Key Collaborations
Università degli Studi del Sannio
Our ongoing collaboration with UniSannio focuses on developing advanced machine learning models for electrical grid optimization and renewable energy integration in urban environments.
TRAIL Ecosystem
As a proud member of the TRAIL (Trusted AI Labs) Ecosystem, we collaborate with leading Belgian AI research groups to develop responsible and innovative machine learning solutions for sustainable energy systems.
Areas of Expertise
Renewable Energy Forecasting
Developing state-of-the-art predictive models for wind and solar power generation to enable efficient integration of renewable energy sources in urban grids.
Energy Load Prediction
Creating accurate short and medium-term electricity demand forecasting systems to optimize grid management and energy market participation.
Grid Optimization
Designing intelligent systems that enhance grid flexibility and reliability through dynamic line rating and adaptive management approaches.
Digital Twin Technology
Building virtual representations of physical energy infrastructure to simulate, optimize, and monitor real-time performance for enhanced decision-making.
Energy Storage Optimization
Developing strategies for optimal utilization of battery storage and other energy storage technologies in conjunction with renewable generation.
Transfer Learning Applications
Applying novel transfer learning techniques to reduce sensor deployment costs and enable affordable large-scale energy infrastructure monitoring.
Research Publications
Improving Ramp Forecasting Accuracy under Class Imbalance [Under Review]
This research addresses the challenge of predicting sudden changes in wind power generation (ramp events) that can disrupt grid stability. Our approach formulates the problem as a multivariate time series classification task and proposes a novel data preprocessing strategy with EasyEnsemble algorithm to handle class imbalance. The method achieved over 85% accuracy and 88% weighted F1 score, offering a practical solution for renewable energy management.
A Novel Dual-CNN Architecture with Adaptive Persistence for Medium-Term Electricity Load Forecasting [Under Review]
This paper introduces a dual-CNN architecture for medium-term electricity load forecasting, designed for 30-hour ahead predictions. The model combines two parallel CNNs processing historical load and meteorological data while leveraging both short-term patterns and long-term seasonal effects. Tested on Belgian total load forecasting, our approach consistently outperformed operational forecasts by 20.47% on average, enhancing day-ahead operational planning capabilities.
PyAWD: A Library for Generating Large Synthetic Datasets of Acoustic Wave Propagation with Devito
This paper introduces PyAWD, a Python library for generating high-resolution synthetic datasets that simulate acoustic wave propagation. The library addresses the challenge of sparse seismic data by enabling the creation of customizable 2D and 3D simulations. Our approach demonstrated that machine learning models trained on this synthetic data can effectively retrieve epicenter locations using minimal sensor data, showing potential for cost-efficient monitoring of energy infrastructure.
Online learning of windmill time series using Long Short-term Cognitive Networks
This research introduces Long Short-term Cognitive Networks (LSTCNs) for forecasting windmill time series in online learning settings. Our model uses a fast, deterministic learning rule that makes it suitable for online learning tasks with large data streams. Compared to traditional approaches, LSTCNs achieved lower forecasting errors while being significantly faster in both training and testing phases, offering a practical solution for real-time monitoring of renewable energy sources.
Measuring Wind Turbine Health Using Fuzzy-Concept-Based Drifting Models
This paper presents novel approaches for monitoring wind turbine health based on fuzzy concept modeling. Unlike existing methods that rely on predictive models to detect anomalies, our approaches use abstract concepts implemented through fuzzy sets to identify performance changes over time. The methodology provides human-interpretable results through linguistic labels, enabling domain experts to easily understand and act on the system's outputs.
A Digital Twin Approach for Improving Estimation Accuracy in Dynamic Thermal Rating of Transmission Lines
This research proposes a Digital Twin approach to improve the accuracy of estimating conductor temperatures in power transmission lines. Our machine learning models significantly outperformed the traditional IEEE 738 standard, reducing the Root Mean Squared Error by 60%. The approach requires fewer features than physics-based models while maintaining accuracy, enabling transmission system operators to optimize line management with more accurate temperature estimations.
DAFT-E: Feature-based Multivariate and Multi-step-ahead Wind Power Forecasting
This paper introduces a novel forecasting approach called Dynamic Adaptive Feature-based Temporal Ensemble (DAFT-E) for predicting wind power generation. The method combines extensive feature engineering, fast feature selection, and an ensemble of computationally inexpensive models to achieve accurate forecasts while maintaining reasonable computational efficiency. Tested on multiple wind farm datasets, DAFT-E outperformed traditional methods including deep learning approaches.
Robust Assessment of Short-Term Wind Power Forecasting Models on Multiple Time Horizons
This research proposes a robust methodology for systematically evaluating various wind power forecasting models across multiple time horizons (1-6 hours ahead). The study presents a comprehensive modeling pipeline and tests various forecasting approaches on data from a wind farm in southern Italy. Our findings show that ensemble forecasting approaches consistently outperformed individual models in terms of accuracy and stability.
