Extreme weather conditions, characterized as high-impact low-probability (HILP) events, pose significant threats to distribution network (DN). Soft open points (SOPs) have emerged for precise power flow regulation and voltage establishment, and hence can be used for post-fault load restoration to improve the DN resilience, incorporating scenarios generated from weather profiles. In this paper, an extreme weather risk-averse planning method for SOPs is proposed. A two-stage scenario-based stochastic programming (SBSP) model is established to minimize expectation and conditional value-at-risk (CVaR) of load shedding and network loss. Due to the integer variables introduced by DN reconfiguration constraints in the operational stage, the classic Benders decomposition algorithm is inapplicable. To address this challenge, we develop a novel generalized Benders decomposition (GBD)-based solution algorithm, designed to improve the computational efficiency on large-scale cases. Lift-and-project (L&P) cutting plane is employed to derive Benders cuts through the convex hull of mixed-integer second-order cone programming (MISOCP) subproblems. Finally, the effectiveness of the proposed method is demonstrated by numerical experiences on 33- and 123-node test DNs.