Abstract: The variation process of anchor bolt (cable) stress exhibits distinct short-term fluctuations and long-term temporal dependencies. However, traditional single prediction models have limited capability in modeling long-term trends and insufficient sensitivity to local fluctuations, often making it difficult to fully capture these complex features. To address this problem, an anchor bolt (cable) stress prediction model based on an Improved Slime Mould Algorithm (ISMA) optimized Temporal Convolutional Network (TCN)-Long Short-Term Memory (LSTM)-Multi-Head Self-Attention (MHSA) architecture is proposed. In the problem of anchor bolt (cable) stress prediction in coal mine roadways, model training often involves complex optimization tasks such as hyperparameter tuning and learning rate selection. To improve the training efficiency and prediction accuracy of the model, ISMA was proposed, which enhanced local search capability by introducing neighborhood search and a dynamic step-size factor. Global search efficiency was improved through integrating an Artificial Bee Colony (ABC) search mechanism, thereby effectively improving the model's ability to escape from local optima. The TCN-LSTM-MHSA model was constructed by using TCN to extract local temporal features, employing LSTM to learn long-term dependencies in the data, and strengthening global temporal modeling through MHSA, thereby enhancing the prediction capability for anchor bolt (cable) stress. During training, ISMA was used to iteratively optimize the learning rate of the TCN-LSTM-MHSA model to improve prediction accuracy and speed. Experimental results showed that: ① Compared with the Slime Mould Algorithm (SMA), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Sparrow Search Algorithm (SSA), the ISMA optimization strategy demonstrated better convergence speed and optimization ability in multiple benchmark function tests. ② In the stress prediction experiment, ablation experiments verified the necessity of TCN, LSTM, and MHSA modules. ③ The ISMA-optimized TCN-LSTM-MHSA model outperformed mainstream prediction models such as BP and GRU in MAE, RMSE, and R² metrics, showing higher prediction accuracy and stability.