Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks.(https://arxiv.org/abs/1703.07015)
Exchange Rate dataset:
stock.shTraffic dataset:
traffic.shSolar-Energy dataset:
solar.shElectricity usage dataset:
ele.sh main.py
location of the data file
show this help message and exit
select the model: LSTNet, CNN, RNN or MHA_Net
window size (history size)
forecasting horizon(step)
number of RNN hidden units each layer
number of RNN hidden layers
number of CNN hidden units (channels)
the kernel size of the CNN layers
The window size of the highway component
num of self attention heads
self attention key dimension
self attention value dimension
gradient clipping limit
upper epoch limit
batch_size
dropout applied to layers (0 = no dropout)
random seed
report interval
path to save the final model'
whether to use cuda device
optimizer method ,default 'adam'
whether to use amsgrad
learning rate
autoregression window size
skiphidden states dimension
whether to use l1 loss function
whether to normalize the data
relu, tanh or sigmoid