Evaluation of Statistical and Deep Learning Methods for Short-Term Weather Forecasting in Semi-arid Regions

Numerical methods represent a powerful tool for weather forecasting. However, they still face various limitations related to energy consumption and the time it takes to run simulations. To overcome these weaknesses, various statistical and deep learning models were developed to combine precision, time, and energy efficiency criteria. In this paper, we evaluated models of both classes for the task of short-term weather forecasting (one day, three days, and one-week forecasts), namely, the autoregressive integrated moving average, the theta method, and fast Fourier transform as statistical models, versus a long short-term memory, neural basis expansion analysis for interpretable time series, and temporal convolutional neural network as deep learning architectures. The dataset used in this study is sourced from the automatic meteorological station installed in the Marrakech region (center of Morocco) covering the period from January 3, 2013, to December 31, 2020, on a half-hour scale. These include air temperature (Ta), air relative humidity (Hr), and global solar radiation (Rg). Before feeding the data to our models, we first used the ERA5-Land Reanalysis data to impute missing values found in our time series. Results show that the TCNN model outperforms the others in terms of the coefficient of determination (R2) and the root mean square error (RMSE).