A method and associated systems for a self-learning energy switch. The switch creates an array of cognitive models for each candidate energy source. Each array returns a probability that its corresponding source is the most cost-effective and operationally suitable energy supplier at that time. Each model in an array contributes to the array's returned probability as a function of a corresponding class of decision-making factors. The system fine-tunes the models by weighting them as functions of extrinsic evidentiary information that may imply future behavior of the decision-making factors and combines each model's returned probabilities to select an optimal energy source. The system then automatically routes power from the optimal source to a consumer's energy-consuming premises. This self-learning procedure repeats indefinitely, continuously tuning the models in response to identifying additional extrinsic evidence and reasons why the system's previous energy selections were either optimal or non-optimal.