In the present thesis, machine learning based pipelines for pressure drop prediction in two-phase flows are proposed, implemented, and optimized. A set of additional models, which require a lower number of features while providing an acceptable accuracy (even if less than the maximum achievable one), are then obtained and presented. The latter procedure facilitates evaluating the relative importance of each feature and assessing the trade-off between the complexity of the utilized model and the obtained accuracy. In this context, the datasets obtained from measurements on diabatic and adiabatic flows (conducted in the Energy Department of Politecnico di Milano) have been utilized. The adiabatic dataset consists of pressure drop measurements on water and water/air flows, in horizontal smooth and micro-finned copper tubes, at various flow conditions. The measurements on diabatic streams instead include pressure drop in R134a and R1234ze flows, which are going through evaporation and condensation processes, along with the corresponding flow conditions. Machine learning based pipelines are first implemented, in which dimensionless parameters are provided as features while the friction factor (for the single-phase case) and two-phase flow multiplier (for the two-phase case) are considered as the targets. Feature selection and pipeline optimization procedures are then applied to each pipeline in order to determine the most promising set of features along with the most accurate machine learning algorithm (and the corresponding tuning parameters). Finally, considering the trade-off between the complexity of the model (in terms of number of utilized features) and the achieved prediction precision, the accuracy of the optimized pipelines with lower number of features is progressively determined and presented. The obtained results demonstrate that utilizing the developed optimal pipelines leads to a significantly higher accuracy compared to the available physical models. As the optimized pipelines will be made publicly accessible, the implemented models offer higher reproducibility, ease of use, and accuracy compared to the physical models that are available in the literature. Moreover, the provided additional pipelines facilitate utilizing a notably lower number of features while achieving an acceptable accuracy.
Nella presente tesi, vengono proposte, implementate e ottimizzate delle pipeline basate sul machine learning per la previsione della caduta di pressione in flussi bifase. Viene poi ottenuta e presentata una serie di modelli aggiuntivi, che richiedono un numero inferiore di parametri, pur fornendo una precisione accettabile (anche se inferiore a quella massima ottenibile). Quest’ultima procedura facilita la valutazione dell’importanza relativa di ciascun parametro e la valutazione del compromesso tra la complessit`a del modello utilizzato e l’accuratezza ottenuta. In questo contesto sono stati utilizzati i dataset ottenuti dalle misure sui flussi diabatici e adiabatici (condotte presso il Dipartimento di Energia del Politecnico di Milano). Il dataset adiabatico `e costituito da misure di perdita di carico su flussi di acqua e acqua/aria, in tubi orizzontali di rame, lisci e micro alettati, a varie condizioni di flusso. Le misure sui flussi diabatici comprendono invece le perdite di carico nei flussi R134a e R1234ze, durante il processo di evaporazione e condensazione, insieme alle corrispondenti condizioni di flusso. Vengono prima implementate delle pipeline basate sul machine learning, in cui i parametri adimensionali sono forniti come parametri di input, mentre il fattore d’attrito (per il caso monofase) e il moltiplicatore bifase (per il caso bifase) sono considerati come obiettivi. Le procedure di selezione dei parameteri e di ottimizzazione della pipeline vengono poi applicate ad ogni pipeline per determinare l’insieme di parametri pi`u promettenti insieme all’algoritmo di machine learning pi`u accurato (e i corrispondenti parametri di tuning). Infine, considerando il trade-off tra la complessit`a del modello (in termini di numero di parametri utilizzati) e la precisione di predizione raggiunta, viene progressivamente determinata e presentata l’accuratezza delle pipeline ottimizzate con un numero inferiore di parametri.I risultati ottenuti dimostrano che l’utilizzo delle pipeline ottimali sviluppate porta ad una precisione significativamente maggiore rispetto ai modelli fisici disponibili. Poich´e le pipeline ottimizzate saranno rese pubblicamente accessibili, i modelli implementati offrono una maggiore riproducibilit`a, facilit`a d’uso e precisione rispetto ai modelli fisici disponibili in letteratura. Inoltre, le pipeline aggiuntive fornite facilitano l’utilizzo di un numero notevolmente inferiore di parametri, pur raggiungendo un’accuratezza accettabile.
Application of machine learning in frictional pressure drop estimation of two-phase flow : a dimensionless approach
ARDAM, KEIVAN
2018/2019
Abstract
In the present thesis, machine learning based pipelines for pressure drop prediction in two-phase flows are proposed, implemented, and optimized. A set of additional models, which require a lower number of features while providing an acceptable accuracy (even if less than the maximum achievable one), are then obtained and presented. The latter procedure facilitates evaluating the relative importance of each feature and assessing the trade-off between the complexity of the utilized model and the obtained accuracy. In this context, the datasets obtained from measurements on diabatic and adiabatic flows (conducted in the Energy Department of Politecnico di Milano) have been utilized. The adiabatic dataset consists of pressure drop measurements on water and water/air flows, in horizontal smooth and micro-finned copper tubes, at various flow conditions. The measurements on diabatic streams instead include pressure drop in R134a and R1234ze flows, which are going through evaporation and condensation processes, along with the corresponding flow conditions. Machine learning based pipelines are first implemented, in which dimensionless parameters are provided as features while the friction factor (for the single-phase case) and two-phase flow multiplier (for the two-phase case) are considered as the targets. Feature selection and pipeline optimization procedures are then applied to each pipeline in order to determine the most promising set of features along with the most accurate machine learning algorithm (and the corresponding tuning parameters). Finally, considering the trade-off between the complexity of the model (in terms of number of utilized features) and the achieved prediction precision, the accuracy of the optimized pipelines with lower number of features is progressively determined and presented. The obtained results demonstrate that utilizing the developed optimal pipelines leads to a significantly higher accuracy compared to the available physical models. As the optimized pipelines will be made publicly accessible, the implemented models offer higher reproducibility, ease of use, and accuracy compared to the physical models that are available in the literature. Moreover, the provided additional pipelines facilitate utilizing a notably lower number of features while achieving an acceptable accuracy.File | Dimensione | Formato | |
---|---|---|---|
Keivan_Ardam_M_Sc_Thesis.pdf
non accessibile
Descrizione: Thesis text
Dimensione
7.97 MB
Formato
Adobe PDF
|
7.97 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/10589/154576