This work proposed an innovative and unified energy forecasting framework, Hierarchical Neural Laplace (HNL) for multi-resolution energy forecasting. Given the desired resolutions, the corresponding forecasts can be seamlessly generated without re-training or post-processing.
Codes for the paper "Seamless and multi-resolution energy forecasting".
Authors: Chenxi Wang, Pierre Pinson, Yi Wang
Python version: 3.8.10
The must-have packages can be installed by running
pip install requirements.txt
All the data for experiments are saved in datasets/
. In experiments/01-EDA.ipynb
, we show the basic plots on both energy data and weather forecasts data.
To reproduce the experiments in the paper, please run
cd experiments/
bash run_experiments.sh
Note: There is NO multi-GPU/parallelling training in our codes.
The results(models) and logs will be saved into new folders i.e. results/
and logs/
under experiments/
.
Then, go into experiments/02-post_analysis.ipynb
for post analysis, including the post-coordination for benchmarks and the comparison of the total consistency error. Please replace the variable pth
in the notebook.
If you also want to have the same figures in the paper, please refer to experiments/03-plot.ipynb
.
An example of seamless multi-resolution forecasts on wind power from Hierarchical Neural Laplace(HNL) and benchmarks.
Package torchlapalace/
is modified based on the open code of Neural Laplace. The rapid development of this work would not have been possible without this open-souce package.