The problem of day ahead load forecasting has been addressed in many forms, particularly, using machine learning approaches. These approaches range from simple linear models to the intrinsically nonlinear multilayer perceptron, which captures the nonlinear relationships between power load and external variables including weather conditions, ambient light and others, as building occupancy for the case of building level forecast. Support Vector Machines and Gaussian Process regression combined with Mercer’s kernels appear to be effective approaches able to model these nonlinear properties and that have, at the same time, a moderate computational burden. The GP approach is taken from a Bayesian perspective that produces a probabilistic model of the forecast, which provides with more information about the goodness of each prediction.   In the research to be presented in this Symposium we introduce a special type of structure that uses multitask Gaussian Processes for day ahead prediction. This structure can forecast an arbitrary number of prediction points using a block algorithm that considers not only the possible nonlinear relationships between the input data and the forecast, but also the relationships between the different prediction points of a single day, and it keeps a uniform detection error along the predicted day. The approach also produces a confidence interval of the prediction error that is in high agreement with the actual error measured a posteriori.   Authors will summarize the theoretical development plus a set of compared experiments in synthetic and real data.