Abstract:
The exponential increase in world population has increased the energy demand.
This has resulted in the accelerated use of more conventional energy resources like
fossil fuels which has caused the exhaustion of these resources. This has also
triggered an increase in pollution thus harming the environment, leading to global
warming. So there is an urgent need of finding alternate energy resources that are
more environment friendly and are to meet out increasing energy demands.
Accordingly, renewable energy is the best option for this purpose. Unambiguously,
wind energy is the most obvious option due to its abundance everywhere and all the
time. The only drawback of using wind as reliable energy resource is its dependence
on natural factors, especially wind speed which depends on climatic conditions and
varies from place to place. The wind turbines harness mechanical energy from the
kinetic energy of wind and convert it into electrical energy. Fortunately, the
accurate estimation of wind speed is possible. The stochastic nature of wind speed
presents a challenging situation in estimation of wind power output. In this research,
the mechanical power of wind turbine (WT) has been estimated using nonlinear
input variables like wind speed (v), angular speed of WT blades (ωr), pitch of blades
(β) and power coefficient (CP). The estimation performed using feed-forward back
propagation neural network (FFBPNN) , recurrent neural network (RNN) and
(ANFIS MODAL). Results are then compared with all networks. Five cases are
considered for neural network which are designed based on number of hidden layers,
different learning rates and activation functions, Both networks are implemented
under similar conditions. The networks are trained using scaled conjugate gradient
(SCG) algorithm. The primary factor used for the performance evaluation of
networks is root mean square error (RMSE) while training time is considered as
secondary factor. While in case of ANFIS cases design on the basis of input ,output
membership function type ,number of input membership function for each input
variable ,in this case primary factor regarding performance evaluation become
(RMSE), The best performance is achieved within NN from FFBPNN using two
hidden layers containing 100 tan-sigmoid (tansig) and 50 log-sigmoid (logsig)
nodes respectively with the RMSE value of 0.49% while as compared with ANFIS
modal best performance achieved using Gaussian input membership function
0.00175429 , 0.17 % using three inputs membership function while linear output
membership function.
