Abstract: International natural uranium price forecasting is vital for nuclear energy industry sustainability. Accurate predictions aid nuclear power enterprises in devising operational strategies and hedging against market volatility, while providing data support for policymakers to enhance energy security and formulate sustainable development plans. However, due to the intricate interplay of market supply-demand dynamics, geopolitical events, and economic policies, traditional econometric and statistical models fall short in capturing the temporal dependencies and complex patterns in uranium price data. Thus, there is a pressing need for advanced predictive models capable of handling the multifaceted nature of uranium price fluctuations. To address these challenges, this study proposed a novel TF-CNN-BiLSTM model, which synergistically combines the self-attention mechanism of Transformer, the local feature extraction capability of convolutional neural network (CNN), and the bidirectional temporal dependency modeling of bidirectional long short-term memory (BiLSTM). The model was trained and tested on a comprehensive dataset of monthly natural uranium spot prices spanning from August 1968 to November 2024, obtained from UxC and Trade Tech. The dataset, totaling 676 data points, encompassed diverse market conditions and volatility regimes, ensuring the model’s robustness and generalizability. Data preprocessing involved cleaning to remove outliers and missing values, normalization using Min-Max scaling to 0, 1, and creating sliding windows with an input width of 3 months and a prediction step of 1 month. The dataset was split into training (80%) and testing (20%) sets to evaluate the model’s performance. The proposed TF-CNN-BiLSTM model was designed with three core modules: 1) a Transformer encoder to capture long-term global dependencies via self-attention mechanisms; 2) a 1D CNN layer to extract localized price fluctuations; 3) a BiLSTM network to model bidirectional temporal relationships. The study meticulously constructed and trained the TF-CNN-BiLSTM model using TensorFlow with the Adam optimizer and mean squared error loss function. The training process was optimized with learning rate reduction and early stopping callbacks to prevent overfitting. The model demonstrates superior performance with an RMSE of 0.044 3, MAE of 0.024 7, and R² of 0.802 0 on the test set, outperforming individual Transformer, CNN, and BiLSTM models as well as their two-component combinations. The results indicate that the TF-CNN-BiLSTM model effectively captures both the long-term trends and short-term fluctuations in natural uranium prices. The Transformer component efficiently models long-range dependencies, the CNN component extracts local features, and the BiLSTM component integrates bidirectional contextual information. This hybrid approach provides a more nuanced understanding of the uranium price dynamics compared to traditional models. The study concludes that the TF-CNN-BiLSTM model offers a powerful tool for international natural uranium price prediction. Its ability to integrate global dependencies, local patterns, and temporal context makes it a promising approach for financial time series forecasting in the energy sector. However, the model’s performance could be further enhanced by incorporating additional relevant features such as geopolitical indicators, economic indices, and policy variables. Future research should focus on expanding the model’s input space and refining its architecture to improve accuracy, especially in periods of market turbulence. Additionally, the model’s interpretability could be enhanced to provide stakeholders with actionable insights into the drivers of uranium price movements. Overall, this study represents a significant advancement in the application of deep learning techniques to the uranium market, offering practical benefits for risk management and strategic planning in the nuclear energy industry.