"""Sun*+ 推論例 (ローカル/オフライン実行用)。 python infer_example.py --station 0658 --start 2026-10-01 --end 2026-10-03 # 観測なしダウンスケーリング (ARCO-ERA5 を自動取得) python infer_example.py --station 0658 --start 2026-10-01 --end 2026-10-03 --nc era5_dir --static static.nc # 取得済み NetCDF を使う 予測対象時刻の前後24時間の余白を自動的に含めて取得します。出力は CSV (時刻 UTC)。""" import argparse, pandas as pd, numpy as np from sunstar_plus import SunstarPlus, era5io ap = argparse.ArgumentParser() ap.add_argument("--station", required=True, help="stations_nationwide.csv の block_no (例 0658=新見)。カンマ区切り可") ap.add_argument("--start", required=True); ap.add_argument("--end", required=True) ap.add_argument("--nc"); ap.add_argument("--static"); ap.add_argument("--out", default="sunstar_plus_pred.csv") ap.add_argument("--known", action="store_true", help="学習に使った地点なら指定 (σ の未知地点拡大を外す)") a = ap.parse_args() st = pd.read_csv("data/stations_nationwide.csv", dtype={"block_no": str}); ids = a.station.split(",") sel = st[st.block_no.isin(ids)].reset_index(drop=True); assert len(sel), "block_no が見つかりません" t0, t1 = pd.Timestamp(a.start) - pd.Timedelta(hours=24), pd.Timestamp(a.end) + pd.Timedelta(hours=24) if a.nc: import glob; times, fields, grid = era5io.load_nc(glob.glob(a.nc + "/20??-??.nc"), a.static); keep = (times >= t0) & (times <= t1) times = times[keep]; fields = {k: v[keep] for k, v in fields.items()} else: times, fields, grid = era5io.fetch_arco(t0, t1) m = SunstarPlus(".") out = m.downscale(times, fields, grid, sel, sigma=True, heldout=not a.known) rows = [] for j, r in sel.iterrows(): d = pd.DataFrame({"time_utc": times, "block_no": r.block_no}) for v, (mean, sg) in out.items(): d[v] = mean[:, j] if sg is not None: d[v + "_sigma"] = sg[:, j] rows.append(d) res = pd.concat(rows); res = res[(res.time_utc >= pd.Timestamp(a.start)) & (res.time_utc <= pd.Timestamp(a.end))] res.to_csv(a.out, index=False); print(res.head(), "->", a.out)