Emanuele Ogliari

The rapid electrification of transportation has significantly increased electric vehicle (EV) charging demand, posing new challenges for energy management systems (EMS). Accurate and reliable EV charging demand forecasting is essential to support flexible and efficient system operation. Probabilistic forecasting, rather than point forecasting, better captures uncertainty and supports more effective energy system control. This study proposes a novel hybrid forecasting framework that integrates deep learning with statistical dependence modeling to enhance uncertainty quantification in EV charging demand forecasting. Building on a previously developed LSTM-based encoder–decoder architecture with an attention mechanism, we extend its point forecasting capability by incorporating quantile regression (QR) to produce probabilistic forecasts. To capture temporal dependencies more effectively, we integrate an empirical copula-based method that models the joint distribution across lagged load values. Furthermore, a Copula-informed QR model is constructed by embedding copula-derived features into the QR framework. Using EV charging data from the a.s.r. Smart Car Park in the Netherlands, the world's largest bi-directional EV charging hub, we evaluate performance over a 24-hour ahead forecast horizon, reflecting the growing importance of day-ahead forecasting in EMS applications. The results demonstrate that the proposed Copula-informed QR model outperforms both independent QR and Copula models, achieving significant improvements in probabilistic metrics, including pinball loss and weighted interval score (WIS), while preserving well-calibrated prediction intervals.

Accurate very short-term Direct Normal Irradiance (DNI) nowcasting is essential for the operational efficiency of Concentrating Solar Power (CSP) plants, where rapid sun–cloud occlusions can lead to significant thermal stress and energy losses. This study proposes a multi-modal nowcasting framework that combines thermal All-Sky Infrared (ASI) image sequences with synchronous meteorological data to predict DNI at 10- and 15-minute horizons. Four advanced deep learning architectures are systematically benchmarked: baseline Convolutional Neural Network (CNN) and Residual Network (ResNet) for deep spatial encoding, CNN-Long Short-Term Memory (CNN-LSTM) for spatiotemporal dynamics, and CNN-Transformer for long-range attention-based dependencies. Evaluated on a rigorously selected 33-day test set, the hybrid CNN-LSTM architecture emerged as the most robust model, achieving a coefficient of determination (R²) of 0.7998 and 0.7688 for the 10- and 15-minute horizons, respectively, and reaching peak daily forecast skill (FS) values exceeding 50%. Crucially, targeted evaluation of high-volatility transients demonstrated the model’s superior ability to proactively anticipate severe irradiance ramps (> 100 W/m²/min). The statistical robustness of these predictive gains over the persistence baseline was strictly validated via the Wilcoxon signed-rank test (p < 10⁻¹⁴⁰). Furthermore, modality comparisons confirmed that thermal infrared imaging effectively mitigates solar glare, providing significantly higher predictive accuracy than standard multi-channel visible-light (RGB) cameras. These findings establish a highly reliable, statistically validated solution for optimizing heliostat aim-point strategies and enhancing real-time solar resource management in CSP facilities.

Compensation networks play a critical role in Resonant Inductive Wireless Power Transfer systems, as their proper design enhances both power transmission and overall efficiency. This paper proposes a systematic approach for sizing the elements of the compensation networks based on a defined set of design targets and a hybrid symbolic–numerical methodology that inherently accounts for the dissipative effects of both the pads and the compensation network components. Three modeling criteria are introduced to quantify the influence of parasitic resistances associated with the circuit elements. Four different compensation topologies, suitable for electric vehicle wireless chargers, are analyzed, each featuring an increased number of circuit elements. Simulation results confirm the effectiveness of the proposed tuning methodology and highlight how dissipative effects impact both the efficiency and dynamic behavior of the compensated coreless transformer.

The rapid expansion of green energy systems has accelerated the adoption of battery energy storage systems (BESS) and increased the demand for advanced battery management system (BMS). While lithium-ion batteries (LIBs) dominate current applications, concerns over resource limitations and cost have motivated growing interest in sodium-ion batteries (SIBs) as a more abundant and economical alternative. State of health (SOH) estimation remains essential for ensuring safe and reliable operation, yet most existing research focuses on LIBs due to the limited availability of high-quality SIB ageing datasets. This study conducted comprehensive ageing tests on LIBs and SIBs, obtained a unified dataset, and analyzed how their health declined over time, along with detailed patterns in their charging behavior, based on incremental capacity analysis (ICA) curve, to understand degradation. Using these features, seven deep-learning models, including deep neural networks (DNN), convolutional neural networks (CNN), long short-term memory (LSTM) networks, gated recurrent units (GRU), and three hybrid CNN–recurrent architectures were developed and evaluated under a Leave-One-Cell-Out (LOCO) scheme. For LIBs, recurrent and CNN–recurrent hybrid models achieved the best accuracy, reaching a minimum mean absolute error (MAE) of 0.8811%. In contrast, simpler architectures such as DNN and CNN were more effective for SIBs, with the best model achieving a minimum MAE of 0.2591%. Transfer learning (TL) from LIB to SIB demonstrated promising feasibility: several model–cell combinations showed unchanged or improved accuracy, although larger degradation-pattern mismatches in some SIB cases led to MAE increases of up to 1.4106%. Overall, the findings highlight shared electrochemical ageing characteristics between LIBs and SIBs and demonstrate the potential of leveraging extensive LIB datasets to support SOH estimation for emerging SIB technologies.

