This paper presents a framework that combines large language models (LLMs) with a multi-agent system (MAS) to automate the translation of natural language instructions into executable network simulations. Designed to improve the usability of traditional simulators, the proposed system enables users, regardless of technical background, to interact with Mininet through intuitive language prompts. The MAS architecture assigns specific tasks to agents, including input parsing, topology analysis, code generation, execution, and result interpretation. A retrieval-augmented generation (RAG) module boosts contextual understanding by accessing authoritative documentation. The framework was tested on benchmark scenarios of increasing complexity, such as connectivity checks, routing optimization, and DDoS mitigation. Results show that LLMs are effective at interpreting intent and refining prompts, while code generation, especially for complex tasks, remains a challenge. Larger models performed better in accuracy and robustness, whereas smaller ones struggled with error handling and refinement. These findings highlight both the promise and current limitations of agentic AI in network simulation, positioning the system as a foundation for more intelligent and accessible tools in education, research, and infrastructure management.

Laser powder bed fusion (L-PBF) is the most popular Additive Manufacturing (AM) process for metals. It builds a 3D object layer-by-layer, by spreading metal powder on top of the previous layer and selectively melting it with a laser. Despite its many advantages, large-scale production may be hampered by the large number of process parameters and the challenges associated with their optimization. We propose an automated parameter selection approach based on process signatures extracted from a parameterized simulation of the process. Specifically, we outline a rapid data-driven simulation method based on Physics-Informed Neural Network (PINN). This approach involves training a neural network to solve the partial differential equation describing the process at varying values of a parameter of interest (for example, the laser power), thus eliminating the need for repeated Finite Elements Method (FEM) simulations. Our preliminary experiments demonstrate the feasibility of our approach.

Wave Energy is one of the renewable sources with greatest potential. Since power coming from waves fluctuates, the grid integration of wave energy involves several power conditioning stages to comply with grid quality requirements. However, to ensure full integration of wave energy in a smart grid scenario and unlock advanced monitoring and control techniques (e.g. Demand/Response), it is crucial to forecast the output power. This work proposes a methodology to forecast in short-term horizons (i.e. 15 min to 240 min) the power delivered to the grid of the Inertial Sea Wave Energy Converter (ISWEC), a device that harnesses wave power through the inertial effect of a gyroscope. Therefore, we designed, optimized and compared the performance of five known machine learning techniques for time series point forecasting: Random Forest, Support Vector Regression, Long Short-Term Memory Neural Network, Transformer Neural Network and 1 Dimensional Convolutional Neural Network. Additionally, we studied the efficacy of downsampling technique aggregating original dataset sampled every 0.1 s in time steps of 1min, 3min, 5min and 15min to compare the performance behaviour of the different machine learning models for these datasets. Furthermore, we implemented Prediction Intervals (PIs) to calculate the inherent uncertainties associated with the previously mentioned machine learning techniques. These PIs were built based on the Non-Parametric Kernel Density Estimator technique. The point forecasting and the PIs results showed that models’ performance improved as the downsampling increased. Moreover, the Random Forest model was the worst-performing in all cases. Finally, none of the other models can be considered the best overall.

With the growth of tourism industry, airplanes have became an affordable choice for medium- and long-distance travels. Accurate forecasting of flights tickets helps the aviation industry to match demand, supply flexibly and optimize aviation resources. Airline companies use dynamic pricing strategies to determine the price of airline tickets to maximize profits. Passengers want to purchase tickets at the lowest selling price for the flight of their choice. However, airline tickets are a special commodity that is time-sensitive and scarce, and the price of airline tickets is affected by various factors.Our research work provides a systematic comparison of various traditional machine learning methods (i.e., Ridge Regression, Lasso Regression, K-Nearest Neighbor, Decision Tree, XGBoost, Random Forest) and deep learning methods (e.g., Fully Connected Networks, Convolutional Neural Networks, Transformer) to address the problem of airfare prediction, by keeping the consumers’ needs. Moreover, we proposed innovative Bayesian neural networks, which represent the first exploitation attempt of Bayesian Inference for the airfare prediction task, to the best of our knowledge. Therefore, we evaluate the performance of our implemented and optimized models on an open dataset. The experimental results show that deep learning-based methods achieve better results on average than traditional ones, while Bayesian neural networks can achieve better performance among the other machine learning methods. However, taking into account both prediction performance and computational time, the Random Forest turns out to be the best choice to apply in this scenario.

