An Improved Hybrid_Stacked Deep Neural Network (HDNN) Model for Enhanced Weather Forecasting

Abstract

Change in climatic conditions is considerably one of the most serious topic confronting the globe today. Weather predictions are based on various temporal and spatial scales along with chaotic dynamics with very high dimensionality domination, which becomes a cause for multiple complex problems in the field. The cutting-edge numerical models with high computational cost are not sufficient for several applications and hence it calls for expansion of work by using Artificial Intelligence to deal with such problems. The current work will look into the possibility of forecasting weather characteristics utilizing the Deep Neural Network (DNN) models. The aim is to find out the capability of DNN in predicting weather conditions. The proposed multiple input single output (MISO) regression model is explored by using well established DNN approaches like Long Short-Term Memory (LSTM) along with the well established Bi-directional LSTM(BiLSTM). Historical weather station data of ten years is being used for this research and also for the purpose of model training. It has been pre-processed to obtain accurate data in desired format. For accurate weather forecasting, the proposed model has also tested utilizing various DL settings and controls and then after performance evaluation is done using different regression metrics like Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Square Error (MSE) and compared with the well-established statistical models like ARIMA, CRNN, LSTM and Bi-LSTM, which were altered as per suitability. To determine the accurate weather forecasting model, comparison research of existing models and proposed weather forecasting model is conducted.