lips-poc/powergrid-benchmark1-tf_leapnet

tf_leapnet_DEFAULT — TensorFlow LeapNet Model

Benchmark1 surrogate model for power grid current flow prediction.
Trained on the l2rpn_case14_sandbox environment to predict line currents (a_or, a_ex) from grid injections and topology.

Architecture: LeapNet (Latent Encoding of Atypical Perturbations) — encoder + main network + decoder, specialised for handling topology changes via a dedicated tau encoding path
Framework: TensorFlow / Keras
Key difference from tf_fc: LeapNet explicitly separates the topology input (attr_tau) from the physical inputs (attr_x), making it better at generalising to unseen topologies.

Folder contents

tf_leapnet/
├── weights.h5                ← trained Keras weights (legacy naming)
├── config.json               ← hyperparameters
├── metadata.json             ← input/output sizes
├── scaler_params.json        ← normalisation statistics
├── losses.json               ← training history
├── tf_leapnet.ini            ← simulator configuration
└── l2rpn_case14_sandbox.ini  ← benchmark configuration

The weight file is named weights.h5 (without the model. prefix) because this model was saved with an older version of Keras. The restore code handles this automatically.

---

Installation

pip install leap-net==0.0.5

leap-net is a separate dependency required only for LeapNet. Without it, the import will fail silently with an ImportError.

---

Usage

import pathlib
import numpy as np
from lips.augmented_simulators.tensorflow_models.powergrid import LeapNet
from lips.dataset.scaler import StandardScaler
from lips.dataset.powergridDataSet import PowerGridDataSet
MODEL_DIR = pathlib.Path("path/to/tf_leapnet_DEFAULT")
# Instantiate and restore
# name="tf_leapnet" + sim_config_name="DEFAULT" → model.name = "tf_leapnet_DEFAULT"
# restore() looks for MODEL_DIR.parent / "tf_leapnet_DEFAULT"
model = LeapNet(
    name="tf_leapnet",
    sim_config_path=MODEL_DIR / "tf_leapnet.ini",
    sim_config_name="DEFAULT",
    bench_config_path=MODEL_DIR / "l2rpn_case14_sandbox.ini",
    bench_config_name="Benchmark1",
    scaler=StandardScaler,
)
model.restore(MODEL_DIR.parent)
# Prepare input data
N = 100  # number of samples
dataset = PowerGridDataSet(
    name="my_data",
    config_path=MODEL_DIR / "l2rpn_case14_sandbox.ini",
    config_name="Benchmark1",
)
dataset.data = {
    "prod_p":      np.random.rand(N, 6).astype(np.float32),   # MW
    "prod_v":      np.random.rand(N, 6).astype(np.float32),   # per-unit voltage
    "load_p":      np.random.rand(N, 11).astype(np.float32),  # MW
    "load_q":      np.random.rand(N, 11).astype(np.float32),  # MVAr
    "line_status": np.ones((N, 20), dtype=bool),               # True = connected
    "topo_vect":   np.ones((N, 57), dtype=np.int32),           # bus assignments (1 or 2)
    # Dummy targets — required by the data pipeline even at inference time
    "a_or":        np.zeros((N, 20), dtype=np.float32),
    "a_ex":        np.zeros((N, 20), dtype=np.float32),
}
# Run inference
predictions = model.predict(dataset)
print(predictions["a_or"].shape)  # (N, 20) — current at origin end of each line (A)
print(predictions["a_ex"].shape)  # (N, 20) — current at extremity end of each line (A)

---

Input / Output

Inputs — 111 features total

Feature	Shape	Unit	Description
prod_p	(N, 6)	MW	Active power output per generator
prod_v	(N, 6)	p.u.	Voltage setpoint per generator
load_p	(N, 11)	MW	Active power demand per load
load_q	(N, 11)	MVAr	Reactive power demand per load
line_status	(N, 20)	bool	Line connection status (1=on, 0=off)
topo_vect	(N, 57)	int	Bus assignment per element (1 or 2)

Internally, LeapNet routes line_status and topo_vect through a separate tau encoding path. This is handled automatically — your input dict format is the same as tf_fc.

Outputs — 40 values total

Feature	Shape	Unit	Description
a_or	(N, 20)	A	Current at the origin end of each line
a_ex	(N, 20)	A	Current at the extremity end of each line