Dynamic thermal response modelling for urban development: an approach integrating a LSTM-AM predictive model with a multi-physical field coupling model

During the initial stage of urban development, accurately capturing the dynamic thermal response characteristics of the development area is helpful to reasonably add building structures in the early planning, thereby improving the spatial thermal environment and optimising thermal management efficiency. This study compared the forward multi-step prediction performance of different predictive models, took the real scenario as the evaluation baseline, and proposed an integrated model based on the LSTM-AM prediction algorithm, the environmental temperature model, and the regional thermal environment numerical calculation. Firstly, the regional meteorological data from 2024 to 2025 were utilised for the training, validation, and testing of different models, to predict the average temperature and the maximum air temperature change amplitude on specific future dates. Secondly, the predicted results were converted into boundary conditions through the environmental temperature model, and input into the numerical model, to further obtain the thermal environment distribution pattern of the region on that day. Finally, the simulated temperature values of the selected test points and the actual measured values were compared to conduct validation. The research results show that the performance of the LSTM-AM predictive model is the best, and the R 2 of its test set is 0.9105, the RMSE is 1.1445, and the MAE is 0.7398, while the RMSEs of the simulated values and experimental values of the selected test points in the numerical model are 0.46 °C, 1.99 °C, and 0.98 °C respectively, indicating that both models maintain high accuracy. This study provides a feasible scheme for the dynamic thermal response simulation in the initial stage of future urban development, so as to be able to scientifically optimise the layout of buildings, infrastructure, and greening within this space.