Official implementation of generative models to compute scenario of renewable generation and consumption on the GEFcom2014 open dataset presented in the paper: A deep generative model for probabilistic energy forecasting in power systems: normalizing flows.
If you make use of this code, please cite our paper:
@article{DUMAS2022117871,
title = {A deep generative model for probabilistic energy forecasting in power systems: normalizing flows},
journal = {Applied Energy},
volume = {305},
pages = {117871},
year = {2022},
issn = {0306-2619},
doi = {https://doi.org/10.1016/j.apenergy.2021.117871},
url = {https://www.sciencedirect.com/science/article/pii/S0306261921011909},
author = {Jonathan Dumas and Antoine Wehenkel and Damien Lanaspeze and Bertrand Cornélusse and Antonio Sutera},
keywords = {Deep learning, Normalizing flows, Energy forecasting, Time series, Generative adversarial networks, Variational autoencoders},
abstract = {Greater direct electrification of end-use sectors with a higher share of renewables is one of the pillars to power a carbon-neutral society by 2050. However, in contrast to conventional power plants, renewable energy is subject to uncertainty raising challenges for their interaction with power systems. Scenario-based probabilistic forecasting models have become a vital tool to equip decision-makers. This paper presents to the power systems forecasting practitioners a recent deep learning technique, the normalizing flows, to produce accurate scenario-based probabilistic forecasts that are crucial to face the new challenges in power systems applications. The strength of this technique is to directly learn the stochastic multivariate distribution of the underlying process by maximizing the likelihood. Through comprehensive empirical evaluations using the open data of the Global Energy Forecasting Competition 2014, we demonstrate that this methodology is competitive with other state-of-the-art deep learning generative models: generative adversarial networks and variational autoencoders. The models producing weather-based wind, solar power, and load scenarios are properly compared in terms of forecast value by considering the case study of an energy retailer and quality using several complementary metrics. The numerical experiments are simple and easily reproducible. Thus, we hope it will encourage other forecasting practitioners to test and use normalizing flows in power system applications such as bidding on electricity markets, scheduling power systems with high renewable energy sources penetration, energy management of virtual power plan or microgrids, and unit commitment.}
}
Note: the reference will be changed if the paper is accepted for publication in Applied Energy
Two libraries are required to implement the normalizing flows models used in the study:
- https://github.com/AWehenkel/Normalizing-Flows -> to access the repositories lib and model
- https://github.com/AWehenkel/UMNN -> to implement the Unconstrained Monotonic Neural Networks normalizing flows
If you make use of the Unconstrained Monotonic Neural Networks code, please cite the paper:
@inproceedings{wehenkel2019unconstrained,
title={Unconstrained monotonic neural networks},
author={Wehenkel, Antoine and Louppe, Gilles},
booktitle={Advances in Neural Information Processing Systems},
pages={1543--1553},
year={2019}
}
Concerning the forecast value assessment: the Python Gurobi library is used to implement the algorithms in Python 3.7, and Gurobi 9.0.2 to solve all the optimization problems.
The GEFcom2014/data folder contains the GEFcom2014 dataset: load, wind, and PV tracks.
The GEFcom2014/models folder contains the generative models in the following folders:
- GAN: generative adversarial networks
- VAE: variational autoencoders
- NFs: normalizing flows (affine and unconstrained monotonic neural networks)
- GC: gaussian copula
- RAND: random
Then, inside each folder, there is a Python file to run the model. For instance, for the NFs:
nf_autoregressive.py For each model, the structure of the Python file is similar.
First, select the track by indicating the track considered:
tag = 'load' # pv, wind, loadSecond, specify the number of scenarios and quantiles to be generated per day:
n_s = 100 # number of scenarios
N_q = 99 # number of quantilesThe hyper-parameters of the models for each track are already specified but can be modified. By default, they are set to the values of the paper.
Finally, by running the Python file, the model is trained and generates scenarios and quantiles over the learning, validation, and testing set.
Note: the function
quantiles_and_evaluation(dir_path=dir_path, s_VS=s_VS, s_TEST=s_TEST, N_q=N_q, df_y_VS=df_y_VS, df_y_TEST=df_y_TEST, name=name, ymax_plf=ymax_plf, ylim_crps=ylim_crps, tag=tag, nb_zones=nb_zones)allows computing some metrics over the validation and testing sets.
The GEFcom2014/forecast_quality folder contains the Python files to evaluate the forecast quality of the generative models. The scenarios to be evaluated are located in the folder GEFcom2014/forecast_quality/scenarios. This folder contains the scenarios per generative model: GAN, GC, NFs, etc.
For instance, GEFcom2014/forecast_quality/scenarios/gan comprises scenarios over the load, wind, and PV tracks over the testing set.
The file:
CRPS_QS_all_tracks.pyallows computing the CRPS, QS, and reliability diagram metrics for all tracks and models.
The file:
CRPS_QS_DM_test.pycomputes DM-statistical test of the CRPS and QS metrics for all models and tracks.
The file:
ES_VS_metrics.py computes the ES, VS multivariate metrics for all tracks and models. It also computes the DM-statistical test.
The file:
scenario_correlations.pycomputes the correlation matrices between scenarios for a given day.
The file:
clf_metric.pycontains the classifier-based metric. Note: scenarios over the learning set are required to fit the classifier. Due to the size of the files, they are not included in the GEFcom2014/forecast_quality/scenarios folder. They need to be generated and included in this folder.
The GEFcom2014/forecast_value/ folder contains the Python files to evaluate the forecast value of the generative models. It uses the retailer energy case study to bid optimally on a day-ahead basis by solving a stochastic optimization problem.
The file:
bidding_retailer.py computes the day-ahead bids for all models and the corresponding profits.
The parameters:
nb_s = 50 # number of scenarios considered per stochastic optimization problem
soc_max = 1 # battery storage capacity
dad_price = 100 # euros /MWh
q_pos = 2
q_neg = 2
pos_imb = q_pos * dad_price # euros /MWh
neg_imb = q_neg * dad_price # euros /MWhare set to the default values used in the paper.
Warning: the Gurobi Python API must be activated to use this file.