Evaluation of an artificial intelligence algorithm for forecasting daily temperature

Abstract

This study aimed to predict daily maximum and minimum temperatures using meteorological data and static time series variable as day of year (DoY) using Artificial Intelligence (AI) techniques. Two models were developed: LSTM (Long Short-Term Memory) (MD) and LSTM (MD&DoY). Both models were trained with eight metrological variables using 37 years of data besides static time-series (DoY) as input variables by using 5 days preceding information. Data were partitioned into three groups as training (60%), testing (20%), and validation (20%), respectively. The results showed that both LSTM-MD and LSTM-MD&DoY could simulate the trend of daily maximum temperature well (MAE<1.20, MSE<2.66, RMSE<1.63, NSE> 0.94 and R2> 0.94), but the latter was able to capture both peak and lower values more accurately. For minimum temperature prediction, the addition of day of year information to meteorological data improved the performance of the LSTM models (MAE<1.49, MSE<4.04, RMSE<2.01, NSE>0.94 and R2> 0.94). Based on the performance error statistics, the LSTM (MD-DoY) model outperformed the LSTM (MD) model for both maximum and minimum temperature predictions. The results suggest that the addition of day of year information could improve the accuracy and precision of maximum and minimum temperature predictions using LSTM models.