Improving reservoir inflow prediction via rolling window and deep learning-based multi-model approach: case study from Ermenek Dam, Turkey

Feizi H., Apaydın H., Sattari M. T., Colak M. S., Sibtain M.

STOCHASTIC ENVIRONMENTAL RESEARCH AND RISK ASSESSMENT, cilt.36, sa.10, ss.3149-3169, 2022 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 36 Sayı: 10
  • Basım Tarihi: 2022
  • Doi Numarası: 10.1007/s00477-022-02185-3
  • Dergi Adı: STOCHASTIC ENVIRONMENTAL RESEARCH AND RISK ASSESSMENT
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, CAB Abstracts, Compendex, Environment Index, Geobase, Index Islamicus, Pollution Abstracts, zbMATH, Civil Engineering Abstracts
  • Sayfa Sayıları: ss.3149-3169
  • Anahtar Kelimeler: Multi-models, Recurrent neural network, Long short-term memory, GRU, BiLSTM, sRNN, Pre-trained neural networks, NEURAL-NETWORKS, ARTIFICIAL-INTELLIGENCE, MODELS, FORECASTS
  • Ankara Üniversitesi Adresli: Evet

Özet

Reliable and accurate prediction models effectively deal with nonlinear and non-stationary characteristics of reservoir inflow and associated variables. Therefore, a Hybrid Deep Learning Inflow Prediction-Rolling Window (HDeepLIP-RW) framework is proposed in this study for efficient inflow prediction. The HDeepLIP-RW framework utilizes hybrid alternatives of four different RNN architectures bidirectional long short-term memory (BiLSTM), gated recurrent unit (GRU), long short-term memory (LSTM), and simple RNN (sRNN) by employing RW technique. Firstly, the conventional split data (CSD) technique is applied to divide the multivariate hydro-meteorological data of Ermenek Dam, Turkey, into training, validation, and testing periods. Afterward, the processed data is employed as input by standalone and hybrid models for inflow prediction. Thirdly, the RW technique is used to divide the data to train and validate the model. Finally, different standalone and HDeepLIP-RW prediction models utilizing the RW technique are also developed to compare the performance of the proposed model with counterparts. The prediction results in terms of statistical metrics and different plots manifest the applicability of models utilizing RW for inflow prediction in contrast to the CSD-built models. Furthermore, the HDeepLIP-RW based BiLSTM-GRU model outperforms all the counterpart models by demonstrating the highest performance and accuracy. The RW-based BiLSTM-GRU model reduces RMSE contrast to the CSD-based best performing by 0.300 for GRU-LSTM during the testing period. Likewise, the prediction accuracy according to NS was increased by 23.37% compared to standalone models and 19.26% according to hybrid methods by splitting the data with the RW method. The overall results are enlightening and substantiate the viability of the BiLSTM-GRU model utilizing the HDeepLIP-RW framework for inflow prediction with an equal opportunity for similar applications.