With the acceleration of economic construction and the sustained growth in demand for raw materials driven by infrastructure development, raw material production enterprises are facing increasingly severe challenges. To maintain stable operations and enhance competitiveness, there is an urgent need for these enterprises to explore effective methods for predicting commodity prices of raw materials. As one of the essential raw materials for infrastructure construction, iron ore exhibits complex and volatile characteristics in its futures price fluctuations, posing significant challenges to price forecasting. Although various prediction methods currently exist, their accuracy and stability still require improvement. Therefore, it aims to explore a more accurate and effective method for forecasting iron ore prices. It proposes an enhanced ensemble learning model that integrates a multi-head convolutional stacked long short-term memory (LSTM) encoder-decoder network with modal signal decomposition and attention mechanisms, based on multivariate time series of iron ore prices. The proposed method first employs modal signal decomposition to reduce the volatility of the price series. Subsequently, a multi-head convolutional network automatically extracts complex spatial features from the multivariate sequence data, while a stacked LSTM (SLSTM) network captures temporal dependencies from historical sequences. Finally, an attention mechanism is applied to identify key features and generate the final prediction results. The experimental results demonstrate that the proposed enhanced ensemble model outperforms other benchmark models without decomposition techniques in terms of both prediction accuracy and stability. The findings possess significant practical implications, enabling mining enterprises to achieve more accurate iron ore price predictions, thereby facilitating sustained high-quality development and operational stability.