Load forecasting is an attractive and complicated application of machine learning theory and algorithms. Continuous efforts have been made from both academics and industry, by using various methods such as Regression, Artificial Neural Network (ANN), Time Series Models like Auto Regressive Moving Average Models (ARMA), Gaussian Process (GP) and Genetic Algorithm (GA). The non-parametric models are not widely used in the forecasting domain, yet the promising results from the recent applications of Gaussian Process have indicated a potential value for this kind of algorithms. In this paper, we describe a very recently proposed machine learning algorithm, Twin Gaussian Process (TGP) and apply it to the load forecasting task. Different from the Gaussian Process Model, the Twin Gaussian Process uses Gaussian Process (GP) priors on both covariance as well as responses, and obtain the output via Kullback-Leibler divergence minimization between two GP modeled as normal distributions over finite index sets of training and testing examples. As a result, TGP is able to account for the correlations of both inputs and outputs. In our case study, TGP is evaluated and compared with other widely used algorithms. And experimental results show that TGP can be a useful tool for load forecasting. © 2012 IEEE.