Research on Ultra-Short-Term Prediction Model of High Temperature in Summer

Abstract

The high temperature weather in summer is one of the extreme meteorological events that frequently occur under the background of global climate change, and high temperature ultra-short-term prediction is of great significance to high temperature disaster prevention and mitigation. Currently, most short-term temperature forecasts are hourly forecasts. This paper proposes a minute-level temperature forecast, which can capture local fluctuations in temperature. This study uses minute-resolution meteorological data and compares and analyzes the prediction results of TimesFM, GRU, LSTM, and MLP. Based on Bayesian optimization for parameter tuning and XGBoost algorithm for feature selection, this study constructs an ultra-short-term prediction model for summer high temperatures, whose performance is comprehensively validated through multiple metrics including $R^2$, MSE, MAE and ACC. An early stop strategy is added to monitor the loss of the verification set during model training to prevent overfitting during model training. We adopt two strategies to build models: (1) using temperature as input. In the task of predicting the next 60 minutes, the BO-GRU performs the best, with an MSE of 0.4705, and an MAE of 0.4335; In the 120-minute ahead prediction task, the BO-GRU model again shows optimal performance with an $R^2$ of 0.9262, an MSE of 1.0763, and an MAE of 0.6804. (2) When using temperature, ground surface temperature, grass surface temperature, minute precipitation, and hourly cumulative precipitation as input features for the prediction of 60 minutes ahead, the three models XGBoost-BO-GRU, XGBoost-BO-LSTM and XGBoost-BO-MLP exhibit reduced prediction errors at 12:00 on August 19th. The research results clearly indicate that the BO-GRU model is more suitable for minute-level temperature prediction tasks.