Assessment of LSTM in sub-hourly urban-microclimate temperature forecasting: The case for specific model selection

Abstract

Near-surface urban temperature can rise by more than 0.5 °C in ten minutes around sunrise, when heat-stress risk builds fastest, yet forecasts for this setting are rarely evaluated by how their accuracy varies through the day. We benchmark a Long Short-Term Memory (LSTM) network against five baselines for one-step-ahead 10-minute prediction at a Local Climate Zone (LCZ) 8 station, using an 844-sample chronological test set. On overall MAE and RMSE the competitive models are indistinguishable; they separate only once errors are split by time of day and by warming or cooling. The LSTM gives the lowest overall RMSE (0.489 °C) and beats persistence in the morning-warming window (06:00–10:00): MAE is 17.4 % lower, over 60 % more than ARIMA's gain and outside the GRU's confidence interval. Permutation importance traces this edge to humidity and dew-point inputs ARIMA lacks. During sunrise warming it lowers MAE by 23 % versus persistence; at the midday peak no model beats persistence. Knowing when deep learning helps matters more than a single average gain, so for this station and season we recommend the LSTM for morning heat-stress alerts, ARIMA for night-time urban-heat-island (UHI) reporting, and persistence otherwise.