More than 180 papers were submitted to the conference. The following is the list of accepted papers to be presented during the conference.
Authors:
Mouna Kaddache, Farouk Mechnane, Abderrahmane Khemis, Said Drid, Boutheyna Hadmer
Abstract:
This paper presents a comparison of Type-1 and Type-2 fuzzy logic controllers (FLCs) utilized in vector control for DFIG-based wind energy systems it presents a brand-new Maximum Power Point Tracking (MPPT) algorithm designed to maximize energy extraction in erratic wind conditions. The study evaluates the controllers based on power output, system stability and dynamic response to wind fluctuations. Simulation results indicate that the Type-2 FLC offer greater robustness and adaptability in uncertain and rapidly changing wind scenarios, while the Type-1 FLC performs reliably under moderate variations. The proposed MPPT algorithm notably enhances energy efficiency and system reliability, highlighting the effectiveness of integrating Type-2 fuzzy control with advanced MPPT techniques in modern DFIG-based wind energy systems.
Authors:
ELMAOUHAB Madjda Ismahen, BOUMEDJOUT Amal, KABLI Fatima
Abstract:
The rapid expansion of the Internet of Things (IoT) has increased network connectivity and cyberattack risks, making traditional intrusion detection systems (IDS) unsuitable due to resource and energy constraints. Recent advances in artificial intelligence, including deep learning, TinyML, and federated learning, have enabled the development of lightweight and adaptive IDS for IoT and edge environments. This survey reviews recent AI-based IDS approaches, focusing on lightweight models, federated and hybrid frameworks, and edge-based solutions. It also discusses their strengths, limitations, datasets, and evaluation metrics, while highlighting key challenges such as data imbalance, adversarial attacks, privacy concerns, and real-world deployment issues. Finally, future research directions toward efficient and scalable IoT security systems are outlined.
Authors:
Khadije Mohamed Yahya, Naima Kalkil, Ghalia Merzougui, Djalal Hedjazi, Rachida Hireche
Abstract:
This work proposes an enhancement to an intelligent traffic light system based on Deep Reinforcement Learning (DRL), specifically the Deep Q-Network (DQN) approach. While the original DQN relied on the “epsilon-greedy” exploration strategy, this study adopts the Boltzmann (Softmax) exploration approach; this allows for more balanced and targeted exploration through a scheduled temperature decay mechanism. Implementing this modification required adapting the agent’s decision-making and reward management modules, as well as adjusting learning parameters, to leverage Softmax exploration. This method fosters a better balance between exploration and exploitation and helps the agent discover more effective traffic light management strategies in dynamic environments. The method was evaluated using SUMO across 10 paired random seeds under moderate (1,000 vehicles) and heavy (4,000 vehicles) traffic conditions. In heavy traffic, Boltzmann exploration reduces the average queue length (AQL) by 55.5%, improves cumulative reward by 75.5%, and decreases the final queue length by 99.8% compared to the epsilon-greedy model. These results demonstrate the effectiveness of guided, efficient probabilistic exploration in stabilizing traffic, alleviating congestion, and supporting sustainable digital and physical infrastructure.
Authors:
Mohamed El Amine GUERBAZI, Samia CHALABI, Hichem BENCHERIF, Mohamed MEGUELLATI , Ines CHABANI, Abdallah MOHAMMEDI, Adel CHINE
Abstract:
This study investigates the effects of flow channel width and electrolyte thickness on the electrochemical performance of a planar Solid Oxide Fuel Cells (SOFCs). A three-dimensional Multiphysics model was developed to simulate coupled fluid flow, species transport, charge conservation, and electrochemical reactions. Results show that increasing the flow channel width from 0.2 to 1.5 mm raises the maximum current density from 1413.4 to 4725.3 A·m⁻², due to improved mass transport and reduced concentration polarization. Reducing electrolyte thickness from 500 μm to 50 μm further enhances current density by lowering ionic resistance and shortening the oxygen-ion transport path. These findings offer practical guidance for optimizing SOFC geometry to improve efficiency and operational stability.
Authors:
Shiven Pradhan, Puja Cholke, Pranjali Sawant, Prajwal Sakharkar, Diya Potdar, Pranav Chaudhari
Abstract:
Industrial places like manufacturing units, warehouses, and processing plants are vulnerable to incidents arising from workman’s exposure to dangerous zones. Existing surveillance techniques are inefficient and utilize single sensor data combined with a wired connection from the sensor to the alarm system. This paper proposes an innovative dual sensor system that employs a VL53L0X laser sensor and an HC-SR04 ultrasonic sensor with a servo scanning mechanism. The monitoring system constantly scans the designated area (up to 180 degrees), comparing and cross-verifying the data from the two sensors. The scanning system enables it to identify obstructions and intrusions in the restricted zone and trigger an alarm. The Arduino circuit board utilizes an alarm circuit that employs a buzzer, light emitting diode, and liquid crystal display port to produce an alarm as a warning signal. Additionally, the microcontroller connects to a Bluetooth module (HC-05) to transmit warning signals to designated personnel via Bluetooth. The proposed design offers a working model based on readily available components. The prototype employs an innovative use of laser and ultrasonic sensors to provide a working model that is efficient and affordable. The system can provide security by detecting hazardous zones and triggering an alarm while also utilizing a directional scanning system to avoid false alarms.
Authors:
Fethi BOURAS, Nadjet GHERAISSA, Fouad KHALDI, Abdallah DOGGA
Abstract:
This study presents a comparative environmental assessment multi-fuels: Hydrogen (H₂), methane (CH₄), biogas, and natural gas under identical operating combustion conditions, using the Specific Emission Rate (SER), Carbon Footprint (CF), and Sustainability Index (SI) as key performance indicators. The results demonstrate that hydrogen achieves the best overall performance, combining a high thermal efficiency of approximately 78.6% with zero CO₂ emissions, positioning it as a highly attractive clean-energy carrier. However, this benefit is partially offset by hydrogen’s elevated NOx production. This drawback necessitates effective mitigation strategies, including combustion optimization or exhaust after-treatment. Conversely, despite being renewable, biogas exhibits higher CF and SER values due to its substantial CO₂ content, thereby diminishing its environmental advantage over hydrogen. Natural gas, employed as a reference, displays intermediate performance in both efficiency and emissions, offering a robust baseline. Collectively, the findings indicate that hydrogen (subject to adequate NOx control) and, to a lesser extent, biogas represent viable pathways toward cleaner energy systems. Nevertheless, their practical deployment requires a balanced consideration of emissions, efficiency, and sustainability, in line with evolving international energy policy frameworks.
