DARNet

Original Implementation of "DARNet: Dual-head Attention Residual Network for Multi-Step Short-Term Load Forecasting"

Clone project

You need to clone the project first.

git clone https://github.com/Mark-THU/DARNet-load-forecasing.git
cd DARNet-load-forecasing

Dependencies

python 3.8.12

You may check and install other dependent packages in requirments.txt.

$ pip install -r requirements.txt

To install Pytorch, you may follow the official instructions here.

Usage

The following examples will show you how to train and test a model on a specific dataset with settings used in this work.

DARNet

$ python main.py --dataset D1 --model DARNet --seq_len 72 --target_len 24 --batch_size 256 \
                 --lr 0.01 --num_epoch 80 --hidden_dim 64 --n_layers 3 --drop_prob 0.5 \
                 --weight_decay 0 --mse_thresh 0.01 --gpu cuda:0 --seed 0

Training losses and evaluation losses will be represented in a figure named DARNet.png, which will be saved in ./figs folder. The trained model's state dict will be saved in the ./models/model folder. Each metrics utilized in the paper will be printed on the screen. Ground-truth values as well as predicted values will be saved in the ./results folder.

BiLSTM

$ python main.py --dataset D1 --model BiLSTM --seq_len 72 --target_len 24 --batch_size 256 \
                 --lr 0.01 --num_epoch 80 --hidden_dim 64 --n_layers 3 --drop_prob 0 \
                 --weight_decay 0 --mse_thresh 0.01 --gpu cuda:0 --seed 0

GRU_CNN

$ python main.py --dataset D1 --model GRU_CNN --seq_len 72 --target_len 24 --batch_size 256 \
                 --lr 0.01 --num_epoch 80 --hidden_dim 128 --n_layers 1 --drop_prob 0 \
                 --weight_decay 0.00001 --mse_thresh 0.01 --gpu cuda:0 --seed 16

Seq2Seq

$ python main.py --dataset D1 --model Seq2Seq --seq_len 72 --target_len 24 --batch_size 512 \
                 --lr 0.01 --num_epoch 80 --hidden_dim 64 --n_layers 2 --drop_prob 0.5 \
                 --weight_decay 0 --mse_thresh 0.01 --gpu cuda:0 --seed 42