Recent advances in deep learning, especially the long short-term memory (LSTM) networks, provide some useful insights on how to tackle time series prediction problems, not to mention the development of a time series model itself for prediction. Runoff forecasting is a time series prediction problem with a series of past runoff data (water level and discharge series data) as inputs and a fixed-length series of future runoff as output. Most previous work paid attention to the sufficiency of input data and the structural complexity of deep learning, while less effort has been put into the consideration of data quantity or the processing of original input data—such as time series decomposition, which can better capture the trend of runoff—or unleashing the effective potential of deep learning. Mutual information and seasonal trend decomposition are two useful time series methods in handling data quantity analysis and original data processing. Based on a former study, we proposed a deep learning model combined with time series analysis methods for daily runoff prediction in the middle Yangtze River and analyzed its feasibility and usability with frequently used counterpart models. Furthermore, this research also explored the data quality that affect the performance of the deep learning model. With the application of the time series method, we can effectively get some information about the data quality and data amount that we adopted in the deep learning model. The comparison experiment resulted in two different sites, implying that the proposed model improved the precision of runoff prediction and is much easier and more effective for practical application. In short, time series analysis methods can exert great potential of deep learning in daily runoff prediction and may unleash great potential of artificial intelligence in hydrology research.

中文翻译：

---

具有时间序列分析方法的深度学习框架用于径流预测

深度学习的最新进展，特别是长短期记忆（LSTM）网络，为如何解决时间序列预测问题提供了一些有用的见解，更不用说开发用于预测的时间序列模型本身了。径流预测是一个时间序列预测问题，以一系列过去的径流数据（水位和排放序列数据）作为输入，而一个固定长度的未来径流序列作为输出。之前的大多数工作都关注输入数据的充分性和深度学习的结构复杂性，而在考虑数据量或原始输入数据的处理方面却投入了较少的精力，例如时间序列分解，可以更好地捕获数据。径流的趋势，或者释放深度学习的有效潜力。互信息和季节趋势分解是处理数据量分析和原始数据处理的两种有用的时间序列方法。在之前的研究的基础上，我们提出了一种结合时间序列分析方法的深度学习模型，用于长江中游的每日径流预报，并通过经常使用的对应模型分析了其可行性和可用性。此外，本研究还探讨了影响深度学习模型性能的数据质量。借助时间序列方法，我们可以有效地获取有关深度学习模型中采用的数据质量和数据量的一些信息。比较实验产生了两个不同的站点，这表明所提出的模型提高了径流预测的精度，并且在实际应用中更加容易和有效。简而言之，时间序列分析方法可以在每日径流预测中发挥深度学习的巨大潜力，并可以在水文学研究中释放人工智能的巨大潜力。