Adopting Renewable Energy Sources (RES) has become a key element in global efforts to combat climate change and decrease dependency on fossil fuels. This transition towards renewable energy is fueling the growing use of microgrids, which provide a decentralized and flexible solution for managing energy. Microgrids, especially when integrated with solar power and other RES, enable localized and efficient energy production that aligns with sustainability goals. Additionally, microgrids can incorporate Battery Energy Storage Systems (BESS), allowing them to store excess energy generated by renewables, thus ensuring a steady supply even when renewable generation is intermittent. The combination of RES and BESS within microgrids boosts energy resilience reduces emissions and supports the broader transition to a decentralized, sustainable energy infrastructure. The inherent variability of renewable energy generation and changing energy demands present significant challenges for traditional energy management systems. This thesis introduces various approaches to address these issues, including empirical modeling, Model Predictive Control (MPC), and various forms of Reinforcement Learning (RL) to improve the efficiency and reliability of microgrid operations. The first phase of this research develops a heuristical model to provide a detailed understanding of the operational dynamics within the microgrid, establishing a foundation for more advanced control strategies. Building on this, MPC is employed to optimize cost-efficiency while ensuring grid stability in the face of uncertainty. While MPC allows for effective real-time decision-making, its performance is often limited by its dependence on accurate system models and forecasts. To address these limitations, the final stage explores different reinforcement learning algorithms. Comprehensive simulation studies based on real microgrid data thoroughly validate the proposed control strategy. The results of the analysis revealed intriguing insights. The heuristic approach, while effective at minimizing power losses, performed poorly in terms of monetary outcomes, making it less ideal for cost-sensitive scenarios. Model Predictive Control (MPC), on the other hand, emerged as the most robust method overall. It provided the best monetary performance, demonstrating superior cost efficiency, but showed limitations in addressing power losses, where it fell short of the heuristic approach. Reinforcement Learning (RL) proved to be the most problematic among the methods evaluated. It failed to meet the imposed constraints, delivering subpar results from a monetary perspective—performing even worse than the heuristic approach—and struggled to provide competitive outcomes in other aspects.
L'adozione delle Fonti di Energia Rinnovabile (FER) è diventata un elemento chiave negli sforzi globali per contrastare il cambiamento climatico e ridurre la dipendenza dai combustibili fossili. Questa transizione verso l’energia rinnovabile sta alimentando la crescente diffusione delle microreti, che rappresentano una soluzione decentralizzata e flessibile per la gestione dell'energia. Le microreti, soprattutto quando integrate con l’energia solare e altre FER, consentono una produzione energetica localizzata ed efficiente, in linea con gli obiettivi di sostenibilità. Inoltre, le microreti possono incorporare Sistemi di Accumulo di Energia a Batteria (BESS), permettendo di immagazzinare l’energia in eccesso generata dalle rinnovabili e garantendo così una fornitura costante anche quando la generazione rinnovabile è intermittente. La combinazione di FER e BESS all’interno delle microreti migliora la resilienza energetica, riduce le emissioni e supporta la transizione verso un’infrastruttura energetica decentralizzata e sostenibile. L’intrinseca variabilità della generazione da fonti rinnovabili e la mutevole domanda di energia rappresentano sfide significative per i sistemi tradizionali di gestione energetica. Questa tesi introduce diverse metodologie per affrontare tali problematiche, tra cui modelli empirici, il Controllo Predittivo Basato su Modello (MPC) e varie forme di Apprendimento per Rinforzo (RL), per migliorare l'efficienza e l'affidabilità delle operazioni nelle microreti. La prima fase di questa ricerca sviluppa un modello euristico per comprendere in dettaglio le dinamiche operative all’interno della microrete, stabilendo una base per strategie di controllo più avanzate. Su questa base, l’MPC viene utilizzato per ottimizzare l’efficienza economica garantendo la stabilità della rete