| 摘要: |
| 入库径流预测对丹江口水库调度以及水资源利用具有重要的指示意义。基于灰狼优化算法（GWO）构建不同的预测模型，开展丹江口水库月入库径流预测研究，并探讨网络结构超参数的选取以及验证GWO全局优化、收敛快的特点。结果表明：灰狼优化的长短期记忆模型（GWO-LSTM）的预测精度和泛化性能优于灰狼优化的人工神经网络模型（GWO-BP）和逐步回归模型，其验证期的均方根误差（RMSE）和纳什效率系数（NSE）平均达到了6.854亿m3和0.969，整体趋势预测较好，峰值捕捉略有不足，可适用于丹江口水库月入库径流预测；模型超参数依据经验取值的其预测结果不如GWO优化的，依据经验取值的其验证期的NSE不足0.5，未达到可接受范围，而且带有一定的偶然性，建议选用具有全局优化特性的优化算法进行超参数选取；验证了GWO算法其全局最优和收敛快的特点，平均在3次迭代后可达到收敛状态。 |
| 关键词: 长短期记忆模型 径流预测 灰狼优化算法 |
| DOI： |
| 分类号: |
| 基金项目:国家重点研发计划课题（2016YFC0400910） |
|
| Research on prediction of inflow to the Danjiangkou reservoir Based on GWO-LSTM |
|
wanglihui, yanghuibin
|
|
Fuzhou University
|
| Abstract: |
| The prediction of runoff into the reservoir has important indications for the operation of Danjiangkou Reservoir and the utilization of water resources. Based on the gray wolf optimization algorithm (GWO), construct different prediction models, carry out the study of reservoir runoff prediction in Danjiangkou, discuss the selection of network structure hyperparameters, verify the characteristics of global optimization and fast convergence of GWO. The results show that the prediction accuracy and generalization performance of the GWO-LSTM model is better than that of the GWO-BP model and the stepwise regression model. The root mean square error (RMSE) and Nash efficiency coefficient (NSE) of its verification period reached 6.854 and 0.969 on average. The overall trend forecast is good, and the peak capture is slightly insufficient, which is suitable for prediction of monthly inflow to the Danjiangkou reservoir; The prediction results of model hyperparameters based on empirical values are not as good as GWO. Based on empirical values, the NSE of the verification period is less than 0.5, which does not reach the acceptable range, and there is a certain contingency. It is recommended to select an optimization algorithm with global optimization characteristics for hyperparameter selection; The global optimality and fast convergence characteristics of the GWO algorithm are verified, and the convergence state can be reached after 3 iterations on average. |
| Key words: long and short-term memory model runoff prediction gray wolf optimization algorithm |
