Assistant Professor Savannah State University, Georgia, United States
Abstract Submission: This study developed a hybrid model consisting of two powerful deep learning algorithms for predicting concentrations of heavy metals in river water. The hybrid model integrated convolutional neural networks (CNN) and long short-term memory (LSTM) for the prediction of heavy metals. The proposed model was trained using a robust dataset from 2013 to 2023 for the Georgia streams. The hybrid CNN-LSTM model was developed to predict different contaminants, such as calcium and magnesium. The model was designed in Python using the TensorFlow platform. The training procedure of the hybrid CNN-LSTM model involved several key steps, such as preprocessing of the data. The data was divided into training and test sets with ratios of 75% and 25%, respectively. Random search cross-validation was used to find the best architecture and optimal hyperparameters for the hybrid CNN-LSTM model. The results showed a successful training procedure for the hybrid CNN-LSTM model. The training and test losses for predicting heavy metals were both in the same range and very close to zero. The hybrid CNN-LSTM model indicated a high accuracy for predicting heavy metals in river water compared to the regular single models. With a coefficient of determination of greater than 0.9 in all simulations and the relative error less than 10%, the CNN-LSTM model outperformed the regular single models. By an integrated data preprocessing approach and reducing the need for extensive manual intervention, the hybrid CNN-LSTM deep model is a highly accurate predictive tool to monitor heavy metals in river water.
