Construct a deep meta-learning model for accurate outdoor thermal comfort prediction and its application

To increase the accuracy of thermal comfort predictions, this study aims to propose an improved thermal comfort analysis model based on deep meta-learning. Multi-source data including meteorological data, urban planning data and crowd feedback data were collected and preprocessed through data cleaning, missing value processing and standardization. A deep meta-learning model combining convolutional neural network (CNN) and long- short-term memory (LSTM) network was developed for feature extraction, based on which a meta-learning module was constructed to realize the rapid migration and adaptation of the model. Strategies of supervised learning and transfer learning were adopted to train the model, after which the model was optimized by using adaptive moment estimation optimizer, and the model performance was evaluated by cross-validation and test set. The results indicate that in both training and test sets, the enhanced deep meta-learning model outperformed the conventional model. The average accuracy rate of the model in thermal comfort prediction reached 92.3%, the precision rate was 89.5%, the recall rate was 90.7%, and the F1 value was 90.1%. All these results were better than the traditional multiple linear regression model (accuracy rate: 85.6%), support vector machine (precision rate: 82.3%), random forest (recall rate: 84.9%) and traditional neural network (F1 value: 85.4%). The model was then applied to a smart city scheme in China, and the thermal comfort score was improved by 22.47%, and the household satisfaction increased by 18.65%, showing good adaptability and application potential.