p in range(82, 89)}, 91: 10} def rh_from(t_c, td_c): a, b = 17.625, 243.04 return np.clip(100 * np.exp(a * td_c / (b + td_c)) / np.exp(a * t_c / (b + t_c)), 0, 100) # ---------------------------------------------------------------- stations def load_stations(csv_path): """data/stations_nationwide.csv (地点メタ + 地形特徴) を読む""" st = pd.read_csv(csv_path, dtype={"block_no": str}) st["region"] = st["region"].astype(int) return st # ---------------------------------------------------------------- ERA5 (Japan grid) class Grid: """ERA5 0.25° 格子の静的場。Grid(static_nc) もしくは Grid.from_arrays(lat, lon, orog_m, lsm, sdor, hv, lv)""" def __init__(self, static_nc=None): if static_nc is None: return st = xr.open_dataset(static_nc).load() self.lat_c = st.latitude.values.astype(np.float64); self.lon_c = st.longitude.values.astype(np.float64) self.orog = (st["geopotential_at_surface"].values / 9.80665).astype(np.float32) self.lsm = st["land_sea_mask"].values.astype(np.float32); self.sdor = st["standard_deviation_of_orography"].values.astype(np.float32) self.hv = st["high_vegetation_cover"].values.astype(np.float32); self.lv = st["low_vegetation_cover"].values.astype(np.float32) @classmethod def from_arrays(cls, lat, lon, orog_m, lsm, sdor, hv, lv): g = cls(); g.lat_c = np.asarray(lat, np.float64); g.lon_c = np.asarray(lon, np.float64) g.orog, g.lsm, g.sdor, g.hv, g.lv = [np.asarray(x, np.float32) for x in (orog_m, lsm, sdor, hv, lv)] return g def points(self, lat, lon, elev): lat = np.asarray(lat, np.float64); lon = np.asarray(lon, np.float64) fy = (self.lat_c[0] - lat) / 0.25; fx = (lon - self.lon_c[0]) / 0.25 iy0 = np.clip(np.floor(fy).astype(int), 1, len(self.lat_c) - 3); ix0 = np.clip(np.floor(fx).astype(int), 1, len(self.lon_c) - 3) dy = np.clip(fy - iy0, 0, 1); dx = np.clip(fx - ix0, 0, 1) return dict(iy0=iy0, ix0=ix0, dy=dy, dx=dx, lat=lat, lon=lon, elev=np.asarray(elev, np.float64)) # ---------------------------------------------------------------- feature helpers (same as V2) def bil(f, p): iy, ix, dy, dx = p["iy0"], p["ix0"], p["dy"], p["dx"] g = lambda y, x: f[..., y, x] return (g(iy, ix) * (1 - dy) * (1 - dx) + g(iy, ix + 1) * (1 - dy) * dx + g(iy + 1, ix) * dy * (1 - dx) + g(iy + 1, ix + 1) * dy * dx) def shift(a, k): out = np.full_like(a, np.nan) if k > 0: out[:-k] = a[k:] elif k < 0: out[-k:] = a[:k] else: out[:] = a return out def roll_mean(a, lo, hi): acc = np.zeros_like(a, dtype=np.float32); n = np.zeros_like(a, dtype=np.float32) for k in range(lo, hi + 1): s = shift(a, k); m = np.isfinite(s); acc += np.where(m, s, 0); n += m return acc / np.maximum(n, 1) def solar_elev(times, lat, lon): doy = times.dayofyear.values.astype(np.float64)[:, None]; hr = (times.hour.values + 0.0)[:, None] g = 2 * np.pi / 365.0 * (doy - 1 + (hr - 12) / 24) decl = 0.006918 - 0.399912 * np.cos(g) + 0.070257 * np.sin(g) - 0.006758 * np.cos(2 * g) + 0.000907 * np.sin(2 * g) eqt = 229.18 * (0.000075 + 0.001868 * np.cos(g) - 0.032077 * np.sin(g) - 0.014615 * np.cos(2 * g) - 0.040849 * np.sin(2 * g)) tst = hr * 60 + eqt + 4 * lon[None]; ha = np.deg2rad(tst / 4 - 180) la = np.deg2rad(lat[None]) sin_e = np.sin(la) * np.sin(decl) + np.cos(la) * np.cos(decl) * np.cos(ha) return