Deep Learning for Multivariate Time Series Forecasting

Mainly referenced paper

Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks.(https://arxiv.org/abs/1703.07015)

Environment

• Python 3.6+
• Pytorch 1.0+
• numpy

Example

1. Exchange Rate dataset:stock.sh
2. Traffic dataset: traffic.sh
3. Solar-Energy dataset:solar.sh
4. Electricity usage dataset: ele.sh

Instruction

main.py

• --data DATA location of the data file
• -h --help show this help message and exit
• --model DATA select the model： LSTNet, CNN, RNN or MHA_Net
• --window WINDOW window size (history size)
• --horizon HORIZON forecasting horizon(step)
• --hidRNN HIDRNN number of RNN hidden units each layer
• --rnn_layers RNN_LAYERS number of RNN hidden layers
• --hidCNN HIDCNN number of CNN hidden units (channels)
• --CNN_kernel CNN_KERNEL the kernel size of the CNN layers
• --highway_window HIGHWAY_WINDOW The window size of the highway component
• -n_head N_HEAD num of self attention heads
• -d_k D_K self attention key dimension
• -d_v D_V self attention value dimension
• --clip CLIP gradient clipping limit
• --epochs EPOCHS upper epoch limit
• --batch_size N batch_size
• --dropout DROPOUT dropout applied to layers (0 = no dropout)
• --seed SEED random seed
• --log_interval N report interval
• --save SAVE path to save the final model'
• --cuda CUDA whether to use cuda device
• --optim OPTIM optimizer method ,default 'adam'
• --amsgrad AMSGRAD whether to use amsgrad
• --lr LR learning rate
• --skip SKIP autoregression window size
• --hidSkip HIDSKIP skiphidden states dimension
• --L1Loss L1LOSS whether to use l1 loss function
• --normalize NORMALIZE whether to normalize the data
• --output_fun OUTPUT_FUN relu, tanh or sigmoid