<p>Demand response is an effective method for demand side management in smart grid. In this paper, we establish the smart grid as a multi-agent system, in which the power supplier acts as the power supply agent (PSAG), the Power Market Scheduling Center acts as the controller agent and users act as the load agents (LAGs), respectively. Then, a social welfare maximization model based on real-time hybrid demand response that combines prices with incentives is proposed to optimize LAGs’ electricity consumption behavior and pursue maximum social welfare, namely the sum of all agents’ welfare. In this model, multi-energy generation sources, uncertainties of LAGs’ electricity consumption and renewable energy sources’ output, and LAGs’ electricity consumption deviation between real electricity consumption and the optimal electricity consumption are also considered. Since the established model ultimately expresses as all LAGs’ utility functions minus the PSAG’s cost, which does not contain the variable of electricity price, it is difficult to obtain the price by solving it directly. Previous works usually solve it by dual optimization method, which requires that the social welfare model must have the specific functional analytic equation. Otherwise, the dual optimization method will not work. Thus, data-driven reinforcement learning is adopted to deal with this challenge by taking the maximum social welfare as the indicator to obtain the electricity price directly via the interaction among multiple agents, which allows that the model may not have the specific analytic equation. Results show that the proposed hybrid demand response strategy is effective in guiding users’ electricity consumption behavior, improves social welfare and has better performance in cutting peak and filling the valley of electricity consumption when compared with the fixed pricing mechanism.</p>