Integration of Particle Swarm Optimization in Bidirectional Memory Networks for Improved Daily Climate Forecasting Accuracy

Abstract

Global climate change has caused rainfall patterns to become increasingly fluctuating and difficult to predict using conventional weather forecasting methods. Accurate daily rainfall prediction is crucial for hydrometeorological disaster mitigation and agricultural sector planning. Although Deep Learning models such as Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) are capable of handling complex time-series data, the manual determination of hyperparameters often results in suboptimal models and entrapment in local optima. This study proposes the integration of the Particle Swarm Optimization (PSO) algorithm to automatically optimize hyperparameters (number of hidden neurons, dropout rate, and learning rate) in LSTM and BiLSTM architectures. The models were evaluated using a multivariate daily climate observation dataset encompassing temperature, humidity, wind speed, and actual rainfall. Experimental results indicate that PSO-based optimization significantly enhances prediction performance compared to baseline models. The PSO-LSTM approach successfully reduced the Root Mean Square Error (RMSE) to 17.59 mm and Mean Absolute Error (MAE) to 9.07 mm, comparable to the performance of PSO-BiLSTM, which achieved an RMSE of 17.59 mm and an MAE of 9.20 mm. These findings prove that automatic parameter tuning using swarm intelligence algorithms can highly optimize sequential neural network architectures in capturing rainfall pattern volatility, making it highly recommended as a foundation for a more accurate early warning system.