Awesome Wireless Traffic Forecasting Library

Intoroduction

• This is an open-source library for wireless traffic forecasting (WTF-Lib), especially for deep wireless traffic analysis.
• In this library, we provide a variety of state-of-the-art short- and long-term wireless traffic forecasting methods.
• We provide a neat code base to evaluate advanced deep time series models or develop your model, which given a strong benchmark.

Leaderboard for WTF

Till November 2023, the top three models for short- and long-term WTF tasks are:

Model Ranking	Long-term Forecasting	Short-term Forecasting	Runtime
🥇 1st	PSLinear	PSLinear	DLinear
🥈 2nd	PatchTST	PatchTST	PSLinear
🥉 3rd	GWNet	FEDformer	STID

Compared models of this leaderboard. ☑ means that their codes have already been included in this repo.

• PSLinear - Progressively Supervision based on Label Decomposition: Towards Long-Term Wireless Traffic Forecasting on Large-Scale Graphs. [arXiv 2023] [Code].

• PatchTST - A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. [ICLR 2023] [Code].

• DLinear - Are Transformers Effective for Time Series Forecasting? [AAAI 2023] [Code].

• GWNet - Graph WaveNet for Deep Spatial-Temporal Graph Modeling. [arXiv 2019][Code].

• STID - Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting. [arXiv 2023][Code].

• DyDgcrn - Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution. [[IEEE TKDE 2021]] (https://ieeexplore.ieee.org/document/9625773) [Code].

• Mvstgn - MVSTGN: A Multi-View Spatial-Temporal Graph Network for Cellular Traffic Prediction. [[IEEE TMC 2021]] (https://ieeexplore.ieee.org/document/9625773) [Code].

• Periodformer - Does Long-Term Series Forecasting Need Complex Attention and Extra Long Inputs? [[arXiv 2023]] (https://arxiv.org/abs/2306.05035) [Code].

• FEDformer - FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting. [ICML 2022] [Code].

• Pyraformer - Pyraformer: Low-complexity Pyramidal Attention for Long-range Time Series Modeling and Forecasting. [ICLR 2022] [Code].

• Autoformer - Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. [NeurIPS 2021] [Code].

• Informer - Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. [AAAI 2021] [Code].

• Reformer - Reformer: The Efficient Transformer. [ICLR 2020] [Code].

• STCNet - Deep Transfer Learning for Intelligent Cellular Traffic Prediction Based on Cross-Domain Big Data. [IEEE JSAC 2019] [Code]

• Transformer - Attention is All You Need. [NeurIPS 2017] [Code].

