A model of power demand represents the foundation of any intelligent Energy Management System, and its accuracy is the key factor determining the performance of such system. In order to improve the accuracy of the modeling process, a multi-model approach based on a Hierarchical Clustering of similar load behaviors is presented. The clustering algorithm joins similar data subsets in groups that are modelled separately using Adaptive Neuro-Fuzzy Inference Systems. Thus, each of the obtained models addresses only the characterization of one behavior, which provides better accuracy than classical approaches based on a single model, in addition to being easier and faster to train. During the training process of the models, an input selection technique based on Genetic Algorithms is proposed to search and select the best combination of inputs. The use of search algorithms allows to reduce the complexity of this task while maintaining the system performance, which represents a significant time saving of expert staff. The proposed approach is validated by means of experimental data from an automotive manufacturing plant. In addition to improving the forecasting accuracy, this methodology automates the segmentation of the load profiles into models and the selection of their inputs, as well as improving parallelization to effectively reduce the computation time.