Abstract: Objectives Traveling Ionospheric Disturbances (TIDs) are wave-like perturbations in ionospheric electron density induced by atmosphere-ionosphere coupling processes, and the associated variations in Total Electron Content (TEC) can significantly affect the stability and reliability of Global Navigation Satellite System (GNSS) high-precision positioning. Existing forecasting methods mainly focus on temporal variations and have limited capability in characterizing the spatial evolution of TIDs. To improve the short-term forecasting performance of ionospheric disturbances, a ConvLSTM-based deep learning framework is proposed for the spatiotemporal prediction of TIDs using sequences of detrended TEC (dTEC) maps over Japan under different solar activity conditions and latitude regions. Methods A short-term forecasting framework based on the Convolutional Long Short-Term Memory (ConvLSTM) network is constructed using an encoder-decoder architecture. By integrating convolutional operations into recurrent neural networks, temporal dependencies and spatial structures in sequential ionospheric maps are simultaneously captured. And the residual learning and skip-connection mechanisms are incorporated to enhance the reconstruction of fine-scale disturbance features. The experimental data are derived from real-time ionospheric maps provided by the National Institute of Information and Communications Technology (NICT) of Japan based on the dense GEONET GNSS network. After preprocessing, the dTEC maps are resized to 256 × 256 pixels. Data from 2020, 2022, and 2024 are processed to represent low, moderate, and high solar activity conditions, respectively. Furthermore, the study region is divided into three latitude bands to comprehensively evaluate the latitude-dependent forecasting performance using SSIM, PSNR, RMSE, and MAE. Results The experimental results show that the proposed framework achieves robust short-term forecasting performance for TIDs in the mid-latitude region. The predicted dTEC maps maintain high structural consistency with observations, with an average SSIM of 0.72 and an RMSE of approximately 0.19 TECu. In terms of solar activity dependence, the best forecasting performance is obtained during the low solar activity year (2020), while larger prediction errors occur during the high solar activity year (2024) due to intense background perturbations. For spatial distribution, the latitude-dependent analysis further indicates that the forecasting performance in the 28°-35°N and 42°-48°N latitude bands is superior to that in the 35°-42°N region. Conclusions The forecasting performance of the proposed ConvLSTM-based framework demonstrates strong adaptability under different solar activity conditions and latitude regions, and the method can effectively preserve the structural characteristics and spatial evolution information of TIDs. The final forecasting accuracy can meet the practical needs of ionospheric disturbance warning and high-precision BDS/GNSS positioning applications.