Abstract: To address the rapid cascading overload problem in active distribution networks caused by the sharp drop of renewable energy output, sharp surge of load demand, and coupling of multi-source uncertainties under extreme heat and calm wind disasters, this paper proposes a risk-constrained safe reinforcement learning preventive control method. First, a cascading overload-constrained Markov decision process model incorporating extreme meteorological features is constructed, where generation rescheduling, network reconfiguration, and load shedding are jointly modeled as control actions to achieve unified characterization of risk and economic cost. Second, a spatiotemporal feature extraction module integrating edge-condition convolutional neural networks and long short-term memory networks is designed to enhance the perception of spatiotemporal correlations between topological coupling and state evolution. Finally, an interior-point policy optimization algorithm is employed to construct a safe deep reinforcement learning (DRL) controller, where logarithmic barrier functions enable implicit satisfaction of safety constraints, thereby avoiding the manual tuning of penalty factors required in the conventional penalty function methods. Monte Carlo simulation results based on the IEEE 33-bus system and a provincial 110 kV distribution system demonstrate that the proposed method reduces the average load loss by 67.3% and decreases the number of line overload occurrences by 82.5%; the constraint violation rate is maintained below 3% for the IEEE 33-bus system and below 4% for the 110 kV practical system, both satisfying the engineering-acceptable threshold of 5%. Meanwhile, the decision time remains at the millisecond level. Compared with DC power flow optimization and standard DQN and A3C algorithms, the proposed method exhibits significant advantages in mitigating cascading overloads, ensuring constraint safety, and handling uncertainties.