File size: 6,808 Bytes
93ed35a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
"""GNN surrogate for fast voltage-feasibility checks on generated scenarios.

The GraphSAGE surrogate (3 layers, 128 hidden) is trained to map per-bus
[wind, solar, load] contributions to per-bus voltage magnitudes on the
IEEE 118-bus topology. We reuse it here to score generated scenarios in
milliseconds — orders of magnitude faster than running pandapower AC PF
inline in the demo.

Headline metric reference: PC-DDPM was evaluated against pandapower AC PF
under both operational [0.85, 1.10] pu and strict ANSI [0.89, 1.05] pu
voltage bounds (`eval_all_models_bounds.py` in pc-ddpm-epec2026). The
surrogate-based check here is a fast proxy of those numbers, not the
authoritative power-flow evaluation.
"""

from __future__ import annotations

from dataclasses import dataclass

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F  # noqa: N812 — PyTorch convention
from huggingface_hub import hf_hub_download

N_BUS = 118
HF_REPO_ID = "jbobym/pc-ddpm-alberta"
DEFAULT_GNN_FILE = "gnn_surrogate.pt"
DEFAULT_GNN_NORM_FILE = "gnn_surrogate_norm.npz"

V_BOUNDS_OPERATIONAL: tuple[float, float] = (0.85, 1.10)
V_BOUNDS_ANSI: tuple[float, float] = (0.89, 1.05)


class SAGELayer(nn.Module):
    def __init__(self, in_dim: int, out_dim: int) -> None:
        super().__init__()
        self.W_self = nn.Linear(in_dim, out_dim, bias=False)
        self.W_neigh = nn.Linear(in_dim, out_dim, bias=True)

    def forward(self, h: torch.Tensor, A_norm: torch.Tensor) -> torch.Tensor:
        neigh = torch.matmul(A_norm, h)
        return F.relu(self.W_self(h) + self.W_neigh(neigh))


class GraphSAGE(nn.Module):
    """Fixed-topology GraphSAGE surrogate. Mirrors EPEC `train_gnn_surrogate.py`."""

    def __init__(
        self,
        in_dim: int = 3,
        hidden_dim: int = 128,
        out_dim: int = 2,
        n_layers: int = 3,
    ) -> None:
        super().__init__()
        dims = [in_dim] + [hidden_dim] * n_layers
        self.sage_layers = nn.ModuleList(
            [SAGELayer(dims[i], dims[i + 1]) for i in range(n_layers)]
        )
        self.out = nn.Linear(hidden_dim, out_dim)

    def forward(self, x: torch.Tensor, A_norm: torch.Tensor) -> torch.Tensor:
        h = x
        for layer in self.sage_layers:
            h = layer(h, A_norm)
        return self.out(h)  # type: ignore[no-any-return]


@dataclass
class SurrogateBundle:
    model: GraphSAGE
    A_norm: torch.Tensor
    X_mean: np.ndarray
    X_std: np.ndarray
    V_mean: np.ndarray
    V_std: np.ndarray
    wind_vec: np.ndarray
    solar_vec: np.ndarray
    load_frac: np.ndarray
    device: torch.device


def load_gnn_surrogate(
    repo_id: str = HF_REPO_ID,
    model_filename: str = DEFAULT_GNN_FILE,
    norm_filename: str = DEFAULT_GNN_NORM_FILE,
    device: str | torch.device | None = None,
) -> SurrogateBundle:
    """Pull weights + norm stats from HF Hub and assemble the bundle."""
    if device is None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    elif isinstance(device, str):
        device = torch.device(device)

    norm = np.load(hf_hub_download(repo_id=repo_id, filename=norm_filename))
    ckpt = torch.load(
        hf_hub_download(repo_id=repo_id, filename=model_filename),
        map_location=device,
        weights_only=False,
    )

    model = GraphSAGE(
        in_dim=3,
        hidden_dim=int(ckpt["hidden_dim"]),
        out_dim=2,
        n_layers=int(ckpt["n_layers"]),
    ).to(device)
    model.load_state_dict(ckpt["model_state"])
    model.eval()

