The management and control of modern power systems is a significant challenge. This is due to the growing penetration of renewable energy sources and the rapid expansion of electric mobility in urban areas. Moreover, the gradual shift from centralized to distributed system architectures requires advanced control and optimization strategies. These strategies must cope with the increased variability and unpredictability of the grid. In this context, and considering the rapid adoption of artificial intelligence and machine learning, this dissertation aims to develop and validate applications of machine learning for power system optimization, control, and forecasting. Three main applications are explored. The first concerns the optimal scheduling of charging and discharging processes for electric vehicles. Reinforcement learning is used to generate effective and realistic control policies. Here, the objectives are to reduce energy costs, maximize revenue from vehicle-to-grid services, and prevent network overloads. The model is also trained using imitation learning on expert data to accelerate learning and improve robustness. The second application addresses the optimal power flow problem. The goal is to determine the network operating point and dispatch that minimize costs while maintaining acceptable voltage profiles and respecting all constraints. This problem is also tackled with a reinforcement learning model designed to manage realistic load and renewable generation profiles in networks with a large number of nodes. An additional study introduces a verification and correction mechanism to ensure all constraints are satisfied, providing feasible and certified solutions. The third application focuses on forecasting models based on neural networks. These models are integrated into a Model Predictive Control system for microgrid management. They improve the accuracy of price predictions and enable the development of economically optimal control policies. All proposed approaches have been extensively tested on standard benchmarks and real datasets. The results show that machine learning techniques can achieve near-optimal performance, ensure operational reliability, and provide fast computations suitable for real-time scenarios. Overall, the thesis offers a comprehensive framework for using machine learning in the control, planning, and forecasting of energy systems.
La gestione ed il controllo dei moderni sistemi elettrici rappresenta una sfida importante, soprattutto a seguito dell’aumento delle fonti energetiche rinnovabili presenti sul territorio e della diffusione della mobilità elettrica nei contesti urbani. Inoltre, il graduale passaggio in molti paesi da un sistema elettrico centralizzato a uno distribuito sta comportando la necessità di elaborare strategie di controllo e ottimizzazione più avanzate, in grado di far fronte alla maggiore imprevidibilità del sistema. In questo contesto, e considerando la rapida diffusione delle tecniche di intelligenza artificiale e machine learning, questa tesi ha come obiettivo quello di sviluppare e testare possibili applicazioni degli algoritmi di machine learning nell’ambito dell’ottimizzazione e controllo di sistemi elettrici e nelle previsioni di serie storiche. In particolare, vengono esplorate tre principali applicazioni. La prima riguarda la pianificazione dei processi di carica e scarica dei veicoli elettrici. Per generare politiche di controllo efficaci e realistiche, viene utilizzato il reinforcement learning, con l’obiettivo di ridurre i costi energetici, massimizzare i ricavi dai servizi di vehicle-to-grid e prevenire sovraccarichi sulla rete. Il modello viene inoltre addestrato mediante imitation learning su dati forniti da esperti, con l’obiettivo di velocizzare l’apprendimento e aumentarne la robustezza. La seconda applicazione riguarda il problema dell’optimal power flow con l’obiettivo di determinare l’assetto della rete e il dispacciamento ottimale, minimizzando i costi e garantendo la qualità della tensione ai nodi nel rispetto dei vincoli di rete. Anche questo problema viene affrontato con un modello di reinforcement learning in grado di gestire profili realistici di carico e generazione rinnovabile in reti ad elevato numero di nodi. Per garantire ulteriormente l’affidabilità operativa, viene studiato anche un meccanismo di verifica e correzione per assicurare che tutti i vincoli operativi siano rispettati, fornendo così una soluzione fattibile e certificata. Infine, vengono sviluppati modelli previsionali basati su reti neurali che sono integrati in un sistema di model predictive control per la gestione di microreti. Questi modelli migliorano l’accuratezza delle previsioni sui prezzi, permettendo l’elaborazione di politiche di controllo ottimali dal punto di vista economico. Tutti gli approcci proposti sono stati ampiamente valutati su benchmark standard e dataset reali, dimostrando che le tecniche di machine learning possono raggiungere prestazioni prossime all’ottimo, garantire affidabilità operativa e consentire calcoli rapidi, adeguati a scenari in tempo reale. Complessivamente, la tesi fornisce un quadro completo per l’impiego del machine learning nel controllo, nella pianificazione e nella previsione dei sistemi energetici.
Advanced machine learning techniques for optimization, control, and forecasting in modern power systems
Rossi, Federico
2025/2026
Abstract
The management and control of modern power systems is a significant challenge. This is due to the growing penetration of renewable energy sources and the rapid expansion of electric mobility in urban areas. Moreover, the gradual shift from centralized to distributed system architectures requires advanced control and optimization strategies. These strategies must cope with the increased variability and unpredictability of the grid. In this context, and considering the rapid adoption of artificial intelligence and machine learning, this dissertation aims to develop and validate applications of machine learning for power system optimization, control, and forecasting. Three main applications are explored. The first concerns the optimal scheduling of charging and discharging processes for electric vehicles. Reinforcement learning is used to generate effective and realistic control policies. Here, the objectives are to reduce energy costs, maximize revenue from vehicle-to-grid services, and prevent network overloads. The model is also trained using imitation learning on expert data to accelerate learning and improve robustness. The second application addresses the optimal power flow problem. The goal is to determine the network operating point and dispatch that minimize costs while maintaining acceptable voltage profiles and respecting all constraints. This problem is also tackled with a reinforcement learning model designed to manage realistic load and renewable generation profiles in networks with a large number of nodes. An additional study introduces a verification and correction mechanism to ensure all constraints are satisfied, providing feasible and certified solutions. The third application focuses on forecasting models based on neural networks. These models are integrated into a Model Predictive Control system for microgrid management. They improve the accuracy of price predictions and enable the development of economically optimal control policies. All proposed approaches have been extensively tested on standard benchmarks and real datasets. The results show that machine learning techniques can achieve near-optimal performance, ensure operational reliability, and provide fast computations suitable for real-time scenarios. Overall, the thesis offers a comprehensive framework for using machine learning in the control, planning, and forecasting of energy systems.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_01_Rossi_Federico.pdf
accessibile in internet per tutti a partire dal 21/12/2026
Descrizione: Advanced Machine Learning Techniques for Optimization, Control, and Forecasting in Modern Power Systems
Dimensione
7.82 MB
Formato
Adobe PDF
|
7.82 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/248357