AI4LS Hackathon #2: Task 1

Introduction

This project involves processing multiple CSV files containing time series data from thousands of measurement water stations in Austria . The data is cleaned, filtered, and merged by geographic coordinates to generate forecasts for a specific geographic location using machine learning models.

The goal is to train models on this data for specific geographic locations and predict the monthly groundwater levels from January 2022 to June 2024.

Groundwater levels are forecasted using 11 variables under 4 headings:

Groundwater:
1- Groundwater level,
2- Groundwater temperature,

Precipitation:
3- Rainfall,
4- Snowfall,

Water source:
5- Source flow rate,
6- Source temperature,
7- Source sediment levels,
8- Source conductivity,

Surface water:
9- Surface water levels,
10- Surface water temperature, and
11- Surface water flow rate.

Setup

You can clone the repository, navigate to the project directory, and install the necessary dependencies by running:

git clone https://github.com/gizemoge/AI4LS_Challenge.git
cd AI4LS2_Task1
pip install -r requirements.txt

Then you can run the script and obtain the forecast (created under groundwater_forecasts directory) by running:

python main.py

Methodology

1. Data Preprocessing

The four datasets from ehyd.gv.at related to groundwater, precipitation, water sources and surface waters in Austria are downloaded into the provided data structure. The folder names are renamed to their English equivalents, and umlaut substitution was applied to necessary subfolders, but no changes are made to the files themselves.

The measurement data is extracted from the CSV files, and a dictionary is created for each variable. In these dictionaries, the keys represent the date (year-month), and the dataframes contain the measurement station IDs as their indices. The coordinates are obtained from the messstellen_alle files.

Processing Data for a Specific Location

487 groundwater level measurement stations are forecasted in this document, though the model can forecast the groundwater levels for any and all of the 3792 groundwater measurement stations (as specified in the messstellen_alle.csv file located in the Groundwater folder).

To train the model for a specific geographic location, first specify the station IDs for that location in the gw_test_empty.csv file. The stations are filtered and processed using the function:

filtered_groundwater_dict, filtered_gw_coordinates = process_and_store_data(
    "datasets_ehyd/Groundwater/Grundwasserstand-Monatsmittel",
    groundwater_all_coordinates, "gw_", station_list)

Merging and Aligning Data Based on Coordinates

The counts of measurement stations per type (e.g., precipitation or surface water) are not consistent. The add_nearest_coordinates_column() function associates each measurement station with its nearest counterparts based on Euclidean distances. For some variables, this document selects three stations for triangulation, while for others, just one is selected, depending on the number of available measurement stations.

In this example, the 864 rain measurement stations are enough in number to be triangulated for each of the 3793 groundwater level measurement stations. Thus the three closest rain measurement station IDs are selected:

data = add_nearest_coordinates_column(rain_coord, 'nearest_rain', 3)

In comparison, there are only 94 water source flowrate measurement stations, and so only one nearest water source flowrate measurement station ID is taken.

2. Data Imputation

Missing data is handled by the nan_imputer() function, which fills the NaN values using monthly averages:

filled_filtered_groundwater_dict = nan_imputer(filtered_groundwater_dict)

3. Feature Engineering

To prepare data for machine learning models, lags and rolling means are added to the datasets using the function add_lag_and_rolling_mean(). Based on our literature review (particularly Sutanto et al., 2024), this function is defined to compute lagged values (1, 2, 3 months) and a rolling mean (over 6 months) for the dataframes. Each dataframe in the filled dictionaries has lagged and rolling mean features added:

df = add_lag_and_rolling_mean(filled_dict['groundwater'])

Two dictionaries are then created:

• new_dataframes is a dictionary storing DataFrames of each measurement station, containing all original and engineered features monthly from January 2022 to June 2024. The keys are the measurement station IDs, and the values are dataframes.
• monthly_dataframes contains dataframes that are monthly snapshots of all measurement stations combined, created by processing the data from new_dataframes. In this dictionary, the keys represent months instead of measurement station IDs.

Data is compiled monthly for the years between 1985 and 2021 in the monthly_dict_85to21 dictionary, allowing for a consolidated view of the data. The data is filtered to start from 1985 instead of 1960 to reduce the amount of imputed synthetic data, which is concentrated in the earlier years.

Lastly, an average correlation feature selection method is applied to identify features that correlate significantly with the target variable, groundwater level. Features with an absolute correlation above a definable threshold (10% by default, 40% in this document) are selected for further analysis.

4. Model Training

The SARIMAX (Seasonal Autoregressive Integrated Moving Average with eXogenous variables) model is used for forecasting as it accounts for both seasonal and non-seasonal factors in the time series data.

From the monthly_dict_85to21 dictionary, 5 years' worth of the data from '2015_01' to '2019_12' are taken as the train set, and 24 months from '2020_01' to '2021_12' as the validation set. Other start and cut-off dates can also be set.

Hyperparameter optimization is performed, aiming to determine the optimal order and seasonal order by calculating the Akaike Information Criterion (AIC) for 10 randomly sampled DataFrames, ultimately printing the best results. This process involves evaluating various parameter combinations to identify the model that provides the best balance between fit and complexity. The sample size can be adjusted as needed.

After hyperparameter optimization, a SMAPE score of 0.15 is obtained for the validation set, and this optimized model is selected for the final forecasting.

model = SARIMAX(
    station_data['Values'],
    exog=station_data.drop(columns=['Values']),
    order=(1, 0, 1),
    seasonal_order=(0, 1, 1, 12)
)
model_fit = model.fit(disp=False)

5. Forecasting Future Values

The model forecasts the next 30 time steps (months) for the given groundwater level measurement stations using the fitted model. The exogenous variables from the last 30 observations are used to improve forecast accuracy.

The forecasted values are compiled into a DataFrame and saved to a CSV file under the groundwater_forecasts directory. The team's submission for this challenge is named forecast_submitted_hydroscope.csv.

When the main script is run, a new CSV file of the forecasted values will be created under the same directory with the current date appended to its name.

References

Sutanto, S. J., Syaehuddin, W. A., & de Graaf, I. (2024). Hydrological drought forecasts using precipitation data depend on catchment properties and human activities. Communications Earth & Environment, 5, Article 118. https://doi.org/10.1038/s43247-024-01295-w