| 王晓玲,轩昕祺,余红玲,李志,张天鸿,倪磊.灌浆功率时序预测的可解释门控循环网络模型[J].水利学报,2025,56(6):791-804 |
| 灌浆功率时序预测的可解释门控循环网络模型 |
| Explainable gated recurrent unit network model for grouting power temporal prediction |
| 投稿时间：2024-05-09 |
| DOI：10.13243/j.cnki.slxb.20240270 |
| 中文关键词: 灌浆功率 门控循环单元 注意力机制 可解释性 自适应噪声完备集合经验模态分解 |
| 英文关键词: grouting power gated recurrent unit explainability attention mechanism complete ensemble empirical mode decomposition with adaptive noise |
| 基金项目:国家自然科学基金项目(52379131);华能集团总部科技项目(HNKJ20-H21TB) |
|
| 摘要点击次数: 41 |
| 全文下载次数: 16 |
| 中文摘要: |
| 灌浆功率是灌浆施工过程控制的重要指标，精准预测灌浆功率有助于优化灌浆施工策略。现有灌浆功率预测大多采用常规的机器学习模型，其浅层结构难以挖掘复杂地质条件、施工布置等多因素耦合影响下的非线性时序特征，导致模型预测精度较低，并且现有模型内部决策机制不透明，模型可解释性较差。针对上述问题，本文提出一种灌浆功率时序预测的可解释门控循环网络(EGRN)模型。首先，采用自适应噪声完备集合经验模态分解方法将灌浆功率时序数据分解为不同频率的本征模态分量，在门控循环单元(GRU)前置多头自注意力机制(MHA)，以有效提取灌浆功率时序数据的多维频率特征，并引入时序注意力机制(TPA)捕捉灌浆功率数据的长期时序依赖关系，提高模型的预测精度。之后结合MHA的权重分配与TPA的滤波器频谱分析结果，挖掘灌浆功率预测模型中关键的频率特征与时序依赖特征，增强模型的可解释性。案例分析结果表明，相较于其它对比模型，本文所提模型具有更高的预测精度，其MAE、RMSE和MAPE平均值降低了24.78%、27.29%和24.99%，R2系数平均提高7.74%，相较传统机器学习算法，本文所提模型可解释性更强，具有更高的透明度与可信度。 |
| 英文摘要: |
| Grouting power is a key index for the control of grouting construction process，and accurate prediction of grouting power helps to optimize the grouting construction strategy.Under the coupled influence of geological conditions，construction arrangement and other factors，the grouting power data presents complex nonlinear temporal series characteristics.However，most of the existing researches in grouting power prediction apply conventional machine learning models，whose shallow structure makes it difficult to effectively extract the complex temporal features of grouting power，resulting in low prediction accuracy.Besides，the decision-making mechanism in the existing models is not transparent，leading to poor model interpretability.According to the above problems，this paper proposes an Explainable Gated Recurrent Unit Neural Network (EGRN) for grouting power temporal prediction.First，a Complete Ensemble Empirical Mode Decomposition with Adaptive Noise method is applied to decompose the grouting power temporal data into different Intrinsic Mode Functions.In order to effectively extract the multi-dimensional frequency features of the grouting power time series data，a multi-head self-attention (MHA) mechanism is embedded at the input side of the Gated Recurrent Unit (GRU).Besides，the temporal pattern attention (TPA) mechanism is introduced to capture the long-term temporal dependence of the grouting power data and improve the prediction accuracy of the model.Further，combining the weight assignment of MHA and the filter analysis results of TPA，the key frequency and temporal dependence features in this model are mined to enhance its interpretability.The analysis results show that the proposed model has higher prediction accuracy compared to other comparative models.Specifically，its MAE，RMSE，and MAPE are reduced by 24.78%，27.29%，and 24.99% on average，and the coefficient of determination (R2) is improved by 7.74% on average.It demonstrates that compared to traditional machine learning algorithms，the proposed model is more interpretable and has higher transparency and credibility. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |