Affiliation:
1. Begum Rokeya University
2. The University of Burdwan
3. King Khalid University
4. Khulna University of Engineering & Technology
5. Bangladesh Atomic Energy Commission
Abstract
Abstract
Rainfall prediction is a fascinating topic, particularly in an urban city experiencing climate change; it is also required for hydrologic system analysis and design. Most real-time rainfall prediction algorithms use conceptual models that simulate the hydrological cycle in a changing climate. However, calibration of “conceptual” or “physically based models” is typically challenging and time-consuming due to the large number of variables and factors. Simpler “artificial neural network (ANN)” predictions may thus seem promising. To this end, this study aimed to evaluate the performance of two of the most commonly used ANN models, namely “Multilayer Perception (MLP)” and “Radial Bias Function (RBF)”, by predicting the rainfall trend patterns in the mega city of Dhaka, Bangladesh. In this perspective, rainfall is considered as a dependent variable and the rest of the parameters are considered as independent variables for predicting the trends of rainfall in this region. In the prediction models, fifteen conditioned atmospheric and meteorological parameters were used, and the multi-collinearity of these parameters was checked by the “Variance Inflation Factor” and “Tolerance” methods. The performance of the ANN models was evaluated by comparing the predicted and residuals and also by using AOC and ROC curves. The importance study from the MLP model revealed that PM10, O3, PM2.5, NOx, and wind speed are the highest causal factors influencing the rainfall changes in Dhaka, Bangladesh. Though both ANN models produced similar robustness in rainfall prediction, results showed that MLP performed well with an AUC value of 0.941, compared to the RBF model with an AUC value of 0.915. Therefore, the application of the MLP model can be suggested as an alternative to predict the pattern of rainfall as well as meteorological and atmospheric variables based on historical recorded datasets.
Publisher
Research Square Platform LLC
Cited by
2 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献