A Hybrid Approach of LSTM and ARIMA for Enhancing Numerical Weather Prediction Accuracy Using Deep Learning

Conventional Numerical Weather Prediction (NWP) models tend to be biased when predicting maximum and minimum temperatures, making them unreliable. As a measure for guaranteeing the effectiveness of models, exploratory data analysis (EDA) that encompasses univariate, bivariate, and multivariate analysis was conducted. Most critical preprocessing operations like missing value handling, outlier removal, skewness correction, and feature scaling were performed in order to enhance model training. Long Short-Term Memory (LSTM), a recurrent neural network with deep learning capabilities, was selected due to its potential to extract intricate temporal relationships in meteorological data. ARIMA, a traditional statistical time-series model, was employed as the baseline for comparison. While ARIMA had lower MAE and MSE, it was not able to capture non-linear trends in temperature fluctuations, while LSTM was able to reduce forecast bias by learning from past trends. In order to enhance forecasting precision, a hybrid technique utilizing LSTM's non-linear pattern detection and ARIMA's power of modelling linear relationships was adopted. Preprocessing techniques, hyperparameter optimization, and bias correction procedures greatly enhanced the precision and reliability of temperature predictions.