A carregar...
Pessoa
Stefenon, Stefano Frizzo
Assistant Professor . Instituto Politécnico de Lisboa, Instituto Superior de Engenharia de Lisboa
8 resultados
Resultados da pesquisa
A mostrar 1 - 8 de 8
- Differentiable neural search architecture with zero-cost metrics for insulator fault predictionPublication . Seman, Laio Oriel; Buratto, William Gouvêa; Gonzalez, Gabriel Villarrubia; Leithardt, Valderi Reis Quietinho; Nied, Ademir; Stefenon, Stefano FrizzoReliable monitoring of high-voltage insulators is critical for maintaining the stability of electrical power systems, particularly under environmental contamination that can lead to flashover. Traditional inspection techniques struggle to anticipate degradation dynamics, while data-driven models often rely on fixed neural architectures that inadequately capture the complex temporal patterns in leakage current signals. This work proposes a Differentiable Neural Architecture Search (DARTS) framework, based on zero-cost metrics, tailored for time series forecasting in insulator monitoring. The method based on DARTS integrates a mixed encoder-decoder design with learnable selection over long short-term memory, gated recurrent units, and transformer components, coupled with a cross-attention bridge featuring temporal bias and gating mechanisms. To ensure efficient architecture exploration, the search leverages metrics such as SynFlow and Jacobian covariance for early candidate screening, followed by a bilevel optimization stage with entropy and diversity regularization. Experiments on real-world leakage current data demonstrate that the discovered architectures outperform manually designed baselines, offering improved forecasting performance.
- Input attention, squeeze and excitation, and spatial transformer of YOLO for fault detection using UAVPublication . Carvalho, João Pedro Matos; Stefenon, Stefano Frizzo; Leithardt, Valderi Reis Quietinho; Seman, Laio Oriel; Yow, Kin-Choong; Santana, Juan Francisco De PazThe detection of faults in insulators is important to guarantee the continuous supply of electricity. To identify faults in these components, various object detection methods based on deep learning have been explored. This paper investigates architectural enhancements to the You Only Look Once (YOLO) framework for fault detection in electrical power grid insulators. Three structural variants are proposed: the Input Attention Transformer (IAT-YOLO) for spatial feature refinement, Squeeze-and-Excitation (SAE-YOLO) modules for channel recalibration, and Spatial Transformer Networks (STN-YOLO) for geometric alignment. Experiments were conducted on a publicly available insulator dataset from Unmanned Aerial Vehicles (UAVs), comprising seven defect categories, including pollution, breakage, and flashover damage. Results demonstrate that STN-YOLO and SAE-YOLO consistently improve generalization and robustness, achieving mAP values of up to 0.995 for specific classes. The findings highlight the effectiveness of integrating attention mechanisms and spatial transformations to enhance YOLO-based detection, contributing to improved automated inspection of the power grid.
- Optimized Harmonic-Multiscale Kolmogorov–Arnold Network denoised by EWT for hydroelectric dam useful volume forecastPublication . Stefenon, Stefano Frizzo; Seman, Laio Oriel; Matos-Carvalho, João Pedro; Nied, Ademir; Gonzalez, Gabriel Villarrubia; Yow, Kin-Choong; Stefenon, Stefano Frizzo; ElsevierAccurate forecasting of the useful volume of a hydroelectric reservoir is essential for energy planning and water resource management. To address this challenge, this paper proposes a novel Harmonic-Multiscale Kolmogorov–Arnold Network forecasting model. The model integrates sinusoidal (Sin) basis functions to model global oscillatory patterns and wavelet (Wav) bases to capture localized transient features within the Kolmogorov Arnold Network (KAN) framework (SinWavKAN). This hybrid representation enhances feature expressiveness by jointly modeling smooth long-term dependencies and short-term irregular dynamics. The input signals are first denoised using the Empirical Wavelet Transform (EWT) to isolate informative modes from noise, and the SinWavKAN’s hyperparameters are optimized via an Adaptive Tree-structured Parzen Estimator (ATPE). An ablation study confirmed the synergistic superiority of the sinusoidal-wavelet combination over other basis functions. Compared to state-of-the-art KAN variants (e.g., fastKAN, PyKAN, WavKAN), the proposed architecture, named in short Opt-EWT-SinWavKAN, demonstrated a reduction in mean absolute error by approximately 85%–93% across forecasting horizons from 1 to 30 days. The results conclusively show that the Opt-EWT-SinWavKAN provides a more reliable approach for hydroelectric reservoir volume forecasting.
