Towards on-board learning for harvested energy prediction