README.md

---
license: mit
---

# PowerDiT Dataset

This repository contains processed power-system graph data for machine learning experiments on power-flow- and optimal-power-flow-related tasks.  
The data is generated with [**gridfm-datakit**](https://github.com/gridfm/gridfm-datakit) based on benchmark grid cases from [**PGLib**](https://github.com/power-grid-lib/pglib-opf), and then preprocessed into graph-structured samples.
It is intended to be stored in the folders examples/data_processed/pf and examples/data_processed/opf. 

## Overview

The dataset consists of graph representations of electric power systems under a range of operating conditions and perturbations.

Each sample corresponds to a power-grid state derived from a benchmark network case, with:

- **nodes** representing buses
- **edges** representing transmission lines
- **node features** encoding electrical state, bus metadata, and generator properties.
- **edge features** encoding electrical line parameters.

The data is split into two different datasets:

- **PF mode (Power Flow)**  
  The samples are merely physically (power-flow) consistent with non-cost-optimal generations.

- **OPF mode (Optimal Power Flow)**  
  The samples are creating using an opf-solver such that generator setpoints are cost-optimal.


## Available data

The processed dataset currently includes the following grid cases and sample counts:

| grid | #buses | #gens | #lines | pf samples | opf samples |
|---|---:|---:|---:|---:|---:|
| case3_lmbd | 3 | 3 | 6 | 1,224,988 | 934,045 |
| case5_pjm | 5 | 5 | 12 | 1,287,400 | 1,214,688 |
| case14_ieee | 14 | 5 | 40 | 1,290,095 | 1,300,186 |
| case24_ieee_rts | 24 | 33 | 76 | 1,290,126 | 1,301,982 |
| case30_ieee | 30 | 6 | 82 | 1,300,680 | 1,296,767 |
| case30_as | 30 | 6 | 82 | 1,300,200 | 1,284,562 |
| case39_epri | 39 | 10 | 92 | 1,300,211 | 1,301,635 |
| case57_ieee | 57 | 7 | 160 | 1,301,222 | 1,273,026 |
| case60_c | 60 | 23 | 176 | 1,302,069 | 1,300,858 |
| case73_ieee_rts | 73 | 99 | 240 | 1,300,405 | 1,301,531 |
| case89_pegase | 89 | 12 | 420 | 654,591 | 1,483,225 |
| case118_ieee | 118 | 54 | 372 | 1,300,742 | 1,301,781 |
| case162_ieee_dtc | 162 | 12 | 568 | 1,046,318 | 181,038 |
| case179_goc | 179 | 29 | 526 | 1,293,617 | 739,821 |
| case197_snem | 197 | 35 | 572 | 1,301,323 | 1,391,101 |
| case200_activ | 200 | 49 | 490 | 1,301,480 | 1,320,560 |
| case240_pserc | 240 | 143 | 896 | 1,011,541 | 260,102 |
| case300_ieee | 300 | 69 | 822 | 1,067,772 | 625,028 |
| case500_goc | 500 | 224 | 1,466 | 1,018,707 | 543,710 |
| case588_sdet | 588 | 167 | 1,372 | 1,427,568 | 622,209 |
| case793_goc | 793 | 214 | 1,826 | 1,414,878 | 405,521 |


## What the data consists of

At a high level, the dataset contains:

- benchmark power-grid topologies
- sampled and perturbed operating conditions
- simulated physical solutions
- graph-structured representations for ML
  
