Sequence Deep Learning Models (LSTM/GRU/TCN) with Data for Streamflow Forecasting
This repo aims to share the models in the paper "A rainfall‐runoff model with LSTM‐based sequence‐to‐sequence learning" and "Distributed long-term hourly streamflow predictions using deep learning–A case study for State of Iowa.". We simply shared one sample data of the catchment of USGS 05389000 Yellow River near Ion, IA in 521_data.csv.
It is the original data with missing values and no pretreatment such as moving average. The precipitation is in the unit of mm. It is obtained from StaveIV and calculated by zonal statistics operation. The ET is the empirical values for Iowa in the unit of mm/month. The discharge data is from the USGS. and in the unit of m3/s.
The whole dataset in the Model 2 will be shared later.
To predict the hourly discharge as good as we can. So, both models are beyond the traditional rainfall-runoff models since we used the current discharge for data assimilation.
This repo provides an end-to-end example from the original data to the results and evaluation. It is created in Google's Colab, working on Python 3 with Tensorflow>=2.0.
Model 1: Encoder-Decoder LSTM (LSTM-seq2seq).
This model have shown strong predictability on the 24 hours predictions as presented in the paper:
Xiang, Z., Yan, J., & Demir, I. (2020). A rainfall‐runoff model with LSTM‐based sequence‐to‐sequence learning. Water resources research, 56(1), e2019WR025326. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2019WR025326
Model 2: Neural Runoff Model (NRM).
To solve the bottle-neck problem of the state vector in model 1, we proposed the second model for the prediction of the future 120 hours.
The model is presented in the paper: Xiang, Z., & Demir, I. (2020). Distributed long-term hourly streamflow predictions using deep learning–A case study for State of Iowa. Environmental Modelling & Software, 104761. https://www.sciencedirect.com/science/article/abs/pii/S1364815220301900
We developed a generalized model with a multi-site structure for hourly streamflow forecast on 125 USGS gauges. Considering watershed-scale features including drainage area, time of concentration, slope, and soil types, the proposed models have acceptable performance and slightly higher model performance than training individual models for each USGS station.
The model is presented in the paper: Xiang, Z., Demir, I., Mantilla, R., & Krajewski, W. F. (2021). A Regional Semi-Distributed Streamflow Model Using Deep Learning. DOI: https://doi.org/10.31223/X5GW3V
This project is developed by the University of Iowa Hydroinformatics Lab (UIHI Lab): https://hydroinformatics.uiowa.edu/.