Abstract: Traditional home energy management system models are mostly based on forecasting or scenario generation and their offline day-ahead optimization is performed through the optimization algorithms. But it is difficult to solve the problems of flexible interactive resource uncertainties such as photovoltaic outputs and load demands. To address the above issue, a low-carbon optimal and real-time management strategy for home energy considering carbon trading under the non- prediction mechanisms is proposed. Under the non-prediction mechanisms, the action state spaces of loads are defined through the deep Q network algorithm with a multi-agent system established for management. Secondly, considering the constraints of load operation conditions and system carbon emissions, a real-time management model of the system is proposed taking as the goal. Solved by the multi-agent system, the model achieves the unified effect of electricity bill carbon trading cost and satisfaction punishment. Ultimately, by comparing with the heuristic algorithm and analyzing different parameters, the generalization and robustness of our real-time management strategy under the non-prediction mechanisms are verified.