With the growing use of renewable energy sources (RES) Distributed Generation (DG) systems are rapidly spreading. Embedding DG in the distribution network may be costly due to grid reinforcements and control adjustments required to maintain the electrical network reliable. Deterministic load flow calculations are usually employed to assess the allowed DG penetration in a distribution network in order to ensure that current or voltage limits are not exceeded. However, these calculations may overlook the risk of limit violations due to uncertainties in the operating conditions of the network. To overcome this limitation, this thesis addresses the problem of the DG penetration level with a Monte Carlo simulation technique that accounts for the variability intrinsic to electric power consumption and production by renewables. The power absorbed by each load of a Medium Voltage (MV) network is characterized by a load variation curve; probabilistic load flow is then used for computing the maximum DG power that can be connected to the MV buses with sufficient margins with respect to electric constraints. Two MV radial distribution networks of different sizes are studied and a comparison is provided between the results of the deterministic load flow (DLF) and probabilistic load flow (PLF) analyses. The repetition of load flow evaluations within the Monte Carlo simulation is computationally expensive for networks of large size and is the bottleneck of the PLA analysis. In this work, we address this issue by implementing a regression model based on machine learning techniques to develop an empirical model of the MV electric distribution network that substitutes the PLF. We assessed three different function estimation methods: Neural network (NN), Support vector Machine (SVM) and Least squares support vector machine (LS-SVM). The DG penetration analysis problem is tackled by the three different learning algorithm and the results are compared. The simulation method based on the integration of Monte Carlo sampling and LS-SVM is identified as the most suitable for the DG penetration analysis of a MV distribution network. A confidence interval for the LS-SVM predictions is computed to account for uncertainties in the system operations and for approximations in the regression model.
Questa tesi è il risultato di una collaborazione tra il Laboratorio Analisi di Segnale e Analisi di Rischio (LASAR) e il Laboratorio di Sistemi Elettrici per l’Energia (EPSlab) entrambi afferenti al Dipartimento di Energia. Per raggiungere gli obbiettivi del pacchetto “20-20-20 al 2020” di riduzione delle emissioni clima-alteranti, il sistema elettrico sarà sottoposto a notevoli innovazioni. La principale di esse riguarda l’impiego massiccio della Generazione Diffusa (GD) a livello della distribuzione in media tensione, da porre in relazione con nuove modalità di gestione delle reti elettriche nella direzione delle Smart Grid, imponendo un passaggio da rete “passiva” ad “attiva”. Nella tesi si effettua l’analisi di penetrazione di GD nella rete di media tensione adottando un approccio probabilistico, indispensabile per assicurarci che le condizioni operative siano rispettate a fronte della variabilità delle condizioni operative dovute all’incertezza sui prelievi di potenza degli utenti. L’obbiettivo è quantificare la massima potenza installabile nei nodi (Nodal Hosting Capacity) di media tensione. A tal fine, si è sviluppato un algoritmo di load flow probabilistico per trattare le problematiche connesse alla variabilità delle configurazioni di carico assunte dalla rete. L’analisi di penetrazione della DG è quindi effettuata per due taglie di rete in media tensione, individuando le peggiori condizioni operative durante l’anno e calcolando la probabilità annua che i limiti operativi vengano superati data la potenza di GD che si è deciso di installare nel nodo. L’algoritmo sviluppato presenta elevati tempi di calcolo per reti di grandi dimensioni a causa del maggior numero di equazioni di load flow da risolvere e del maggior numero di vincoli da verificare. Pertanto, la seconda parte della tesi è rivolta a ridurre i costi computazionali dell’analisi di penetrazione mediante la sostituzione delle equazioni di load flow con algoritmi di apprendimento automatico, cosiddette tecniche di machine learning. Tre tipi di apprendimento sono stati confrontati tra di loro le reti neurali (NN), le Support Vector Machine (SVM) e le Support Vector Machine ai minimi quadrati (LS-SVM). Dopo il confronto è stato sviluppato un metodo basato su Least Squares Support Vector Machines (LS-SVM). Il risultato è una notevole riduzione dei tempi computazionali. L’accuratezza dei risultati è assicurata tramite la quantificazione dagli intervalli di confidenza sulle massime potenze installabili.
Probabilistic analysis of distributed generation by Monte Carlo simulation and support vector machine
GIORGI, LIVIO
2010/2011
Abstract
With the growing use of renewable energy sources (RES) Distributed Generation (DG) systems are rapidly spreading. Embedding DG in the distribution network may be costly due to grid reinforcements and control adjustments required to maintain the electrical network reliable. Deterministic load flow calculations are usually employed to assess the allowed DG penetration in a distribution network in order to ensure that current or voltage limits are not exceeded. However, these calculations may overlook the risk of limit violations due to uncertainties in the operating conditions of the network. To overcome this limitation, this thesis addresses the problem of the DG penetration level with a Monte Carlo simulation technique that accounts for the variability intrinsic to electric power consumption and production by renewables. The power absorbed by each load of a Medium Voltage (MV) network is characterized by a load variation curve; probabilistic load flow is then used for computing the maximum DG power that can be connected to the MV buses with sufficient margins with respect to electric constraints. Two MV radial distribution networks of different sizes are studied and a comparison is provided between the results of the deterministic load flow (DLF) and probabilistic load flow (PLF) analyses. The repetition of load flow evaluations within the Monte Carlo simulation is computationally expensive for networks of large size and is the bottleneck of the PLA analysis. In this work, we address this issue by implementing a regression model based on machine learning techniques to develop an empirical model of the MV electric distribution network that substitutes the PLF. We assessed three different function estimation methods: Neural network (NN), Support vector Machine (SVM) and Least squares support vector machine (LS-SVM). The DG penetration analysis problem is tackled by the three different learning algorithm and the results are compared. The simulation method based on the integration of Monte Carlo sampling and LS-SVM is identified as the most suitable for the DG penetration analysis of a MV distribution network. A confidence interval for the LS-SVM predictions is computed to account for uncertainties in the system operations and for approximations in the regression model.File | Dimensione | Formato | |
---|---|---|---|
2011_12_Giorgi.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
2.47 MB
Formato
Adobe PDF
|
2.47 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/10589/39262