Multi-reservoir systems management is complex because of the uncertainty on future events and the variety of purposes, usually conflicting, of the involved actors. An efficient management of these systems can prevent political crisis and conflicts between the stakeholders and can often improve resource allocation with respect to the historical management, providing a win-win strategy. Bellman Stochastic Dynamic Programming (SDP) is the most famous among the many proposed approaches to solve this Optimal Control (OC) problem. Unfortunately, Stochastic Dynamic Programming is affected by the curse of dimensionality: computational effort increases exponentially with the complexity of the considered system (i.e. number of reservoirs), and the problem rapidly becomes intractable. This thesis proposes a novel approach for the solution of the OC problem. The core idea is using the Artificial Neural Networks (ANNs) for designing neural policies which well approximate the optimal multi-reservoir network management policy obtained by an open-loop approach. A well trained ANN can adequately reproduce the relationship between input and output, providing an efficient control policy. The dataset used for the neural policy identification is composed by a large number of synthetic inputs and release series, generated under the assumption of deterministic knowledge of the future inflows. ANNs replicate the optimal management but with a closed-loop control system using only the current information, to take the decision in real-time. For this reason, they are effective real-time management policies. The proposed methodology is applied to the Nile River basin for two principal objectives: minimization of the irrigation water deficit and maximization of the hydropower production.
La gestione di sistemi multi-serbatoio è un’attività complessa in quanto gli eventi futuri non sono noti con certezza e poiché gli attori presenti hanno obiettivi diversi, spesso conflittuali. Un’efficiente gestione di questi sistemi è utile a prevenire crisi politiche e conflitti tra i portatori di interesse e spesso comporta una migliore allocazione delle risorse rispetto alla gestione storica, accrescendo la soddisfazione di tutte le parti in gioco. Sono stati proposti molti approcci per la soluzione di questo problema di controllo ottimo: il più famoso è la ben nota Programmazione Dinamica Stocastica (SDP) di Bellman. Sfortunatamente, la SDP risente della maledizione della dimensionalità: il costo computazionale cresce esponenzialmente con la complessità del sistema considerato (nel nostro caso con il numero di serbatoi presenti), e quindi il problema diventa rapidamente irrisolvibile. Questa tesi propone un approccio originale per la soluzione del problema di controllo ottimo, basato sull'idea di utilizzare le reti neurali artificiali (ANN) al fine di progettare una politica neurale in grado di approssimare in modo adeguato la politica ottima per la gestione della rete di serbatoi. Una ANN ben addestrata può riprodurre in modo soddisfacente la relazione tra ingresso e uscita, fornendo un’efficiente politica di controllo. Il dataset usato per l’identificazione della politica neurale è composto da un gran numero di serie sintetiche, generate calcolando la sequenza di controllo ottima in anello aperto, assumendo una conoscenza deterministica degli afflussi futuri. Le reti neurali sono in grado di replicare la gestione ottima così calcolata, ma in anello chiuso, basandosi cioè solo sull'informazione corrente (e non futura) e possono quindi essere impiegate in tempo reale per l'effettiva gestione della rete. La metodologia proposta è stata applicata al sistema del bacino del Nilo, considerando due obiettivi principali: minimizzare il deficit idrico relativo all'irrigazione e massimizzare la produzione di energia idroelettrica.
A neural approach for multi-reservoir system operation
SANGIORGIO, MATTEO
2015/2016
Abstract
Multi-reservoir systems management is complex because of the uncertainty on future events and the variety of purposes, usually conflicting, of the involved actors. An efficient management of these systems can prevent political crisis and conflicts between the stakeholders and can often improve resource allocation with respect to the historical management, providing a win-win strategy. Bellman Stochastic Dynamic Programming (SDP) is the most famous among the many proposed approaches to solve this Optimal Control (OC) problem. Unfortunately, Stochastic Dynamic Programming is affected by the curse of dimensionality: computational effort increases exponentially with the complexity of the considered system (i.e. number of reservoirs), and the problem rapidly becomes intractable. This thesis proposes a novel approach for the solution of the OC problem. The core idea is using the Artificial Neural Networks (ANNs) for designing neural policies which well approximate the optimal multi-reservoir network management policy obtained by an open-loop approach. A well trained ANN can adequately reproduce the relationship between input and output, providing an efficient control policy. The dataset used for the neural policy identification is composed by a large number of synthetic inputs and release series, generated under the assumption of deterministic knowledge of the future inflows. ANNs replicate the optimal management but with a closed-loop control system using only the current information, to take the decision in real-time. For this reason, they are effective real-time management policies. The proposed methodology is applied to the Nile River basin for two principal objectives: minimization of the irrigation water deficit and maximization of the hydropower production.| File | Dimensione | Formato | |
|---|---|---|---|
|
2016_12_Sangiorgio.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
9.62 MB
Formato
Adobe PDF
|
9.62 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/10589/131628