An autoencoder-based stacked LSTM transfer learning model for EC forecasting

The aim of this study is to assess, monitor, and analyzed water quality parameters recorded from reservoirs in of Saudi Arabia for Electrical Conductivity (EC) forecasting. Water quality parameters play a key role in the lives of humans, plants, and animals depending on those reservoirs for survival. In this paper, we evaluate the performance of a novel autoencoder stacked LSTM transfer learning (TL + aeSLSTM) model for EC forecasting. Although various factors such as pH, EC, CaCO 3 , Turbidity, Salinity, Sodium (Na + ), Potassium (K + ), Magnesium (Mg), Calcium (Ca2 + ), Fluoride (F − ), Chloride (Cl − ), Bromide (Br − ), Nitrate (NO 3− ), and Sulfate ( SO_4^2- ) affects EC forecasting, in this paper we utilized Total Dissolved Solids (TDS) and Na as our source dataset for EC forecasting. In order to select the most appropriate or suitable transfer learning model network for EC forecasting, a trail-by-error approach was employed where parameters e.g., hidden layers, number of neurons, batch-size, epochs, optimizers, and activation functions were varied and applied to each model. To assess the performance of the proposed model to the baseline model, we employed the most common evaluation metrics ie. MSE, MAE, and MAPE. Superior performance by the novel TL + aeSLSTM over the state-of-the-art models illustrates the capability of the model to utilize both the advantage of the autoencoder, stacked LSTM, and transfer learning capability for EC forecasting. The results show that the least value of train MAE, train MSE, train RMSE, test MAE, test MSE, test RMSE when Na is used as the source dataset for EC forecasting by the TL + aeSLSTM are 248.044, 519.292, 22.788, 206.114, 212.004, and 14.560, respectively.