Authors:
Fayrouz Difallah, Lyamine Guezouli
Abstract:
Federated learning allows distributed intrusion detectors to train without transmitting raw training records, but neither privacy guarantees nor communication savings follow automatically. This study evaluates federated LSTM and multilayer perceptron (MLP) classifiers on UNSW-NB15 as a generic network-security benchmark relevant to future Internet of Vehicles (IoV) studies. Ten simulated clients use the same 175,341 training records under IID and capacity-constrained binary label-skew partitions. Twenty runs cover two architectures, two partitions, and five random seeds, with 20 communication rounds and five local epochs per round. Attack-class F1 is 0:8630:003 and 0:863 0:005 for LSTM under IID and non-IID allocation, respectively; MLP achieves 0:899 0:004 and 0:882 0:012. The LSTM uses one record per input and does not exploit temporal sequences. Across 20 rounds, bidirectional float32 model exchange costs 34.629 MiB per client for LSTM and 2.051 MiB for MLP, compared with 2.609 MiB for one upload of numerical training features. The contribution is a controlled empirical comparison and explicit accounting of accuracy, heterogeneity, and model-payload costs. Vehicular validity, formal privacy, and energy savings remain unverified.
Authors:
Dario Antonelli, Yunhan Zhang, Antonio Maffei, Bo Yang
Abstract:
This paper discusses the development and implementation of a Cognitive Digital Twin (CDT) for an autonomous industrial mobile manipulator in a manufacturing environment. The CDT integrates sensor data (motion, camera), extended task descriptions, and Machine Learning (ML), to enhance perception, decision-making, and path planning capabilities. To address the unpredictability of battery life during industrial operation, a data-driven endurance prediction model is developed to estimate mission-level Remaining Useful Cycles (RUC). The proposed RUC metric provides a stable and task-consistent representation of residual operational capability under non-linear battery discharge behavior. By mapping cycle telemetry to RUC using CatBoost regression, the framework provides accurate and interpretable endurance prediction under the investigated operating conditions.
Authors:
Maria D. R-Moreno, Rafael Yllana, Jose L. Aguilar
Abstract:
The machine learning community has traditionally prioritized predictive performance, often overlooking the computational and environmental cost of the infrastructure on which models are executed. This work investigates whether highperformance computing (HPC) provides a favorable trade-off between predictive performance, execution time, and operational CO2 emissions compared with execution on a standard personal computer (PC). Across classification, time-series forecasting, and a real-world hospital admissions case study, we jointly evaluate model and infrastructure choices through predictive– environmental Pareto analysis. Results show that greater computational capacity does not necessarily translate into better predictive performance: PC configurations account for most Pareto-efficient solutions, while HPC frequently reduces execution time at the cost of substantially higher operational emissions. Moreover, model selection can have a greater impact on the predictive–environmental trade-off than infrastructure selection, as illustrated by the hospital case study, where Prophet-PC outperforms more computationally intensive alternatives while maintaining low operational emissions. These findings highlight the need to evaluate predictive quality, execution time, and environmental cost jointly when selecting both machine learning models and computing infrastructure.
Authors:
Abid Ali, Sari Bilal, CHOUABA Seif Eddine Allah
Abstract:
Smart microgrids with renewable energy sources are being deployed more widely, raising new challenges around cybersecurity and reliable operation. This paper proposes a machine learning-based framework for detecting and classifying cyber-attacks in a grid-connected PV–Battery Energy Storage System (BESS) microgrid. A MATLAB/Simulink model generates operating data under normal conditions and two attack scenarios: False Data Injection (FDI) and Delay Attack. The signals are processed in MATLAB Diagnostic Feature Designer, where relevant features are extracted and selected before classification. Several classifiers are evaluated in Classification Learner using 5- fold cross-validation. The proposed approach distinguishes normal operation from attacked conditions with high accuracy. Among the models tested, the selected classifier gives the best overall performance, suggesting that feature-based machine learning is a promising direction for this type of cyber-attack detection problem.
Authors:
Slimane FIZAZI, Mohamed El Amine GUERBAZI, Adel CHINE, Abdallah MOHAMMEDI, Ines CHABANI, Alaeddine ZEREG, Nadhir LEBAAL
Abstract:
Proton Exchange Membrane Fuel Cells (PEMFCs) are promising energy conversion systems for sustainable power applications; however, their performance and reliability are strongly influenced by flow-field geometry and manufacturing accuracy. This paper presents a comprehensive numerical and tolerance-based evaluation of a Single-Variable Tapered Channel (SVTC) flow-field design and its hydraulic and electrochemical performance. A threedimensional, steady-state, non-isothermal computational model incorporating coupled mass, momentum, species, energy, and electrochemical transport is developed and validated against experimental polarization data. Six taper ratios, defined by the short-to-long base length ratio (R = ℓ/L), are investigated to quantify their effects on pressure drop, net power density, and flow uniformity. Results indicate that increased tapering reduces total pressure losses by up to 21% while maintaining nearly constant gross power density, leading to an improvement of approximately 0.84% in net power density. Tolerance sensitivity analysis identifies the long base length as the most critical geometric parameter affecting performance and flow distribution.
Authors:
Fethi BOURAS, Abdallah DOGGA, Fouad KHALDI, Nadjet GHERAISSA
Abstract:
This study investigates the extraction of castor seed kernel oil and its subsequent conversion into biodiesel to evaluate its potential as a renewable alternative to conventional fossil fuels. The exploration of castor plants is motivated by two key factors: the widespread availability of castor trees in southern Algeria (El Oued region) and the high oil yield (~50 wt.%) of the seed kernels. Castor methyl ester (biodiesel) was synthesized via alkali-catalyzed transesterification with methanol, achieving a conversion yield of approximately 64%. Comprehensive physicochemical characterization was conducted for both the extracted oil and the produced biodiesel, evaluating kinematic viscosity, iodine value, cetane number, density, flash point, and Fourier-transform infrared (FTIR) spectroscopy to identify characteristic functional groups. The findings confirm that Algerian castor seed kernels constitute a viable, sustainable feedstock for biofuel production, supporting global efforts toward renewable energy transition.
Authors:
Sahli Zahir, Djabali Chabane, Sayah Samir
Abstract:
The Internet of Things (IoT) plays an important role in the efficient management of photovoltaic (PV) systems by enabling real-time data acquisition. In this work, an IoT-based platform using an ESP32 microcontroller is developed to collect electrical parameters, such as voltage and current, as well as environmental variables, including temperature and solar irradiance. The collected data are transmitted to a web-based application for real-time visualization and analysis. In parallel, the optimal sizing and placement of PV units in the real 112-bus radial distribution network are determined using the Reptile Search Algorithm (RSA), resulting in an improved voltage profile and reduced active power losses compared with other methods. The developed web application further facilitates the control and visualization of the various PV operating parameters.
Authors:
Abderraouf Bouakkaz, Mahmoud Bensaci, Nabil Talbi, Adel Lahsasna, Salim Haddad, Antonio J Gil Mena
Abstract:
Accurate photovoltaic (PV) power prediction is important for the planning, operation, and management of renewable energy systems, especially under changing weather and climatic conditions. However, the lack of long-term measured PV power data in many regions represents a big challenge for the development and evaluation of reliable prediction models. In this study, the performance of PV power prediction in different Algerian regions is investigated using machine learning techniques. The developed models are trained using three years of hourly meteorological data obtained from the National Solar Radiation Database (NSRDB) for ten locations distributed across different climatic zones in Algeria. The prediction performance is evaluated using statistical indicators such as the Root Mean Square Error (RMSE) and the coefficient of determination (R²). In addition, a comparative analysis is carried out to examine the impact of climatic diversity on the accuracy of the considered models. The obtained results show that the proposed models can accurately estimate PV power output for different locations. The results also highlight the influence of climatic conditions on PV generation characteristics and demonstrate the capability of machine learning techniques to capture regional variations in solar energy resources.
Authors:
Houssam Deboucha, Said Aissou, Elyazid Amirouche, Ali Berboucha, Kaci Ghedamsi, Bessam Deboucha, Saad Mekhilef
Abstract:
Reinforcement learning (RL)-optimized stealthy false data injection (FDI) attacks pose a critical cyber threat to modern cyber-physical microgrids, driven by the digitalisation of power systems that requires information exchange among power subsystems, by bypassing the security measures, observer residuals. In this work. A novel stealthy FDI attack detection framework based on denoising data-driven is proposed. First, Weak Sparse Identification of Nonlinear Dynamics (WSINDy) for state estimation with local test function is derived to reveal hidden physics violations and provide physics-informed residual vector. Second, low-pass filter is applied to extract the attack signature from high-frequency switching ripple, and finally, dual-mode robust statistical thresholding based on median absolute deviation (MAD) and adaptive sliding-window threshold to distinguish malicious data from a grid events. The proposed framework detection of stealthy FDI attacks is evaluated on simulated DC microgrid system under a sophisticated RL attack using MATLAB/ Simulink. The simulations results prove the framework’s effectiveness and resilience in detecting optimal RL stealthy FDI attacks, achieving a True Positive Rate (TPR) of 95.08% with a detection delay of 100ms and preventing false alarms. The proposed framework offers a promising solution against intelligent cyberattacks on microgrids.
Authors:
Anes OUALI, Mouloud ATMANI, Yani-Athmane BENNAI
Abstract:
Precision agriculture relies increasingly on continuous environmental data to support informed decision-making in the face of climate variability and resource constraints. The Internet of Things (IoT), combined with Low Power Wide Area Networks (LPWAN) technologies such as LoRaWAN (Long Range WAN), offers a promising approach to large-scale environmental monitoring in rural and urban environments. However, commercial LoRaWAN network providers often impose restrictive duty cycles, limited message volumes, and cloud-based data handling, raising concerns regarding cost, scalability, and data ownership for small-scale agricultural deployments. This paper presents the design and implementation of a low-cost, privately deployed IoT system for environmental data collection in precision agriculture, built around a single-channel LoRaWAN gateway and a fully open-source software stack. The system was deployed and operated end-to-end, continuously collecting, transmitting, decoding, storing, and visualizing temperature and humidity readings from two sensor nodes, while also capturing communication quality metrics for every received transmission. The results demonstrate the feasibility of building a reliable, self-hosted LoRaWAN infrastructure from low-cost components, offering a practical and economical alternative to commercial gateway solutions for small-scale precision agriculture applications.
Authors:
Hamimid Ilyes, Maaref Zineb, Lefkir Mira
Abstract:
Stable periodic-frequent pattern mining augments periodic-frequent discovery by tolerating bounded periodicity violations through the lability measure. In crime hotspot analysis, this stability view complements frequency-only maps by revealing areas that recur regularly at a weekly scale while remaining robust to occasional disruptions. However, fixed calendar partitions such as monthly slices suffer from boundary effects: periods are reset at segment edges, which can bias periodicity and lability estimates and prevent continuous evolution analysis. This paper proposes a sliding-window framework that mines stable periodicfrequent patterns on overlapping 90-day windows shifted every 10 days (maxPer = 7, maxLab = 10, minSupRatio = 30%), enabling continuous temporal evolution of periodic-frequent crime hotspots without a monthly reset. Using the obtained mining outputs on a real felonies/misdemeanors dataset (579 daily transactions, 22 areas), we (i) characterize persistent and emerging hotspots across 49 windows, (ii) describe output continuity using Jaccard similarity between consecutive windows, and (iii) analyze gradual periodicity transitions using single-run MaxLab levels. Results show an always-strict core hotspot (region 4) across all windows, a highly persistent near-strict hotspot (region 2), intermittent behavior for region 6 with a transient high-lability episode (MaxLab = 10), and a three-episode emergence of region 10 that stabilizes from high lability (MaxLab = 6) to strict periodicity (MaxLab = 0). Consecutive outputs have mean Jaccard 0.8714 (median 0.8661), influenced by window overlap. An 11-configuration sensitivity study examines the robustness of these findings. This supports interpretable monitoring for sustainable public-safety analytics.
Authors:
Mohamed El Amine GUERBAZI, Adel CHINE, Abdallah MOHAMMEDI, Riadh KHEBAOUET, Alaeddine ZEREG, Nadhir LEBAAL
Abstract:
The geometry of the CCS has a significant impact on the hydraulic behavior and electrochemical performance of proton exchange membrane fuel cells (PEMFCs). In this study, a three-dimensional computational fluid dynamics (CFD) analysis is conducted to investigate five different CCS geometries: rectangular, inverted triangular, inverted semicircular, inverted trapezoidal, and inverted semi-elliptical. The objective is to assess how these configurations affect pressure drop and the overall performance of the fuel cell. The numerical model was validated through mesh independence analysis and comparison with published experimental data, demonstrating good agreement. The hydraulic and electrochemical performances of the different configurations were assessed in terms of pressure drop, output power density, pumping power density, and net power density. The results show that the inverted triangular configuration achieved the highest output power density of 3068.80 W m⁻², but also produced the largest pressure drop (753.80 Pa), leading to increased pumping losses. In contrast, the rectangular configuration exhibited the lowest pressure drop (606.00 Pa) and pumping power density (78.10 W m⁻²), resulting in the highest net power density (2986.20 W m⁻²). These findings demonstrate that reducing pressure drop is essential for enhancing the overall efficiency of PEMFCs and identify the rectangular CCS as the most effective configuration among those investigated.
Authors:
Abid Ali, Sari Bilal, CHOUABA Seif Eddine Allah
Abstract:
Cyber-physical systems (CPS) constitute the vital infrastructure of modern technology, tightly coupling the cyber and physical domains. However, their increasing interconnectedness exposes them to sophisticated cyberattacks and operational failures, posing significant risks to human life and economic stability. Consequently, prompt and accurate anomaly detection has become a critical security imperative. This paper provides a comprehensive narrative review of recent advancements in artificial intelligence (AI) and machine learning (ML) methodologies designed for CPS anomaly detection from 2020 to 2026. By categorizing the literature into foundational and recent operational periods, we evaluate the performance of key architectures—including Autoencoders, CNNs, and Transformers—across standard industrial benchmarks such as SWaT and WADI, and Transformers across standard industrial benchmarks such as SWaT and WADI. The strengths and limitations of each approach are critically analyzed, with a specific focus on trade-offs between precision, recall, and real-time computational complexity. Finally, this review identifies critical open research gaps, such as explainability weaknesses, class imbalance, and the lack of robust cross-domain generalization, offering strategic directions to guide future research toward physics-informed and resilient CPS security frameworks.
Authors:
Z. Noumri, H. Bencherif, S. Kouda, Q. Aouni, E. Chebaki
Abstract:
Thin-film Cu₂ZnSnS₄ (CZTS) solar cells have attracted considerable attention owing to their earth-abundant, environmentally friendly, and low-cost constituent elements. In this work, an improved analytical model for CdS/CZTS thinfilm solar cells is developed by extending the conventional analytical approach through the incorporation of a single-diode current-voltage model together with bulk Shockley–Read–Hall (SRH) and surface/interface recombination mechanisms. The proposed model is first validated against experimental results, showing excellent agreement with the reported device performance, yielding Voc = 661 mV, Jsc = 19.41 mA/cm², FF = 65.95%, and η = 8.46%, compared with the experimental values of 661 mV, 19.5 mA/cm², 65.8%, and 8.4%, respectively. Following model validation, the effects of absorber thickness and acceptor concentration on the photovoltaic performance are systematically investigated. The simulation demonstrates that increasing the acceptor concentration significantly improves the open-circuit voltage and fill factor through enhanced junction electric field and reduced saturation current, whereas excessive absorber thickness increases carrier recombination, resulting in reduced current density and conversion efficiency. An optimized CdS/CZTS solar cell is obtained with Voc = 714 mV, Jsc = 22.71 mA/cm², FF = 66.61%, and η = 10.33%, corresponding to an efficiency improvement of 18% compared with the experimentally reported device. These results demonstrate the capability of the proposed analytical model as an efficient design and optimization tool for high-performance CZTS thin-film solar cells.
Authors:
Abdelmeraim Raziqa, Djallel Eddine Boubiche, Sabrina Boubiche
Abstract:
Machine learning-based intrusion detection systems (IDS) for Distributed Network Protocol 3 (DNP3) in Supervisory Control and Data Acquisition (SCADA) networks often report high accuracy, while overlooking class-level performance, data leakage, and model explainability. This paper presents a related reimplementation and protocol extension of IDM-DNP3 using the same public CICFlowMeter dataset. We compare XGBoost with a fully connected deep neural network under a leakage-safe evaluation protocol that excludes identifier features and incorporates entity-disjoint and leaveone- campaign-out testing. We also assess per-class performance and apply SHAP to interpret model predictions. XGBoost outperforms the neural network, achieving 98.93% accuracy and 0.9835 macro-F1 under random splitting, compared with 87.36% and 0.8411, respectively. Performance declines substantially under entity-disjoint and campaign-level evaluation, highlighting limited generalization to unseen traffic. These findings demonstrate the importance of leakageaware evaluation, class-level analysis, and explainability when assessing DNP3 intrusion detection for smart grid SCADA systems.
Authors:
Z. Noumri, H. Bencherif, S. Kouda, Q. Aouni, E. Chebaki
Abstract:
In this work, an analytical optimization of antireflection coatings (ARC) for perovskite solar cells is presented, focusing on minimizing optical losses over the spectral range of 300–1100 nm. A Transfer Matrix Method (TMM)-based model incorporating interference and absorption effects is employed to evaluate the impact of ARC configuration (single-layer, bi-layer, and double bi-layer), material selection, and thickness optimization on the average reflectance (%) . The study considers multiple transport layers, including SnO₂, WO₃, C60, and NiOx, combined with ARC materials such as SiO₂, MgF₂, Al₂O₃, TiO₂, and ZnO. Results demonstrate that optimized lowindex single-layer ARC structures, particularly MgF₂ and SiO₂, achieve near-zero reflectance ( ≈ .% ) across all configurations. While bi-layer designs provide comparable performance through graded refractive index profiles, double bi-layer structures do not yield consistent improvement and increase structural complexity. The findings highlight that precise thickness control and appropriate material selection are sufficient to achieve optimal anti-reflection performance, providing practical guidelines for efficient optical design in perovskite solar cells.
Authors:
Alaeddine ZEREG, Mounir aksas, Mohamed Taher bouzaher, Mohamed El Amine GUERBAZI GUERBAZI, ines chabani, Nadhir LEBAAL
Abstract:
This study investigates the influence of the helical pitch angle on the start-up behavior and aerodynamic performance of a Savonius wind turbine operating under the low-tomoderate wind conditions of urban and semi-urban areas. Four rotors with pitch angles of 0, 60, 90, and 120 are simulated at a free-stream velocity of 6 m/s using three-dimensional transient CFD in ANSYS Fluent, coupled with a six-degree-of-freedom solver that lets each rotor accelerate freely under the aerodynamic load. The numerical model is validated against measured free-running rotor speeds, with deviations below 7% between 3 and 6 m/s. The results show that the 90 rotor accelerates fastest during start-up, whereas the 60 rotor reaches the highest rotational speed at the end of the simulated period, 56% above that of the conventional rotor, and the highest power coefficient, about 3.2 times that of the conventional rotor. Pitch angles beyond 60 reduce the aerodynamic power, and the 120 rotor exhibits a marked performance loss. A moderate helical twist of about 60 therefore provides the best overall performance, while a twist of 90 favors a rapid start-up.
Authors:
Chemmami Abderrezaq, Guezouli Lyamine, Djallel Eddine Boubiche, Mohammed Islam Nouri, Houssem Eddine Hoggui, Smair Selsabi
Abstract:
Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated lesion localization can support screening workflows by providing spatial evidence for pathological findings rather than an image-level label alone. This paper presents the lesion-localization component of a larger DR analysis prototype: a standard YOLOv8s object detector trained, as a single-class lesion detector, on a manually annotated subset of the APTOS 2019 dataset at 1024×1024 pixel resolution. We describe the annotation protocol, the detector configuration, and the evaluation methodology, and report lesion-level and imagelevel performance, placed in the context of published YOLObased fundus lesion detectors. On a 64-image internal validation subset, the detector achieved a mAP@0.5 of 0.546, mAP@0.5:0.95 of 0.190, box precision of 0.633, and per-box recall of 0.539. In a binary image-level analysis (DR-positive defined as APTOS grade ≥1; 30 positive, 34 negative images), the detector flagged 30/30 positive and 1/34 negative images at a confidence threshold of 0.25 (sensitivity 1.00, 95% Wilson CI 0.89–1.00; specificity 0.97, CI 0.85–0.99). Because the same subset was used for early stopping and the class balance is far above screening prevalence, these figures are optimistic internalvalidation estimates, not evidence of clinical screening safety.
Authors:
Salah Eddine SAADI, Ziyad BOUCHAMA, Randa BABOUCHE
Abstract:
Achieving predefined-time convergence in step-up DC-DC converters without violating physical duty cycle constraints remains an open challenge due to non-minimum phase dynamics. This paper proposes a novel Adaptive Predefined-time Synergetic Control (PSC) strategy for a boost converter discharging a battery into a completely unknown load. To bypass non-minimum phase limitations, the synergetic macro-variable is formulated using a powerbalance and integral tracking error framework. Furthermore, an adaptive observer is designed to estimate the unknown grid load, with its convergence guaranteed via Lyapunov analysis. Crucially, by dynamically initializing the observer states, the initial macro-variable is forced to zero. This design completely eliminates the reaching phase, averts predefined-time peaking, and mathematically guarantees the duty cycle remains strictly bounded. Comparative MATLAB simulations validate the proposed controller’s rapid convergence, high tracking precision, and robust load disturbance rejection.
