Results and discussion
In this study, different algorithms (Artificial Neural Network, Linear Regression and M5’Rules) were used.
As seen in Table 6, the optimal result from these algorithms was obtained using the ANN. The usual way that
a trained ANN is used is to feed new unknown patterns to the network and get the results. Instead of applying
this methodology a new approach is developed here that make the use of ANN dedicated software unnecessary.
Thus, in order to calculate the COP and f values of AWRS, mathematical formulations are derived from
the resulting weights and the activation functions used in the ANN. As the statistical results obtained from
both the training and testing of the ANNs were extremely good in both cases it is believed that the results thus
obtained would be accurate.
Mathematical formulations derived from the ANN model are presented here. In the following formulas the
coefficients of the input parameters are used to evaluate the Ei (summation function of neuron i) and Fi (activation
function of neuron i). These coefficients represent the weight values of the summation function of each
neuron belonging to the hidden layer of the trained network. For this purpose, and for the case of ANN model
used for circulation ratio (f) and COP prediction, five pairs of equations are required as the neural network
model has 5 hidden neurons. The activation function chosen is the log-sigmoid as shown from the Fi function.
In the output neuron, two summation functions are used as two output parameters exist which correspond to
COP and f. In the ANN model, the inputs of the network are the evaporator temperature (TE), absorber temperature
(TA), condenser temperature (TC) generator temperature (TG), poor solution concentration (Xp) and
rich solution concentration (Xr) and the output are circulation ratio (f) and COP. In order to calculate the
COP and circulation ratio values of AWRS, the following equations are derived