Scholarlink

<p>In the field of forecasting, Transformer models have shown outstanding performance across various prediction tasks. However, they face challenges such as high computational complexity and a tendency to overlook the local trend characteristics of sequences when handling multimodal time-series data. To address these limitations, this paper proposes a novel LSTPencoder model based on dual-architecture parallel computing and prediction information sharing. This model integrates a gating mechanism with the Transformer encoder, where the gating module precisely captures the trend dependencies between adjacent sequences. Meanwhile, the multi-head attention mechanism simultaneously extracts global encoded features from different subspaces. Through a unique sharing mechanism, the model achieves a comprehensive fusion of both local and global features. Additionally, the study introduces the AOAAO algorithm to optimize the model’s hyperparameters, aiming to achieve the best forecasting results. Based on real flow prediction experiments at the Tangnaihai Hydrological Station and the Dongjiang Hydropower Station, the results show that, compared to other models, the LSTPencoder model improved the coefficient of determination (R²) by an average of 44.79% and 14.07%, respectively. This model not only demonstrates superior forecasting performance but also offers a more effective explanation of the complex causal relationships in runoff sequences, providing an efficient new method for hydrological forecasting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating LSTM and Transformer for Improved Daily Runoff Prediction: A Parallel Computing Approach

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Dong-mei Xu,
  • Yang-hao Hong,
  • Wen-chuan Wang,
  • Miao Gu,
  • Jun Wang,
  • Yan-wei Zhao,
  • Hong-fei Zang

摘要

In the field of forecasting, Transformer models have shown outstanding performance across various prediction tasks. However, they face challenges such as high computational complexity and a tendency to overlook the local trend characteristics of sequences when handling multimodal time-series data. To address these limitations, this paper proposes a novel LSTPencoder model based on dual-architecture parallel computing and prediction information sharing. This model integrates a gating mechanism with the Transformer encoder, where the gating module precisely captures the trend dependencies between adjacent sequences. Meanwhile, the multi-head attention mechanism simultaneously extracts global encoded features from different subspaces. Through a unique sharing mechanism, the model achieves a comprehensive fusion of both local and global features. Additionally, the study introduces the AOAAO algorithm to optimize the model’s hyperparameters, aiming to achieve the best forecasting results. Based on real flow prediction experiments at the Tangnaihai Hydrological Station and the Dongjiang Hydropower Station, the results show that, compared to other models, the LSTPencoder model improved the coefficient of determination (R²) by an average of 44.79% and 14.07%, respectively. This model not only demonstrates superior forecasting performance but also offers a more effective explanation of the complex causal relationships in runoff sequences, providing an efficient new method for hydrological forecasting.