Authors:
Chaima Bensaci, Abderraouf Bouakkaz
Abstract:
Industrial air compressors are essential equipment in manufacturing systems, where unexpected component failures may increase maintenance costs, reduce operational reliability, and interrupt production processes. To improve condition monitoring and support predictive maintenance strategies, accurate fault diagnosis is required. This paper presents a machine learning framework for failure probability estimation and maintenance risk assessment of industrial air compressors using multi-sensor measurements. A Random Forest (RF) classifier is employed to performs fault classification of exhaust valve and estimates the probability of failure, which is subsequently converted into qualitative maintenance risk levels to facilitate maintenance decisionmaking. In addition, feature importance analysis is applied to identify the operating parameters that have the greatest influence on exhaust valve degradation. The obtained results demonstrate that the proposed framework provides accurate fault diagnosis achieved 100% accuracy, with an interpretable probability-based maintenance assessment, making it suitable for intelligent condition monitoring and predictive maintenance of industrial air compressors.
Authors:
Ismail Boucherit, Sid Ahmed Lachenani, and Hamza Kheddar
Abstract:
Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is essential for safe operation and condition-based maintenance. However, many existing methods provide only point estimates without quantifying prediction reliability, while random cycle-level data splits often produce overly optimistic results. This paper proposes a cross-battery RUL prediction framework that combines tree-based ensemble models with conformal uncertainty quantification. Dischargecycle data from NASA batteries B0005, B0006, B0007, and B0018 are used. The initial capacity is defined as the median of the first five discharge cycles, with state of health (SOH) computed relative to this reference. End of life is identified when three consecutive cycles reach an SOH of at most 80%. A total of 35 causal features are extracted from capacity, voltage, current, temperature, and degradation trends. Models are trained on B0005 and B0006, calibrated on B0007, and evaluated on the unseen B0018 battery. B0007 is used only for conformal calibration, with no model selection on the calibration or test battery. Conformalized quantile gradient boosting regression achieves the lowest MAE of 5.349 cycles and the lowest RMSE of 7.643 cycles, with R2 = 0:872. Robustness experiments are repeated over 30 random realizations per perturbation level. Although all nominal 90% conformal intervals cover the test observations, their wide ranges indicate conservative uncertainty estimates across batteries. Across 12 rotating train–calibrate–test assignments, learned-model mean MAE ranges from 20:783 +- 13:921 to 23:505 +- 16:447 cycles and mean PICP from 0.613 to 0.818. These results support useful unseen-cell prediction while revealing battery-dependent generalization and uncertainty transfer in this four-cell dataset. Index Terms—lithium-ion batteries, remaining useful life, state of health, cross-battery prediction, tree ensembles, conformal prediction, uncertainty quantification, battery prognostics
Authors:
Saifi Salma, Sellami Mohammed
Abstract:
The recent adoption of edge artificial intelligence (Edge AI) in mobile and embedded systems has increased the demand for computationally efficient and sustainable AI solutions, shifting toward deep learning models that deliver high performance while remaining efficient in computation and memory usage. Lightweight convolutional neural networks offer an effective balance between accuracy and computational cost and are designed for edge deployment. Post-training quantization (PTQ) has emerged as a practical model compression technique; however, its influence on deploymentoriented characteristics, including inference time and robustness, remains insufficiently explored for lightweight CNNs. This study presents a comparative analysis of the MobileNetV3 and EfficientNet-B0 lightweight CNN architectures, using INT8 post-training dynamic quantization for resource-constrained inference on the CIFAR-10 dataset. The proposed quantization method is evaluated in terms of accuracy, model size, inference latency, and robustness to Gaussian Blur, Brightness variation, and Contrast Shift to examine the trade-offs between predictive performance and sustainable and efficient AI deployment on resourceconstrained edge devices.
Authors:
Rafik Dembri, Mourad Kezai, Fouaz Ghedjati
Abstract:
The increasing penetration of converterinterfaced renewable energy sources has reduced the inertia of modern smart grids, making frequency and voltage regulation increasingly challenging. This paper investigates the capability of a Doubly-Fed Induction Generator (DFIG)-based wind energy conversion system to support grid stability through coordinated active- and reactive-power control. Frequency support is achieved through de-loaded operation and controlled rotor kinetic-energy release, while voltage support is provided by coordinated reactive-power injection from the Rotor-Side Converter (RSC) and Grid-Side Converter (GSC) within the modeled operating constraints. Three simulation scenarios evaluate frequency regulation, PCC voltage support, and the coordinated operation of the DFIG with conventional synchronous generators. The results show improved frequency recovery, reduced voltage deviations, and reduced transient support requirements from conventional generation. These findings demonstrate the potential of DFIG-based wind turbines as active grid-supporting resources for enhancing the stability and resilience of future smart grids.
Authors:
Zahra Meziani, Mourad Rahmani, Zohir Dibi
Abstract:
In the context of sustainable and intelligent energy systems, flexible solar cells are of particular interest due to their small dimensions, lightweight nature, and ease of transportation and integration into various applications. These cells predominantly employ transparent electrodes (TE) based on indium tin oxide (ITO). However, the high indium content of ITO contributes to the increased cost of these electrodes, thereby motivating research into alternative materials such as graphene. In a first study, we developed a theoretical model of a graphene-based transparent electrode (GTE) to assess its potential as an alternative to conventional ITO-based electrodes. In a second study, we theoretically investigated the behavior of the GTE in a Schottky junction solar cell (SJSC). These two studies provide the theoretical basis for the present work, in which we extend this approach to the modeling of a flexible graphene silicon-based solar cell (FG/Si SC). This article proposes a simulation model using artificial neural networks (ANN). The behavior of the FG/Si SC was interpreted using the ANN model, allowing us to predict the dark J-V curve of the FG/Si SC. From this curve, we determined the profile parameters, and subsequently we can calculate the quality parameters using the appropriate equations, highlighting optimal values for silicon substrate doping (ND) and thickness (tSi), as well as the work function of graphene (WF) and its number of layer (N). The resulting device parameters were then used to evaluate J-V curve simulations under illumination for twelve various structures, spotlighting the quality parameters of each structure: short-circuit current density (JSC), open-circuit voltage (VOC), fill factor (FF) and efficiency (n). We identified an optimal structure with the quality parameters set at: (JSC, VOC, FF, n) = (23.02 mA/cm², 0.50 V, 51.09%, 5.88%). These findings underscore the potential of FG/Si SC as a promising contender in the field of flexible solar cell.
