DeepHydro

Code for "Forecasting the Evolution of Hydropower Generation"

Prerequisites

Python >= 3.6

PyTorch >= 1.1.0

Numpy >= 1.15.1

Install torchdiffeq from https://github.com/rtqichen/torchdiffeq.

Visualization

The distributions of latent representation along with the process of learning temporal dependencies in our CL-RNN model:

The process of transforming latent representation via continuous normalizing flow:

Dateset

We use two different types datasets, namely DGS(large-scale, 1/1/2017--31/12/2018) and PDS(small-scale, 1/1/2017--31/12/2018), to demonstrate DeepHydro performs the best against other baselines. The data of last 11 weeks (77 days) of the year are used for testing, and the rest for training. The more details descriptions can be obtained in the paper. Due to company policy, our data will be announced in a few months later. Here, we present the figure that shows the power stations distribution of Dadu River:

Prediction

To better observe the visualization of predcition, we randomly select the data of one week on DGS dataset and plot the predicted results of DeepHydro and ground truth for comparison:

Experiments on different datasets

• Experiment on dataset DGS-P1:

  python train.py --batch_size 128 --lr 1e-4 --ext True --target_year 2017 --model "DeepHydro" --dataset "DGS"

• Experiment on dataset DGS-P2:

  python train.py --batch_size 128 --lr 1e-4 --ext True --target_year 2018 --model "DeepHydro" --dataset "DGS"

• Experiment on dataset PDS-P1:

  python train.py --batch_size 128 --lr 1e-4 --ext True --target_year 2017 --model "DeepHydro" --dataset "PDS"

• Experiment on dataset PDS-P2:

  python train.py --batch_size 128 --lr 1e-4 --ext True --target_year 2018 --model "DeepHydro" --dataset "PDS"