DCN-Bench v1
A labelled thermal-hydraulic District Cooling Network dataset for GNN-based fault diagnosis: 3,016 scenarios on one closed-loop DCN topology, with node/edge/graph/temporal fault labels and full physical ground truth.
Generated by dcngen — a
simulation-based, physics-gated fault-scenario generator that couples an
EPANET/WNTR hydraulic solve to a custom Lagrangian thermal transport layer,
injects labelled faults, and validates every scenario against eight physics
invariants before release.
Download
The release ships as a single archive (~3.2 GB compressed, 3.7 GB unpacked, 21,114 files — one archive is far kinder than 21,114 downloads).
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="asgersong/dcn-bench-v1",
filename="dcn-bench-v1.tar.gz",
repo_type="dataset",
)
tar xzf dcn-bench-v1.tar.gz
plan.json and run_report.json are also uploaded unpacked, so you can read
the manifest and the generation accounting without downloading the archive.
What is in it
| Tier | Horizon | Steps | Normal | Leak | Fouling | Bypass |
|---|---|---|---|---|---|---|
| bulk | 24 h | 144 | 1000 | 800 | 550 | 450 |
| week | 7 d | 1008 | 80 | 50 | 40 | 30 |
| month | 30 d | 4320 | 6 | 5 | 3 | 2 |
| total | 1086 | 855 | 593 | 482 |
Network hanoi_8GB_1Y (DiTEC-WDN Hanoi, mirrored into supply + return):
31 junctions ×2 plus 2 plant headers, 34 pipes ×2, 31 ETS crossovers.
Timestep 600 s. 950 of 1,000 static parameter draws exercised.
Master seed 42 (PCG64, numpy 2.5.1).
2,880 of 3,016 scenarios pass the physics gate (95.5 %), uniformly across classes (95.0–96.5 %). The 136 failures are kept, with their verdicts, so including or excluding them is your choice rather than one made for you.
Per scenario: node_dynamics, edge_dynamics, node_static, edge_static,
plant_dynamics, labels (Parquet) and card.json — the latter carrying the
master seed, full config echo, RNG identity, the plan row, the gate verdict
with per-rule metrics, and the scenario-level fault label.
Three things to know before training
1. Drop the plant leak channels. In a closed loop make_up_flow is
identically zero unless water is escaping, and under a leak it equals the leak
flow exactly and noise-free. Measured across 200 scenarios: clean records peak
at 8.9e-16 m³/s (machine zero), leaks bottom out at 8.9e-02 — a gap of ~13
orders with no overlap. Any threshold inside it detects every leak timestep
with no false positives. Drop make_up_flow and leak_flow, or degrade them
through a sensor model, before reporting any leak result.
2. Split by uid, not at random. Scenarios share static draws, and every
fault class visits all 31 consumer sites, so a random split puts the same fault
location on both sides.
3. severity is not one quantity. It shares a column but means a UA
multiplier for fouling (smaller is worse), a short-circuit fraction for bypass,
and a fraction of network design flow for leaks (larger is worse in both).
Validation
- Physics gate on every scenario — eight rules (mass balance, first-law energy closure, temperature envelope, velocity self-consistency, pressure rails, ETS reverse flow, unmet budget, convergence budget). Worst first-law residual across all 3,016 scenarios: 4.23e-16.
- External cross-check — an independent pandapipes rebuild of the steady transport fields agrees to 0.001 K on node temperatures and 7e-6 on the plant energy balance (implying total pipe heat gain agrees to 0.07 %).
- Reproducibility, measured — regenerating a published scenario from (master seed, uid) reproduces the plan row, loads, statics, fault label, mask and gate verdict bitwise; solved fields drift 2.1e-12 K / 1.8e-12 m / 2.2e-13 m³/s over a 144-step march (WNTR iterates in memory-address order, so bitwise solved fields are unattainable — compare physically, never by hash).
Known limitations
Stated at length in dataset-card.md inside the archive. In brief: one
topology; fully observed (no sensor sparsity, noise or sensor faults); single
fault per scenario; no weather coupling; ground temperature constant within a
scenario; fouling deliberately over-represented at 20 % for class balance, well
above its real-world incidence; and the inherited DiTEC pipe friction is
outside the physical range (Hanoi Hazen–Williams C ≈ 1576–2477 against ≈ 130
for real pipe), which shapes every hydraulic fault signature.
Provenance and licence
Topology and static parameters are inherited from
DiTEC-WDN
(rugds / University of Groningen, CC-BY-4.0, DOI
10.57967/hf/6341); everything else — the
supply/return mirror, the thermal field, the loads, the faults and the labels —
is generated by dcngen.
Access is granted individually while the accompanying MSc thesis is under examination; the open release (with its own licence, honouring the upstream attribution) follows publication. Until then, please do not redistribute.
Generator, tests, ADRs and the full design record: https://github.com/asgersong/dcngen
- Downloads last month
- 9