To enhance nuclear safety, an autonomous anomaly response system was proposed, which integrated a deep reinforcement learning (DRL) framework with a surrogate model to replace computationally expensive nuclear power plants (NPPs) simulation codes. The surrogate model, constructed with gated recurrent unit (GRU) and temporal convolutional network (TCN) blocks, extracts both temporal and spatial features from time-series data, significantly reducing computational demands while accurately approximating reactor behavior under various anomaly conditions. By integrating the surrogate model with DRL, the system accelerates decision-making and ensures a balance between short-term anomaly mitigation and long-term system stability. A case study using the high temperature gas cooled reactor (HTGR) analytical code, ACCORD, demonstrates the effectiveness of the proposed system in restoring reactor states from anomalies to near-normal operation, offering a promising approach to enhancing autonomous anomaly response in NPPs.