Abstract: Disturbances in centrifugal cascade systems manifest as flow or pressure perturbations, typically exhibiting abnormal pressure fluctuations. Taking staged centrifugal cascades as an example, dynamic hydraulic analysis reveals that disturbances propagating toward the depleted-end units cause pressure oscillations that amplify progressively downstream. During operational transitions, dedicated personnel must be arranged for real-time monitoring and necessary intervention to maintain the feed main pressure of Unit 01 within its safe operating range. Accurate determination of pressure trends in Unit 01’s feed main requires both prolonged disturbance propagation durations and operators’ extensive experience. Conventional monitoring approaches predominantly rely on threshold-triggered alarms control systems and operator experience-driven judgments. These methods cannot promptly capture subtle pressure variations during disturbances’ incipient stages, while the nonlinear, multivariable-coupled nature of cascade systems creates intricate parameter interdependencies. Consequently, operators face difficulties in timely disturbance detection, predicting propagation patterns, and assessing impact scopes. Due to the lack of means to predict pressures in a centrifugal cascade system, it is difficult for operators to predict pressure changes instantly and accurately. This limitation hinders timely response to potential disturbances impacting system stability and safety. To address this challenge, an improved Transformer model specifically designed for effective pressure prediction in centrifugal cascade systems was proposed. The characteristic relation of the input sequence was captured by a two-dimensional convolution layer, and the information at the adjacent time steps was amalgamated by a one-dimensional convolution layer. Crucially, to overcome the computational inefficiency of the standard Transformer, especially for long sequences common in pressure monitoring, a mechanism of multi-head sparse self-attention (MHSSA) was employed. This MHSSA mechanism sparsifies the attention weight matrix to significantly reduce the computational complexity and improve the computational efficiency. Furthermore, to enhance the temporal coherence and accuracy of the generated pressure sequences, autoregressive decoding was adopted. By simulating the pressure change caused by mis-operation of an electric regulating valve on the main pipe for the depleted material, the pressure prediction effect of the model within 60 s was verified. The average absolute error (MAE), the mean absolute percentage error (MAPE) and the root-mean-square error (RMSE) of the improved Transformer model for pressure prediction are 0.42, 1.5% and 0.48, respectively. Compared with the traditional Transformer model, these metrics are reduced by 59.6%, 57.1% and 61.9%, demonstrating the effectiveness of the proposed improvements. Furthermore, the proposed model (F₄) exhibits superior prediction accuracy compared to other baseline models, including a standard LSTM (F₁), an encoder-decoder LSTM (F₂), and the traditional Transformer (F₃), across all evaluation metrics. The significant performance gains highlight the model’s capability in capturing complex spatio-temporal dependencies in cascade system pressure dynamics, providing a valuable tool for real-time monitoring and early anomaly detection.