A new deep learning architecture with inductive bias balance for transformer oil temperature forecasting

Journal of Big Data

Table 2 MAE metric for the best models for every window size and dataset tested

Dataset	Window	ARIMA	Informer^a	CNN^a	HLNet	NBEATS	SCINet	SRCNet
ETTh1	24	0.290	0.247	0.172 (±0.030)	0.197 (±0.014)	0.201 ( ±0.012)	\(\underline{0.186 (\pm 0.006)}\)	0.138 (±0.010)
	48			\(\underline{0.167(\pm 0.015)}\)	0.193 (±0.019)	0.210 (±0.043)	0.166 (±0.021)	0.137 (±0.006)
	72			\(\underline{0.139 (\pm 0.013)}\)	0.202 (±0.007)	0.152 (±0.009)	0.149 (±0.013)	0.131 (±0.003)
	96			\(\underline{0.144 (\pm 0.007)}\)	0.205 (±0.014)	0.149 (±0.013)	0.146 (±0.010)	*0.129 (±0.001)
	120			0.156 (±0.012)	0.192 (±0.007)	0.15 (±0.018)	*0.138 (±0.010)	0.131 (±0.004)
ETTh2	24	0.433	0.240	0.205 (±0.008)	0.333 (±0.016)	0.197 (±0.003)	\(\underline{0.198 (\pm 0.006)}\)	\(\underline{0.198 (\pm 0.006)}\)
	48			\(\underline{0.198 (\pm 0.002)}\)	0.339 (±0.011)	0.201 (±0.003)	0.201 (±0.009)	0.195 (±0.005)
	72			0.195 (±0.004)	0.334 (±0.010)	0.202 (±0.002)	0.196 (±0.009)	0.189 (±0.003)
	96			0.200 (±0.005)	0.338 (±0.003)	0.207 (±0.003)	\(\underline{0.197 (\pm 0.004)}\)	*0.193 (±0.002)
	120			0.201 (±0.005)	0.339 (±0.004)	0.206 (±0.002)	\(\underline{0.198 (\pm 0.004)}\)	0.194 (±0.005)

^aThese results have been obtained from the Liu et al. work. Note that the concept of input window does not exists in ARIMA, which explains why there are repeated values in each window. In addition, randomness does not affect ARIMA which justifies there is no standard deviation