| """Smoke-validate data.py, graph.py, distributions.py against the real data.""" |
| import sys |
| sys.path.insert(0, "/work") |
|
|
| import numpy as np |
| import torch |
|
|
| from heatingforecast import data, graph |
| from heatingforecast.distributions import JohnsonSU, QUANTILE_LEVELS |
|
|
| print("=== data ===") |
| wide, n_q3 = data.load_wide("/work/epronet_full_20260427.csv") |
| print(f"wide: {wide.shape}, dropped {n_q3} quality!=1 rows") |
| print(f"range: {wide.index.min()} .. {wide.index.max()}") |
| print(f"NaNs before bridging: {int(wide.isna().sum().sum())}") |
| wide, obs_mask = data.bridge_gaps(wide) |
| print(f"NaNs after bridging: {int(wide.isna().sum().sum())}, " |
| f"obs_mask mean {obs_mask.mean().mean():.4f}") |
| tr, va, te = data.splits(wide) |
| for name, m in [("train", tr), ("val", va), ("test", te)]: |
| share = wide[m].isna().mean() |
| print(f"{name}: NaN share per consumer: max {share.max():.2f} " |
| f"({share.idxmax()}), median {share.median():.4f}") |
| feats = data.exog_features("/work/weather/weather_memmingen_hourly.csv", |
| "/work/weather/holidays_bavaria.csv", wide.index) |
| print(f"exog: {feats.shape}, NaN {int(feats.isna().sum().sum())}") |
| tr, va, te = data.splits(wide) |
| print(f"split slots: train {tr.sum()}, val {va.sum()}, test {te.sum()}") |
| print(f"per-consumer kWh median: min {wide.median().min():.3f}, " |
| f"max {wide.median().max():.2f}") |
|
|
| print("\n=== graph ===") |
| g = graph.load_graph("/work/Memmingen_pressure.yaml", |
| "/work/graph/consumer_mapping.csv") |
| print(f"nodes {g['n_nodes']}, edges {g['edge_index'].shape[1]}, " |
| f"edge_attr {g['edge_attr'].shape}") |
| print("consumers ordered:", g["consumers"][:6], "...") |
| assert g["n_nodes"] == 42 and g["edge_attr"].shape[0] == g["edge_index"].shape[1] |
| assert g["consumers"] == list(wide.columns), "graph consumers != wide columns" |
|
|
| print("\n=== JohnsonSU ===") |
| torch.manual_seed(0) |
| loc = torch.randn(5) |
| d = JohnsonSU(loc=loc, scale=torch.rand(5) + 0.5, skew=torch.randn(5) * 0.5, |
| tail=torch.rand(5) + 0.3) |
| q = d.quantile(torch.tensor(QUANTILE_LEVELS, dtype=torch.float32)) |
| assert torch.all(torch.diff(q, dim=0) > 0), "quantiles not monotone" |
| med = d.quantile(torch.tensor([0.5])) |
| assert torch.allclose(med[0], loc - d.scale * torch.sinh(d.skew / d.tail), |
| atol=1e-5), "median formula mismatch" |
| s = d.rsample(torch.Size([200_000])) |
| assert torch.isfinite(d.log_prob(s)).all() |
| emp = s.mean(dim=0) |
| print("mean: closed form vs empirical:", |
| [f"{a:.3f}/{b:.3f}" for a, b in zip(d.mean.tolist(), emp.tolist())]) |
| print("OK: monotone quantiles, median formula, finite log-prob, mean matches") |
|
|