In recent times, the consistent growth of wave energy makes it one of the most promising forms of renewable energy. Due to the intermittency and non-stationary nature of waves, the grid integration of these renewable energy sources involves a series of complex power conditioning stages to deliver grid electric power that meets the corresponding quality standards. Furthermore, to enable optimal management and operation of a smart grid power system, forecasting the wave power delivered to the grid is essential. In this paper, we present a novel approach based on Long Short-Term Memory Neural Network to forecast the wave power delivered to the grid of a Wave Energy Converter (WEC) - the ISWEC, which is a device able to harvest sea energy by exploiting the inertial effect of a gyroscope - in short-time horizons (e.g. 1min). The data for the analysis was obtained from a simulator that combines a model of the ISWEC device and the power conditioning grid integration for this particular WEC. In addition, to investigate the effectiveness of downsampling, we compared the performance behavior of the raw dataset and downsampled versions of it. The results showed that as the downsampling increases, so does the forecasting accuracy: the forecasting performance of the raw dataset returned the worst results, while the one of the dataset with the biggest downsampling studied returned the best.

Electric vehicles (EVs) have been globally recognized as a reliable alternative to fossil fuel vehicles. The core component of an electric vehicle is its rechargeable battery pack. However, there still needs to be large-scale publicly available EV data to investigate and distribute effective solutions to monitor the conditions of the EV’s battery pack. Hence, we propose an EV simulator that generates EV battery pack internal signals starting from the input driving cycle. The simulated data resemble the behavior of a multi-cell EV battery pack undergoing the user’s utilization of the EV. The simulated data include vehicle speed, voltage, current, State of Charge (SOC), and internal temperature of the battery pack. The virtual-EV model simulator, including the battery pack subsystem, has been tuned using real-world EV data-sheet information. The battery pack embeds thermal and aging models for further realism, influencing the output signals given the environmental temperature and the battery’s State of Health (SOH). The data generated by the virtual EV simulator have been validated with real EV data signals sampled by an equivalent real-world EV. The data comparison yields a minimum R2value of 0.94 and a Root Mean Squared Error not higher than 2.74V for the battery pack’s voltage and SOC, respectively.

Li-ion batteries have become the most important technology for electric mobility. One of the most pressing chal-lenges is the development of reliable methods for battery state-of-health (SOH) diagnosis and estimation of remaining useful life. In electric mobility scenario, battery capacity degradation prediction is crucial to ensure service availability and life duration. This research work provides a comprehensive comparative analysis of neural networks for a data-driven approach suitable for SOH estimation on single cells, stressed under laboratory conditions. For this purpose, different neural networks (i.e., LSTM, GRU, 1D-CNN, CNN-LSTM) are trained and optimized on NASA Randomized Battery Usage dataset. Experimental results demonstrate that data-driven neural networks generally performed well SOH estimation on single cells. In detail, the 1D-CNN best predicts SOH and has the lowest variance in the output. The LSTM have the highest variance in estimating SOH, while GRU and CNN-LSTM tend to overestimate and underestimate the value of SOH, respectively.

In recent years, the research about energy waste and CO2 emission reduction has gained a strong momentum, also pushed by European and national funding initiatives. The main purpose of this large effort is to reduce the effects of greenhouse emission, climate change to head for a sustainable society. In this scenario, Information and Communication Technologies (ICT) play a key role. From one side, advances in physical and environmental information sensing, communication and processing, enabled the monitoring of energy behaviour of buildings in real-time. The access to this information has been made easy and ubiquitous thank to Internet-of-Things (IoT) devices and protocols. From the other side, the creation of digital repositories of buildings and districts (i.e. Building Information Models - BIM) enabled the development of complex and rich energy models that can be used for simulation and prediction purposes. As such, an opportunity is emerging of mixing these two information categories to either create better models and to detect unwanted or inefficient energy behaviours. In this paper, we present a software architecture for management and simulation of energy behaviours in buildings that integrates heterogeneous data such as BIM, IoT, GIS (Geographical Information System) and meteorological services. This integration allows: i) (near-) real-time visualisation of energy consumption information in the building context and ii) building performance evaluation through energy modelling and simulation exploiting data from the field and real weather conditions. Finally, we discuss the experimental results obtained in a real-world case-study.