Abstract
The stable operation of key parameters in natural gas compressor units, including exhaust pressure, exhaust temperature, and shaft power, is essential for ensuring the safety, reliability, and efficiency of gas transmission systems. To address challenges associated with multi-source heterogeneous data, such as redundancy, noise, uncertainty, and complex temporal dependencies, this study proposes an intelligent trend prediction model based on Multi-Source Data Fusion (MSDF-ITPM). The framework consists of three main components: data preprocessing, multi-source data fusion, and an improved TCN-GRU prediction module. The preprocessing stage enhances data quality by removing outliers, imputing missing values, and applying normalization. Subsequently, a fusion strategy integrating one-dimensional convolutional neural networks (1D-CNN), Dempster–Shafer evidence theory, and an attention mechanism is employed to achieve robust uncertainty handling and adaptive feature integration. The prediction module combines Temporal Convolutional Networks (TCN) for capturing long-term temporal patterns with Gated Recurrent Units (GRU) for modeling short-term dynamics, while residual connections improve feature propagation and learning efficiency. The proposed model is evaluated using real operational data from natural gas compressor units under forecasting horizons of 1, 3, 5, and 10 steps and compared with ARIMA, BP, LSTM, GRU, and TCN models. Experimental results demonstrate that MSDF-ITPM consistently achieves superior prediction accuracy. For the 10-step forecasting task, it reduces MAE, RMSE, and MAPE by 33.0%, 32.0%, and 32.2%, respectively, compared with the best-performing baseline (TCN), while achieving faster convergence with an acceptable computational overhead. These results demonstrate the model's effectiveness for intelligent condition monitoring, predictive maintenance, and safety warning in natural gas compressor stations.