Stefano Frizzo Stefenon

Convolutional Neural Networks (CNNs) and transformer architectures offer strengths for modeling temporal data: CNNs excel at capturing local patterns and translational invariances, while transformers effectively model long-range dependencies via self-attention. This paper proposes a hybrid architecture integrating convolutional feature extraction with a multi-head attention backbone to enhance multivariate time series forecasting. The CNN module first applies a hierarchy of one-dimensional convolutional layers to distill salient local patterns from raw input sequences, reducing noise and dimensionality. The resulting feature maps are then fed into the prediction model, which applies multi-head attention to capture both short- and long-term dependencies and to weigh relevant covariates adaptively. We evaluate the proposed CNN-transformer-based model on a hydroelectric natural flow time series dataset. Experimental results demonstrate that the proposed model outperforms well-established deep learning models, with a mean absolute percentage error of up to 2.2%. The explainability of the model is obtained by a proposed SHapley Additive exPlanations (SHAP) and Multi-Head Attention Weights (MHAW). Our novel architecture, named CNN-Transformer-SHAP-MHAW, is promising for applications requiring high-fidelity and multivariate time series forecasts.

Fault identification in Nuclear Power Plants (NPPs) is critical for ensuring operational safety, reliability, and efficiency. Traditional diagnostic methods often rely on physical models and expert systems, which may struggle to capture the complex dynamics of transient events. To overcome these limitations, this paper proposes an optimized stacked Graph Attention Network (GAT) for fault detection in NPPs by modeling the complex interdependencies among system components as graphs. Transient operational data are transformed into graph representations, where nodes correspond to system variables, and edges capture physical relationships. The architecture of the proposed model is optimized using a Heteroscedastic and Evolutionary Bayesian Optimization (HEPO), ensuring the use of the best configuration. The proposed GAT-based model, hypertuned by HEPO, is trained to recognize patterns associated with both normal and faulty transient conditions, including sensor anomalies and actuator failures. Based on synthetic data generated from the Personal Computer Transient Analyzer (PCTRAN), the proposed model achieved results above 0.96 for accuracy, precision, recall, and F1-score in a statistical analysis.

Deep 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.

Accurate 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.

Outdoor ceramic insulators are critical components of power transmission and distribution overhead lines, providing essential mechanical support and electrical insulation for high-voltage conductors. Their reliable operation is increasingly challenged by material aging, physical defects, and surface contamination, which can lead to flashover events and subsequent power outages. In recent years, machine learning (ML) and deep learning (DL) techniques have emerged as powerful, scalable, and data-driven tools for automated condition assessment of outdoor ceramic insulators, offering faster, more objective, and more consistent alternatives to traditional manual inspection practices. This paper presents a comprehensive and structured review of state-of-the-art ML/DL-based approaches for the diagnosis and condition assessment of outdoor ceramic insulators. The review contributes by organizing the field along three complementary dimensions: (i) foundational ML/DL methodologies and their adaptation to non intrusive insulator monitoring; (ii) comparative analysis of image and radiation based sensing modalities, including infrared, ultraviolet, acoustic, and radio frequency techniques, together with multimodal data fusion strategies; and (iii) a critical examination of dataset limitations, validation practices, and common misconceptions such as overfitting and insufficient field evaluation that explain contradictory results across studies. Emerging trends, including unmanned aerial vehicle (UAV)-based inspection, synthetic data generation, and autonomous intelligent monitoring, are synthesized to identify open challenges and best practices toward reliable, interpretable, and scalable artificial intelligence (AI)-driven inspection systems.

This 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.

The growing complexity of thermal power generation systems demands advanced forecasting solutions capable of integrating data analysis, model selection, and interpretability. This study proposes a modular large language model (LLM) agent framework for time series forecasting, designed to operate locally and interactively through natural language instructions. The framework incorporates a domain‐specific time series agent that was developed to automate data preprocessing, anomaly detection, and forecasting tasks using neural and statistical models. Experiments demonstrated the agent's capacity to autonomously conduct end‐to‐end analyses, achieving accurate forecasts with minimal user intervention. The PatchTST model, automatically selected by the agent, yielded the lowest mean squared error among evaluated methods. Results highlight the potential of LLM‐based agents to enhance transparency, usability, and reproducibility in energy forecasting pipelines.