Authors:
Abderahmane LOUIZINI, Abdelkader GARMAT, Kamel GUESMI
Abstract:
Path tracking in mobile robotics often comes down to one stubborn problem: fixed-lookahead controllers struggle to handle both tight corners and open stretches simultaneously. This paper addresses that gap with the Curvature-Adaptive Pure Pursuit (CAPP) controller, designed for a differential-drive robot guided by an A* global planner. Rather than committing to a single lookahead distance, CAPP reads the local path shape using the Menger curvature formula and adjusts both the lookahead distance and the robot’s speed on the fly. When the path bends sharply, the robot slows and looks closer ahead; on open straights, it extends its horizon and moves faster. The system runs over an occupancy grid and triggers replanning automatically when obstacles enter the robot’s vicinity. Unlike curvature-adaptive methods designed for high-speed autonomous vehicles, which rely on complex dynamic modeling, offline system identification, or computationally expensive model predictive control (MPC), CAPP employs a purely analytical, closed-form adaptation law. This eliminates the need for offline tuning and reduces computational overhead to under 4 ms per cycle. Testing in PyBullet across three navigation scenarios showed that CAPP achieves a 28% cross-track error reduction compared to standard Pure Pursuit, a difference confirmed by a two-sample Welch’s t-test, responds quickly to replanning demands, and completes every run without a single collision.
Authors:
Dikra-Nadjah Chouchane, Saïd Drid, Samir Bouslimani
Abstract:
This paper presents a hybrid PV/battery system that guarantees the continuous power delivery to a resistive load under variable solar irradiation conditions. A Maximum Power Point Tracking (MPPT) algorithm is used to extract maximum power from PV panels and a battery storage unit takes care of the difference between generation and demand by charging and discharging operations. For simplicity, the load is modelled as an equivalent resistive element rather than the detailed PEM electrolyzer model. The system is implemented and simulated in MATLAB/Simulink under a dynamic irradiation profile, to simulate a sudden cloud passage, so that the dynamic response, power balance and battery State of Charge (SOC) behaviour can be evaluated. The results show that the power from the PV source, battery and load are well coordinated under varying irradiation conditions.
Authors:
Oussama SEMASSEL, Raziqa ABDELMERAIM, Mohamed El Amine GUERBAZI , Djallel Eddine BOUBICHE, Samir ATHMANI
Abstract:
Hydrogen fuel cell electric vehicles (FCEVs) represent a promising alternative for sustainable transportation due to their high energy efficiency, rapid refueling, and zero tailpipe emissions. However, their development and large-scale deployment remain constrained by fuel cell performance, hydrogen consumption, component degradation, and energy management under varying driving conditions. Recent advances in artificial intelligence (AI) have opened new perspectives for addressing these challenges through data-driven prediction, optimization, and control. This paper provides an overview of AI applications in hydrogen fuel cell vehicles, with particular emphasis on fuel cell performance prediction, energy management, thermal and water management, degradation assessment, and fault diagnosis. Machine learning and deep learning methods are examined in relation to their ability to estimate fuel cell voltage and power output, optimize operating parameters, and predict component lifetime. Their application to hybrid energy management is also discussed, particularly in coordinating power demand between the fuel cell stack and battery system under dynamic driving cycles. In addition, current limitations related to data availability, model reliability, computational cost, and real-time implementation are addressed. The review outlines the potential of AI-assisted approaches to improve fuel cell efficiency, system durability, and operational reliability, while identifying research needs for their effective integration into hydrogen-powered transportation.
Authors:
Mohamed El Amine GUERBAZI, Abdallah MOHAMMEDI, Adel CHINE, Abir BOUCETTA, Youcef SAHLI, Hichem BENCHERIF
Abstract:
Flow-channel geometry significantly affects the electrochemical and hydraulic performance of proton exchange membrane fuel cells (PEMFCs). This study presents a three-dimensional computational fluid dynamics (CFD) analysis of three configurations: the Reference Straight Channel (RSC), Central Constriction Channel (CCC), and Central Expansion Channel (CEC). Simulations were performed using ANSYS Fluent under steady-state, non-isothermal, single-phase conditions. The configurations were compared in terms of polarization behavior, power density, pressure drop, net power density, and the distributions of hydrogen, oxygen, water vapor, and temperature. Mesh independence was examined, and the model predictions were compared with published numerical and experimental results, showing deviations of approximately 10.5% and 13.3%, respectively. The results indicate that the CCC configuration achieves the highest power density, approximately 0.310 W/cm², and a net power density of 0.306 W/cm², with a pressure drop of 363.73 Pa. In contrast, the CEC configuration exhibits a pressure drop of 2631.69 Pa and a net power density of 0.221 W/cm². The RSC configuration reaches a net power density of 0.298 W/cm². The findings demonstrate that channel geometry influences reactant distribution and hydraulic losses. Under the investigated conditions, the CCC design provides the best balance between electrical output and pumping requirements.
Authors:
Montassir Aberkane, Ilyes Saadna, Abdesselem Beghriche, Djallel Eddine Boubiche
Abstract:
Anticipating lane changes is important for proac-tive motion planning in automated driving. This paper investi-gates lane-change intention prediction at multiple prediction horizons using raw NGSIM trajectories from two freeways. A vehicle-grouped evaluation framework is developed using causal trajectory features and leakage-aware training and validation procedures. The results show that XGBoost provides strong pre-diction performance across horizons and retains useful discrimi-nation across traffic periods and a second freeway. Predictive cues also shift from lateral dynamics near the lane-boundary crossing to adjacent-lane gaps at earlier horizons. A central find-ing is that non-causal trajectory smoothing can introduce sub-stantial look-ahead bias, with offline processing increasing 2-s F1 from about 0.72 to 0.99 on both freeways. This finding highlights the importance of causal feature computation and explicit report-ing of trajectory-processing procedures in LCIP studies. Calibra-tion and conformal prediction further support the assessment of probability reliability and prediction uncertainty.
Authors:
Nousseiba GUIDOUM
Abstract:
Manual fruit sorting is slow, costly, and prone to human error, which is a real problem in post-harvest facilities that must inspect large volumes of produce every day. This paper compares three deep learning models : a conventional Convolutional Neural Network (CNN), MobileNetV2, and EfficientNetB0, for the automatic classification of fresh and rotten fruits. The models were trained and tested on a public dataset with six classes (fresh apples, fresh bananas, fresh oranges, rotten apples, rotten bananas, and rotten oranges), using 10,901 training images and 2,698 test images resized to 224×224 pixels. Performance was measured with accuracy, precision, recall, F1-score, and confusion matrices. Additionally, the number of trainable parameters was used as an indicator of computational cost. The CNN reached a test accuracy of 95.59%, MobileNetV2 reached 99.41%, and EfficientNetB0 reached 99.85%. EfficientNetB0 achieved the highest observed accuracy in this single-run experiment, while MobileNetV2 achieved a comparable result with about 78% fewer total parameters than the CNN and 42% fewer than EfficientNetB0. Both transfer models have the same number of trainable parameters. MobileNetV2 is therefore a promising candidate for low-cost sorting equipment, pending direct measurements of inference latency and energy consumption. These results provide an initial empirical baseline for lightweight, pretrained models in automated fruit quality inspection.
