Prediction of relative humidity using soft computing techniques

Accurate and reliable prediction of relative humidity is of great importance in all fields concerning global climate change. Aim of the present study was to evaluate the performance of soft computing techniques viz. ANN and LSTM for daily morning and evening relative humidity forecasting. Both models were trained using eight meteorological parameters and meteorological with static time series as days of the year. The results indicate that the LSTM models tended to underestimate peak values and overestimate lower values for both morning and evening relative humidity. However, the use of static time series improved the prediction of lower values. On the other hand, the ANN models performed well and closely predicted the observed values for both scenarios. The performance error statistics showed that the LSTM models had poor performance with negative NSE values (-0.33 to 0 and 36 to 61), lower KGE values (0.42 to 0.51 and 0.68 to 0.73), and negative PBias values (-6.17 to -4.41 and -5.11 to -1.95) for models trained, tested and validated using meteorological data and meteorological data with static time series. Moreover, the ANN models exhibited very good performance with NSE > 0.99, KGE > 0.98, -0.52 < PBias < -0.03, and R² > 0.99 for both scenarios. Overall, it can be concluded that the LSTM model showed limitations in accurately predicting both morning and evening relative humidity, while the ANN model demonstrated excellent performance in estimating evening relative humidity.