Abstract: Long-Sequence forecasting aims to model and predict future long-term time series trends by leveraging historical knowledge and patterns and has many practical applications in various industries. To fully utilize long-time series industrial data characteristics, this paper presents an improved self-attention mechanism suitable for modeling and forecasting long sequence industrial data. Our model builds a new embedding representation learning module, combined with the pooling operations, and uses the generative inference for long-range dependency modeling and time-series signal prediction. Compared with the previous self-attention-based method, the proposed model effectively solves the problems of insufficient prediction accuracy and high training cost in long sequence prediction. Our model significantly improves long-sequence water level prediction accuracy and efficiency compared with other benchmark methods. Experiments conducted on the real-world water level data from a large-scale hydropower station proved the superior performance of the proposed model in terms of both effectiveness and efficiency over existing state-of-the-art models.