np.rad2deg(np.arcsin(np.clip(sin_e, -1, 1))) class StaticSet: """per-station static arrays derived from the station table + ERA5 static grid""" def __init__(self, st, grid): self.st = st; self.grid = grid self.p = grid.points(st.lat.values, st.lon.values, st.elevation_m.values) self.terr = {v: st["t_" + v].values.astype(np.float32) for v in TERRAIN_VARS} self.r10 = st.t_elev_mean_r10.values.astype(np.float32); self.r40 = st.t_elev_mean_r40.values.astype(np.float32) def build_features(times, E, ss, sl=None): """times (T,), E (T,H,W) dict covering times, ss StaticSet. returns F (T,S,nf), names, base dict (T,S). Contexts use +-24 h shifts: caller must pad by >=24 h and crop.""" p = ss.p; g = ss.grid; S = len(p["lat"]); T = len(times) names, cols = [], [] def add(n, a): names.append(n); cols.append(np.asarray(a, np.float32).reshape(T, S) if np.ndim(a) == 2 else np.broadcast_to(np.asarray(a, np.float32)[None], (T, S))) I = {k: bil(v, p) for k, v in E.items()} t2, d2, u10, v10 = I["2m_temperature_c"], I["2m_dewpoint_temperature_c"], I["10m_u_component_of_wind"], I["10m_v_component_of_wind"] u100, v100 = I["100m_u_component_of_wind"], I["100m_v_component_of_wind"] tp = np.maximum(I["total_precipitation"], 0); ssrd = I["surface_solar_radiation_downwards"] rh = rh_from(t2, d2); spd10 = np.hypot(u10, v10); spd100 = np.hypot(u100, v100) elev = p["elev"]; orog_i = bil(g.orog, p); dz = (elev - orog_i).astype(np.float32) base = dict(t2m=t2, u10=u10, v10=v10, rh=rh, tp=tp, spd10=spd10, d2m=d2) for n, a in [("t2m", t2), ("d2m", d2), ("u10", u10), ("v10", v10), ("u100", u100), ("v100", v100), ("spd10", spd10), ("spd100", spd100), ("rh", rh), ("dd", t2 - d2), ("sp", I["surface_pressure_hpa"]), ("msl", I["mean_sea_level_pressure_hpa"]), ("tp", tp), ("ssrd", ssrd / 3600.0), ("fdir", I["total_sky_direct_solar_radiation_at_surface"] / 3600.0), ("tcc", I["total_cloud_cover"]), ("lcc", I["low_cloud_cover"]), ("mcc", I["medium_cloud_cover"]), ("hcc", I["high_cloud_cover"]), ("tcwv", I["total_column_water_vapour"]), ("cape", np.sqrt(np.maximum(I["convective_available_potential_energy"], 0))), ("skt", I["skin_temperature_c"]), ("skt_t2", I["skin_temperature_c"] - t2), ("gust", I["instantaneous_10m_wind_gust"]), ("shear", spd100 - spd10), ("t2m_lapse", t2 - 0.0065 * dz[None])]: add(n, a) add("wdir_s", u10 / np.maximum(spd10, 1e-3)); add("wdir_c", v10 / np.maximum(spd10, 1e-3)) for n, a in [("t2m", t2), ("skt", I["skin_temperature_c"]), ("u10", u10), ("v10", v10), ("spd10", spd10), ("tcc", I["total_cloud_cover"]), ("rh", rh), ("ssrd", ssrd / 3600.0), ("tp", tp)]: ks = (-3, -1, 1, 3) if n not in ("t2m", "rh") else (-6, -3, -1, 1, 3, 6) for k in ks: s = shift(a, k); add(f"{n}_d{k:+d}", s - a if n not in ("tp", "ssrd", "tcc") else s) add("t2m_mean24", roll_mean(t2, -12, 12) - t2); add("t2m_prev24", roll_mean(t2, -24, -1) - t2) tpa = tp add("tp_sum_p3", sum(shift(tpa, -k) for k in (1, 2, 3))); add("tp_sum_f3", sum(shift(tpa, k) for k in (1, 2, 3))); add("tp_sum_p12", roll_mean(tpa, -12, -1) * 12) add("tcc_mean6", roll_mean(I["total_cloud_cover"], -3, 3)); add("ssrd_mean6", roll_mean(ssrd / 3600.0, -3, 3)) tm = np.stack([shift(t2, k) for k in range(-12, 13)], 0); add("t2m_max24", np.nanmax(tm, 0) - t2); add("t2m_min24", np.nanmin(tm, 0) - t2); del tm iy, ix = p["iy0"], p["ix0"] def nb(f, fn): patch = np.stack([f[:, iy + a, ix + b] for a in (0, 1) for b in (0, 1)], 0); return fn(patch) tpf = np.maximum(E["total_precipitation"], 0) add("tp_nbmax", nb(tpf, lambda a: a.max(0))); add("tp_nbmin", nb(tpf, lambda a: a.min(0))) wide = lambda f, fn: fn(np.stack([f[:, np.clip(iy + a, 0, f.shape[1] - 1), np.clip(ix + b, 0, f.shape[2] - 1)] for a in (-1, 0, 1, 2) for b in (-1, 0, 1, 2)], 0)) add("tp_wmean", wide(tpf, lambda a: a.mean(0))); add("tp_wmax", wide(tpf, lambda a: a.max(0))) add("cape_wmax", wide(np.sqrt(np.maximum(E["convective_available_potential_energy"], 0)), lambda a: a.max(0))) t2f = E["2m_temperature_c"]; add("t2m_nbspread", nb(t2f, lambda a: a.max(0) - a.min(0))) add("t2m_gy", bil(np.gradient(t2f, axis=1), p)); add("t2m_gx", bil(np.gradient(t2f, axis=2), p)) add("tcc_wmean", wide(E["total_cloud_cover"], lambda a: a.mean(0))) add("elev", elev); add("dz", dz); add("orog_era", orog_i); add("lsm", bil(g.lsm, p)); add("sdor", bil(g.sdor, p)); add("hv", bil(g.hv, p)); add("lv", bil(g.lv, p)) for v in TERRAIN_VARS: add("terr_" + v, ss.terr[v]) add("elev_mean_r10", ss.r10); add("elev_mean_r40", ss.r40) hh = times.hour.values[:, None] * np.ones((1, S)); doy = times.dayofyear.values[:, None] * np.ones((1, S)) add("hr_s", np.sin(2 * np.pi * hh / 24)); add("hr_c", np.cos(2 * np.pi * hh / 24)); add("doy_s", np.sin(2 * np.pi * doy / 365.25)); add("doy_c", np.cos(2 * np.pi * doy / 365.25)) add("hr_utc", hh) sel = solar_elev(times, p["lat"], p["lon"]); add("sun_elev", sel) asp_s = ss.terr["aspect_sin"][None]; asp_c = ss.terr["aspect_cos"][None]; sl_ = ss.terr["slope_deg"][None] add("down_slope_flow", (u10 * asp_s + v10 * asp_c) * sl_) add("flow_x_tpi2k", spd10 * ss.terr["tpi_2km"][None]); add("flow_x_tpi10k", spd10 * ss.terr["tpi_10km"][None]) add("night_x_tpi2k", np.maximum(-sel, 0) * ss.terr["tpi_2km"][None]); add("night_x_dz", np.maximum(-sel, 0) / 90 * dz[None]) n_v2 = len(names) # ------------- Sun*+ 追加特徴 (全国版) ------------- st = ss.st for n, col in [("lat", "lat"), ("lon", "lon"), ("region", "region"), ("t_elev_max_r10", "t_elev_max_r10"), ("t_elev_min_r40", "t_elev_min_r40"), ("t_sea_frac_20km", "t_sea_frac_20km"), ("t_sea_frac_50km", "t_sea_frac_50km"), ("t_sea_frac_150km", "t_sea_frac_150km")]: add("x_" + n, st[col].values) add("x_dem_minus_stn", (st.t_dem_pt_m.values - st.elevation_m.values)) # 地形 x 気象 の追加相互作用 (沿岸・斜面) cd = ss.terr["coast_dist_km"][None] add("x_coast_x_sun", np.clip(cd, -20, 20) * np.maximum(sel, 0) / 90) # 日射と海岸距離 (海風) add("x_coast_x_spd", np.exp(-np.abs(cd) / 10) * spd10) # 海岸付近の風 add("x_t2m_dem", t2 - 0.0065 * (st.t_elevation_m.values - orog_i)[None]) # DEM標高での気温補正 add("x_valley_night", np.maximum(-sel, 0) / 90 * ss.terr["tpi_10km"][None]) add("x_cold_pool", np.maximum(-sel, 0) / 90 * (st.t_elev_min_r40.values[None]) * (1.0 / (1 + spd10))) add("x_abs_lat_x_doy", np.abs(st.lat.values[None] - 36) * np.cos(2 * np.pi * (doy - 15) / 365.25)) F = np.stack(cols, 2).astype(np.float32) return F, names, base, n_v2