Table 3: Structure of ANN model.

No. of hidden layers

A combination of input-hidden-output layer with respective number of neurons is used to defined the ANN model [16,17]. For instance, 5-8-1 for a given ANN model means that input layer has 5 neurons, hidden layer has 8 neurons while, output layer has 1 neuron.

No. of neurons in a given layer

The required number of neurons in the input and output layer depends on the given problem. However, the number of neurons in the neurons must be chosen carefully. If the number of neurons are not selected optimally, it will lead to under- or over-fitting in the given dataset [18]. Various studies have been carried that suggested to choose the number of neurons in between 5~10 [19-28].

Activation function

It is a non-linear transformation on given the input signal. In other words, it makes a decision to activate and deactivate a given neuron. On the basis of its performance, it is considered a vital part of an ANN model. It is important to mention that a network without having an activation function behaves like a model for linear regression, thus, cannot handle the complicated tasks [29]. Some commonly used activation functions are:

ReLU x  = max 0,x (1)

Tanh x  =  2 1 -  e  2x  - 1 (2)

Sigmoid x  =  1 1 -  e  x (3)

Loss estimation function

The loss function is usually determined using the real-world problem, carries the interpretation real-time data. The root means square (RMS) and absolute mean error (AME) are two commonly used ways to estimate the difference between the predicted vector and the target value [8]. Their mathematical expressions are given below:

RMS =  i = 1 n x i  -  x t 2 n (4)


AME =  i = 1 n x i  -  x t n (5)

where i a sample index, xi is a predicted value and xt is the target value.