The measurement reliability of nanofilm humidity sensors in extreme industrial environments is significantly degraded by temperature-induced drift, resulting in pronounced uncertainty across wide temperature ranges and dynamically varying conditions. To mitigate this issue, a hybrid temperature compensation framework is proposed by integrating a Kolmogorov-Arnold Network (KAN) with an enhanced metaheuristic optimization strategy. Specifically, a multi-strategy enhanced sand cat swarm optimization (MESCSO) algorithm is developed to improve parameter optimization performance of KAN. The proposed method incorporates a dual-trajectory hybrid attack mechanism, a nest memory decay strategy, and a dynamic leveraged opposition-based learning strategy, collectively enhancing global exploration capability while maintaining population diversity and convergence stability. Based on the optimized parameters, the KAN-based nonlinear mapping model is established to characterize the coupled temperature–humidity response. Experimental results demonstrate that the proposed approach achieves a compensation accuracy of ±0.55% RH over the temperature range of 10 °C to 50 °C, validating its effectiveness and robustness for temperature drift correction in complex operational environments.