hydromodel

image image

image

What is hydromodel

Hydromodel is a python implementation for common hydrological models such as the XinAnJiang (XAJ) model, which is one of the most famous conceptual hydrological models, especially in Southern China.

Not an official version, just for learning

This is a Python console program (no graphic interface now). It is still developing.

How to run

Install

We provided a pip package. You can install it with pip:

# create python environment
$ conda create -n hydromodel python=3.10
$ conda activate hydromodel
# install hydromodel
$ pip install hydromodel

If you want to run the model as a developer, you can clone the repository

# fork hydromodel to your GitHub, and clone it to your computer
$ git clone <address of hydromodel in your github>
# move to it
$ cd hydromodel
# if updating from upstream, pull the new version to local
$ git pull
# create python environment
$ mamba env create -f env-dev.yml
# if mamba is not installed:
# $ conda install -c conda-forge mamba
# activate it
$ conda activate hydromodel

Prepare data

You can use the CAMELS dataset (see here to prepare it) to run the model.

If CAMELS is used, you can skip this step.

If the warning “hydro_setting.yml not found” is shown, you can create a hydro_setting.yml file in the user directory. In windows, it is usually “C:/Users/". In Linux, it is usually "/home/". The content of the file is as follows:

# Update the following with your own settings, if you don't have, you can ignore it
minio:
  server_url: ''
  client_endpoint: ''
  access_key: ''
  secret: ''
local_data_path:
  root: ''
  datasets-origin: ''
  datasets-interim: ''
  basins-origin: ''
  basins-interim: ''
postgres:
  server_url: ''
  port: 5432
  username: ''
  password: ''
  database: ''

To use your own data to run the model, you need prepare the data in the required format.

NOTE: We currently only support calibrating one basin at a time, although the following instructions cover multiple basin information, we have not thoroughly tested it.

We provide some transformation functions in the “scripts” directory. You can use them to transform your data to the required format.

But you still need to do some manual work before transformation. Here are the steps:

  1. Put all data in one directory and check if it is organized as the following format: 
    your_data_directory_for_hydromodel/
# one attribute csv file for all basins
├─ basin_attributes.csv
# one timeseries csv file for one basin, xxx and yyy are the basin ids
├─ basin_xxx.csv
├─ basin_yyy.csv
├─ basin_zzz.csv
├─ ...

    basin_attributes.csv should have the following columns:

    id     name  area(km^2)
xxx  Basin A         100
yyy  Basin B         200
zzz  Basin C         300

    basin_xxx.csv should have the following columns:

    time  pet(mm/day)  prcp(mm/day)  flow(m^3/s)  et(mm/day)  node1_flow(m^3/s)
2022-01-01 00:00:00            1                 10                 13                 16                 19
2022-01-02 00:00:00            2                 11                 14                 17                 20
2022-01-03 00:00:00            3                 12                 15                 18                 21

    The sequence of the columns is not important, but the column names should be the same as the above. No more unnecessary columns are allowed. For time series csv files, et and node1_flow are optional. If you don’t have them, you can ignore them. The units of all variables could be different, but they cannot be missed and should be put in () in the column name.

  2. Use prepare_data.py – run the following code to transform the data format to the required format: 
    $ python prepare_data.py --origin_data_dir <your_data_directory_for_hydromodel>

Run the model

To run calibration with CAMLES dataset, you can use the following code:

# just an example the hyper-parameters of the model and the algorithm should be tried by yourself
$ python calibrate_xaj.py --data_type camels --data_dir "C:/Users/wenyu/OneDrive/data/camels/camels_us" --exp expcamels001 --cv_fold 2 --warmup 365 --period 2007-01-01 2014-01-01 --calibrate_period 2007-01-01 2014-01-01 --test_period 2007-01-01 2014-01-01 --basin_id 01439500 06885500 08104900 09510200 --model "{\"name\": \"xaj\", \"source_type\": \"sources5mm\", \"source_book\": \"HF\"}" --algorithm "{\"name\": \"SCE_UA\", \"random_seed\": 1234, \"rep\": 10, \"ngs\": 10, \"kstop\": 5, \"peps\": 0.1, \"pcento\": 0.1}" --loss "{\"type\": \"time_series\", \"obj_func\": \"RMSE\", \"events\": null}"

To use your own data, run the following code:

