In the CNN 5-bus power-flow surrogate, pairing model capacity with sufficient training length improves voltage accuracy, and line flows are the more demanding test¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? The 200-epoch, 240-filter voltage model achieved test MAE below 10^-3 pu, while both 50-epoch models incurred more than an order-of-magnitude larger error. → Junjie Yin 2026 q2 i? The line-flow model reproduced active flows across all six lines with test MAE about 14 MW and reactive flows with test MAE about 2 MVAr. → Junjie Yin 2026 (2)

Evidence¶

Junjie Yin 2026¶

Junjie Yin, Buxin She, Xinyu Feng, Fangxing (Fran) Li. (2026). Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework. arXiv preprint. https://arxiv.org/abs/2608.02599 doi:10.48550/arxiv.2608.02599

q2 · i? · design · r2

Numerical evaluation of three CNN voltage-model variants on held-out test operating points of a PJM 5-bus system (Monte-Carlo power-flow data, 100 samples). The 200-epoch/240-filter model reached "test MAE below10−3 pu"; lightly trained variants erred far more.

"the model trained for 200 epochs with 240 filters reproduces this profile almost exactly (test MAE below10−3 pu), including the dip at the PQ bus, whereas the two lightly trained 50-epoch models incur more than an order- of-magnitude larger error"

Junjie Yin 2026 (2)¶

Junjie Yin, Buxin She, Xinyu Feng, Fangxing (Fran) Li. (2026). Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework. arXiv preprint. https://arxiv.org/abs/2608.02599

q2 · i? · design · r2

Evaluation of the line-flow CNN variant (Model 4) on unseen test operating points. It captured active flows with "test MAE≈14MW" and reactive flows with "test MAE≈2MV Ar," with the largest residual at the most heavily loaded line.

"with moderate per-line deviations (test MAE≈14MW on flows spanning roughly±240MW). The reactive flows are likewise captured (test MAE≈2MV Ar)"

Discussion¶

Related Claims¶

• Deep learning techniques such as CNNs and RNNs better detect cheating from visual cues than traditional techniques, but each carries stated trade-offs — related
• With identical architectures and budgets, a PINN more closely reproduced swing-equation rotor angle and speed trajectories than an ordinary neural network