Ημερολόγιο Εκδηλώσεων

Περιγραφή:
ΠΟΛΥΤΕΧΝΕΙΟ ΚΡΗΤΗΣ Σχολή Ηλεκτρονικών Μηχανικών και Μηχανικών Υπολογιστών Πρόγραμμα Μεταπτυχιακών Σπουδών ΠΑΡΟΥΣΙΑΣΗ ΜΕΤΑΠΤΥΧΙΑΚΗΣ ΕΡΓΑΣΙΑΣ ΑΓΓΕΛΟΥ ΑΓΓΕΛΙΔΑΚΗ με θέμα Παραγοντικές Στοχαστικές Διαδικασίες Markov για Βέλτιστες Αποφάσεις Αγοραπωλησίας Ενέργειας στο Έξυπνο Δίκτυο Ηλεκτροδότησης Factored MDPs for Optimal Prosumer Decision Making in the Smart Grid Εξεταστική Επιτροπή Επίκουρος Καθηγητής Γεώργιος Χαλκιαδάκης (επιβλέπων) Επίκουρος Καθηγητής Ευτύχιος Κουτρούλης Αναπληρωτής Καθηγητής Μιχαήλ Λαγουδάκης Abstract Tackling the decision-making problem faced by a prosumer (i.e., a producer that is simultaneously a consumer) when selling and buying energy in the emerging smart electricity grid, is of utmost importance for the economic profitability of such a business entity. In this thesis, we model, for the first time, this problem as a factored Markov Decision process (MDP). Our model successfully captures the main aspects of the business decisions of a prosumer corresponding to a com- munity microgrid of any size. Moreover, it includes appropriate sub-models for prosumer production and consumption prediction. Employing this model, we are able to represent the problem compactly, and to provide an exact optimal solution via dynamic programming—notwithstanding its large size. In addition, we show how to use approximate MDP solution meth- ods for taking decisions in this domain, without the need of discretizing the state space. Specifically, we employ fitted value iteration, a sampling-based approxi- mation method that is known to be well behaved. By so doing, we generalize our factored MDP solution method to continuous state spaces. Our experimental simulations verify the effectiveness of our approach. They show that our exact value iteration solution matches that of a state-of-the-art method for stochastic planning in very large environments, while outperforming it in terms of computation time. Furthermore, we evaluate our approximate solution method via using a variety of basis functions over different state sample sizes, and comparing its performance to that of our exact value iteration algorithm. Our approximation method is shown to exhibit stable performance in terms of accu- mulated reward, which for certain basis functions reaches 90% of that gathered by the exact algorithm.