DEVELOPMENT OF MULTI-VERSE OPTIMIZER IN ARTIFICIAL NEURAL NETWORK FOR ENHANCING THE IMPUTATION ACCURACY OF DAILY RAINFALL OBSERVATIONS

This research study introduces an innovative approach to infill missing rainfall data by combining the Multi-Verse Optimizer (MVO) with a feedforward neural network (FNN) to form the Multi-Verse Optimizer Feedforward Neural Network (MVOFNN). This approach aims to overcome the limitations of conventional training algorithms for artificial neural networks (ANN), which often get stuck in local optima. MVOFNN is compared against the conventional Levenberg-Marquardt Feedforward Neural Network (LMFNN) and a promising data mining-based imputation approach, the Regularized Expectation Maximization (RegEM) algorithm, to assess its reliability and feasibility in reconstructing missing daily rainfall data. The comparison was conducted by reconstructing 20% of artificially missing daily rainfall data for Kuching Third Mile Station. Optimal hyperparameters for the ANN models were determined through trial-and-error combined with 5-fold cross-validation approaches. Model performance was evaluated using the correlation coefficient and mean absolute error. The results revealed that all imputation models achieved high correlation predictions within the range of 0.9769 to 0.9797. RegEM demonstrated the best performance among the imputation approaches, followed by MVOFNN and LMFNN. While MVOFNN did not outperform the others in imputation performance, it showcased robustness, reliability, and feasibility in predicting missing daily rainfall data.