SafeLeak-RCD — TsFile format

This repository is a conversion to TsFile format of SafeLeak-RCD: Residential Residual Current Decomposition Benchmark (single-phase electrical-safety monitoring / NILM).

• Original dataset: haayan/safeleak-rcd
• Paper: Physics-Regularized Conditional Flow Matching for Branch-Conditioned Residual Current Decomposition in Electrical Safety Monitoring
• License: CC-BY-NC-4.0 (per the original dataset)

Task

Given: the aggregate residual current (total_residual_current), the aggregate active power (total_power), and the target branch's power hint, predict the selected branch's residual current. Single-phase residential, 12 branches, 1-minute target interval, 7 entity-level panels, entity-level disjoint train/validation/test = 5/1/1.

Files

The conversion preserves the benchmark's three-way split — one CSV per TsFile:

File	rows	devices (segment_id)
train.tsfile	104,835	15 (5 entities × {base, variant_1, variant_2})
validation.tsfile	7,091	1
test.tsfile	11,991	1

validation/test contain only real observations (synthetic_variant=0); train includes the synthetic augmented variants.

TsFile structure

• TAG (device dimension) = segment_id. The base and 2 synthetic variants of the same entity share one time axis; (segment_id, timestamp) is measured to be unique with no duplicates, whereas (entity_id, timestamp) has ~70k duplicates in train, so segment_id is used as the device to guarantee monotonic time within each device.
• Time: parsed from the original timestamp (2024-08-01 00:10:00 text, 1-minute interval) into INT64 milliseconds.
• FIELD (28 columns):
  • total_residual_current, total_power (DOUBLE)
  • branch_1_power / branch_1_current … branch_12_power / branch_12_current (24 columns, DOUBLE)
  • synthetic_variant (INT64, 0=real base, 1/2=synthetic variant)
  • entity_id (STRING, owning entity)

Conversion Notes

• Conversion path: script (scripts/converters/safeleak_rcd.py), written throughout with the TsFile Java tool's schema mode.
• Time precision: ms.
• Dropped column: the original timestamp text column is dropped after parsing into Time (INT64 milliseconds); the time info is fully preserved, only the format changes from string to millisecond integer. No other columns are dropped.
• Only the three CSVs of the default config benchmark_split were converted. The original repo's processed_entities/ per-entity bundles (base/variant/combined) overlap with the benchmark data and were not converted this time.

Reading example

from tsfile import TsFileReader
reader = TsFileReader("train.tsfile")
for name, schema in reader.get_all_table_schemas().items():
    print(name, [c.get_column_name() for c in schema.get_columns()])