anche in condizioni di incertezza. Sebbene l’MPC consenta un’efficace presa di decisioni in tempo reale, le sue prestazioni sono spesso limitate dalla dipendenza da modelli e previsioni di sistema accurati. Per affrontare tali limitazioni, la fase finale esplora diversi algoritmi di apprendimento per rinforzo. Studi di simulazione approfonditi basati su dati reali di microreti validano in modo esaustivo le strategie di controllo proposte. I risultati dell’analisi hanno rivelato intuizioni interessanti. L’approccio euristico, pur essendo efficace nel minimizzare le perdite di potenza, ha mostrato scarse prestazioni dal punto di vista economico, rendendolo meno adatto a scenari sensibili ai costi. L'MPC si è invece rivelato il metodo più robusto in assoluto. Ha garantito le migliori prestazioni economiche, dimostrando un’elevata efficienza nei costi, ma ha mostrato limiti nella gestione delle perdite di potenza, risultando inferiore rispetto all’approccio euristico. Il RL si è dimostrato il più problematico tra i metodi valutati. Non è riuscito a rispettare le limitazioni imposte, producendo risultati scarsi dal punto di vista economico, persino peggiori dell’approccio euristico, e ha faticato a fornire risultati competitivi anche in altri aspetti.
Comparative analysis of optimization techniques for microgrid energy management
SCARANELLO, VINCENZO
2023/2024
Abstract
Adopting Renewable Energy Sources (RES) has become a key element in global efforts to combat climate change and decrease dependency on fossil fuels. This transition towards renewable energy is fueling the growing use of microgrids, which provide a decentralized and flexible solution for managing energy. Microgrids, especially when integrated with solar power and other RES, enable localized and efficient energy production that aligns with sustainability goals. Additionally, microgrids can incorporate Battery Energy Storage Systems (BESS), allowing them to store excess energy generated by renewables, thus ensuring a steady supply even when renewable generation is intermittent. The combination of RES and BESS within microgrids boosts energy resilience reduces emissions and supports the broader transition to a decentralized, sustainable energy infrastructure. The inherent variability of renewable energy generation and changing energy demands present significant challenges for traditional energy management systems. This thesis introduces various approaches to address these issues, including empirical modeling, Model Predictive Control (MPC), and various forms of Reinforcement Learning (RL) to improve the efficiency and reliability of microgrid operations. The first phase of this research develops a heuristical model to provide a detailed understanding of the operational dynamics within the microgrid, establishing a foundation for more advanced control strategies. Building on this, MPC is employed to optimize cost-efficiency while ensuring grid stability in the face of uncertainty. While MPC allows for effective real-time decision-making, its performance is often limited by its dependence on accurate system models and forecasts. To address these limitations, the final stage explores different reinforcement learning algorithms. Comprehensive simulation studies based on real microgrid data thoroughly validate the proposed control strategy. The results of the analysis revealed intriguing insights. The heuristic approach, while effective at minimizing power losses, performed poorly in terms of monetary outcomes, making it less ideal for cost-sensitive scenarios. Model Predictive Control (MPC), on the other hand, emerged as the most robust method overall. It provided the best monetary performance, demonstrating superior cost efficiency, but showed limitations in addressing power losses, where it fell short of the heuristic approach. Reinforcement Learning (RL) proved to be the most problematic among the methods evaluated. It failed to meet the imposed constraints, delivering subpar results from a monetary perspective—performing even worse than the heuristic approach—and struggled to provide competitive outcomes in other aspects.File | Dimensione | Formato | |
---|---|---|---|
Executive_Summary___Scuola_di_Ingegneria_Industriale_e_dell_Informazione___Politecnico_di_Milano (5).pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
464.39 kB
Formato
Adobe PDF
|
464.39 kB | Adobe PDF | Visualizza/Apri |
Thesis (10).pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
2.75 MB
Formato
Adobe PDF
|
2.75 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/10589/231235