Adaptive local learning techniques for multiple-step-ahead wind speed forecasting
This research applies adaptive local learning methods to the problem of multi-step ahead wind speed forecasting. It aims to improve prediction accuracy by using instance-based techniques that can adapt to the non-stationary nature of wind patterns.
Time series prediction for energy-efficient wireless sensors: Applications to environmental monitoring and video games
This paper explores the use of time series prediction techniques to enhance energy efficiency in wireless sensor applications. It presents case studies in environmental monitoring and interactive systems (video games), showing how prediction can reduce data transmission needs.
A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition
This paper provides a comprehensive review and empirical comparison of various strategies for multi-step ahead time series forecasting. Using data from the NN5 forecasting competition, it evaluates the performance of methods like recursive, direct, and MIMO strategies.
Recursive multi-step time series forecasting by perturbing data
This paper proposes a novel approach for recursive multi-step time series forecasting. The method involves perturbing the input data during the recursive prediction process to potentially improve robustness and accuracy, particularly for longer horizons.
Conditionally dependent strategies for multiple-step-ahead prediction in local learning
This work introduces conditionally dependent strategies within the local learning framework for multi-step ahead prediction. It proposes methods that account for the dependencies between predictions at different future time steps, improving accuracy over independent predictions.
Multiple-output modeling for multi-step-ahead time series forecasting
This paper focuses on multiple-output modeling techniques for multi-step-ahead time series forecasting. It explores strategies like the MIMO (Multiple-Input Multiple-Output) approach where a single model predicts multiple future values simultaneously, contrasting it with other methods.
Long-term prediction of time series by combining Direct and MIMO strategies
This research investigates strategies for long-term time series prediction by comparing and combining Direct and Multi-Input Multi-Output (MIMO) approaches. The study aims to identify effective methods for forecasting far into the future.
Long term time series prediction with multi-input multi-output local learning
This work explores the use of Multi-Input Multi-Output (MIMO) strategies within a local learning framework for long-term time series prediction. It investigates how instance-based methods can simultaneously predict multiple future steps.
Adaptive model selection for time series prediction in wireless sensor networks
This paper addresses the challenge of resource constraints in wireless sensor networks (WSNs) by proposing an adaptive model selection strategy for time series prediction. The goal is to optimize prediction accuracy while minimizing computational and communication costs in WSNs.
Predicting stock markets in boundary conditions with local models
This study investigates the application of local modeling techniques for predicting stock market behavior, particularly under challenging boundary conditions or market extremes. It explores the effectiveness of instance-based learning in financial time series forecasting.
Local learning techniques for modeling, prediction and control
This PhD thesis provides a comprehensive overview of local learning techniques and their application to modeling, prediction (especially time series), and control problems. It details methodologies and theoretical foundations for instance-based machine learning approaches.
A multi-steap ahead prediction method based on local dynamic properties
This paper proposes a method for multi-step ahead time series prediction that leverages local dynamic properties of the data. The approach aims to capture time-varying patterns by focusing on localized model estimation.
Local learning for iterated time series prediction
This research further develops the concept of local learning for multi-step ahead time series forecasting using an iterated strategy. It focuses on building models based on nearby data points in the input space to make sequential predictions.
Lazy learning for iterated time series prediction
This paper introduces a lazy learning approach specifically designed for iterated time series prediction. It investigates how local, instance-based methods can be effectively applied to predict multiple steps ahead in time series data by iteratively using previous predictions.
Simulation and forecasting in intermodal container terminal
This work explores the use of simulation and forecasting techniques to model and predict operations within intermodal container terminals. It aims to improve efficiency and planning in logistics by applying computational methods to terminal processes.
Current Projects
Transfer Learning for Smart Grid Monitoring
Development of transfer learning methodologies to reduce the number of required temperature sensors in power transmission lines. Our approach enables cost-effective implementation of Dynamic Thermal Rating systems across urban grid infrastructure, enhancing capacity while maintaining safety.
Adaptive Forecasting for Renewable Integration
Implementation of ensemble-based forecasting systems that combine multiple prediction models to improve accuracy and stability for wind and solar energy production. Our forecasting systems adapt to changing environmental conditions and energy consumption patterns in urban settings.
Virtual Power Plant Simulator
Development of a Virtual Power Plant simulation platform that generates synthetic weather, solar, and wind data, along with grid load and market prices. The system implements a training pipeline for time series forecasting and optimization of battery management strategies for day-ahead market participation.
Get in Touch
If you have a project you'd like to discuss or any questions, feel free to reach out to us:
Prof. Gianluca Bontempi
Co-director, Machine Learning Group, ULB
Email: gianluca.bontempi@ulb.be
Dr. Natalia García-Colín
Research and Innovation Manager, Machine Learning Group, ULB
Email: natalia.garcia.colin@ulb.be