Model orthogonalization and Bayesian forecast mixing via Principal Component Analysis