Abstract Water-level fluctuation (WLF) forecast plays an essential role in the regulations and management of the Yangtze River’s navigable rivers. Traditional methods for navigable rivers’ forecasts are often time-consuming and cost-inefficient, often requiring constant updates and are univariate in nature. This study utilized one of the Yangtze River’s navigable rivers, Nanjing navigable river, to perform forecasts employing deep learning techniques. For the first time, Nanjing navigable river’s WLF data was exploited to make multi-step univariate and multivariate time series forecasts, employing a deep learning technique, multilayer perceptron (MLP). The data consisted of 3545 days worth of WLF measurements from ten different navigable rivers and was allocated properly without data leakage into training and testing. MLP was built to take time series sequence inputs from up to ten navigable rivers on a daily basis and output fast, accurate, stable, and reliable short-term forecasts and undeviating trends for long-term forecasts. The models were also constructed in an adaptive way so they can self-update daily based on the daily measurements. The naive model was constructed as a baseline model to measure the improvement and the validity of MLP models. Univariate and multivariate adaptive models were then constructed based on the augmented data by a popular data-augmentation method, rolling windows. Hyperparameters of MLP models were optimized based on relevant testing and large scale grid-search. Both short-term and long-term forecasts showed promises of MLP model in time series forecast for navigable river water-level problems and they achieved at least 27.6 % percent lower in root mean square error (RMSE) compared to the baseline. Since the total time required for the trained models to produce forecasts and self-update is at most 15 min on a local quad-core Intel Core i5 10th generation (2.0 GHz) and because of the nonnegotiable performance of the models, it was concluded that MLP is a fast, cost-efficient, and accurate alternative technique for navigable rivers WLF forecasts.

中文翻译：

---

基于自适应多层感知器的通航河流水位波动长短期预测研究

摘要 水位波动（WLF）预报在长江通航河流的整治和管理中发挥着重要作用。用于通航河流预测的传统方法通常耗时且成本低效，通常需要不断更新并且本质上是单变量的。本研究利用长江的通航河流之一——南京通航河流，利用深度学习技术进行预测。首次利用南京通航河流的WLF数据进行多步单变量和多变量时间序列预测，采用深度学习技术，多层感知器（MLP）。数据包括来自 10 条不同通航河流的 3545 天 WLF 测量值，并正确分配，没有数据泄漏到训练和测试中。MLP 旨在每天从多达 10 条通航河流中获取时间序列序列输入，并为长期预测输出快速、准确、稳定和可靠的短期预测和不变的趋势。这些模型还以自适应方式构建，因此它们可以根据每日测量结果每天自我更新。朴素模型被构建为基线模型，以衡量 MLP 模型的改进和有效性。然后通过流行的数据增强方法滚动窗口基于增强数据构建单变量和多变量自适应模型。基于相关测试和大规模网格搜索优化了 MLP 模型的超参数。短期和长期预测都显示了 MLP 模型在可通航河流水位问题的时间序列预测中的前景，并且至少达到了 27。与基线相比，均方根误差 (RMSE) 降低了 6%。由于在本地四核英特尔酷睿 i5 第 10 代 (2.0 GHz) 上，训练模型生成预测和自我更新所需的总时间最多为 15 分钟，并且由于模型的性能不可协商，因此得出的结论是MLP 是一种快速、经济且准确的可通航河流 WLF 预测替代技术。