Long-term Wireless Traffic Forecasting (LWTF) Benchmark

Methods	Metric	C2TM					Milano-All				CBSD
		4	5	6	7	8	24	36	48	72	24	36	48	72
PSLinear-MS	MSE	9.057	9.161	9.298	9.431	9.389	0.633	0.775	0.907	1.184	1.641	1.772	1.822	1.955
	MAE	0.171	0.174	0.171	0.317	0.169	0.249	0.269	0.285	0.301	0.625	0.644	0.646	0.660
PSLinear-STL	MSE	9.084	9.183	9.328	9.447	9.400	0.708	0.870	1.062	1.366	1.644	1.759	1.821	1.965
	MAE	0.174	0.172	0.172	0.169	0.167	0.275	0.294	0.315	0.343	0.633	0.649	0.653	0.669
PatchTST	MSE	9.226	9.34	9.509	9.504	9.561	0.662	0.832	0.929	1.222	1.87	2.036	2.113	2.084
	MAE	0.195	0.197	0.201	0.182	0.192	0.254	0.273	0.282	0.302	0.706	0.736	0.735	0.697
STID	MSE	9.213	9.769	9.625	9.925	9.56	0.73	0.84	0.985	1.248	2.066	2.038	1.986	2.416
	MAE	0.375	0.339	0.353	0.509	0.387	0.367	0.382	0.377	0.404	0.744	0.738	0.706	0.797
Periodformer	MSE	11.584	11.663	11.822	11.883	11.818	1.341	1.457	1.583	1.892	5.139	5.491	5.335	5.515
	MAE	1.001	0.98	0.975	0.938	0.923	0.571	0.539	0.542	0.54	1.213	1.251	1.231	1.235
DLinear	MSE	12.702	12.58	12.99	13.457	12.945	0.898	1.031	1.106	1.338	2.414	2.576	2.584	2.7
	MAE	0.516	0.464	0.494	0.456	0.556	0.359	0.37	0.392	0.403	0.781	0.805	0.806	0.815
FEDformer	MSE	11.587	11.631	11.679	11.763	11.544	1.836	1.836	1.961	2.274	6.048	6.368	6.474	6.434
	MAE	0.969	0.943	0.914	0.881	0.825	0.734	0.734	0.762	0.772	1.424	1.469	1.472	1.474
Autoformer	MSE	11.68	11.721	11.828	11.94	11.832	1.796	2.652	2.571	2.617	6.107	6.345	6.089	6.078
	MAE	1.03	1.01	0.994	0.981	0.951	0.733	0.999	0.955	0.898	1.453	1.472	1.432	1.412
Informer	MSE	9.401	9.513	9.634	9.894	9.724	20.268	20.236	20.292	20.394	15.156	15.125	15.196	15.219
	MAE	0.581	0.581	0.582	0.67	0.577	2.137	2.111	2.098	2.079	2.365	2.362	2.37	2.372
Transformer	MSE	9.373	9.474	9.612	9.796	9.712	20.049	20.019	20.113	20.243	14.622	15.097	15.137	15.156
	MAE	0.584	0.583	0.582	0.639	0.58	1.988	1.948	1.941	1.906	2.339	2.351	2.352	2.392
Mvstgn*	MSE	9.087	9.195	9.321	9.433	9.395	1.372	1.443	2.656	2.928	3.001	2.442	3.013	3.442
	MAE	0.352	0.36	0.357	0.329	0.337	0.524	0.49	0.559	0.731	0.813	0.783	0.861	0.907
GWNet*	MSE	9.178	9.288	9.431	9.443	9.42	1.21	1.603	2.151	3.413	1.893	2.06	2.074	2.225
	MAE	0.61	0.617	0.624	0.409	0.412	0.48	0.526	0.564	0.573	0.714	0.76	0.739	0.754
Dgcrn*	MSE	9.086	9.198	9.398	9.527	9.478	-	-	-	-	-	-	-	-
	MAE	0.464	0.458	0.577	0.559	0.539	-	-	-	-	-	-	-	-

Note that the method marked by * connot process the LWTF tasks, which are reimplemented by ourselves.

Dataset

Here is a summary of supported datasets.

1. C2TM

2. Milano

Usage

1. Install Python 3.8. For convenience, execute the following command.

pip install -r requirements.txt

2. Prepare Data. You can obtain the well pre-processed datasets from [Google Drive], [Tsinghua Cloud], or [Baidu Drive], Then place the downloaded data in the folder./dataset.

3. Train and evaluate model. We provide the experiment scripts for all benchmarks under the folder ./scripts/. You can reproduce the experiment results as the following examples:

# C2TM
bash run_c2tm.sh

# Milano-All
bash run_milano_all.sh

# Milano-SMS
bash run_milano_sms.sh

# Milano-Call
bash run_milano_call.sh

# Milano-Internet
bash run_milano_internet.sh

4. Develop your own model.

• Add the model file to the folder ./models. You can follow the ./models/Transformer.py.
• Include the newly added model in the Exp_Basic.model_dict of ./exp/exp_basic.py.
• Create the corresponding scripts under the folder ./scripts.

Visualization

Place figures here.

Citation

If you find this repo useful, please cite our paper.

Come soon!