Multi-Site Monthly Generation¶

Single-site generators ignore spatial dependencies between gages. KirschGenerator is a nonparametric bootstrap that preserves cross-site correlations using Cholesky decomposition.

Note: Run this notebook from the examples/ directory. Figures are saved to figures/02_multisite/ and data outputs to outputs/02_multisite/ (both are gitignored).

In [1]:

from pathlib import Path

FIG_DIR = Path("figures") / "02_multisite"
OUT_DIR = Path("outputs") / "02_multisite"
SHARED_DIR = Path("outputs") / "shared"
FIG_DIR.mkdir(parents=True, exist_ok=True)
OUT_DIR.mkdir(parents=True, exist_ok=True)
SHARED_DIR.mkdir(parents=True, exist_ok=True)

from pathlib import Path FIG_DIR = Path("figures") / "02_multisite" OUT_DIR = Path("outputs") / "02_multisite" SHARED_DIR = Path("outputs") / "shared" FIG_DIR.mkdir(parents=True, exist_ok=True) OUT_DIR.mkdir(parents=True, exist_ok=True) SHARED_DIR.mkdir(parents=True, exist_ok=True)

Fit and generate¶

In [2]:

import synhydro

Q_daily = synhydro.load_example_data()
Q_monthly = Q_daily.resample("MS").sum()    # all sites, monthly

gen = synhydro.KirschGenerator()
gen.fit(Q_monthly)
ensemble = gen.generate(n_realizations=50, n_years=30, seed=42)

import synhydro Q_daily = synhydro.load_example_data() Q_monthly = Q_daily.resample("MS").sum() # all sites, monthly gen = synhydro.KirschGenerator() gen.fit(Q_monthly) ensemble = gen.generate(n_realizations=50, n_years=30, seed=42)

Verify spatial correlations¶

The key diagnostic for a multi-site generator is whether pairwise correlations match the historical record. Side-by-side heatmaps make this easy to assess:

In [3]:

from synhydro.plotting import plot_spatial_correlation

fig, axes = plot_spatial_correlation(
    ensemble,
    observed=Q_monthly,
    timestep="monthly",
)
fig.savefig(FIG_DIR / "spatial_correlation.png", dpi=150, bbox_inches="tight")

from synhydro.plotting import plot_spatial_correlation fig, axes = plot_spatial_correlation( ensemble, observed=Q_monthly, timestep="monthly", ) fig.savefig(FIG_DIR / "spatial_correlation.png", dpi=150, bbox_inches="tight")

Save for reuse¶

Later tutorials reuse this monthly ensemble. Save it to the shared outputs directory so they can load it instead of regenerating:

In [4]:

ensemble.to_hdf5(SHARED_DIR / "kirsch_monthly_ensemble.hdf5", compression="gzip")

ensemble.to_hdf5(SHARED_DIR / "kirsch_monthly_ensemble.hdf5", compression="gzip")

Next steps¶

• Monthly-to-daily disaggregation - Tutorial 03
• Monthly-to-daily pipeline - Tutorial 08
• Quantitative validation - Tutorial 05
• Algorithm details - Kirsch Bootstrap

---

Previous: Quickstart (Thomas-Fiering) | Next: Disaggregator (Nowak)