基于权重修正和DRSN-LSTM模型的向家坝下游水位多时间尺度预测

Multi-time scale prediction for Xiangjiaba’s downstream water level based on weight correction and DRSN-LSTM model

为从大量水位影响因子中提取重要特征实现水位的高效、精准预测,提出改进的深度残差收缩网络与长短时记忆网络的混合模型用于多时间尺度水位预测。选取水位、流量、电站出力等数据拟合为高维特征输入的形式便于网络提取水位变化的动态特征。利用新的半软阈值函数消除深度残差收缩网络的恒定偏差并降低水文数据中的噪声干扰。根据预测误差,采用新构建的误差权重修正函数配合交叉熵函数对水位影响因子进行权重修正。阿基米德优化算法被用于调整长短时记忆网络的参数。将新模型应用于向家坝水电站下游水位的多时间尺度预测,结果表明,该方法的预测精度比现有CNN-LSTM、SVR等模型分别提高47%、61%,预测效率分别提高57%、20%,其短期、中期和长期的最大预测误差为0.09 m、0.14 m、0.31 m,证明模型在多时间尺度的水位预测中取得良好的精度和效率。此外,考虑支流流量后的预测误差最高可降低0.03m,证明模型对回水顶托等复杂水文的适应性,研究成果为洪水预测和城市雨洪预警提供新思路。

In order to extract important features from a large number of water level influencing factors for efficient and accurate water level prediction, an improved hybrid model of deep residual shrinkage network and long and short time memory network is proposed for multi-time scale water level prediction. Data such as water level, flow rate, and power plant output are selected to be fitted to form a high-dimensional feature input form to facilitate network extraction of dynamic features of water level changes. A novel semi-soft threshold function is used to eliminate the constant bias of the depth residual shrinkage network and reduce the noise interference in the hydrological data. According to the prediction error, the newly constructed error weight correction function with cross entropy function is adopted to weight the water level influence factor. The Archimedes optimization algorithm is used to adjust the parameters of the long and short term memory network. The new model is applied to the multi-time scale prediction of the Xiangjiaba Hydropower Station′s downstream water level, and the results show that the prediction accuracy of the method is 47% and 61% higher than the CNN-LSTM and SVR, respectively, and the prediction efficiency is 57% and 20% higher, respectively, and its maximum prediction errors are 0.09 m, 0.14 m, and 0.31 m in short-term, medium-term, and long-term, which proves that the model achieves decent accuracy and efficiency in multi-time scale water level prediction. In addition, the prediction error can be reduced by up to 0.03 m after considering the tributary flow, which proves the adaptability of the model to complex hydrology such as backwater jacking, and the results provide new ideas for flood prediction and urban flood warning.