應用時空深度學習模型建立中尺度紊流熱傳系統數值模擬之替代模型

Abstract

本研究主旨在應用深度學習技術加速電池熱管理系統最佳化設計流程。透過建立圖像資料集，設計並訓練神經網路，建立可以替代傳統數值模擬方法的溫度場、速度場分佈預測模型。此替代模型以低於2 %之均方誤差，預測任意入風口條件對物理場之影響，耗時僅需數值模擬0.1 %以下。 大型電池儲能系統需要熱管理系統維持穩定的工作溫度，避免效率下降與潛在運維風險。傳統上以實驗與數值模擬為驗證熱管理系統設計成效的主要方法。然而，針對大型儲能系統，上述二方法往往曠日廢時，甚至經濟上不可行。本研究應用神經網路模型於數值模擬設計流程，縮短設計驗證耗時。首先，自動化數值模擬蒐集擬穩態資料後，測試以單層全連接層預測不同入風口風速下電池最高溫與最大溫差之變化。結果顯示，少量的資料造成神經網路過擬合問題嚴重，易於因起始條件改變、邊界條件超出訓練資料範圍而失效，推論結果違反物理法則。 為解決資料量不足問題，本研究以暫態模式重建資料集，並基於ConvLSTM建構可捕捉時空特徵的深度學習模型。訓練後的模型經質性、量性及長時距誤差傳播分析，證得與傳統數值模擬相當的預測準確度，均方誤差在各測試資料中最高不超過2 %，而二者運算速度差4300倍以上，在17秒內推論出需時20小時以上的數值模擬結果。此時間差打破設計迭代與實時監控的數值模擬時間瓶頸，未來應用於熱點分析、設計驗證、自動化熱管理調控等場景，皆具極高的價值。

This study applies deep learning techniques to accelerate the optimization design process for battery thermal management systems. By establishing image datasets and designing neural networks, prediction model for temperature and velocity field distributions was developed, serving as substitutes for traditional numerical simulation methods. The surrogate model predicts the effects of various inlet air conditions on physical fields with a root mean square error below 2%, while requiring less than 0.1% of the computational time needed for numerical simulations. Large-scale battery energy storage systems depend on thermal management systems to maintain stable operating temperatures, preventing efficiency losses and potential operational hazards. Traditional validation of thermal management system designs relies primarily on experimental testing and numerical simulation. However, for large-scale energy storage systems, these approaches are often prohibitively time-consuming and economically impractical. This study integrates neural network models into numerical simulation design workflows to reduce design verification time. Initially, automated numerical simulations collected quasi-steady-state data to test single-layer fully connected networks for predicting maximum battery temperature and temperature differential variations under different inlet air velocities. Results revealed that limited datasets caused severe overfitting in neural networks, making them prone to failure when initial conditions changed or boundary conditions exceeded training data range, leading to physically unrealistic inference results. To overcome data limitations, this study reconstructed the dataset using transient simulations and developed a ConvLSTM-based deep learning model capable of capturing spatiotemporal characteristics. Following comprehensive qualitative, quantitative, and long-term error propagation analyses, the trained model demonstrated prediction accuracy equivalent to traditional numerical simulations. Mean square errors remained below 2% across all test datasets, while achieving computational speeds over 4,300 times faster than conventional methods, generating results equivalent to 20-hour numerical simulations within 17 seconds. This substantial time reduction eliminates numerical simulation bottlenecks that constrain design iteration and real-time monitoring capabilities. Future applications in thermal hotspot analysis, design validation, and automated thermal management control represent significant opportunities for practical implementation.