This paper uses the COF-MED-LSTM model, an advanced approach integrating multi-modal event-driven data for predicting crude oil futures prices. By leveraging distributed multi-thread crawlers for extensive data collection and employing the transformer's multi-head attention mechanism for feature extraction, this research overcomes the challenges of sampling heterogeneity between structured historical data and unstructured news information. The event-driven memory mechanism addresses data dimensionality and sampling issues, enhancing prediction accuracy. Empirical analysis across various periods demonstrates its superiority over traditional and advanced algorithms. This study contributes to crude oil futures theories and suggests further exploration into the indirect effects of news on related futures, highlighting the potential for model optimization and a deeper market understanding.