In the context of energy transformation and a growing supply of renewable energy, additional flexibility of transmission lines is required. Dynamic Line Rating provides a solution to this question by allocating additional capacities in the power lines depending on the actual weather conditions. Modern implementations of Dynamic Line Rating are typically based on costly on-line sensors or unreliable weather stations mounted on pylons. This paper proposes a novel solution to the Dynamic Line Rating, by implementing a tiny machine learning approach suitable for integration into the sensor and a real-time decision algorithm for the detection of potential new allocation spots. The proposed forecasting methodology is based on a self-learning radial basis function model, that that forecasts the temperature of the conductor 5-, 10-, and 15 minutes into the future. The model can be accommodated on a cheap microcontroller device and is suitable for on-device self-calibration at the desired retraining frequency (RF). The forecasting results are promising, achieving mean absolute errors below 0.3 °C in the best test scenarios, a clear improvement over a naive persistence benchmark, and competitive performance when compared with the enhanced proposed Echo State Network (ESN). Subsequently, an algorithmic non-parametric Change Detection Test is executed on top of the forecaster to determine the possible points of change for dynamic allocation of the Over Head Line capacity, qualitatively and quantitatively demonstrating the potential capacity expansion compared to traditional Static Line Rating.

Low-voltage ride-through (LVRT) capability is essential for grid-connected photovoltaic (PV) systems, especially as rising renewable integration challenges grid stability during voltage disturbances. Existing LVRT methods often target isolated control functions, leading to limited system resilience. This paper presents a unified control strategy integrating DC-link voltage regulation, reactive power injection, and overvoltage mitigation using a coordinated fuzzy logic framework. The proposed architecture employs a cascaded control structure comprising an outer voltage loop and an inner current loop with feed-forward decoupling, synchronized via a Synchronous Reference Frame Phase-Locked Loop (SRF-PLL). At its core is a dual-input, single-output Fuzzy Logic Controller (FLC), featuring optimized membership functions and dynamic rule-based logic to manage multiple control objectives during grid faults. The proposed FLC-based unified LVRT controller for grid-tied PV system was implemented and validated for both symmetrical and asymmetrical fault conditions in MATLAB/Simulink 2023b platform. The proposed FLC-based LVRT controller achieves voltage sag compensation of 97.02% and 98.4% for symmetrical and asymmetrical faults, respectively, outperforming conventional PI control, which achieves 94.02% and 96.5%. The system maintains a stable DC-link voltage of 800 V and delivers up to 78% reactive power support during faults. Fault detection and recovery are completed within 200 ms, complying with Bangladesh grid code requirements. This integrated fuzzy logic approach offers a significant advancement for enhancing grid stability in high-renewable environments and supports reliable renewable utilization, and more sustainable grid operation in developing regions.

Accurate state of charge (SOC) estimation remains a challenge in lithium‐ion battery management systems (BMSs) due to the cells’ complex, nonlinear internal dynamics, and their high sensitivity to operating conditions. While most existing data‐driven studies rely on a core input configuration of voltage ( V ), current ( I ), and temperature ( T ), these models often struggle with temperature‐induced variations, which typically require complex architectures. This study addresses the challenge of thermal dependency through an innovative feature engineering approach designed to maintain a low computational cost suitable for real‐time applications. Recognizing that internal resistance ( R ) inherently reflects both the cell’s temperature and aging state, we propose replacing T with R as the third input feature, resulting in an [I, V, R] configuration. We systematically compared the performance of four estimation models: DE‐long short‐term memory (LSTM), DE‐gated recurrent unit (GRU), LSTM‐unscented Kalman filter (UKF), and GRU‐UKF, using both the traditional [ I , V , T ] and the proposed [ I , V , R ] input sets. The results validate the superiority of the R ‐based configuration, which led to significant reductions in mean absolute error (MAE) by 46.02%, 43.31%, 14.02%, and 35.85%, and in root mean square error (RMSE) by 48.95%, 39.87%, 21.90%, and 23.95% for the DE‐LSTM, DE‐GRU, LSTM‐UKF, and GRU‐UKF models, respectively. This confirms that substituting T with R captures nonlinear temperature and aging effects more effectively while maintaining low computational complexity.