- End-to-end decomposition-aware neural architecture search with sparse expert transformers for thermal energy forecastingPublication . Seman, Laio Oriel; Stefenon, Stefano Frizzo; Matos-Carvalho, João Pedro; Nied, Ademir; Yow, Kin-Choong; Stefenon, Stefano Frizzo; IEEEThis study presents a unified framework for thermal energy forecasting that integrates signal decomposition, Neural Architecture Search (NAS), and sparse transformer modeling. The method employs a hyperparameter-optimized hierarchical Variational Mode Decomposition (VMD) to enhance signal representation through adaptive noise suppression and mode separation. A differentiable preprocessing module, referred to as HeadComposer, performs joint search over normalization, decomposition, and convolutional transformations within a bilevel optimization procedure. The forecasting backbone consists of a transformer architecture incorporating gated residual connections for adaptive token selection and Mixture-of-Experts (MoE) feedforward layers for conditional computation. Experiments are conducted on real-world thermal power generation data comprising approximately 10,000 test samples, using an input look-back window of 96 time steps and a multi-step prediction horizon of 24 steps ahead. Results demonstrate consistent improvements over state-of-the-art baselines, achieving a mean absolute percentage error of 5.94%, corresponding to reductions of approximately 12-15% relative to competing methods. The proposed design demonstrates the effectiveness of combining decomposition-based preprocessing (VMD and HeadComposer) with NAS-driven architecture adaptation and sparse expert routing (based on MoE) for modeling non-stationary thermal generation processes.
- Multi-scale convolutional Chebyshev Kolmogorov-Arnold networks using conformal prediction for photovoltaic power forecastingPublication . Takara, Lucas de Azevedo; Souza, Rodrigo Clemente Thom de; Mariani, Viviana Cocco; Coelho, Leandro Dos Santos; Stefenon, Stefano Frizzo; Stefenon, Stefano Frizzo; IEEEAccurate photovoltaic (PV) power forecasting remains challenging due to nonlinear dynamics, rapid weather variations, and multi-scale temporal dependencies. This paper proposes ConvChebyKAN, a hybrid deep learning architecture that integrates multi-scale causal convolutions, Chebyshev polynomial parameterized Kolmogorov Arnold networks (KANs), and conformal prediction to improve both deterministic and probabilistic forecasting performance. The multi-scale convolutional module captures short-term fluctuations and long-term temporal dependencies, while the Chebyshev-based KAN improves nonlinear representation capability and numerical stability. A multi-objective hyperparameter optimization strategy based on Pareto efficiency and TOPSIS ranking is used to jointly optimize forecasting accuracy, prediction interval sharpness, and coverage reliability. The model is evaluated using real photovoltaic generation and meteorological data from four seasonal datasets and forecasting horizons from 1 to 48 hours ahead. ConvChebyKAN reaches an MAE of 149.13 kWh, RMSE of 227.34 kWh, R2 of 98.2%, and sMAPE of 6.21%, outperforming KAN and other machine learning models. These results show that the proposed architecture effectively captures multi-scale temporal patterns while providing reliable uncertainty estimates, supporting PV forecasting applications in power system operation and energy management.
