Interpretable Multivariate Time Series Forecasting with Temporal Attention Convolutional Neural Networks

This repository contains the official implementation for the models described in Interpretable Multivariate Time Series Forecasting with Temporal Attention Convolutional Neural Networks. These include a Temporal Convolutional Network (TCN), as well as the proposed Temporal Attention Convolutional Network (TACN) that combines a TCN with an attention mechanism.

If you find this work helpful in your research, consider citing our paper:

@INPROCEEDINGS{pantiskas2020tacn,  
author={L. {Pantiskas} and K. {Verstoep} and H. {Bal}},  
booktitle={2020 IEEE Symposium Series on Computational Intelligence (SSCI)},   
title={{Interpretable Multivariate Time Series Forecasting with Temporal Attention Convolutional Neural Networks}},   
year={2020},  
volume={},  
number={},  
pages={1687-1694},  
doi={10.1109/SSCI47803.2020.9308570}}

Requirements

The code is written in Python 3.7.7 and has the following dependencies for the training and evaluation notebooks:

• tensorflow==2.1.0
• tensorflow-addons==0.8.2
• tqdm==4.46.0
• seaborn==0.10.1
• scipy==1.4.1
• scikit-learn==0.22.1
• pygam==0.8.0
• pydotplus==2.0.2
• pandas==1.0.3
• numpy==1.18.1
• matplotlib==3.1.3
• ipywidgets==7.5.1
• eli5==0.10.1

To install requirements:

pip install -r requirements.txt

Training

To train a model from the paper, either execute the suitable training Jupyter notebook or convert it to Python script with:

jupyter nbconvert --to script NotebookName.ipynb

and run it with:

python NotebookName.py

The scripts require interactive input.

Evaluation

To evaluate a model from the paper, it is suggested to execute the suitable evaluation Jupyter notebook in order to easily view the resulting graphs apart from the metrics. The evaluation notebooks load the pre-trained models in their respective folder by default, so if you want to load your own trained models you will have to edit the location in the suitable cell in the notebooks.

Pre-trained Models

The pre-trained models with the parameters described in the paper are in their respective Weights folders. Each folder contains 10 .h5 weight files, which correspond to 10 trained instances of the same model with different seeds.

Results

Our model is evaluated in the task of multi-step forecasting on the following datasets:

• Air Quality variables
• Water quality variables

Below we can see the performance metrics of our method and the baselines:

Air Quality Dataset
		Baseline	Proposed Model
CO	RMSE	1.56 +- 0.10	1.32 +- 0.04
	MAE	1.1 +- 0.07	0.93 +-0.03
Benzene	RMSE	7.59 +- 0.31	7.02 +- 0.52
	MAE	5.69 +- 0.27	4.85 +- 0.25
NOX	RMSE	244.43 +- 22.63	214.49 +- 10.31
	MAE	180.62 +- 15.56	152.25 +- 7.46
NO2	RMSE	52.35 +- 13.29	48.57 +- 1.53
	MAE	41.83 +- 12.66	36.94 +- 1.18
Training time (sec)		477 +- 7	524 +- 11
Water Quality Dataset
Temperature	RMSE	0.59 +- 0.07	0.50 +- 0.02
	MAE	0.39 +- 0.04	0.33 +- 0.02
Training time (sec)		4554 +- 143	5310 +- 262

Training and evaluation random seeds

Air Quality Baseline		Air Quality Proposed Model		Water Quality Baseline		Water Quality Proposed Model
Seed	Experiment Id	Seed	Experiment Id	Seed	Experiment Id	Seed	Experiment Id
2650481	3af068ef-70de-4a71-86e2-e3a07b2a8b20	2650491	36c93c9e-3105-47f5-9e85-5beb23e94e68	2639467	5aa4c32f-3a25-49b8-997b-1bdb4af37939	2639478	b4b94f1c-6d65-459c-adb8-389cbd0e97c5
2650482	60a8f6f3-f528-4e00-94ad-b4a963457319	2650492	25d6d2d6-2265-4431-88a2-bc49831eb8c6	2639468	80c4e454-726a-405f-9e83-f524f20939a9	2639479	3a4e66b9-5244-4bc0-9076-a3eca47d5fe4
2650483	4ed1d1c2-476d-4741-a980-55e830c8c79b	2650493	492bc644-16a5-41b9-b226-20c2dbefce13	2639469	d1ef568b-174d-4f9e-97b6-6813c4270a5c	2639480	78337ffe-d26e-4e7b-9497-0bd0cb9c2974
2650484	01e02876-b021-4215-9111-4122b3519778	2650494	6cd2a5a8-35f7-431a-a671-7c54aaf36ac3	2639470	317ff62b-8bcd-40c1-811d-b5ba67d2ab7c	2639481	c6c30ea5-33ab-4900-bc0e-d2fd98c1bcab
2650485	8a62a6c5-8570-474a-aa4a-a758573778bb	2650495	a54c34b0-69ee-43d2-b900-61e0fb0d8228	2639471	83f3dcf9-491a-4c20-8c39-064c65163d35	2639482	db7b10b4-4b52-4a26-97ca-56186eb924a1
2650486	6e1973ff-8fcc-47ce-a28e-41c7daa661d2	2650496	f43a22dd-ac6f-4db9-a5b4-17d4a7ef9c5b	2639472	cef6c672-f8eb-4916-9221-82051633b99c	2639483	f08332bc-d654-4219-a7c1-e0e6854fb2b5
2650487	3a6c674e-abcf-4879-bd8c-3f893769c9f0	2650497	f62368ea-9cc9-4fd6-8f7a-a28b6bf3b4b4	2639473	7e31a1d4-5fed-4924-939e-b2cde7fdf96b	2639484	7fba0620-13b4-4306-8a2a-6c82b025a8fe
2650488	a5697c28-0c05-466b-a17a-5de349cc156a	2650498	7c4c024c-f347-4027-9c19-1264357ec174	2639474	350cbc49-a71e-4336-839b-2be9ed889eca	2639485	67f95b23-e4a5-4b5b-9512-6c65e3918545
2650489	35801255-51d1-42a3-86fe-b3efd7094b58	2650499	393c6468-7a41-455f-8fdd-70cd674b3b11	2639475	6847ae60-80d8-4580-bb3e-10ee1e9ccaf3	2639486	b82aad09-1631-4540-9058-3c6eff69511e
2650490	db1722ce-f7ab-499b-b865-6165fe73cc3b	2650500	510e465d-c041-4fb3-b76c-f514fde218ae	2639476	b6108ad8-4a30-49d8-aaad-4d8d1c40ad37	2639487	bcb62682-39db-483e-aa24-b9afa625d99e

The random seed that was used for all evaluation experiments is 45112.

Visualization examples

Temperature attention distribution pH attention distribution Dissolved oxygen attention distribution