JointDemandForecasting.py

A framework for multi-step probabilistic time-series/demand forecasting models. Code for "Conditional Approximate Normalizing Flows for Joint Multi-Step Probabilistic Forecasting with Application to Electricity Demand". If you find this repository useful, please cite the following:

@article{jamgochian2022conditional,
	author = {Arec Jamgochian and Di Wu and Kunal Menda and Soyeon Jung and Mykel J. Kochenderfer},
	title = {Conditional Approximate Normalizing Flows for Joint Multi-Step Probabilistic Forecasting with Application to Electricity Demand},
	journal = {arXiv:2201.02753 [cs]},
	year = {2022}
}

File stucture

• JointDemandForecasting contains the source code for the package
• datasets contains datasets pertaining to experiments.
  • datasets/raw contains raw datasets
  • datasets/processed contains processed datasets
  • datasets/process_utils contains utilities for processing utilities
  • datasets/README.md contains instructions for sourcing at processing all datasets.
• experiments contains any experiments conducted in the submitted manuscript.

Environment

A conda environment for exact experiment reproduction can be set up and uses with:

conda create -y --name jdf python==3.7
conda install -y --name jdf --file requirements.txt
conda activate jdf
...
conda deactivate

To enable CUDA capabilities, rerun:

conda install --name jdf pytorch cudatoolkit=10.2 -c pytorch

Running Experiments

To run experiments from the associated paper:

1. Install and activate the environment.
2. Follow directions in the datasets folder to download and process the OPENEI dataset.
3. Navigate the the experiments/arxiv/ folder and run the appropriate scripts:
   • Toy Density Estimation:
     • the best hyperparameter config from our experiments is saved in experiments/arxiv/results/simple_square_best_hyperparams.json
     • to run the experiment from the paper with this config, run python simple_experiment.py --test
     • to retune hyperparameters based on a grid search with a grid defined in-file, run python simple_experiment.py --tune
   • Electricity Demand Application
     • with, for example, python ConditionalModels.py
       • ARMA, IFNN, IRNN models are in IterativeModels.py, CG and CGMM are in ConditionalModels.py, MOGP is in MultiOutputGP.py, JFNN is in JFNN.py, JRNN is in JRNN.py, and CANF is inCANF.py
       • pick appropriate setting (location, input length, output length, etc.) by uncommenting appropriate line
     • our multi-trial results, as well as run-rejection for outlier trials can be seen in experiments/arxiv/results (no CANF trials were rejected)
     • plots can be generated with python plots.py