[关键词]
[摘要]
南水北调中线工程通常以闸前常水位调度运行,而水位在闸门调控影响下多数处于非平稳状态,探索其变 化规律对于监测数据和研究方法均有一定限制和要求。监测数据方面,针对大量的高频监测数据选取均值滤波、 滑动平均值滤波、递推中位值平均滤波法、滑动小波变换进行数据预处理,提高数据质量、增强数据预测的可行 性。研究方法层面,以 BP 神经网络模型和长短期记忆(long?short-term?memory,LSTM)网络模型为主体框架,以水 动力模型的模拟数据为辅助支撑,对比单神经网络在不同工况下的预测效果,输出水动力-神经网络组合预测结果。 结果表明:数据预处理是数据分析和预测的必要环节,高频数据滤波处理再预测可以提高数据预测的精度;均值 滤波、递推中位数均值滤波方式对数据预处理的效果最好,指标合理时滤后决定系数(R2)精度均超过 0.95,且均 方根误差(RMSE)和平均绝对误差(MAE)不超过 0.02,准确性高;基于滤后数据进行模型构建,通过对比数据驱动 模型和数据-机理双重驱动模型的计算结果,R2维持在 0.98 附近,RMSE、MAE 维持在 0.01 左右,耦合模型具有更 好的稳定性和准确性。
[Key word]
[Abstract]
The Middle Route of South-to-North Water Transfers Project has made a significant contribution to mitigating the water scarcity challenges prevalent in the central and northern regions of China, and in the process of the project scheduling and operation, it usually operates in accordance with the normal water level before the gate. Under the influence of gate regulation, the water level before and after the gate is in a non-stationary process most of the time, so exploring its regular changes has certain limitations and requirements on the monitoring data and research methods. To enhance the precision of water level forecasting within the ambit of the South-to-North Water Transfers Project, the monitoring data and research methodology are improved respectively, with a view to obtaining better prediction results.For a large amount of high-frequency monitoring data, mean filter, sliding mean filter, recursive median mean filter, and sliding wavelet transform are selected for data preprocessing to improve the data quality and enhance the feasibility of data prediction. The data prediction framework leverages two primary neural network models, namely the BP neural network model and LSTM neural network model, and the hydrodynamic model simulation data as the auxiliary support, and selects the pre-gate, and post-gate water level, openness, and flow rate data of the upstream gate itself as the model input factor for prediction. The predictive output factor pertains to the upstream water level of the gate within the subsequent 2 h. The assessment of predictive performance is predicated upon key indicators, namely the coefficient of determination, root-mean-square error, and average absolute error. Indicators compare the single neural network prediction results and network-hydrodynamic combination prediction results and analyze the accuracy and stability of the prediction results.The accuracy of data prediction can be improved after data filtering preprocessing of high-frequency data, and the selection of appropriate water-level data filtering methods can significantly improve the prediction effect. After filtering the data can more clearly show the water level before and after the gate, flow change rule, and the minimum frequency of training data can be selected for 15 minutes of data for filtering and data processing. Constructing BP and LSTM neural networks based on the monitoring data, a comparative analysis is conducted encompassing the number of gate inputs, temporal scales, and data filtering methodologies. The investigation reveals the following insights: The prediction results under the hourly data can already reach the optimal state; The number of gates can be 2 or 3 gates to reach the optimal state, which is related to the frequency of the data; Comparing the prediction results after a variety of filtered data, the recursive median-mean filtering algorithm is the best, and the mean filtering is the worst. The sliding wavelet transform has the worst effect, so it is suggested that the filtering methods are mean filtering and recursive median-mean filtering. The combined hydrodynamic-neural network prediction results are better than the single network prediction results.The computational outcomes prove that high-frequency data preprocessing is a necessary part of data analysis, and the suggested filtering methods mean filter and recursive median-mean filter can be applied in the water condition data processing of the water transfer project. These two filtering methods yield commendable outcomes in data processing. The neural network model necessitates tailoring to the specific parameters corresponding to distinct objects and varying temporal cycles, and after adjusting the parameters, it can better reflect the data change process in a period of time, and the prediction effect is better. Different neural network models have different prediction characteristics, but the prediction accuracy is higher under the condition of sufficient high correlation data. Moreover, the data mechanism dual-drive model can play the advantages of the hydrodynamic model and neural network model at the same time, and the prediction accuracy is higher.
[中图分类号]
[基金项目]