ROBUST WEATHER FORECASTING USING ISLANDPARALLEL HARRIS HAWKS OPTIMIZER WITHCONVOLUTIONAL LONG SHORT TERM MEMORY

Abstract

Convolutional Long Short-Term Memory (CLSTM) networks
and the Island Parallel Harris Hawks Optimizer (IP-HHO) are integrated in this study to improve the precision and effectiveness of weather forecasting. Despite their intricacy, standard Numerical Weather Prediction (NWP)  models sometimes fail because of their high processing requirements and sensitivity to beginning circumstances. The IP-HHO, a nature-inspired
optimization algorithm, optimizes the hyperparameters of the CLSTM network, significantly improving predictive performance. With a Mean Absolute Error (MAE) of 1.40°C, Root Mean Squared Error (RMSE) of 1.75°C, and Mean Absolute Percentage Error (MAPE) of 4.7%, the IPHHO-CLSTM approach outperformed other models using a dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF). These results mark a substantial improvement over traditional LSTM models. The error distribution centred around zero indicates minimal bias and
high consistency in forecasting. The model’s performance varied across weather conditions, excelling in clear conditions while highlighting challenges in extreme weather scenarios. This research has significant implications for agriculture, disaster management, energy, and transportation sectors, offering a robust tool for better decision-making and operational
efficiency. The study also contributes to computational meteorology by validating the synergy between advanced optimization algorithms and Deep Learning (DL) techniques, suggesting pathways for future research to further enhance weather forecasting models.