- 學位論文
使用人工類神經網路再現彭羅賓森流體之熱力學性質與相平衡
Reproducing Thermodynamic Properties and Phase Equilibrium of Peng-Robinson Fluids Using Artificial Neural Network
摘要
本研究欲建立一套人工類神經網路模型(Artificial Neural Network, ANN)，以準確預測純流體(fluids)的熱力學性質(thermodynamic properties)和相平衡。在人工類神經網路模型的架構中，需要提供大量數據用以訓練其中神經元之參數，因此我們在本研究中選擇使用彭羅賓森狀態方程式(Peng-Robinson Equation of State, PR-EOS)計算大量的純流體熱力學數據，以供後續的模型訓練。以PR-EOS計算純流體熱力學數據的方法大致為：在包含液體、氣體和超臨界流體範圍的對比溫度和對比壓力下，給定合理範圍並隨機亂數分布的大量PR-EOS參數如(a,b,ω)或是(Tc,Pc,ω)，如此便可透過PR-EOS計算各種熱力學數據如臨界點、蒸氣壓(Pvap)、標準狀態沸點(Tb)等，若再給定定壓熱容(Cp)還可計算其他熱力學狀態函數如焓、熵和自由能等。 在本研究中，我們特別想要了解ANN模型是否可用於預測處於不同相狀態的純流體性質，並辨認純流體的相變化(phase transition)。經本研究發現，簡單的ANN模型無法準確預測純流體在全相圖範圍的某些熱力學狀態函數；然而，若對純流體的熱力學數據進行分類預處理，將純流體數據根據其相狀態的不同進行分類，再對各類別分別建立ANN模型並且訓練，便能大幅提升模型預測準確性，降低模型預測值與PR-EOS計算值之間約25%~73%左右的誤差。我們更進一步研究使用機器學習(machine learning)方法，對純流體數據自動進行相狀態的分類，本研究應用一種名為k平均聚類(k-means clustering)的非監督式學習演算法進行分類，便可對純流體數據達到約95%以上的分類準確度。因此綜合以上結果，我們藉由結合k平均聚類以及ANN訓練模型，便可自動化預測純彭羅賓森流體在全相圖範圍中的各種熱力學性質。
並列摘要
In this work we develop reliable artificial neural network (ANN) models for the prediction of thermodynamic properties and phase equilibrium of pure fluids. In order to provide a large data set for training the models, we use the Peng-Robinson equation of state (PR-EOS) to generate thermodynamic properties of pure fluids, including critical point, vapor pressure (Pvap), normal boiling temperature (Tb) and other thermodynamic state functions (enthalpy, entropy, and free energy, etc.) by giving a large number of randomly distributed fluids parameters in PR-EOS such as (a,b,ω) or (Tc,Pc,ω) with specific constant pressure heat capacities (Cp) under specified ranges of environmental factors: reduced temperature and pressure (Tr,Pr). One point of particular interest is whether ANN can recognize phase transition and correctly predict the properties of a fluid in different phases (i.e., vapor, liquid, and supercritical). It is found that a simple ANN fails to model some thermodynamic state functions of the fluid over the whole phase diagram; however, with preprocessing, i.e., classification of data into different groups based on their phase differences, significantly enhanced accuracies of prediction and lowered deviations by about 25%~73% between the predicted and PR-EOS calculated values are observed. We further examine the possibility of machine learning for classification of the data into proper phases. For this purpose, we apply the unsupervised learning algorithm called k-means clustering, which provides phase classification with accuracy higher than 95%. Therefore, combining k-means clustering and ANN allows for prediction of the thermodynamic properties of Peng-Robinson fluids over the whole phase diagram.