Weather Forecasting Using Recurrent Neural Networks and Vector Autoregressive Model: A Comprehensive Analysis of Time Series Data from Rohingya Camps and Control Areas

The study aimed to generate weather forecasts for the Rohingya camps in Ukhia and Teknaf by employing a Recurrent Neural Network (RNN) with a Long Short-Term Memory (LSTM) layer alongside a Vector Autoregressive (VAR) model. It utilized six meteorological variables-temperature, humidity, precipitation, surface pressure, wind speed, and wind direction-sourced from NASA’s POWER project. The analysis focused on two sets of areas: control regions in Ukhia and Teknaf that are not impacted by the Rohingya presence, and target regions in the same locations where the Rohingya reside. The findings revealed notable differences in prediction accuracy between the models.The RNN with LSTM demonstrated superior accuracy in control areas, likely due to its ability to capture complex, nonlinear patterns. In contrast, the VAR model outperformed the RNN with LSTM in target areas, suggesting that weather patterns in these regions exhibit more linear relationships. These findings underscore the importance of selecting forecasting models based on the underlying structure of meteorological data to improve prediction accuracy. Dhaka Univ. J. Sci. 73(2): 93-99, 2025 (July)