Deep reinforcement learning is increasingly applied to the optimal scheduling of integrated energy systems (IES). To reduce training episodes, speed up convergence, and improve sample efficiency, an improved DDPG algorithm is proposed, incorporating multiple environment instances and feature score-based experience sampling. Multiple environment interactions provides richer experience and accelerates convergence. Meanwhile, data features are quantified, and sampling is guided by feature scores to enhance sample utilization. Simulation results validate the method’s effectiveness in improving convergence speed and sample efficiency.