Univariate-Time-Series-Prediction-using-Deep-Learning

0. Overview

This repository provides Univariate Time Series Prediction. It supports;

• using various deep learning models including DNN, CNN, RNN, LSTM, GRU, and Attentional LSTM.
• using single-step and multi-step prediction.

The dataset used is Appliances Energy Prediction Data Set and can be found here.

1. Quantitative Analysis

According to the table below, CNN using 1D Convolutional layer outperformed the other models on single-step time series prediction.

Model	MAE↓	MSE↓	RMSE↓	MPE↓	MAPE↓	R Squared↑
DNN	29.3038	3673.7921	57.0114	-15.6800	26.4763	0.3820
CNN	27.5182	3614.1634	56.1604	-11.2039	23.7301	0.4057
RNN	29.1327	3627.1491	56.7243	-16.2193	26.9323	0.3809
LSTM	29.6157	3575.5541	56.4002	-16.7178	27.9683	0.3771
GRU	29.0402	3564.9701	56.2790	-16.9984	26.9390	0.3872
Attentional LSTM	28.9658	3603.0751	56.3838	-16.8199	26.3129	0.3898

According to the table below, DNN outperformed the other models on multi-step time series prediction.

Model	MAE↓	MSE↓	RMSE↓	MPE↓	MAPE↓	R Squared↑
DNN	31.3555	2913.6521	49.3946	-16.7329	29.1459	0.1775
CNN	32.9762	2893.2201	49.5900	-21.7513	32.3016	0.1206
RNN	32.9153	2951.9055	50.0931	-20.7460	32.2081	0.1223
LSTM	32.8141	2955.5278	50.1237	-20.5471	32.0873	0.1191
GRU	33.0092	2927.5575	49.9503	-21.2869	32.5345	0.1177
Attentional LSTM	32.2182	2920.8744	49.7972	-19.1188	30.8223	0.1347

2. Qualitative Analysis

It definitely suffers from the typical lagging issue. Also, I averaged multi-step for plotting thus it looks to be smoothed.

3. Run the Codes

1) Train

If you want to train Attention LSTM,

python main.py --model 'attention'

If you want to train with multi-step with time step of 5,

python main.py --model 'attention' --multi_step True --output_size 5

2) Test

python main.py --model 'attention' --mode 'test'

To handle more arguments, you can refer to here.

Development Environment

- Windows 10 Home
- NVIDIA GFORCE RTX 2060
- CUDA 10.2
- torch 1.6.0
- torchvision 0.7.0
- etc