Authors:
Cheyma ARIBI, Ali AILLANE, Kamel GUESMI, Ahmed BENDIB
Abstract:
This paper proposes a Dual Integral Synergetic Direct Power Control (DIS-DPC) strategy, optimized via the Mirage Search Optimization (MSO) algorithm, to address the steady-state power ripple, variable switching frequency, and parameter sensitivity inherent in conventional Direct Power Control for DFIG-based systems. The controller is free of an internal current loop and removes the steady-state error and the cross-coupling by embedding a dual integral macro-variable into a continuous synergetic control law. The comparative study carried out with relevant approaches and under different wind speeds and parametric uncertainties shows that the proposed approach achieves better tracking dynamic performance. In addition, the proposed approach also significantly reduces the active and reactive power ripple compared with the classical and relevant techniques.
Authors:
Khadidja Ait Meddour, Abdelkarim Ait Meddour, Dalila Kessira, Karim Akilal
Abstract:
Sustainable artificial intelligence is becoming increasingly important for precision agriculture, where crop stress detection must run reliably on resource-constrained field systems. Most existing approaches rely on deep learning models that are computationally heavy, data-hungry, and difficult to interpret, limiting their deployment on Unmanned Aerial Vehicles (UAV) edge hardware. This paper proposes a lightweight, trainingefficient framework for soybean disease classification from aerial canopy imagery, contributing to sustainable computer vision by offering a computationally efficient alternative to conventional deep network pipelines. The method uses two HSV color masks, with Hue boundaries optimized through a Firefly-based search, to classify canopy images as healthy, rust, or mosaic. The non-zero pixel count under each mask forms a two-dimensional feature vector, and a multinomial logistic regression model, trained across all three classes together, predicts the final class. Evaluated on the MH-SoyaHealthVision UAV dataset, the proposed framework achieves a test accuracy of 97.08% with a weighted-averaged F1- score of 0.97. Compared with classical texture-based baselines (GLCM and HOG, each combined with logistic regression) and a fine-tuned MobileNetV2 model, the proposed framework achieves higher accuracy than both classical baselines. Compared with MobileNetV2, it has a model size approximately 11,776 times smaller, while its accuracy is only 2.92% lower. Its low computational requirements may contribute to reduced energy consumption, supporting the sustainability of UAV-based fieldmonitoring systems.
Authors:
Omar Bait
Abstract:
Cooking food is one of the daily basis practices and requirements of all human beings. Mainly, edible food is automatically prepared using conventional resources, such as non–solid fuels of natural gas or LPG for making. At present, although cooking has taken another way to solar heat utilization, still the problem of solar energy fluctuation, which hampers the accomplishment of cooking items in time. This investigation suggests incorporate which called nano–enhanced phase change material, namely paraffin (RT–35HC)–based system. Thus, the chosen phase change media (PCM) was improved using metallic nanoparticles of aluminum (Al). The latter owns fascinating properties when it is mixed with the PCM at a certain concentration to observe the thermal behavior of the mixture and its effect on the cooker function. Results are based on showing several parameters related to figures of merits (F1 and F2) and energy efficiency, ηth. It was found that the cooker at stagnation test reveals a first figure of merit of ~0.10 °C m2/W, while the other second one was as ~0.191 °C m2/W (at sensible test), when the cooker is charged with nano–PCM. Likewise, the thermal efficiency has also been forecasted to be about 14% for the solar cooker device incorporating nano–PCM.
Authors:
CHENNA Abdelbassette, Samir Athmani, Djallel Eddine Boubiche, Mohamed Houache
Abstract:
A battery setpoint must lie inside a feasible interval fixed by the state of charge, the converter rating and the import limit. Where the interval is slack, the optimal command is one subtraction: the net power imbalance. Where it binds, the optimum lies at an end of the interval instead. This paper measures whether a large language model (LLM) finds the second. One open-weight 7 B model, CPU-served, was asked for the setpoint on 96 dwelling-scale states under six specifications of the task. Half the states were drawn so that the net-power balance cannot be optimal. On those 48, none of the four specifications that ask only for an answer puts a command on the optimum once the sign is counted. Printing an interval moves feasibility, but not the median error. A fifth prints the true interval and is exact on 2 of 48. A sixth also asks the model to compute the imbalance, the endpoints and the cost of each candidate, and it splits the draw in two. Where the answer is a single subtraction it is exact on 44 of 48, with sign agreement 47 of 48. Where the answer lies at a bound it is exact on 11 of 48, and all eleven sit at that bound. The model performs the subtraction when asked, and still does not perform the projection. A model-free rule performs both on all 96 states, with no model call. Respecting a feasible set is not computing the optimum inside it.
Authors:
Abdelmadjid Benimam, Ahmed Cherif Mazari, El bahi Toubal
Abstract:
Plant disease detection is often dominated by image classification, where a disease class is assigned based on feature analysis. Segmenting a region in a leaf image is more precise and provides the explainability that classification might miss. We investigate how knowledge distillation affects smaller deployable models such as TinyVit and YOLOv8, using a small variant of the DINOv3 ViT-S/16 model as teacher. We conduct a two-stage benchmark on the PlantSeg dataset and the Kaggle leaf disease dataset. Results showed that distillation provided a consistent improvement for TinyVit and a marginal improvement for YOLOv8 on both datasets under intra-dataset evaluation. However, improvements in cross-dataset settings were not conclusive for either student model. We close with concrete guidance on which deployable model to choose and when distilling DINOv3 earns its extra training stage.
Authors:
Ibtihel Rezaiguia, Habiba Drias, Naila Aziza Houacine
Abstract:
Wildfire risk prediction in fire-prone Mediterranean regions is hindered by the scarcity of comprehensive, ready-touse spatio-temporal datasets integrating environmental, climatic, and anthropogenic drivers of fire occurrence. This paper presents a reproducible spatio-temporal dataset for wildfire risk modeling in the Wilaya of Tizi Ouzou, Algeria, combining static variables (topography, land cover, soil properties, population, historical fire frequency) with dynamic daily variables (NDVI, soil moisture, meteorological conditions, Fire Weather Index components, and FIRMS fire detections) over a uniform 3 km × 3 km grid spanning 2010–2025. A tailored spatio-temporal gap-filling strategy yields a complete dataset of over 2.2 million observations with no missing values. Diagnostic analysis reveals strong seasonal and interannual fire patterns consistent with the Mediterranean fire regime, with Fire Weather Index components emerging as the most discriminative predictors. This dataset provides a robust, standardized foundation for developing and evaluating predictive wildfire risk models in a region where such resources remain scarce.
