To improve the accuracy of power load forecasting under extreme climate conditions, this study utilizes daily maximum load and meteorological observations from Linyi City, Shandong Province, during 2022-2023. The temporal characteristics of load and its meteorological drivers are analyzed, and meteorological-sensitive load is separated using the least squares method. The Temperature-Humidity Index (THI) is introduced to quantify thermal comfort effects, highlighting the critical role of temperature-humidity interactions. A temperature-zoning prediction model is developed using 18℃ as the threshold to partition temperature zones, with unified Support Vector Machine (SVM) modeling applied across all zones. Results indicate that: 1) Daily load exhibits a summer-winter dual-peak and spring-autumn dual-valley structure; 2) Temperature and THI dominate, with THI-load correlation reaching 0.942 under hot-humid conditions in August; 3) For the whole temperature range, the SVM 8-factor model achieves R2=0.955 and mean relative error (MRE)=2.54%; in the temperature-zoning model, the low-temperature zone (≤18℃) achieves MRE=2.39% and the high-temperature zone (>18℃) achieves MRE=2.52%, reducing average error by 3.3% compared to the full-range model and achieving 97.54% prediction accuracy. These findings demonstrate that the temperature-zoning model decouples thermodynamic response mechanisms, providing enhanced adaptability and precision under extreme climate conditions and offering a new technical pathway for refined power system dispatching.