- Diffusion-augmented YOLO26-Swin cascaded framework with hybrid SHAP-CAM for autonomous power grid inspectionPublication . Stefenon, Stefano Frizzo; Matos-Carvalho, João Pedro; Mariani, Viviana Cocco; Coelho, Leandro dos Santos; Yow, Kin-Choong; Stefenon, Stefano Frizzo; Springer NatureDeep learning-based autonomous inspection of power grid insulators is challenged by data imbalance and model opacity. This paper presents an end-to-end solution integrating advanced data synthesis, detection, classification, and explainability. First, a conditional diffusion model generates realistic synthetic fault images to balance the dataset. A two-stage architecture based on You Only Look Once version 26 (YOLO26) extra-large and Shifted windows (Swin)-V2-B, called YOLO26-Swin, fine-tuned with Bayesian optimization, performs robust insulator detection and then fault classification. Finally, a novel SHapley Additive exPlanations with Class Activation Mapping (SHAP-CAM) method provides intuitive visual explanations for model predictions. Extensive experiments validate our framework’s superiority: it achieves an F1-score of 0.98149 and a mean Average Precision (mAP)@[0.5] of 0.98951, exceeding leading detection and classification models. This work highlights the efficacy of diffusion models for data augmentation in critical infrastructure and advances the interpretability of vision-based inspection systems.
- Patch-based transformer with mixture of experts and conformal inference for curtailment time series forecastingPublication . Heidrich, Mikhail Zimmer; Reis, Mari Aurora Favero; Yamaguchi, Cristina Keiko; Seman, Laio Oriel; Stefenon, Stefano Frizzo; Stefenon, Stefano Frizzo; IEEEThe growth of variable renewable energy sources has increased the occurrence of curtailment in power systems, particularly in regions affected by transmission constraints and limited operational flexibility. Reliable forecasting of curtailment time series, together with well-calibrated uncertainty estimates, is therefore essential to support operational planning, market decisions, and risk management. This paper proposes a curtailment forecasting framework that combines a high-capacity neural time series model with distribution-free conformal prediction to jointly address accuracy and uncertainty quantification. The forecasting backbone is a Transformer-based encoder-decoder architecture enhanced with reversible instance normalization to mitigate distribution shift, patch-based tokenization to reduce sequence length and improve temporal representation, and a Mixture of Experts feed-forward module to enable conditional computation and improved generalization. The proposed method is evaluated against a wide range of state-of-the-art models, including NBEATS, NBEATSx, NHITS, PatchTST, LSTM, GRU, TFT, Informer, and Autoformer, across multiple forecasting horizons. The model achieves an MSE of 0.0059 for a horizon equal to 3h, corresponding to a 13.2% reduction relative to the second-best methods. For a horizon equal to 5h, the MSE is reduced to 0.0119, representing a 29.6% improvement, while for a horizon equal to 10, the model attains an MSE of 0.0461, outperforming NHITS by 22.9%. Ablation analysis confirms the importance of patch embedding, the Mixture of Experts mechanism, and reversible instance normalization. The conformal prediction module achieves an empirical coverage of 91.27% for a nominal 90% level, with adaptive and practically useful interval widths.
- Decomposition-driven Mamba state space models with expert routing for wind curtailment forecastingPublication . Seman, Laio Oriel; Stefenon, Stefano Frizzo; Matos-Carvalho, João Pedro; Stefenon, Stefano Frizzo; ElsevierWind power curtailment, known in Brazil as constrained-off, has intensified with the rapid expansion of renew able generation, driven by transmission operational constraints. Accurate forecasting of curtailment is essential for grid planning, compensation mechanisms, and infrastructure optimization. However, constrained-off time series exhibit strong non-stationarity, multi-scale behavior, and long-range dependencies that challenge conventional forecasting models. This paper proposes a decomposition-driven Mamba state space framework with Mixture-of-Experts (MoE) routing for wind curtailment forecasting. The approach integrates Entropy-Guided Variational Mode Decomposition (EG-VMD) for adaptive trend and noise separation, Reversible Instance Normalization to mitigate distribution shift, multi-scale seasonal-trend modeling, and a sparse expert-based Mamba backbone with linear-time complexity. In benchmark comparisons against state-of-the-art forecasting models, the proposed MoE-Mamba reduces Mean Squared Error (MSE) by up to 27%, Root Mean Squared Error by 14.6%, and Mean Absolute Error by 16.1% relative to the strongest baselines, while improving upon a single-expert Mamba by 15.6% in MSE. The results highlight the effectiveness of decomposition-enhanced State Space Models with expert routing for scalable, long-horizon forecasting in renewable power systems.