The present thesis aims at determining the most promising set of dimensionless features and the optimal machine learning based pipeline for data-driven heat transfer estimation of R134a condensing and evaporating flows in horizontal micro-finned tubes. The models are trained and validated using the experimental data obtained from the experiments (previously conducted in the energy department of Politecnico di Milano) that were performed on the mentioned flows, undergoing various flow conditions within 3 different configurations of micro-fin tubes. A set of dimensionless parameters, which are the representatives of the corresponding governing physical phenomena, are first determined for each data point and are then employed as the input features in the implemented pipelines. The Nu number (calculated based on the average two-phase flow heat transfer coefficient and the liquid thermal conductivity) is instead considered as the estimation target. Subsequently, while utilizing the Random Forest as the benchmark machine learning algorithm, a feature selection procedure is implemented and performed to find the most promising set of features for the condensation and evaporation cases. A genetic algorithm based optimization procedure is next carried out to find the optimal machine learning algorithm (and its corresponding hyper-parameters) for each pipeline. Through conducting a forward feature combination algorithm, the contribution of each feature to the achieved accuracy of the obtained optimal pipeline was also demonstrated, aiming at representing the trade-off between the model’s complexity and the resulting achieved performance. In addition, the physical models, that have been proposed in the literature with the same objective, were reviewed. Then, the accuracy offered by the obtained optimal pipeline, for the condensation and evaporation cases, were compared with the one obtained using the corresponding most promising physical model. Analyzing the feature selection results reveals that (for both of the evaporation and condensation cases), among all of the considered features, only 7 parameters are proved to be contributing to the achieved accuracy. It is also shown that utilizing the obtained optimal pipeline, which is being trained employing the selected features, leads to notably higher performance (MARD (determined for the test set) of 4.98% for the condensation case and 4.4% for the evaporation case) compared to the corresponding most promising empirical models available in the literature (resulting in the MARD of 35.2% for the condensation case and 22.42% for the evaporation case). Furthermore, as the datasets and the obtained optimal pipelines will be made publicly accessible as open-source software, the proposed approach also offers a higher reproducibility and ease of use.
La presente tesi mira a determinare l’insieme piu` promettente dei parametri adimensionali e la pipeline ottimale basata sul machine learning per la stima data-driven del trasferimento di calore di flussi di R134a in condensazione ed evaporazione in tubi orizzontali micro alettati. I modelli sono addestrati e validati utilizzando i dati sperimentali ottenuti dagli esperimenti (precedentemente condotti nel dipartimento di energia del Politecnico di Milano) che sono stati eseguiti sui flussi menzionati, sottoposti a varie condizioni di flusso all’interno di 3 diverse configurazioni di tubi a microalette. Un set di parametri adimensionali, che sono i rappresentanti dei corrispondenti fenomeni fisici, sono prima determinati per ogni punto (del dataset) e sono poi impiegati come parametri di input nella pipeline implementata. Il numero Nu (calcolato sulla base del coefficiente medio di trasferimento di calore del flusso bifase e della conducibilit`a termica del liquido) `e invece considerato come obiettivo della stima. Successivamente, utilizzando la Random Forest come algoritmo di machine learning di riferimento, una procedura di selezione dei parametri viene implementata ed eseguita per trovare il set di parametri piu` promettenti per i casi di condensazione ed evaporazione. Una procedura di ottimizzazione basata su un algoritmo genetico viene poi eseguita per trovare l’algoritmo di machine learning ottimale (e i suoi iper-parametri corrispondenti) per ogni pipeline. Attraverso la conduzione di un algoritmo di combinazione dei parametri, `e stato anche dimostrato il contributo di ogni parametro all’accuratezza raggiunta della pipeline ottimale ottenuta, con l’obiettivo di rappresentare il trade-off tra la complessit`a del modello e le prestazioni ottenute. Inoltre, i modelli fisici che sono stati proposti in letteratura con lo stesso obiettivo sono stati rivisti e l’accuratezza offerta della pipeline ottimale ottenuta, per i casi di condensazione ed evaporazione, sono stati confrontati con quella ottenuta utilizzando il corrispondente modello fisico piu` promettente. L’analisi dei risultati della selezione dei parametri rivela che (sia per per il caso di evaporazione che quello per la condensazione), tra tutti parametri considerati, solo 7 parametri hanno dimostrato di contribuire all’accuratezza raggiunta. Viene anche mostrato che l’utilizzo della pipeline ottimale ottenuta, che viene addestrata utilizzando i parametri selezionati, porta a prestazioni notevolmente piu` elevate (MARD (determinato per il set di test) del 4,98% per il caso di condensazione e 4,4% per il caso di evaporazione) rispetto ai corrispondenti modelli empirici piu` promettenti disponibili in letteratura (risultanti in MARD del 35,2% per il caso di condensazione e 22,42% per il caso di evaporazione). Inoltre, poich ́e i set di dati e le pipeline ottimali ottenute saranno resi pubblicamente accessibili come software open-source, l’approccio proposto offre anche una maggiore riproducibilita` e facilita` d’uso.
Machine learning based heat transfer estimation of evaporating and condensing R134a flow in micro-finned tubes : determination of the most promising dimensionless feature set
MILANI, SHAYAN
2019/2020
Abstract
The present thesis aims at determining the most promising set of dimensionless features and the optimal machine learning based pipeline for data-driven heat transfer estimation of R134a condensing and evaporating flows in horizontal micro-finned tubes. The models are trained and validated using the experimental data obtained from the experiments (previously conducted in the energy department of Politecnico di Milano) that were performed on the mentioned flows, undergoing various flow conditions within 3 different configurations of micro-fin tubes. A set of dimensionless parameters, which are the representatives of the corresponding governing physical phenomena, are first determined for each data point and are then employed as the input features in the implemented pipelines. The Nu number (calculated based on the average two-phase flow heat transfer coefficient and the liquid thermal conductivity) is instead considered as the estimation target. Subsequently, while utilizing the Random Forest as the benchmark machine learning algorithm, a feature selection procedure is implemented and performed to find the most promising set of features for the condensation and evaporation cases. A genetic algorithm based optimization procedure is next carried out to find the optimal machine learning algorithm (and its corresponding hyper-parameters) for each pipeline. Through conducting a forward feature combination algorithm, the contribution of each feature to the achieved accuracy of the obtained optimal pipeline was also demonstrated, aiming at representing the trade-off between the model’s complexity and the resulting achieved performance. In addition, the physical models, that have been proposed in the literature with the same objective, were reviewed. Then, the accuracy offered by the obtained optimal pipeline, for the condensation and evaporation cases, were compared with the one obtained using the corresponding most promising physical model. Analyzing the feature selection results reveals that (for both of the evaporation and condensation cases), among all of the considered features, only 7 parameters are proved to be contributing to the achieved accuracy. It is also shown that utilizing the obtained optimal pipeline, which is being trained employing the selected features, leads to notably higher performance (MARD (determined for the test set) of 4.98% for the condensation case and 4.4% for the evaporation case) compared to the corresponding most promising empirical models available in the literature (resulting in the MARD of 35.2% for the condensation case and 22.42% for the evaporation case). Furthermore, as the datasets and the obtained optimal pipelines will be made publicly accessible as open-source software, the proposed approach also offers a higher reproducibility and ease of use.| File | Dimensione | Formato | |
|---|---|---|---|
|
2021_04_Milani.pdf
non accessibile
Descrizione: Thesis text
Dimensione
5.26 MB
Formato
Adobe PDF
|
5.26 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/10589/173436