The first part of the present thesis is focused on estimation of characteristic and performance of buildings through analyzing the corresponding consumption profiles. The data set has been obtained from two case studies: thermal consumption data of buildings in a hospital complex along with thermal and electrical consumption of buildings in a university campus. The corresponding climatic condition information were obtained from near-by weather stations and on-site measurements. The building characteristics and performance of these buildings were obtained from previously conducted auditing procedures. In the first step, temporal features that are quantitative measures, which represent different types of trends in profiles, are extracted from historical thermal and electrical consumption data along with the corresponding climatic condition profiles. The above-mentioned generated features are next provided as inputs to machine learning algorithms which are implemented in order to predict the heat transfer coefficient and energy performance index of buildings. In order to optimize the machine learning based pipeline, a feature selection methodology is first implemented. In this procedure, considering a state-of-the-art machine learning algorithm, the most promising features, which result in the most accurate prediction, are determined. Next, a genetic algorithm based optimization procedure is conducted to investigate the optimal machine learning algorithm and the corresponding tuning parameters for each pipeline. The obtained results demonstrate that, by employing the proposed methodology, energy auditing procedures, albeit with a lower accuracy, can be carried out in a notably faster and cheaper manner. In the second part of the thesis, a machine learning based pipeline is implemented and optimized for predicting the heating and cooling load of residential buildings being provided 8 geometrical properties. The result of this procedure demonstrated by implementing a feature selection (which selects only 5 parameters) and determining the optimal machine learning algorithm, an elevated accuracy (R2 score of 0.991) can be achieved.
La prima parte della presente tesi è incentrata sulla stima delle caratteristiche e delle prestazioni degli edifici attraverso l'analisi dei corrispondenti profili di consumo. Il set di dati è stato ottenuto da due casi: i dati sui consumi termici degli edifici di un complesso ospedaliero e i consumi termici ed elettrici degli edifici di un campus universitario. Le corrispondenti informazioni sulle condizioni climatiche sono state ottenute da stazioni meteorologiche vicine e da misurazioni in loco. Le caratteristiche costruttive e le prestazioni di questi edifici sono state ottenute da procedure di diagnosi energetica precedentemente condotte. Nella prima fase, le caratteristiche temporali che sono misure quantitative, che rappresentano diversi tipi di tendenze dei profili, vengono estratte dai dati storici dei consumi termici ed elettrici e dai corrispondenti profili delle condizioni climatiche. Le suddette caratteristiche generate vengono poi fornite come input agli algoritmi di Machine learning che vengono implementati al fine di prevedere il coefficiente di scambio termico e l'indice di prestazione energetica degli edifici. Al fine di ottimizzare la pipeline basata sull' Machine learning, viene prima implementata una metodologia di selezione delle caratteristiche. In questa procedura, considerando un algoritmo di machine learning allo stato dell'arte, vengono determinate le caratteristiche più promettenti, che si traducono nella previsione più accurata. Successivamente, viene condotta una procedura di ottimizzazione basata su algoritmi genetici per studiare l'algoritmo di Machine learning ottimale e i corrispondenti parametri di regolazione per ogni pipeline. I risultati ottenuti dimostrano che, utilizzando la metodologia proposta, le procedure di auditing energetico, anche se con una precisione inferiore, possono essere eseguite in modo notevolmente più veloce ed economico. Nella seconda parte della tesi, viene implementata e ottimizzata una pipeline basata sull'apprendimento automatico per prevedere il carico di riscaldamento e raffreddamento degli edifici residenziali fornendo 8 proprietà geometriche. Il risultato di questa procedura ha dimostrato che implementando una selezione di elementi (che seleziona solo 5 parametri) e determinando l'algoritmo di machine learning ottimale, è possibile ottenere un'elevata precisione di previsione (punteggio R2 di 0,991).
Machine learning based building characteristics and performance estimation through analyzing consumption profiles
SHAJU, ARUN
2017/2018
Abstract
The first part of the present thesis is focused on estimation of characteristic and performance of buildings through analyzing the corresponding consumption profiles. The data set has been obtained from two case studies: thermal consumption data of buildings in a hospital complex along with thermal and electrical consumption of buildings in a university campus. The corresponding climatic condition information were obtained from near-by weather stations and on-site measurements. The building characteristics and performance of these buildings were obtained from previously conducted auditing procedures. In the first step, temporal features that are quantitative measures, which represent different types of trends in profiles, are extracted from historical thermal and electrical consumption data along with the corresponding climatic condition profiles. The above-mentioned generated features are next provided as inputs to machine learning algorithms which are implemented in order to predict the heat transfer coefficient and energy performance index of buildings. In order to optimize the machine learning based pipeline, a feature selection methodology is first implemented. In this procedure, considering a state-of-the-art machine learning algorithm, the most promising features, which result in the most accurate prediction, are determined. Next, a genetic algorithm based optimization procedure is conducted to investigate the optimal machine learning algorithm and the corresponding tuning parameters for each pipeline. The obtained results demonstrate that, by employing the proposed methodology, energy auditing procedures, albeit with a lower accuracy, can be carried out in a notably faster and cheaper manner. In the second part of the thesis, a machine learning based pipeline is implemented and optimized for predicting the heating and cooling load of residential buildings being provided 8 geometrical properties. The result of this procedure demonstrated by implementing a feature selection (which selects only 5 parameters) and determining the optimal machine learning algorithm, an elevated accuracy (R2 score of 0.991) can be achieved.File | Dimensione | Formato | |
---|---|---|---|
ArunShaju_10554547_thesis.pdf
non accessibile
Descrizione: Thesis text
Dimensione
7.55 MB
Formato
Adobe PDF
|
7.55 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/10589/145604