This project contains necessary data and codes to reproduce our results in our proposed aggregation and learning framework.

Project description
Requirements
ANNs training
License

This project contains the source codes and post-processing pipelines of the methodology and results presented in our article. The major parts of this repository contain the following:

• Graph generator: a graph generator using graph-theory based processes to generate District Heating Networks like graphs. A paper which dives into the details of this graph generator will be presented in future conference. Graphic representation and manipulation are realized using Networkx, a python package built for creation, manipulation and study of complex networks structures.

• Clusters learning: we provide in this repository the source codes to randomly select clusters of heat consumers within District Heating Networks using random-walks and the necessary machine learning pipeline codes to train different architectures of Artificial Neural Networks models and replace the clusters. The machine learning models are created using Keras, an open-access and straightforward to-use deep learning library built on top of TensorFlow. Not defined hyperpameters use default values provided by Keras. By default, TensorFlow performs training on gpu if available. Training processes have been realized using on the High Performance Computer Jean Zay.

• Hybrid simulation results: we also provide post-processing jupiter notebooks to retrieve the results of the hybrid simulation presented in our paper. Physical and Hybrid modeling codes are not available in this repository as they will be made available within an in-progress journal paper presenting in all details the physical components. However, our paper presents the major physical equations involved in these modeling codes. Further details can be shared at request.

To install all required packages, enter the following line commands at the project source folder

python -m pip install -r requirements.txt

Physical simulation results, needed for ML models training, case study networks topology and selected clusters json files can be downloaded from open-access Mendeley Data. Simulation folder ARTICLE_dhn_data must be available at the project source folder for training codes to work.

Selected clusters are found in considered_clusters_1.json and considered_clusters_2.json refereed as set 1 and set 2 for each of the case study network. Clusters are identified by the parent DHN (INTEGER), the set (INTEGER) and a key (STRING). Considered ANNs achitectures are [GRU, MLP, CNN] with two versions each. Further details can be found in our paper. For instance, to train a cluster denoted with key ("a") belonging to the network (DHN 1) and from the set 1, the following command may be used.

python train.py --model=gru --model_version=1 --cluster_key="a" --cluster_dhn_id=1 --cluster_set_id=1

Data and codes in this project are protected under the European Union Public Licence (EUPL) v1.2. For more information see LICENSE.