
International Journal of Control and Automation
Vol.8, No.5 (2015)
Copyright ⓒ 2015 SERSC 327
dimension. Checking whether wind power time series have chaotic properties needs to
compute the maximum Lyapunov index of the time sequence of [27], which is only
available modeling phase space reconstruction method when it is bigger than zero.
Using a certain wind farm 2011-1-1 0:00 to 2011-1-1 23:45 (a resolution of 15
minutes), a total of 1440 sets of data as the training sample, calculate the wind speed time
series phase space reconstruction parameters as Table 1.
Table 1. Wind Speed Time Sequence Parameters in a Wind Farm
3.2 GRNN Prediction Model based on Chaos Phase Space Reconstruction
Generalized regression neural network (GRNN) is a branch of RBF neural network, its
way to get the relationship of the data is different from the interpolation and fitting, it can
change network directly under the same structure through sampling or data by calculated
and don't need to recalculate the parameters. Unlike typical BP network, simulation effect
of NN network is better than BP neural network. It has better prediction effect, fast
calculation and stable results. Only though a simple smoothing parameter, don't need
training cycle process [28-29].
GRNN is a feed forward neural network model based on nonlinear regression theory.
Unlike typical BP network, it approaches by activation of neurons function, that is, the
function of input vector value approximates by a function of neuron vector which
corresponding with its neighborhoods map to form it, GRNN network is composed of
input layer, hidden layer and output layer, as shown in Figure 1.
The hidden layer radial base layer, adopting Gaussian transformation function to
control the hidden layer output, thereby inhibiting activation of output units, in the input
space, Gaussian function is symmetrical about acceptance domain. The network output
impact by input neurons exponentially attenuate varies with the distance between the
input vector. In GRNN network, each training vector has a corresponding radial neurons
in the hidden layer, neurons in hidden layer to store each training vector. When a new
vector enters the network, the distance between the new vector and each unit weight
vector in the hidden layer can be calculated by the next type:
(3)
Where X is input vector, R is the dimension of X,
is the number of hidden layer
unit,
is the unit weight vector of hidden layer, dist is the distance between the input
vector and weight vector. The Gaussian function output of hidden layer accord to the
following formula:
1 1 2
exp[ (|| || ) ]a dist b
(4)
Distance can be adjusted by the type of computer:
(5)