Publication

Abstract : Modelling, predicting, and forecasting hydrological phenomena like groundwater have been one of the prominent applications of artificial intelligence techniques. Using Multi-Layer Perceptron and Deep Neural Networks (DNN), a deep neural multi-model ensemble (DBM2E) framework has been introduced to model groundwater levels. Three training approaches—Levenberg–Marquardt, Bayesian Regularization (BR), and Scaled Conjugate Gradient, are utilized to capture the groundwater fluctuations over 140 sites in Kerala, India. Dynamic variables used are the Normalized Difference Vegetation Index, soil moisture, rainfall, temperature, and lag-1 groundwater levels, which span the years 1999–2020 (22 years) and the months of January, April, August, and November. By bringing together these three learning algorithms, the multi-model ensemble (MME) average approach improves upon the conventional methods. Approximations are also enhanced using DNN having at most three layers, with varying numbers of neurons in each layer. For each location and month, the most effective method is chosen from the three learning algorithms, four combinational techniques, and nine neural network models using a developed performance index. On a monthly average, 31 locations improved with the use of MME averaging, whereas 48 locations improved with the use of DNN. The BR method was observed to be the most effective in the comparison studies, while single-layer networks produced the best modelling results overall. Multi-model combinations with BR produced superior outcomes than its other counterparts. The research offers inference in different forms of multi-models and DNN, which yield sufficient results than compromising with a single model and one hidden layer network.