A method of forecasting a resource load for consumption at a site, e.g., an electrical load of a site. The method includes receiving historical load data and historical ambient condition data that are time series data pertaining to a site and generating additional data from the received data. The method includes building a best sub-model for each of multiple forecast intervals. The building includes clustering in parallel training portions of the historical load data and the additional data, training possible sub-models using the clustered training portions, verifying forecasted loads output from the possible sub-models against verification portions of the historical load data and the additional data, and determining a first subset of parameters for the best sub-model based upon accuracy of the forecasted loads. The method includes forecasting a resource load at the site for each of the forecast intervals using an ensemble of the best sub-models.