With advancements in intelligent computing for financial analysis, effectively integrating diverse data sources is crucial for improving digital currency price prediction accuracy. This paper introduces CSTNet (Conv Spatio-Temporal Network), a novel multi-source data fusion framework enhanced by Transformer self-attention, specifically designed for cryptocurrency price prediction. CSTNet extracts key features from structured data (e.g., historical prices, technical indicators) and unstructured data (e.g., news texts). Its architecture uniquely merges localised convolutional feature extraction with global attention and event-driven signals, enhancing robustness to market shifts and improving predictive accuracy. Tested on the highly volatile Bitcoin market of 2024, experiments demonstrate that CSTNet combined with Transformer significantly outperforms traditional models (e.g., LSTM, Bi-LSTM, LSTM + Transformer) in metrics like MSE and RMSE, and exhibits enhanced robustness and accuracy under extreme market conditions. This study highlights the critical role of advanced intelligent computing in digital currency price prediction and provides a viable approach for applying multi-source heterogeneous data in financial forecasting.