The Personalized Thermal Comfort Prediction Using an MH-LSTM Neural Network Method

Table 6

Overview of performance of the experimental model.

Model name	RMSE			MAE			MRE
	S1	S2	S3	S1	S2	S3	S1	S2	S3
FFNN	1.1677	0.9440	0.2604	1.0228	0.8246	0.2631	0.9510	0.7901	0.2317
RNN	0.9032	0.7678	0.2256	0.7347	0.5606	0.1774	0.6518	0.5277	0.1499
Single-head LSTM	0.9648	0.8028	0.2379	0.7670	0.6035	0.1917	0.6883	0.5209	0.1473
Multi-head LSTM	1.0940	0.8642	0.2225	0.7305	0.5552	0.1620	0.6673	0.5452	0.1458

Strategy 1 (S₁): Training with collective dataset, Strategy 2 (S₂): Training with collective dataset and transfer learning with individual dataset, Strategy 3 (S₃): Training with individual dataset. Values indicated in bold highlight the best performance metrics achieved across the different training strategies for the multi-head LSTM model. The significance of these bolded metrics (RMSE: 0.2225, MAE: 0.1620, MRE: 0.1458) under Strategy 3 indicates the highest efficiency and accuracy in prediction when the model is trained with individual datasets. This underscores the effectiveness of using a more tailored approach in dataset training for improving model performance in these specific metrics.