Developing a novel layer network structure for a LSTM model to predict mean monthly river streamflow

Abstract

In this research, novel innovative DDN layer network structures by hybridizing double-LSTM model with an addition layer (+) (i.e., 2LSTM and 2LSTM + layer network models) are developed purposefully to enhance prediction performance of the mean monthly Maroon River streamflow (MRSF_m) in Iran from October 1987 to September 2017. For doing so, to select the most effective parameters on MRSF_m, the Pearson’s correlation coefficient (PCC) and Cosine amplitude sensitivity (CAS) as features selection process are carried out for potential meteorological variables in the study area (i.e., average monthly temperature (T_m), evaporation (ET_m), and precipitation (P_m)) and target (MRSF_m). The results show that T_m and ET_m have an insignificant influence on MRSF_m, thus, only P_m is used as the most effective input variable in predicting MRSF_m. Due to a well-balanced network model’s structural outline in the suggested novel hybrid 2LSTM + model, it accordingly yields to a suitable total learnable parameter (TLP) compared to ordinary standalone LSTM and GRU as the benchmark models developed in the similar meta-parameters. This model under the optimal meant meta-parameters tuned i.e., state activation functions (SAF) = tanh-softsign, numbers of hidden neurons (NHN) = 75, dropout rate (P-rate) = 0.5, performs best among the models with an R² of 0.68, NSE of 0.63, PBIAS of 41%, KGE of 0.79, and RMSE of 19.24 m³/s. Comparatively, a standard gated recurrent units (GRU) and LSTM as benchmark models using the optimal scenario generate the following results: R² are 0.57 and 0.67, NSE are 0.53 and 0.61, PBIAS are 109 and 49%, KGE are 0.63 and 0.79, and RMSE are 21.11 and 19.32 m³/s, respectively. Generally, in all models, in the equal NHN, rising P-rate value reduces convergence time.