Authors:
Nadjib Titaouine, Abdelhadi Ouazene, Mohamed Amir Berghouti, Djallel Eddine Boubiche
Abstract:
Physics-informed neural networks (PINNs) offer a promising approach to traffic-flow modeling and state estimation, but often struggle to capture sharp traffic shock waves. This study investigates how optimization strategies affect PINN accuracy on the Poisson, Heat, and Burgers equations. Warm-started L-BFGS consistently outperforms first-order optimizers, while adaptive loss weighting (AW), cosine learning-rate annealing, and residual-adaptive resampling (RAR) improve first-order training on smooth problems but degrade performance on Burgers. To address this limitation, we propose a staged hybrid optimizer that applies AW and RAR between L-BFGS stages with a frozen objective, curvature resets, and rollback safeguards. The method reduces error on Burgers while maintaining comparable performance on Poisson and Heat. On a viscous Lighthill–Whitham–Richards (LWR) model with Greenshields flux, it also improves the prediction of upstream-moving congestion fronts. These findings highlight the importance of scheduling adaptive interventions appropriately when training PINNs for problems involving sharp solution gradients.
Authors:
AISSA GHADDAB, AMRANI HICHEM, KHALLAF BOUZGOU, Beddiaf Raouf, FEDAOUI KAMEL
Abstract:
This work investigates the structural, electronic, mechanical, thermodynamic, and optical properties of Cs2KXH6 (X = Al, Ga, In) double-perovskite hydrides using first-principles density functional theory (DFT). All three compounds adopt a cubic structure with the Fm-3m space group and exhibit favorable formation energetics. Their phonon spectra show no imaginary frequencies, while their elastic constants satisfy the Born stability criteria, supporting their dynamical and mechanical stability. Electronic-structure calculations using PBE and HSE06 reveal semiconducting behavior, and optical calculations indicate strong absorption in the ultraviolet region. The calculated gravimetric hydrogen-storage capacities are 1.789, 1.588, and 1.420 wt% for Cs2KAlH6, Cs2KGaH6, and Cs2KInH6, respectively. These results provide a theoretical assessment of their stability and functional properties while indicating that further refinements are needed to enhance their hydrogen-storage capacities.
Authors:
Imad MAHAYA, Ali BEHLOUL, Khaled HAMOUID
Abstract:
Differential privacy in federated learning is evaluated along two axes, privacy and utility. On battery-powered edge devices a third decides whether a configuration is deployable at all, and it is not reported.We measure the energy cost of examplelevel differential privacy applied locally at each client, with NVML hardware counters differenced around each federated round, sweeping the privacy budget on UCI-HAR, whose clients are the individual study participants, and on CIFAR-10. The cost separates in two. A fixed per-round overhead from persample gradients and clipping costs +51% on HAR and +34% on CIFAR-10, with no dependence on ” that a permutation test over seeds resolves (p = 0:27 and p = 0:79); everything budgetdependent acts through the number of rounds required and is the larger term. Measured as energy-to-target-accuracy, privacy is free at 0:40 accuracy on HAR for e +- 1, 5:7 +- at 0:70, and unbounded at 0:85, which no budget we tested reaches at any expenditure. Because the cost is paid in joules it inherits the local carbon intensity: the premium attributable to privacy alone is 106–148 gCO2eq per 1000 runs on Algeria’s grid against 5:7–7:9 g in Norway, a quantity the privacy-utility framing cannot express. We also report a measurement hazard: running the privacy accountant’s noise calibration inside the timed loop reproduced an entire apparent e trend.
Authors:
Badis Lekouaghet, Fathi Brioua
Abstract:
Reliable battery operation near the lower stateof- charge (SOC) limit is essential in electric vehicles (EVs), where inaccurate voltage prediction can cause premature power restriction, reduced usable energy, or unsafe discharge. The accuracy of a battery model depends on the accurate identification of the electrical and open-circuit-voltage parameters, which becomes more difficult at low SOC. This region may have low accuracy and consistency for the conventional methods while the metaheuristic optimisation (MO) is a good alternative to solve complex nonlinear and multimodal problems. This paper explores the use of Dream Optimisation Algorithm (DOA) for parameter identification of a second-order equivalent circuit model (ECM) under low-SOC operating conditions. Experimental data from a High Dynamic Profile (HDP) used to evaluate the model under rapidly varying current conditions. DOA’s performance is compared with two recent MOs reported for ECM parameter identification, Osprey Optimisation Algorithm (OOA) and FATA Morgana Algorithm (FATA), and with the popular Grey Wolf Optimiser (GWO) as a classical benchmark. A detailed statistical assessment is carried out considering the voltage estimation accuracy, repeatability analysis, box-plot distributions and convergence behaviour. The results show that DOA gives the best voltage reconstruction and improves numerical stability and convergence rate towards good parameter solutions.
Authors:
Rima Hadjadj, Mohammed Boulesbaa
Abstract:
This paper presents the design and numerical analysis of a two-dimensional photonic crystal (PhC) temperature sensor based on germanium (Ge). The proposed structure consists of a periodic array of circular air holes embedded in a Ge background and incorporates a defect cavity coupled to input and output waveguides. The photonic band structure is analyzed using the plane-wave expansion (PWE) method, while the temperature dependent spectral response is investigated using the finitedifference time-domain (FDTD) method. Temperature sensing is achieved through the thermo-optical response of germanium, which modifies the effective refractive index and consequently shifts the cavity resonance wavelength. Temperature sensitivity and refractive-index sensitivity are defined and evaluated separately. Numerical convergence and comparison with previously reported photonic-crystal temperature sensors are also considered. The results indicate that the proposed Ge-based photonic-crystal cavity has potential for compact integrated optical temperature sensing.
Authors:
Mourad Meguellati, samir bouslimani, Amira Douaa Hamidatou, Aicha Abdelatif
Abstract:
This project presents the design and implementation of a laboratory-scale solar simulator for photovoltaic (PV) module testing under controlled environmental conditions. The system reproduces indoor solar irradiance using an adjustable artificial light source, enabling electrical characterization independently of outdoor weather conditions. A monitoring and data acquisition system measures voltage, current, and output power, automatically generating current–voltage (I–V) and power–voltage (P–V) curves. Experimental results demonstrate the simulator’s practicality and repeatability in evaluating PV performance. By minimizing environmental fluctuations, the platform improves measurement consistency and reliability. Its flexible, cost-effective design supports photovoltaic research and education, with potential for improved irradiance control and automated testing.
Authors:
MENANI Karima, MEDJBEUR Lemia, ACHOURI Fouzia
Abstract:
The problem of closed-loop control of nonlinear mobile robot systems is investigated in this paper. Our main contribution lies in the control of these systems, A nonlinear mathematical model of the mobile robot is developed, and a Terminal Synergetic Control (TSC) strategy is proposed to achieve accurate and robust tracking of desired trajectories. The proposed controller combines the robustness of synergetic control with the fast convergence characteristics of terminal control, providing improved tracking performance in the presence of system nonlinearities and disturbances. The stability and convergence of the closed-loop system are rigorously analyzed using Lyapunov stability theory. Numerical simulations are performed to evaluate the effectiveness of the proposed control strategy. The results demonstrate that the TSC method ensures accurate trajectory tracking, fast convergence of tracking errors, and robust closed-loop performance. These results confirm the potential of the proposed approach for high-performance control of nonlinear mobile robotic systems.