WLCR-SEA Predictor
WLCR-SEA (Window-Local Context Representation with Seasonal Expert Attention) forecasts cellular traffic using only the target cell's request history and frozen model assets. Each request contains an ordered 336-hour history and observation mask. The model forecasts the next 24 hours for four indicators:
- uplink active users;
- downlink active users;
- average used downlink PRBs;
- average used uplink PRBs.
The two PRB fields are counts of average used physical resource blocks, not utilization percentages. WLCR-SEA routes between eight seasonal experts with hard availability masking, reliability-aware Entmax routing, and a bounded residual.
Public five-model predictor
The checkpoints/ directory contains the five members used by the public
predictor. Each checkpoint stores schema_version, experiment_version, its
seed, selected configuration and epoch, a frozen (24, 4) training prior, and
the CPU state_dict. SHA256SUMS records their integrity hashes.
The artifact filenames retain the paper's internal experiment identifier so that the release can be mapped exactly to research manifests:
| File | Seed | Selected configuration | Epoch |
|---|---|---|---|
A6_mixed_aug_seed42.pt | 42 | d16_h32_lr1e3_delta025 | 67 |
A6_mixed_aug_seed43.pt | 43 | d32_h64_lr5e4_delta050 | 40 |
A6_mixed_aug_seed44.pt | 44 | d16_h32_lr1e3_delta025 | 65 |
A6_mixed_aug_seed45.pt | 45 | d32_h64_lr5e4_delta050 | 56 |
A6_mixed_aug_seed46.pt | 46 | d32_h64_lr5e4_delta050 | 65 |
The reader-facing name is WLCR-SEA five-model ensemble. Its output is the arithmetic mean of the five member forecasts after inverse transformation to linear traffic space.
Load the same predictor as the Demo
Clone the
implementation repository,
install requirements-demo.txt, and use the verified ensemble loader:
from demo.model_loader import load_ensemble
from demo.runtime import run_forecast
ensemble = load_ensemble()
result = run_forecast("request.csv", ensemble=ensemble)
forecast = result.prediction # shape: (24, 4), linear traffic space
The public runtime pins model revision
eb4447f4ebab,
verifies every checkpoint SHA-256, and checks model metadata before loading.
Request parsing, observation-mask handling, five-member aggregation, and audit
export all use the same code path as the public Demo.
Evaluation snapshot
On the project's reported holdout workflow, the five-model predictor reports
macro-cell WAPE 0.177612, pooled WAPE 0.184915, and macro-indicator WAPE
0.195511. These are different aggregations of one dataset and protocol, and
they are not general deployment guarantees.
Intended use and limitations
This release supports research and reproducibility for request-local cellular
traffic forecasting. It is not a general-purpose pretrained time-series model
and should be validated on the target network before operational use. Training
data is not included, and the source dataset's terms are separate from this
model release. PyTorch .pt files use pickle-based serialization; load only
trusted files and verify their checksums.
Routing weights describe internal allocation. They are not causal effects, feature attributions, or calibrated uncertainty estimates.
License
Apache License 2.0. See LICENSE.