基于EMD-LSTM模型的右江水库日入库流量预测
Inflow Prediction of Youjiang Reservoir Based on EMD-LSTM model

入库流量预测可为水电站水库防汛抗旱、水位控制等提供必要依据。然而传统物理建模方法预测入库流量对数据资料要求较高，在资料相对匮乏地区难以应用。针对入库流量序列非线性非平稳以及原始数据包含噪声等问题，将经验模态分解(empirical mode decomposition, EMD)良好的滤波特性和长短期记忆(long-short-term memory, LSTM)神经网络强大的非线性映射能力相结合，用于右江水库日入库流量预测。首先采用EMD方法对入库流量序列进行分解，然后对分解的各分量分别建立LSTM进行预测，最后加和重构得到预测结果，同时采用单一LSTM作为对比。在右江水库日入库流量预测中，EMD-LSTM预测结果平均绝对误差(MAE)为29.46 m³/s，平均相对误差(MAPE)为11.06%，确定性系数(DC)为0.95，各评价指标均优于单一LSTM，表明EMD-LSTM在右江水库日入库流量预测中具有很好的适用性，明显提高了日入库流量预测精度，对今后右江水库安全经济运行具有重要指导意义。
The prediction of inflow can provide necessary basis for flood prevention, drought relief and water level control in hydropower reservoirs. However, traditional prediction methods based on physical models have a high demand for data and are difficult to apply when data is relatively scarce. The inflow series has characteristics of nonlinearity and nonstationary. And the original data contains some noise. To solve these problems, empirical mode decomposition (EMD) and long-short-term memory (LSTM) neural network were combined to predict Youjiang Reservoir daily inflow, which made full use of the good filtering property of EMD and the powerful nonlinear mapping ability of LSTM. Firstly, the EMD was applied to decompose the inflow series, then the LSTM was built separately for each component from EMD for prediction, and finally each prediction result was superposed to obtain the final prediction result. Meanwhile, a single LSTM was used as comparison. The results indicated that the EMD-LSTM model exhibited better than the single LSTM in terms of the average absolute error (MAE = 29.46 m³/s), the average relative error (MAPE = 11.06%) and the coefficient of certainty (DC = 0.95). Based on the above results, EMD-LSTM has good applicability and obviously improves the prediction accuracy, which is of great significance to guide the safe and economic operation in the future.