Neural network analysis of fin-tube refrigerating heat exchanger with limited experimental data

Arturo Pacheco-Vega, Mihir Sen, K. T. Yang, Rodney L. McClainInternational Journal of Heat and Mass Transfer 2001 · DOI

TL;DR

This paper introduces a novel cross-validation methodology to estimate the error bounds of Artificial Neural Networks (ANNs) used for modeling refrigerating heat exchangers, especially when trained on limited experimental data. This approach helps identify regions where more data is needed, improving the reliability of ANN predictions in complex thermal systems.

What Problem Does This Paper Address?

Heat exchangers, particularly those used in refrigeration and air-conditioning, are incredibly complex. Their performance is influenced by intricate geometries and physical phenomena like condensation and frost formation, making them difficult to model using traditional first-principles approaches. Manufacturers typically rely on experimental data to create correlations for design and selection. However, collecting comprehensive experimental data across all possible operating conditions is prohibitively expensive and time-consuming. This leads to situations where ANNs, a powerful tool for modeling such complex systems, are trained on limited datasets. When ANNs are used to predict performance outside the range of their training data, or in regions where data is sparse, their accuracy can degrade significantly. The problem this paper addresses is how to reliably estimate the prediction error of these ANNs when trained on limited data, and how to guide future data collection efforts.

Methods

The authors employ Artificial Neural Networks (ANNs), specifically a feedforward architecture with a backpropagation algorithm for training. The core innovation lies in their proposed methodology for error estimation, which is based on a variant of the cross-validation technique.

Here's how it works:

  1. Training with Leave-One-Out: For a dataset of M experimental points, the ANN is trained M times. In each training iteration, M-1 data points are used for training, and the remaining single data point is "left out."
  2. Calculating Cross-Validation Error (Scv): After training, the ANN predicts the value for the left-out data point. The percentage error between this prediction and the actual experimental value (Scv) is calculated. This Scv value quantifies how important that specific data point was for the overall model's accuracy. A high Scv indicates the point is crucial and its absence significantly impacts predictions.
  3. Modeling the Error Surface: The M Scv values are then used to create a "neural network model of the error surface." This model can then estimate the expected error for new, unseen data points.
  4. Validation: The methodology is validated by selecting specific test experiments and comparing the estimated error from their model against the actual error of the ANN when predicting those test cases.

Datasets

The study utilizes a dataset of 38 experimental runs provided by the manufacturer, Tyler Refrigeration Corp. These experiments cover various multi-row, multi-column fin-plate type heat exchangers with staggered tubes. The experiments varied parameters such as length (L), fin spacing (δ), number of rows (Nrow), number of columns (Ncol), number of circuits (Ncir), and tube-center distances (xa, xb). The input variables to the ANN included mass flow rate of air (ma), inlet air temperature (Tindb), inlet humidity ratio (win), inlet refrigerant temperature (Tin), and the scaled geometrical parameters. The output variable was the total heat rate (Q).

Key Insights and Ideas

  • ANNs can model complex thermal phenomena with surprising accuracy, even with limited data: The paper demonstrates that a well-trained ANN can correlate experimental data with errors comparable to the measurement uncertainty itself, outperforming traditional correlations.
  • Data scarcity is a critical bottleneck for ANN reliability: The performance of an ANN is directly tied to the quantity and distribution of training data. Insufficient or poorly distributed data leads to significant prediction errors, especially in regions far from the training set.
  • Cross-validation is a powerful tool for error estimation in data-scarce ANN applications: The proposed cross-validation methodology provides a robust way to estimate the upper bound of prediction errors, offering valuable insights into the reliability of the ANN model.
  • Identifying "data-poor" regions is crucial for practical applications: The error estimation technique can pinpoint areas where the ANN is likely to perform poorly due to a lack of supporting experimental data, guiding manufacturers on where to conduct new experiments.

Key Results

  • The ANN model, trained on 38 experimental data points, achieved a root-mean-square error of less than 1.5% when predicting the total heat rate, approaching the experimental uncertainties.
  • The proposed cross-validation error estimation method consistently provided an upper bound for the actual ANN prediction errors. For three validation tests, the estimated errors were 0.761%, 23.3%, and 48.78%, while the actual errors were 0.112%, 6.19%, and 14.21%, respectively. This indicates the method is conservative and reliable.
  • The error estimation procedure, visualized as a bar diagram, showed error values ranging from 0% to 60%, highlighting specific data points that were more influential or poorly represented in the training set.

Limitations and Future Work

  • Limited Data: The primary limitation acknowledged is the inherent scarcity of experimental data, which is a common issue in industrial applications. While the methodology addresses this by estimating errors, it doesn't magically create more data.
  • Generalization Beyond Training Domain: The paper notes that ANNs are likely to perform poorly when used for predictions outside the range of available training data. The error estimation method helps quantify this risk.
  • Future Work: The authors suggest that their methodology can directly help manufacturers plan new experiments by identifying specific regions where more data is needed to improve ANN predictions and overall model reliability. Further research could explore more advanced ANN architectures or data augmentation techniques to further enhance performance with limited data.

Summary updated

BibTeX
@article{pachecovega2001neural,
  author = {Arturo Pacheco-Vega and Mihir Sen and K. T. Yang and Rodney L. McClain},
  title = {{Neural network analysis of fin-tube refrigerating heat exchanger with limited experimental data}},
  journal = {International Journal of Heat and Mass Transfer},
  year = {2001},
  doi = {10.1016/s0017-9310(00)00139-3},
}
Figure