Step 3: Scaled-up Domain and Interactions of Heat Plumes

<p>This dataset serves as training data for modeling the temperature field emanating from open-loop groundwater heat pumps.</p>

<p>The dataset was simulated in 2D with Feflow using cut-outs from <a href="https://www.lfu.bayern.de/geologie/gepo/index.htm"> interpolated hydrogeological measurements of the Munich, Germany, region</a>. Heat pump locations are realistic positions, and extraction rates are adapted to fit the available groundwater.<br>
To prepare the data for machine learning, it was transformed from unstructured to structured data (2560x2560 cells) with Python. Inputs consist of hydrogeological parameters such as hydraulic conductivity [m/d], log(conductivity), hydraulic head [m a.s.l.], hydraulic head gradient, transmissivity [m^2/d], aquifer thickness [m], and operational pump parameters such as maximum flow rate [m^3/d]. All hydrogeological parameters were extracted prior to heat pump operation. Additionally, the Darcy velocities in the x- and y-directions [m/d] before heat pump operation can be used (and are already part of the inputs), as they can be easily estimated from other input fields.<br>
Heat pumps are operated under seasonal load as depicted in `prepared/normed_flow_injection_series.npy` and `prepared/temperature_injection_series.npy`. The former contains the normalized flow rates of the heat pumps, while the latter contains the corresponding injection temperatures.</p>

The prepared data is split into training and validation data. Each contains a set of inputs (`inputs_unnormed`) and labels (`labels_unnormed`). The overall normalization information is contained in `general/properties_info_normalization.yaml`. The order of inputs is given by `index_orig` in the info-yaml.<br>
The train and validation splits are mere suggestions and can be chosen differently.</p>

The test data is hidden for final evaluation and only contains the last time step of the temperature label in a format of (N, C, H, W) with N being the number of samples, C the number of channels (1 in this case), and H and W the spatial dimensions of the grid.