Weight and Bias Initialization of ANN for Load Forecasting using Cuckoo Search Algorithm

Abstract

Artificial Neural Network (ANN) is used for electricity load forecasting for quite a time. ANN uses backpropagation for computing the gradient of the cost function. One obvious way to initialize weights and biases is to use Gaussian independent random variables, which is normalized to have zero mean and unit standard deviation. Issue with this kind of initialization is that it an exceptionally wide Gaussian distribution, not strongly peaked by any means. Another way to initialization of weights and biases for an ANN with n_in input weights would be to use random Gaussian variables with zero mean and 1/√n_in deviation. Case study utilizes half hourly electricity load data from five states in Australia to predict 48 hours ahead electricity load. In this paper a multi-objective cuckoo search algorithm is utilized for weights and biases initialization for quicker learning. The results show that the convergence time using proposed algorithm has reduced considerably as compared to Gaussian distribution initialization generally used in ANN.