Titolo della tesi: Machine Learning for the Sustainable Energy Transition: a Data-Driven Perspective along the Value Chain from Manufacturing to Energy Conversion
According to the special report Global Warming of 1.5°C of the IPCC, climate action is not only necessary but more than ever urgent. The world is witnessing rising sea levels, heat waves, events of flooding, droughts and desertification resulting in the loss of lives and damage to livelihoods, especially in countries of the Global South. To mitigate climate change and commit to the Paris agreement, it is of the uttermost importance to reduce greenhouse gas emissions coming from the most emitting sector, namely the energy sector. To this end, large-scale penetration of renewable energy systems into the energy market is crucial for the energy transition toward a sustainable future by replacing fossil fuels and improving access to energy with socio-economic benefits. With the advent of Industry 4.0, Internet of Things technologies have been increasingly applied to the energy sector introducing the concept of smart grid or, more in general, Internet of Energy. These technologies are steering the energy sector towards more efficient, reliable, flexible, resilient, safe, and sustainable solutions with huge environmental and potential social benefits. To realize these solutions, new information technologies are required and among the most promising possibilities are Artificial Intelligence and Machine Learning which in many countries have already revolutionized the energy industry. This thesis presents different Machine Learning algorithms and methods for the implementation of new strategies to make renewable energy systems more efficient and reliable. It presents various learning algorithms, highlighting their advantages and limits, and evaluates their application for different tasks in the energy context. In addition, different techniques are presented for the preprocessing and cleaning of time series data, nowadays collected by sensor networks mounted on every renewable energy system. With the possibility to install large numbers of sensors that collect vast amounts of data, it is vital to detect and remove irrelevant, redundant, or noisy features, and alleviate the curse of dimensionality, thus improving the interpretability of predictive models, speeding up their learning process, and enhancing their generalization properties. Therefore, this thesis discusses the importance of dimensionality reduction in sensor networks mounted on renewable energy systems and, to this end, presents two novel unsupervised algorithms. The first approach maps time series data in the network domain through visibility graphs and uses a community detection algorithm to identify clusters of similar variables and select representative parameters. This method can group both homogeneous and heterogeneous physical parameters, even when related to different functional areas of a system. The second approach proposes the Combined Predictive Power Score, a method for feature selection with a multivariate formulation that explores multiple sub-sets of expanding variables and identifies the combination of features with the highest predictive power over specified target variables. This method proposes a selection algorithm for the optimal combination of variables that converges to the smallest set of predictors with the highest predictive power. Once the combination of variables is identified, the most relevant parameters in a sensor network can be selected to perform dimensionality reduction. Data-driven methods open the possibility of supporting strategic decision-making, resulting in a reduction of operation and maintenance costs, machine faults, repair stops, and spare parts inventory size. Therefore, this thesis presents two approaches in the context of predictive maintenance to improve the lifetime and efficiency of the equipment based on anomaly detection algorithms. The first approach proposes an anomaly detection model based on Principal Component Analysis that is robust to false alarms, can isolate anomalous conditions, and can anticipate equipment failures. The second approach has at its core a neural architecture, namely a Graph Convolutional Autoencoder, which models the sensor network as a dynamical functional graph by simultaneously considering the information content of individual sensor measurements (graph node features) and the nonlinear correlations existing between all pairs of sensors (graph edges). The proposed neural architecture can capture hidden anomalies in wind farms even when the turbines continue to deliver the power requested by the grid and can anticipate equipment failures. Since the model is unsupervised and completely data-driven, this approach can be applied to any wind turbine equipped with a SCADA system. When it comes to renewable energies, the unschedulable uncertainty due to their intermittent nature represents an obstacle to the reliability and stability of energy grids, especially when dealing with large-scale integration. Nevertheless, these challenges can be alleviated if the natural sources or the power output of renewable energy systems can be forecasted accurately, allowing power system operators to plan optimal power management strategies to balance the dispatch between intermittent power generation and load demand. To this end, this thesis proposes a multi-modal spatio-temporal neural network for multi-horizon wind power forecasting. In particular, the model combines high-resolution Numerical Weather Prediction forecast maps with turbine-level SCADA data and explores how meteorological variables on different spatial scales, together with the turbines’ internal operating conditions impact wind power forecasts. The world is undergoing a third energy transition with the main goal of tackling global climate change through decarbonization of the energy supply and consumption patterns. This is not only possible thanks to global cooperation and agreements between parties, power generation systems advancements, and Internet of Things and Artificial Intelligence technologies but also necessary to prevent the severe and irreversible consequences of climate change that are threatening life on the planet as we know it. This thesis is intended as a reference for researchers that want to contribute to the sustainable energy transition and are approaching the field of Artificial Intelligence in the context of renewable energy systems.