    A_norm = torch.as_tensor(norm["A_norm"], dtype=torch.float32, device=device)
    return SurrogateBundle(
        model=model,
        A_norm=A_norm,
        X_mean=norm["X_mean"].astype(np.float32),
        X_std=norm["X_std"].astype(np.float32),
        V_mean=norm["V_mean"].astype(np.float32),
        V_std=norm["V_std"].astype(np.float32),
        wind_vec=norm["wind_vec"].astype(np.float32),
        solar_vec=norm["solar_vec"].astype(np.float32),
        load_frac=norm["load_frac"].astype(np.float32),
        device=device,
    )


def _build_node_features(
    scenarios_flat: np.ndarray,
    bundle: SurrogateBundle,
) -> np.ndarray:
    """Map (B, 3) global scenarios to (B, 118, 3) per-bus features."""
    X_norm = (scenarios_flat - bundle.X_mean) / bundle.X_std
    feat = np.stack(
        [
            np.outer(X_norm[:, 0], bundle.wind_vec),
            np.outer(X_norm[:, 1], bundle.solar_vec),
            np.outer(X_norm[:, 2], bundle.load_frac),
        ],
        axis=2,
    )
    return feat.astype(np.float32)


@torch.no_grad()
def predict_voltages(scenarios: np.ndarray, bundle: SurrogateBundle) -> np.ndarray:
    """Run the surrogate on `(N, 3, T)` scenarios; return `(N, T, 118)` voltages in pu."""
    if scenarios.ndim != 3 or scenarios.shape[1] != 3:
        raise ValueError(f"scenarios must be (N, 3, T); got {scenarios.shape}")

    n_scenarios, _, n_steps = scenarios.shape
    flat = scenarios.transpose(0, 2, 1).reshape(-1, 3)
    feat = _build_node_features(flat, bundle)
    x = torch.as_tensor(feat, device=bundle.device)
    pred = bundle.model(x, bundle.A_norm)
    v_norm = pred[..., 0].cpu().numpy()
    v_pu = v_norm * bundle.V_std + bundle.V_mean
    out: np.ndarray = v_pu.reshape(n_scenarios, n_steps, N_BUS).astype(np.float32)
    return out


def feasibility_check(
    scenarios: np.ndarray,
    bundle: SurrogateBundle,
) -> dict[str, float | int]:
    """Score scenarios under operational and ANSI voltage bounds.

    Returns the per-scenario feasibility rate (a scenario is feasible iff
    every bus stays in bounds for every hour). The two bound regimes match
    `eval_all_models_bounds.py` in pc-ddpm-epec2026; absolute numbers may
    differ from the AC-PF headline because the GNN is a learned surrogate.
    """
    v_pu = predict_voltages(scenarios, bundle)
    op_lo, op_hi = V_BOUNDS_OPERATIONAL
    a_lo, a_hi = V_BOUNDS_ANSI

    op_feasible = ((v_pu >= op_lo) & (v_pu <= op_hi)).all(axis=(1, 2))
    a_feasible = ((v_pu >= a_lo) & (v_pu <= a_hi)).all(axis=(1, 2))

    n_total = int(scenarios.shape[0])
    return {
        "n_total": n_total,
        "operational_feasible": int(op_feasible.sum()),
        "ansi_feasible": int(a_feasible.sum()),
        "operational_pct": float(op_feasible.mean() * 100),
        "ansi_pct": float(a_feasible.mean() * 100),
    }


def traffic_light(operational_pct: float) -> str:
    """Bucket operational feasibility into a green/yellow/red badge.

    Headline metric is 100% under operational bounds, so any drop is
    surprising. Bucket strictly: 100 → green, 90–99 → yellow, <90 → red.
    """
    if operational_pct >= 100.0:
        return "green"
    if operational_pct >= 90.0:
        return "yellow"
    return "red"