Time Series Prediction Using Complex-Valued Legendre Neural Network with Different Activation Functions

In order to enchance the flexibility and functionality of Legendre neural network (LNN) model, complex-valued Legendre neural network (CVLNN) is proposed to predict time series data. Bat algorithm is proposed to optimize the real-valued and complex-valued parameters of CVLNN model. We investigate performance of CVLNN for predicting small-time scale traffic measurements data by using different complex-valued activation functions like Elliot function, Gaussian function, Sigmoid function and Secant function. Results reveal that Elliot function and Sigmoid function predict more accurately and have faster convergence than Gaussian function and Secant function.