Attention-Based Gated Recurrent Graph Convolutional Network for Short-Term Traffic Flow Forecasting

Table 3

The prediction performance of different model on the HW-ENG dataset.

Methods	15 min			45 min			1 h			Training time (s/epoch)
	MAE	RMSE	MAPE (%)	MAE	RMSE	MAPE (%)	MAE	RMSE	MAPE (%)
HA	76.94	115.62	39.80	95.03	142.18	49.9	104.96	156.04	56.10	—
SVR	57.06	102.21	27.70	78.38	133.03	38.4	89.22	149.64	43.40	—
ARIMA	35.5	57.34	17.70	70.57	115.26	31.6	88.73	142.65	40.00	—
MLP	30.29	48.51	15.10	49.62	77.80	27.3	59.01	90.58	34.00	4
GRU	29.7	47.46	15.20	43.08	68.86	21.0	50.79	78.75	28.60	12
T-GCN	30.47	48.42	16.00	49.46	77.54	26.6	54.43	83.71	29.70	33
HGCN	29.23	46.97	15.30	46.56	72.93	23.8	53.43	84.21	25.60	11
ASTGCN	26.98	43.25	15.70	45.70	73.08	21.8	39.57	60.03	22.80	120
AGCRN	28.72	45.23	17.80	35.64	57.57	19.5	35.92	58.81	19.40	105
AGRGCN	26.06	42.23	14.80	33.12	54.27	17.5	31.16	52.2	16.30	12

The bold values indicate the best performance metrics in the performance comparision.