
Abstract Bergeron interfaces decouple terminal equations in real-time co-simulation by exchanging delayed travelling-wave variables. Communication, synchronization, and buffering may impose a delay different from the physical propagation delay of the target line, so the terminal admittance no longer matches it over frequency. This paper formulates the mismatch as terminal-admittance compensation for a delay-scaled Bergeron interface. A low-order rational residual is fitted from sampled admittance data while preserving the explicit delay. The mutual target is shaped in frequency, fitted by vector fitting, passivated by residue perturbation, and checked again after delay reconstruction. If a local passivity deficit remains, a passive diagonal admittance supplies the missing dissipation. For the representative case, the fixed-step realization reduces the 50 Hz terminal-admittance errors from approximately 67 % to 11–14 % while retaining a positive passivity margin after discretization.
Abstract Accurate photovoltaic power forecasting is essential for integrating solar energy into electrical grids, yet most high-accuracy models remain “black boxes” that do not reveal the mechanism of their decisions. This study addresses the short-term forecasting of photovoltaic DC power output at the OEDI station, at a five-minute resolution, through a systematic comparison of six algorithms belonging to three families: linear models (Ridge and SVR), tree-based ensembles (XGBoost and Random Forest), and neural networks (MLP and LSTM). Alongside accuracy evaluation, the SHAP method was applied to interpret each model’s behaviour and reveal the contribution of meteorological and temporal features to its predictions. The results showed that five of the six models achieved normalized errors within the range conventionally regarded as excellent (nRMSE < 10 %), with the LSTM marginally above this threshold at 10.47 %, and with a slight superiority for the MLP. Notably, this relatively simple model surpassed the more time-series-specialized LSTM network, indicating that architectural complexity does not automatically translate into higher accuracy at this short horizon. The SHAP analysis revealed a structural divergence among the model families in their handling of physical relationships: The attributions of the tree-based and neural models exhibited the non-linear pattern expected from thermal efficiency decline, while the linear models reflected their structure in straight relationships. The consistency of the feature ordering across families suggests that the attributions are driven predominantly by the structure of the data rather than by the inductive bias of any individual algorithm. This study contributes to combining predictive accuracy with interpretive transparency, strengthening confidence in deploying machine learning models in energy systems.
Zusammenfassung Es werden Modelle und Methoden entwickelt, mit denen sich Resonanzeffekte in Hochstromwand-lern bereits vor der Fertigung berechnen und analysieren lassen. Analytische und semi-analytische Feldberechnungsverfahren liefern die Ersatzelemente eines Ersatzschaltbilds. Zusammen mit Netz- und Bürdenmodell entsteht als Systembeschreibung ein reguläres Deskriptorsystem, aus dessen verallgemeinerten Eigenwerten die Resonanzfrequenzen abschätzt werden. Kennzahlen aus den Eigenvektoren eines Zustandsraummodells gleicher Dynamik zeigen, welche Induktivitäten und Kapazitäten die Resonanzen bestimmen. Messungen bis 40 kHz bestätigen das Modell mit maximal 13,4 % Abweichung bei den Resonanzfrequenzen.
Abstract Modern transmission networks face increasing protection challenges due to renewable energy integration, bidirectional power flows, high-impedance faults (HIFs), and potential failures of local protection devices, motivating the development of intelligent wide-area protection schemes. This paper presents a distributed real-time framework for wide-area fault detection and localization using synchrophasor measurements. Optimal Phasor Measurement Unit (PMU) placement is achieved using Binary Integer Linear Programming (BILP) to ensure complete network observability with the minimum number of PMUs. A Virtual PMU implemented in the OPAL-RT HYPERSIM real-time simulator streams synchrophasor measurements via the IEEE C37.118 protocol over User Datagram Protocol (UDP) to a separate computer, where a multilayer Artificial Neural Network (ANN) and a Support Vector Machine (SVM) perform real-time inference for fault detection and faulted-line identification. The framework is evaluated on the IEEE 9-bus system under symmetrical, asymmetrical, and high-impedance fault conditions (up to 90 Ω in testing, 80 Ω in training). The ANN reaches 98.68 % ± 0.10 % test accuracy (mean ± SD, 5 training runs) with zero false positives; the SVM reaches 98.37 % ± 0.06 %. The estimated average inception-to-decision latency is ≈9.4 ms (8.33 ms average PMU frame-waiting time, ∼0.5 ms estimated communication/IEEE C37.118 processing, ∼0.5 ms estimated feature extraction, and 0.083 ms measured Artificial Neural Network (ANN) inference), with a worst-case PMU frame-waiting time of 16.67 ms at 60 frames/s well within the 50–150 ms fault duration range in the dataset. The trained models are not inherently topology-adaptive; however, topology-aware retraining on additional samples recovers both models to at least 98.7 % accuracy across all evaluated topology-change scenarios. SHapley Additive exPlanations (SHAP) analysis identifies current magnitude features at THE three PMU buses as the dominant fault-discriminating features, consistent with (though not an independent statistical proof of) the BILP placement strategy. These results demonstrate that combining optimized PMU placement, real-time synchrophasor communication, and explainable machine learning enables fast, accurate, and transparent wide-area protection suitable for next-generation transmission systems.