High-Resolution Short-Range Weather Forecasting Using Data Driven BDCL-Net Forecasting Model

Abstract

Recent developments in computational techniques have significantly increased interest in probabilistic weather forecasting. Despite these advancements, accurately predicting weather remains a complex task because atmospheric processes are highly nonlinear and involve numerous interacting variables. Even minor variations in the initial atmospheric conditions can result in considerable differences in forecast outcomes, making reliable prediction a challenging problem. Conventional forecasting approaches, including persistence methods, climatology-based models, linear regression, Markov models, and Auto-Regressive Integrated Moving Average (ARIMA), have demonstrated limited effectiveness in capturing the complex temporal relationships present in weather data. Artificial intelligence (AI)-based methods have surfaced as a viable substitute for weather prediction in order to overcome these constraints. These approaches are capable of learning complex nonlinear patterns directly from historical observations while providing increased computational effectiveness and forecasting accuracy. Motivated by these advantages, the present research proposes a deep learning-based forecasting framework trained on real-world meteorological datasets. The proposed Bidirectional Drop Connect-Regularized LSTM Network (BDCL-Net) has been developed using meteorological information gathered from Jaipur, Delhi, and Chandigarh. The effectiveness of the proposed model is assessed through comprehensive experimental evaluation using multiple performance measures, including the Coefficient of Determination (R²), Mean Squared Error (MSE), and Root Mean Square Error (RMSE), to determine its forecasting accuracy and overall predictive capability.