Hunan Province, located in central China, features a terrain dominated by mountains and hills, with plains enclosed by mountains on three sides. The region experiences a subtropical monsoon climate, with frequent high-temperature events during summer, particularly in the peak summer months of July and August. Research indicates that a rising trend in extreme heat events in Hunan, with the southeastern region experiencing the highest occurrence. Accurate fine-scale temperature forecasting remains a key challenge in regional weather prediction, while effective forecasting and timely warnings of severe weather are essential for disaster prevention and mitigation. Unlike short-term weather forecasts, extended-range forecasts (10—30 days) provide a longer decision-making window, allowing government authorities to implement proactive measures to enhance public safety and reduce disaster losses. However, current temperature forecasting studies in Hunan primarily focus on nowcasting and short-term model corrections, with limited research on extended-range forecasting. Furthermore, existing extended-range high-temperature forecasts in Hunan largely rely on sub-seasonal to seasonal (S2S) models, which often exhibit insufficient accuracy. Therefore, developing a dedicated forecasting model for extended-range high-temperature forecasting is crucial. The study aims to develop an extended-range forecasting model for heat wave events in Hunan Province during the peak summer period (July-August). The model integrates physical predictors derived from S2S model temperature forecasts and their corrections with a convolutional neural network (CNN) approach to enhance forecasting skill. Daily maximum temperature data from 97 meteorological stations in Hunan Province (1999—2022) and S2S model outputs from ECMWF and NCEP are utilized. Physical forecast factors are extracted from temperature and circulation forecast products using singular value decomposition (SVD) and the spatiotemporal projection model (STPM). These factors are then integrated into a CNN-based high-temperature prediction model (HTPM). Additionally, the maximum temperature forecasts from the S2S models undergo bias correction, and the corrected forecasts are combined with predictions from the HTPM to create an ensemble forecasting scheme. This approach aims to enhance the stability and accuracy of regional high-temperature forecasts. Results indicate that while the original S2S model forecasts exhibit low predictive skill, bias correction significantly improves their performance, though false alarm rates remain high. The CNN-based high-temperature forecasting model trained on ECMWF S2S data (HTPM-ECS2S) and NCEP S2S data (HTPM-NCEPS2S) effectively capture high-temperature events, demonstrating improved forecasting skill. The ensemble scheme successfully integrates multiple model outputs, further enhancing forecast accuracy and reliability.