Spaces:
Running
GENCO for Corrective Grid Planning- Neural Initialization, Solver Workloads and Verified Outcomes
Author: Bikash Sah
Affiliation: Chair of Interconnected Automation Systems, University of Siegen, Germany
Contact: Bikash.Sah@uni-siegen.de
Results viewer: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning
Full report: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning/blob/main/report/report.pdf
Code: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning/tree/main/code
Frozen inputs and configuration: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning/tree/main/protocol
Checkpoint manifest: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning/blob/main/code/workload_checkpoint_lock.json
Complete final raw evidence: https://huggingface.co/spaces/opti-bikash/genco-corrective-planning/resolve/main/evidence/final_feedback.zip
Idea:
My starting point was the use of EV flexibility to support critical loads during severe grid conditions. Initial technical discussions led to a broader research question: how much can GENCO contribute to the time required to obtain a verified corrective operating plan?
I developed a workflow in which pretrained GENCO supplies initial states to a reserve-limited AC corrective problem. The optimization first serves critical load, then remaining load, and finally minimizes the operating objective. The study examines the interaction between neural depth, task alignment, initialization, CPU scheduling and numerical verification.
The report documents how the original resilience idea developed into this benchmark and formulates a subsequent extension to EV energy scheduling and electrothermal constraints.
Method:
The final comparison uses 231 fresh demand-contingency instances, 10 methods, and 10 balanced repetitions, giving 300 campaigns and 23,100 online evaluations.
IEEE 14-, 30- and 118-bus systems are evaluated with IPOPT and CPU MadNLP. Each solver receives five workflows:
- A full AC power-flow seed.
- The known preventive operating state.
- Optional GENCO PF inference with 12 layers.
- Optional GENCO PF inference with four layers.
- Optional GENCO OPF inference with 12 layers.
All methods use the same FIFO scheduler and seven solver workers. The first seven cases begin immediately. A GPU producer supplies later seeds when they are ready; other dispatches use the preventive state. Each unsuccessful corrective attempt receives one preventive-state retry within the same solver engine.
PF4 applies four layers of the fixed PF checkpoint and is an approximate computation. CUDA graph execution is checked against eager execution for each variant. The GridFM graphkit implementation provides inference, and the pinned DataKit power-balance implementation provides an additional numerical check.
Results:
All final campaigns executed, and the subsequent numerical recheck passed. It checked 307 frozen files, 23,562 case records including references, and 70,986 attempted solver-stage records.
There were 23,000 accepted, service-matched online plans and 22,950 economically qualified plans.
| Grid | Strongest classical control | Median quality-qualified wait (s) | Best neural workflow | Median quality-qualified wait (s) |
|---|---|---|---|---|
| IEEE 14 | MadNLP, preventive | 0.252121 | MadNLP, optional OPF12 | 0.258634 |
| IEEE 30 | MadNLP, preventive | 0.594578 | MadNLP, optional OPF12 | 0.596435 |
| IEEE 118 | IPOPT, preventive | 5.240312 | IPOPT, optional PF4 | 5.371996 |
The neural workflows did not satisfy the prespecified end-to-end speedup rule against the strongest controls. PF4 shortened producer elapsed time by approximately 1.75 to 1.90 times in the IPOPT workloads and increased the number of consumed neural seeds. The resulting workload time depended strongly on the quality of the known preventive initialization and the speed of CPU correction.
Reported warm-campaign times include prediction, queueing, correction, verification, record output and the producer tail. Loading, compilation and qualification are recorded separately.
Two 118-bus MadNLP scenarios reached the cost-stage iteration limit in both allowed attempts. Across five workflows and ten repetitions, this produced 100 unresolved evaluations. A third scenario exceeded the economic allowance, producing 50 additional economic failures. All these outcomes are retained.
The waiting endpoint includes every designated case and assigns the full deadline to an unqualified case. Economic qualification requires service matching and an operating objective no more than max(0.0001, 0.001 × |reference objective|) above the local reference. The accompanying CSV identifies every affected case and its solver status.
Additional interpretation of the frozen results:
The report revision dated 26 October 2026 before the deadline of 1300 CET, adds descriptive analyses of the existing V10 records. Experimental inputs, checkpoints, correction code, acceptance thresholds, and headline outcomes are unchanged.
The first analysis separates producer time, prediction availability, and the final planning outcome. Because the first seven cases start classically, neural delivery has a fixed maximum share of all cases. On IEEE 118, every eligible dispatch received a GENCO seed for all three variants and both correction engines. Even with this complete coverage, each neural workflow had a longer economic-qualified wait than its same-engine preventive control in every matched repetition. The full figure retains all three grids and all neural workflows.
The second analysis explains the role of final plan quality. For the IEEE 118 preventive controls, median warm campaign duration was 9.364 s with MadNLP and 10.072 s with IPOPT. Economic-qualified waits were 11.270 s and 5.240 s, respectively. MadNLP's three unqualified cases per workload contribute 8.571 s under the unchanged 300 s deadline. The stacked bars retain every method and every unsuccessful outcome. The iteration-limited records remain unresolved under the benchmark's acceptance rule; they do not establish physical infeasibility of the contingencies.
The report also displays all ten matched repetition differences for each neural workflow. These are repeated timings of fixed workloads, not additional independent demand samples. Derived CSVs, the figure-reproduction script, and revision notes are available in report/analysis. The original report is retained alongside the revised report.
Reproducibility and hardware:
The experiments ran on OMNI at the University of Siegen, using AMD EPYC 7452 CPUs and Tesla V100-PCIE-16GB GPUs. Each grid task requested eight physical CPU cores, 32 GB RAM and one GPU. Seven cores served correction workers; one served the coordinator and optional prediction producer. Final correction ran on CPUs in Float64.
The recorded software includes Python 3.11.2, PyTorch 2.7.1+cu118, CasADi 3.7.2, Julia 1.10.10, ExaModels 0.12.1 and MadNLP 0.9.2.
The release preserves node and GPU identities, CPU affinity, thread settings, observed package versions, model hashes, checkpoint revisions, frozen demands, all final attempts and logs. It includes readable source, the Julia dependency manifest, numerical recheck scripts, figure-generation scripts and a static results viewer.
The development ledger enumerates 2,094 retained workload campaigns. Earlier summaries and archive digests document the progression. Repeated campaigns are timing repetitions over specified inputs. Complete raw evidence for the final comparison is embedded; earlier summary extracts are preserved separately.
Limitations and next work:
The final cases contain connected single-line contingencies and fresh demands on topology families known during development. Overlap with GENCO training data is unknown. The reference consists of accepted local IPOPT solutions from two starts. Global optimality is unestablished.
EV dynamics, cold-weather electrothermal limits, island energization and transient security remain subsequent application studies. Exact timing reproduction depends on hardware, runtime libraries and scheduling; a complete transitive Python wheel lock was not captured.
The immediate next experiments will examine the three 118-bus MadNLP outliers, followed by task-specific corrective prediction, larger grids and time-coupled EV support. The report provides the resource equations and evaluation conditions for those extensions.
The released results establish the measured contribution and limitations of the present GENCO workflow. Attribution and software assistance are documented in the report.




