類神經網路應用於三重微流道之散熱分析

Application of Neural Network on Heat Dissipation Analysis of Three-Layer Microchannels

10.6840/cycu201600087

吳彥廷

三重微流道之散熱分析 ； Application of Heat Dissipation Analysis of Three-Layer Microchannels

中原大學機械工程學系學位論文

2016年

碩士

許政行

繁體中文

本研究的目的是探討三重微流道散熱性能之最佳化，依照參考論文中之模型將其三維度的尺寸等比例放大，並藉由無因次分析此熱傳模型，推導出各項相關的無因次參數，再將固定各無因次參數值所求得的熱傳參數值帶入Ansys Fluent 進行數值模擬分析，以求出各模型的溫度差、壓力差…等熱傳性能。 利用倒傳遞類神經網路原理，將數據庫匯入Matlab 的 nntool程式之中，訓練網路系統，藉此系統模擬的輸出以達到同時最佳化溫差及壓力差的結果。 本論文使用Ansys Fluent軟體做數值模擬分析，使用原模型的設計，並將其三維度的尺寸階級等比例放大至1.8倍，產生10組模型，以固定各無因次係數之設定數值，推算相關的熱傳參數，並代入Ansys Fluent軟體進行模擬分析，結果得到相同的無因次熱傳數值，這結果驗證所使用的無因次參數是足以描述本研究的物理特性。 後續使用類神經網路訓練網路系統，比較藉由無因次分析配合Ansys Fluent 模擬所得以及 Matlab nntool 訓練後所得之最佳化數值，以確定神經網路之訓練成效。 最後再藉由改變各項熱傳參數，首先變化熱通量，帶入類神經系統中，得到測試結果與Ansys Fluent 模擬誤差百分比約介於為1~3%，其次則是改變流道入口水流的速度，測試結果之誤差百分比約為1%，以此驗證類神經網路於訓練完成後在最佳化參數之性能。並且比較執行兩種模擬分析方法所花費的時間，由Ansys Fluent 模擬得到的結果，花費時間約為10分鐘，而由訓練過後的類神經網路分析得到結果，花費時間只需要2秒鐘即可得到結果。

In this study we focus on the performances optimization for a three-layer micro channel, which was based on the model used in the previous study. With the help of dimensional analyses and Back-Propagation Network, we could predict and optimize the performances, pressure drop and temperature difference, of the proposed multi-layer micro channel. Firstly, the dimensional analyses were used in the present study. A dimensional analysis was performed by choosing the main geometry characteristics and important variables related to the heat transfer of the micro channel. Then by executing the dimensional analyses, the key dimensionless parameters of the present model were obtained. In order to verify the completeness of the dimensional analysis, we proportionally enlarge the geometry dimensions of our base model step-equally to a ratio 1.8 (which produces 10 sets of sub-models); then fix the values of dimensionless parameters obtained from pervious stage and calculate the corresponding values of heat transfer-related parameters for each sub-model respectively. By conducting simulations of these sub-models, the results showed that the values of their dimensionless heat transfer parameter were equal to each other as expected, which indicated sufficient dimensionless parameters were included in the present study. Secondly, by applying Back-Propagation Network and predicting technique, i.e. Matlab nntool, for the neural fitting of this model, ten sets of data from previous studies were used to fulfill the network training. After the training, the network system can perform the optimization for heat transfer-related parameters in the base model case. Finally, the trained Back-Propagation Network system was used to predict the sub-models’ cases by changing the heat transfer-related dimensionless parameters in equal difference. The differences of heat transfer rates and pressure drops obtained by the network were within 1-3% and 1%; respectively, in comparison to the values computed by the Ansys Fluent; the method also save time in analyses 300 times faster than Ansys Fluent.(10 minutes vs. 2seconds )

1. 蔡瑋楓(2017)。帶翅種子氣動力特性應用在風力機葉片之仿生設計。中原大學機械工程學系學位論文。2017。1-92。