Abstract: Accurate forecasting of ionospheric total electron content(TEC) is critically important for ensuring the reliable performance of Global Navigation Satellite System(GNSS) and supporting various space weather applications, particularly in the context of increasing reliance on satellite-based technologies and the growing awareness of space weather impacts on technological systems. This study addresses the complex characteristics of TEC sequences—including their inherent nonlinearity, non-stationary behavior, and complicated spatiotemporal dependencies—by developing an advanced hybrid neural network model that strategically integrates the complementary strengths of temporal convolutional network(TCN), Transformer, and long short-term memory (LSTM) architectures in a carefully designed hierarchical framework. Specifically, the model employs TCN to effectively extract local multi-scale features from TEC sequences, utilizes the self-attention mechanism of Transformer to capture long-range global dependencies that are characteristic of ionospheric phenomena, and incorporates LSTM networks to integrate temporal dynamics and maintain memory of past states, thereby creating a comprehensive modeling approach that addresses multiple aspects of TEC variability simultaneously. The experimental investigation was conducted using extensive Global Ionosphere Map (GIM) data from Center for Orbit Determination in Europe(CODE), with a focused regional analysis over China to evaluate model performance across diverse space weather conditions spanning both solar minimum and maximum periods. Comprehensive evaluation results demonstrate that during the solar minimum of 2019, our proposed model achieved a remarkable RMSE of 0.55 TECU, representing substantial error reductions of 37%, 40%, and 15% compared to the standalone LSTM, Transformer, and combined Transformer-LSTM(TL) models respectively, while during the more challenging solar maximum of 2023, the model maintained strong performance with an RMSE of 1.82 TECU and corresponding error reductions of 43%, 44%, and 20% against the same baseline models, clearly indicating its superior accuracy and robustness across different solar activity regimes. Particularly noteworthy is the model's consistent performance during different intensity levels of geomagnetic storms occurring in the high solar activity period, where it successfully maintained high forecasting accuracy despite the disturbed space weather conditions, thereby demonstrating its practical applicability and reliability under extreme space weather scenarios that typically pose significant challenges for ionospheric forecasting. The proposed hybrid architecture therefore represents a significant advancement in neural network solutions for ionospheric TEC forecasting, providing both high accuracy and robust performance across varying space weather conditions while effectively addressing the complex spatiotemporal characteristics of TEC variations. The demonstrated consistency during both solar minimum and maximum periods, coupled with the maintained performance during geomagnetic disturbances, suggests strong potential for operational implementation in space weather forecasting systems and practical applications in GNSS positioning, navigation, and related technological systems that are vulnerable to ionospheric variability.