### Node features

| # | Feature | Name | PF-PQ | PF-PV | PF-REF | OPF-PQ | OPF-PV | OPF-REF | SE-PQ | SE-PV | SE-REF |
|---:|---|---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| 0 | P_d | Active load demand | I | I | I | I | I | I | M | M | M |
| 1 | Q_d | Reactive load demand | I | I | I | I | I | I | M | M | M |
| 2 | Q_g | Reactive generation | I | O | O | I | O | O | M | M | M |
| 3 | V_m | Voltage magnitude | O | I | O | O | O | O | M | M | M |
| 4 | V_a | Voltage angle | O | O | I | O | O | I | O | O | O |
| 5 | PQ flag | PQ bus indicator | I | I | I | I | I | I | I | I | I |
| 6 | PV flag | PV bus indicator | I | I | I | I | I | I | I | I | I |
| 7 | REF flag | Slack bus indicator | I | I | I | I | I | I | I | I | I |
| 8 | V_m^min | Min. voltage magnitude | -- | -- | -- | I | I | I | I | I | I |
| 9 | V_m^max | Max. voltage magnitude | -- | -- | -- | I | I | I | I | I | I |
| 10 | Q_g^min | Min. reactive generation | -- | -- | -- | I | I | I | I | I | I |
| 11 | Q_g^max | Max. reactive generation | -- | -- | -- | I | I | I | I | I | I |
| 12 | G_s | Shunt conductance | -- | -- | -- | I | I | I | I | I | I |
| 13 | B_s | Shunt susceptance | -- | -- | -- | I | I | I | I | I | I |
| 14 | V_n (kV) | Nominal voltage | -- | -- | -- | I | I | I | I | I | I |

### Generator features

| # | Feature | Name | PF-PQ/PV | PF-REF | OPF | State Est. |
|---:|---|---|:---:|:---:|:---:|:---:|
| 0 | P_g | Active generation | I | O | O | M |
| 1 | P_g^min | Min. active generation | -- | -- | I | I |
| 2 | P_g^max | Max. active generation | -- | -- | I | I |
| 3 | c_0 | Cost constant term | -- | -- | I | I |
| 4 | c_1 | Cost linear coefficient | -- | -- | I | I |
| 5 | c_2 | Cost quadratic coefficient | -- | -- | I | I |

### Edge features

| # | Feature | Name | Power Flow | OPF | State Est. |
|---:|---|---|:---:|:---:|:---:|
| 0 | P_e | Active power flow | O | O | M |
| 1 | Q_e | Reactive power flow | O | O | M |
| 2 | Y_ff,r | From-from admittance (real) | I | I | I |
| 3 | Y_ff,i | From-from admittance (imag.) | I | I | I |
| 4 | Y_ft,r | From-to admittance (real) | I | I | I |
| 5 | Y_ft,i | From-to admittance (imag.) | I | I | I |
| 6 | tap | Transformer tap ratio | I | I | I |
| 7 | theta_min | Min. angle difference | -- | I | I |
| 8 | theta_max | Max. angle difference | -- | I | I |
| 9 | rate_A | Thermal rating (MVA) | -- | I | I |

- I = input feature
- O = output / target feature
- M = measured feature
- -- = not used


## Data source

The samples have been generated using **Gridfm-datakit** on the [**Future Technologies Partition**](https://www.nhr.kit.edu/userdocs/ftp/) on four intel sapphire rapids nodes. 
The underlying network cases (topologies and initial load and generation profiles) are based on **PGLib** benchmark systems, the perturbation framework is **Gridfm-datakit**.
The preprocessing from **Gridfm-graphkit** has been adapted to allow for larger dataset sizes.

## Perturbations applied during data generation

The data generation process in **gridfm-datakit** applies perturbations to create diverse operating conditions and topologies.

The perturbation types include:

### 1. Load perturbations
Load levels are randomly varied using a global scaling factor and local noise.

### 2. Topology and admittance perturbations
N-k perturbations are used to randomly cut off buses or edges. We used k=1.
Further, the resistance and reactance parameters are randomly scaled during the perturbation process.

### 3. Generation perturbations
The generators are perturbed by changing their cost coefficients. In OPF-mode, the generator setpoints are recalculated with an opf-solver
after topology and load perturbations have been applied. In PF-mode, an OPF-solver is only used BEFORE the perturbations. 

## Preprocessing

The raw power-system scenarios are preprocessed into a disk-backed heterogeneous graph format for efficient training and random access.

