To address the intensive computational demands triggered by sudden sea conditions, such as typhoon variations, in weather routing services, this study aims to resolve the strict time-window constraints and multi-dimensional heterogeneous resource contention in cloud-edge collaborative scheduling. It seeks to achieve a dynamic trade-off among multiple objectives, including server-side task timeliness, server node energy consumption, and load balancing.The event-driven scheduling process is modeled as a semi-Markov decision process (SMDP), and a preference-based multi-objective hierarchical reinforcement learning framework is proposed. First, to tackle the large and sparse joint action space, a Preference-Based Multi-Objective Hierarchical Reinforcement Learning (PBMO-HRL) algorithm for weather routing task scheduling is proposed, decoupling the decision-making into a two-layer policy: "task selection" and "node allocation". Second, dynamic action masking and precise expectation evaluation are introduced to eliminate gradient variance in the discrete action space. Furthermore, to address the variable step-size characteristics, a time-aware discount factor and an idle-state transition merging mechanism are incorporated to mitigate value overestimation. Finally, a dynamic preference manager based on Logit space is designed to smoothly adjust preferences according to online congestion and energy consumption metrics. In a discrete-event simulation environment, a multi-preference evaluation protocol based on a simplex grid demonstrates that the framework converges on Pareto metrics, including hypervolume (HV) and expected utility metric (EUM). Compared to the Earliest Deadline First (EDF) and Energy-greedy baselines, PBMO-HRL reduces energy consumption by approximately 16.6% and 5.6%, and decreases SLA violation rates by about 20.5% and 28.8%, respectively. The proposed framework effectively manages bursty workloads, ensuring the service level agreements (SLA) of core weather routing tasks while achieving adaptive multi-objective optimization in non-stationary environments.