Fast and accurate node-level hosting capacity estimation in distribution networks using graph neural networks
1UiT, The Arctic University of Norway, Tromsø, Norway · 2Ricerca sul Sistema Energetico – RSE S.p.A., Milano, Italy · 3DISI, University of Trento, Trento, Italy
Electric Power Systems Research · Volume 264 · Article 113871 · March 2027
In one sentence. HCGNN combines graph message passing with feeder-path electrical features to estimate every node’s hosting capacity in milliseconds, without running a new AC-OPF for each node.
Problem
Distribution system operators need node-level hosting capacity estimates to determine where new distributed generation or loads can connect without violating voltage or thermal limits. AC optimal power flow provides accurate values but must solve a nonlinear, non-convex problem for every node, making repeated assessments and interactive hosting-capacity maps computationally expensive.
Main contributions
- HCGNN formulates hosting capacity estimation as node-level graph regression on radial distribution networks.
- Path-based resistance, reactance, loading, and structural descriptors expose upstream electrical dependencies to a compact GIN model.
- An open dataset contains 7,188 synthetic Italian medium-voltage networks and approximately 1.43 million nodes and edges labeled through AC-OPF.
- Accuracy, feasibility, inference time, data efficiency, solver robustness, and out-of-distribution behavior are evaluated separately.
Method
Each bus receives local structural features plus statistics computed along its shortest path to the feeder, including cumulative impedance and line-capacity information. Three Graph Isomorphism Network layers with 128 hidden units refine these representations, and a multilayer perceptron outputs one hosting-capacity value per node. Training targets come from offline AC-OPF simulations; deployment requires only a non-iterative forward pass.
Experimental setting
- Networks
- 7,188 connected radial feeders with 68–559 nodes each, representative of Italian medium-voltage grids.
- Target
- Node-level hosting capacity under ±5% voltage limits and line thermal constraints, computed with Pandapower AC-OPF.
- Split
- 70% training, 15% validation, and 15% testing; five deterministic runs with different random seeds.
- Baseline
- Linear DistFlow hosting-capacity estimation through direct inversion of a linearized power-flow model.
- Metrics
- MAE, MAPE, tolerance-aware feasibility rate, and inference time; IEEE 123-bus and 33-bus tests probe robustness and domain shift.
Key results
On the in-domain test set, HCGNN lowers MAE by about 30% and MAPE by about 20% relative to Linear DistFlow. GPU inference is approximately 4.6 times faster than the linear baseline and about 11,000 times faster than AC-OPF. With a 1% tolerance, 93% of predictions do not exceed the AC-OPF value by more than that margin.
| Measure | HCGNN | Reference |
|---|---|---|
| MAE [p.u.] ↓ | 0.0705 ± 0.0083 | Linear DistFlow: 0.1010 |
| MAPE [%] ↓ | 2.17 ± 0.23 | Linear DistFlow: 2.73 |
| Inference speed ↑ | ≈11,000× faster | AC-OPF |
| Feasibility rate at 1% tolerance ↑ | 93% | Strict feasibility: 60% |
When this work is relevant
The approach targets interactive hosting-capacity maps, rapid distributed-energy-resource connection studies, large-scale scenario analysis, and operational decision support where many grid nodes or network variants must be assessed repeatedly.
Limitations
The study assumes radial, balanced single-phase networks and omits existing load and generation. Training uses synthetic medium- and large-scale Italian feeders, and accuracy degrades under strong domain shift: on the unseen IEEE 33-bus network, HCGNN has higher error than Linear DistFlow. Voltage-critical cases form only about 25% of the data. Meshed grids, unbalanced multiphase systems, live operating profiles, physical constraints in the loss, and interpretability remain future work.
Citation
Persistent identifier: https://doi.org/10.1016/j.epsr.2026.113871