In this paper, Artificial neural network training by genetic algorithm is applied to modeling the governor-turbine system using a big system test data. Three different data sets are used for training, testing and validation. The validation data used to stop GA training to prevent data over fitting. The studied system is a standard nonlinear thermal device implemented for stabilizing the system frequency due to load changes. Genetic algorithm is used to finding the optimum
values for the weights and bias matrix of the fixed structure neural network, which provide the minimum error. Varying the model of the network can be fulfill by affecting the connections between neurons and the number of hidden layers. Different number of hidden layers with the same number of neurons are applied to determine the best structure for governor
control problem. A comparative study is used for different neural network structures using mean square error and the time of optimization to achieve the best hidden layer frame. The studied system is represented by a set of nonlinear differential and algebraic equations. MATLAB software is implemented for solving the proposed system equations.