为解决因水库数据采集设备能力有限、水文数据不全导致预测水库水位时预测精度较低的问题，以四岭水 库每小时水位监测数据为例，提出基于嵌入式-门控循环单元（Embedding-gated?recurrent?unit，Embedding-GRU）的 水库水位预测模型，即利用 Embedding 方法将单维降雨量数据升维至多维数据，扩大降雨的气候特征，结合 GRU 算法进行水库水位预测。将该模型与传统深度学习算法长短期记忆（long?short-term?memory，LSTM）、门控循环单 元（gated?recurrent?unit，GRU）、双向门控循环单元（bidirectional?recurrent?neural?network，BiGRU）这 3 种模型对比， 结果显示：Embedding-GRU 模型的预测效果均优于其他传统模型，平均绝对误差 EMA和均方根误差 ERMS分别平均 下降 19.6% 和 7.7%，并且在预测次日水库水位的应用场景中决定系数 R2能够达到 0.98937。结果表明：该模型耦 合多种算法，扩大单变量的气候特征，具有较高预测精度和泛化能力。相较传统模型，基于 Embedding-GRU 的水 库水位预测模型能够对缺少温度、气压、风速、蒸发量等监测数据的水库进行可靠度较高的预测，适用水库范围 更广，为水库日常运维、除险加固提供参考。

The prediction of reservoir water level is of great significance in the daily operation and management of reservoir, the reinforcement of dam, the mitigation of flood disaster, and the protection of people's life and property safety. However, with the change of global temperature, the frequency of extreme weather increases and the uncertainty of abnormal rainfall increases, which lead to the lagging of reservoir level prediction methods relying on traditional engineering hydrology. Due to the high practicability of deep learning algorithms used in various fields, there are a few examples of domestic and foreign scholars using artificial intelligence to predict water levels. In order to make up for the shortcomings of single artificial intelligence model, some scholars also used the neural network model coupling method to study water level prediction, and a small number of scholars input a single variable to predict water level. The above research shows that it is feasible to use the coupled model for water level prediction, and the advantages of multiple models complement each other, and the prediction accuracy is improved to different degrees compared with the previous single model. Considering various practical factors, the monitoring data of water level of Siling Reservoir was taken as an example and the coupling prediction model of water level of reservoir was put forward based on Embedding GRU on the condition that there was only a single characteristic rainfall, in order to provide a reference for realizing the high-precision prediction of water level with single characteristics. According to the rainfall scale of the data set and the largest rainfall in the history of Zhejiang Province, the training parameter rainfall scale sets of the Embedding stage is defined as {500,550,600,650,700,750} with the accuracy of mm×10?1. In order to study the optimal parameter setting, the range of feature dimension setting was extended to {2,3,4,5,6} on the premise of adopting the control variates. The ERMS indicator was selected for this experiment. To further validate the predictive performance and generalization ability of the Embedding GRU model, an experiment was conducted based on the total daily rainfall to predict the next day's reservoir water level. The comparison algorithm is still LSTM, GRU, and BiGRU, with a total of 1826 sets of data with a data volume of 5 years.Compared with other existing artificial intelligence models of reservoir water level, the prediction accuracy is higher and the scope of reservoir is wider. In the comparative experiment of predicting the next hour's water level, the prediction ability of the four models was excellent, and they could fit the real water level data relatively accurately, which shows that the method of predicting the reservoir water level by deep learning algorithm is effective and feasible. By comparing of prediction accuracy of four models, the experiment proved that GRU algorithm is better than LSTM in prediction effect, and the embedding method can further effectively reduce the prediction error and improve the prediction accuracy of the model.It is the Embedding method that enlarges the features between rainfall and climate, coupled with lightweight deep learning algorithm GRU to predict reservoir water level. Conclusions are as follows:(1)The prediction accuracy of the Embedding GRU model is obviously better than that of LSTM, GRU, BiGRU and other single deep learning models.(2)Embedding parameters in the Embedding GRU model shall be determined by comparative test according to the actual data set.(3)The Embedding-GRU model has excellent performance in predicting different period of multiple times within 7 days, and has good prediction effect and generalization ability, which fully proves the effectiveness of the model.