# you can change the algorithm parameters:
$ python calibrate_xaj.py --data_type owndata --data_dir "C:/Users/wenyu/OneDrive/data/biliuhe" --exp expbiliuhe001 --cv_fold 1 --warmup 720 --period "2012-06-10 00:00" "2022-08-31 23:00" --calibrate_period "2012-06-10 00:00" "2017-08-31 23:00" --test_period "2017-09-01 00:00" "2022-08-31 23:00" --basin_id 21401550 --model "{\"name\": \"xaj\", \"source_type\": \"sources5mm\", \"source_book\": \"HF\"}" --param_range_file "C:/Users/wenyu/OneDrive/data/biliuhe/param_range.yaml" --algorithm "{\"name\": \"SCE_UA\", \"random_seed\": 1234, \"rep\": 10, \"ngs\": 10, \"kstop\": 5, \"peps\": 0.1, \"pcento\": 0.1}" --loss "{\"type\": \"time_series\", \"obj_func\": \"RMSE\", \"events\": null}"
# for advices of hyper-parameters of sceua, please see the comment of the function 'calibrate_xaj.py'

NOTE: For the parameter range in the param_range_file file. You can copy it from “hydromodel/models/param.yaml” of this repo and put it anywhere you want. Then you can modify the parameter range in the file. The parameter range is used to limit the parameter space of the hydromodels. If you don’t provide the file, the default parameter range will be used.

Then you can evaluate the calibrated model with the following code:

# $ python evaluate_xaj.py --exp expcamels001
# for your own data
$ python evaluate_xaj.py --exp expbiliuhe001

See the results

Run the following code to see the results of the evaluation:

# $ python visualize.py --exp expcamels001
# for your own data
$ python visualize.py --exp expbiliuhe001

You will see the results in the example directory.

Why does hydro-model-xaj exist

When we want to learn about the rainfall-runoff process and make forecasts for floods, etc. We often use classic hydrological models such as XAJ as a baseline because it is trusted by many engineers and researchers. However, after searching the website very few repositories could be found. One day I happened to start learning Python, so I decided to implement the model with Python. Previous commits for hydro-model-xaj have some errors, but now at least one executable version is provided.

Actually open-source science has brought a great impact on hydrological modeling. For example, SWAT and VIC are very popular now as they are public with great performance and readable documents; as more and more people use them, they become more stable and powerful. XAJ is a nice model used by many engineers for practical production. We need to inherit and develop it. I think hydro-model-xaj is a good start.

What are the main features

We basically implement the formula in this book – 《流域水文模拟》

Other reference Chinese books：

More English references could be seen at the end of this README file.

The model mainly includes three parts:

For the first part, we use an evaporation coefficient K (ratio of potential evapotranspiration to reference crop evaporation generally from Allen, 1998) rather than Kc (the ratio of potential evapotranspiration to pan evaporation) because we often use potential evapotranspiration data from a system like GLDAS, NLDAS, etc. But it doesn’t matter, when you use pan evaporation, just treat K as Kc.

For the second part, we provide multiple implementations, because, for this module, formulas in different books are a little different. One simplest version is chosen as a default setting. More details could be seen in the source code directly now. We provide four versions, two versions from two books.

For the third part – routing module, we provide different ways: the default is a common way with recession constant ( CS) and lag time (L) shown in the figure; second (You can set the model’s name as “xaj_mz” to use it) is a model from mizuRoute to generate unit hydrograph for surface runoff (Rs -> Qs) , as its parameters are easier to set, and we can optimize all parameters in a uniform way.

We provide two common calibration methods to optimize XAJ’s parameters:

Now the model is only for one computing element (typically, a headwater catchment). Soon we will provide calibration for multiple headwater catchments. To get a better simulation for large basins, a (semi-)distributed version may be needed, and it is not implemented yet. The following links may be useful:

We also provide a differentiable version of XAJ, which is based on the PyTorch framework.

The idea comes from this paper: From calibration to parameter learning: Harnessing the scaling effects of big data in geoscientific modeling by Tsai et al. (2021). We use the same structure as the original XAJ model but replace the original Numpy code with PyTorch. Then we can use the automatic differentiation technique and stochastic gradient descent algorithms to optimize all parameters. The advantage of this method is that we can use the same code to optimize many basins at once and use big data to improve the model performance. Generally, with the native parallel computing ability of PyTorch, the differentiable version is faster than the original version without any parallel processing. The differentiable version is in torchhydro.

Other implementations for XAJ:

How to contribute

If you want to add features for hydro-model-xaj, for example, write a distributed version for XAJ, please create a new git branch for your feature and send me a pull request.

If you find any problems in hydro-model-xaj, please post your questions on issues.

References

This site is open source. Improve this page.