Internal Value Alignment in Large Language Models through Controlled Value Vector Activation