基于LSTM-Informer模型的乌东德水库水位多步长预测研究

Multi-step prediction of water level in Wudongde Reservoir based on LSTM-Informer model

摘要

针对目前深度学习算法在水库水位预测中预见期较短的问题，构建基于LSTM-Informer模型的乌东德水库水位预测模型，预测未来6、12、24、48和96个步长的水库水位，并与LSTM和Informer模型的预测结果进行比较。结果表明：当预测步长不大于12个时，3种模型均能较好模拟水库水位且性能差异不大，当预测步长大于12个时，3种模型的性能表现为LSTM-Informer>Informer>LSTM,LSTM-Informer模型在96个步长时的RMSE和MAE分别为0.147和0.120,LSTM-Informer模型在24、48和96步长的RMSE分别比LSTM低25%、46%和62%,MAE分别比LSTM低23%、40%和47%,组合模型LSTM-Informer能较好地解决水库水位长时间序列预测问题。

Abstract

In order to solve the problem of short foresight period of deep learning algorithms in reservoir water level prediction, the LSTM-Informer reservoir water level prediction model is constructed with Wudongde Reservoir as an example, which predicts the reservoir water level in the future for 6, 12, 24, 48, and 96 steps, and compares the prediction result with those of the LSTM and Informer models. The results show that when the prediction step size is not more than 12, all three models can simulate the reservoir level well and the performance difference is not obvious, when the prediction step size is more than 12, the performance of the three models is LSTM-Informer>Informer>LSTM, and the RMSE and MAE of the LSTM-Informer model at 96 steps are 0.147 and 0.120 respectively, and the RMSE of the LSTM-Informer model is 25%, 46% and 62% lower than LSTM, and the MAE is 23%, 40% and 47% lower than LSTM, respectively. The combined model LSTM-Informer can solve the long time series reservoir level prediction problem better.

关键词

LSTM-Informer模型 / 乌东德水库 / 水位预测

Key words

LSTM-Informer model / Wudongde Reservoir / water level prediction

基金资助

中国长江电力股份有限公司科研项目(Z542302007)