Abstract:
The Tarkwa mining area, which forms part of the Ankobra River basin has over 100 years of
history of mining, predominantly gold and manganese. Two main rivers that drain the area are
the Huni River to the north and the Bonsa River to the south. Over the past decade, there has
been a proliferation of boreholes and hand-dug wells in the area, in response to the growing
deterioration of the surface water that served as a source of water supply for the Tarkwa
Township. Geophysical investigations have been conducted in the area to locate features that
are related to groundwater occurrence, and the results of these individual studies are stored in
the archives of the mining companies. In this thesis, these individual results are collated and
re-analysed as one composite document to create a more unified database for a better
hydrogeophysical investigation.
The overall aim of this PhD thesis is to integrate available historical ERI geophysical data to
develop a conceptual groundwater potential model and also develop a hybrid groundwater
level prediction model using signal decomposition and an optimised extreme learning
machine.
Three different signal decomposition techniques: wavelet transform (WT), empirical wavelet
transform (EWT) and variational mode decomposition (VMD) were initially applied to
hydrometeorological data to denoise input data (precipitation, temperature and evaporation).
Afterwards, the denoised signals were used as input and fed to the self-adaptive differential
evolutionary extreme learning machine (SaDE-ELM) algorithm to create the resulting hybrid
models of WT-SaDE-ELM, EWT-SaDE-ELM and VMD-SaDE-ELM.
Prior to the development of the hybrid groundwater level model, the impact of data partition
on groundwater level prediction was explored. This was done by testing five partition
percentages using the traditional hold-out cross-validation, popularly known as the “train-test”
data partition. The partitions used are 50-50, 60-40, 70-30, 80-20 and 90-10 representing
training and testing percentages in each case. These partitions were tested on four different
traditional artificial neural network methods, namely: Backpropagation Neural Network
(BPNN), Radial Basis Function Neural Network (RBFNN), Generalised Regression Neural
Network (GRNN) and Group Method of Data Handling (GMDH). Statistical indicators of
Root Mean Square Error (RMSE), Normalised Root Mean Square Error (NRMSE) and Mean
Absolute Percentage Error (MAPE) was used to make an objective evaluation.
Results indicate that the hybrid models created performed excellently, as compared with very
potent metaheuristic methods in literature like particle swarm optimisation - artificial neural
network (PSO-ANN), genetic algorithm – artificial neural network (GA-ANN) and SaDE ELM. Thus, the superiority, consistency and efficacy of the hybrid models are attributed to
the application of the signal decomposition on the input variable coupled with its application
to an optimised extreme learning machine