Signal Temporal Logic (STL) has emerged as an expressive language for reasoning intricate planning objectives. However, existing STL-based methods often assume full observation and known dynamics, which imposes constraints on real-world applications. To address this challenge, we propose a hierarchical planning framework that starts by constructing the Value Function Space (VFS) for state and action abstraction, which embeds functional information about affordances of the low-level skills. Subsequently, we utilize a neural network to approximate the dynamics in the VFS and employ sampling based optimization to synthesize high-level skill sequences that maximize the robustness measure of the given STL tasks in the VFS. Then those skills are executed in the low-level environment. Empirical evaluations in the Safety Gym and ManiSkill environments demonstrate that our method accomplish the STL tasks without further training in the low-level environments, substantially reducing the training burdens.