### Raw inputs

The preprocessing pipeline starts from three parquet tables per dataset split:

- `bus_data.parquet`
- `gen_data.parquet`
- `branch_data.parquet`

Each row belongs to a specific `scenario`, and each scenario corresponds to one graph sample.

### Scenario validation and filtering

Before graph construction, scenarios are validated for structural consistency.

In particular, scenarios are removed if their bus indices are invalid, i.e. if:

- bus IDs are duplicated, or
- bus IDs are not consecutive from `0` to `N-1`

This ensures that each scenario can be converted into a well-formed graph with a consistent node indexing scheme.

### Feature extraction

The preprocessing selects fixed feature subsets for buses, generators, and branches.

#### Bus features
Bus node features include electrical state, bus type indicators, and operating constraints:

- `Pd`, `Qd`
- `Qg`, `Vm`, `Va`
- `PQ`, `PV`, `REF` flags
- `min_vm_pu`, `max_vm_pu`
- `min_q_mvar`, `max_q_mvar`
- `GS`, `BS`
- `vn_kv`

#### Generator features
Generator node features include operating quantities and cost parameters:

- `p_mw`
- `min_p_mw`, `max_p_mw`
- `cp0_eur`, `cp1_eur_per_mw`, `cp2_eur_per_mw2`
- `in_service`

#### Branch features
Branch edge features include power-flow quantities, admittance terms, and branch constraints.

For each physical branch, the preprocessing creates **two directed edges**:
- one forward edge (`from_bus -> to_bus`)
- one reverse edge (`to_bus -> from_bus`)

The forward and reverse edges receive direction-specific attributes, e.g.:
- forward: `pf`, `qf`, `Yff_*`, `Yft_*`
- reverse: `pt`, `qt`, `Ytt_*`, `Ytf_*`

Shared branch attributes include:
- `tap`
- `ang_min`, `ang_max`
- `rate_a`
- `br_status`

### Bus-level aggregation of generator limits

Reactive power limits from generators are aggregated onto buses before graph construction.

For each `(scenario, bus)`, the preprocessing sums:
- `min_q_mvar`
- `max_q_mvar`

across all generators connected to that bus, and merges these totals into the bus feature table.

This enriches bus nodes with bus-level reactive capability information.

### Target construction

Task labels are derived from subsets of the selected features.

- **Bus targets** are taken from the leading subset of bus features
- **Generator targets** are taken from the leading subset of generator features
- **Branch targets** are stored separately as edge-level flow quantities

The exact interpretation of inputs and outputs depends on the downstream task (e.g. PF, OPF, state estimation, reconstruction).

### Heterogeneous graph construction

Each scenario is converted into a `HeteroData` graph with:

- **bus nodes**
- **generator nodes**
- **bus-to-bus edges** for physical branches
- **generator-to-bus edges**
- **bus-to-generator edges**

This yields a heterogeneous graph representation that preserves both the electrical network structure and generator attachment structure.

### Storage format

Instead of storing one file per scenario, the processed dataset is written into a small number of large NumPy memory-mapped arrays (`memmap`).

This avoids metadata overhead on large cluster file systems and enables efficient random access.

The processed directory contains arrays such as:

- `bus_x.npy`, `bus_y.npy`
- `gen_x.npy`, `gen_y.npy`
- `edge_index.npy`, `edge_attr.npy`, `edge_y.npy`
- `gen_bus_ei.npy`, `bus_gen_ei.npy`

along with an `index.pt` file containing per-scenario offsets.

Using these offsets, a single scenario can be loaded without materializing the full dataset in memory.

### Runtime loading and normalization

When a sample is accessed:

1. the corresponding slices are read from the memmap arrays,
2. a PyTorch Geometric `HeteroData` object is reconstructed,
3. edge indices are converted into PyG format,
4. feature normalization is applied at access time.

Normalization is therefore **not baked into the stored arrays**, but applied dynamically during dataset loading.