The selection of parameters may be dependent on the problems at hand since the physical significance of the parameters are usually related to the problems. However, you can choose a set of different parameters from literature, for example, no. of hidden layers, no. of nodes per hidden layer, learning rate, etc.
Most important issues in ANN is to have good and reliable training, testing and validation data. Normally, in case of supervised ANN you have to which algorithm to use. A two-layer feed-forward network with sigmoid hidden neurons (e.g. 10 neurons) and linear output neurons (e.g. 12 neurons), can fit multi-dimensional mapping problems arbitrarily well, given consistent data and enough neurons in its hidden layer. In addition, the training algorithm plays an important role you can use either "Levenberg-Marquardt", "Bayesian regularization", or "scaled conjugate gradient". The training algorithm selection depends on certain constraints such as lack of memory in this case use the last one.The second training algorithm work best with noisy data but takes more time. The first one takes more memory but less time.Finally, number of iterations plays a crucial role in either ANN converge toward global or local optimal solution as long run can result in over-fitting problem.
The division of data used into training, testing and validation are important and also two-layer feed-forward network is important. You may try ANFIS or some other soft computing models, these have better results than only ANN.