Boosting Performance of Recurrent Neural Networks via Task-Oriented Modifications

IŞIK Ş., Isil Y., ANAGÜN Y.

Sustainable Aviation, Springer Nature, ss.224-229, 2026 identifier

  • Yayın Türü: Kitapta Bölüm / Araştırma Kitabı
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1007/978-3-032-07678-6_40
  • Yayınevi: Springer Nature
  • Sayfa Sayıları: ss.224-229
  • Anahtar Kelimeler: Compressor Map, Compressor Performance Parameter, Deep Learning, GRU, LSTM, Mass Flow Rate
  • Eskişehir Osmangazi Üniversitesi Adresli: Evet

Özet

This paper presents some of the improvements on different types of Recurrent Neural Networks (RNNs). The LSTM and GRU models were applied to three distinct scenarios, including time-series prediction, multi-label classification, and binary classification. The compressor maps were used to examine the performance of generating predictions with RNNs. The results demonstrate that employing an exponential function significantly enhances the accuracy and predictive capability of models by the smoothing output of the forgetgate. The models’ performance was additionally evaluated using the MNIST dataset and the heart failure dataset. For the task of predicting heart failure (a binary classification problem), we achieved accuracy results of 86.97% and 88.59% using the original and improved versions of LSTM models, respectively. Moreover, the highest R-squared values are 0.9912 and 0.9905 for predicting the corrected mass flow rate.