Abstract
Three artificial intelligence approaches - K-nearest neighbor (KNN), artificial neural network (ANN), and extreme learning machine (ELM) - are used for the seasonal forecasting of summer monsoon (June-September) and post-monsoon (October-December) rainfall from 2011 to 2016 for the Kerala state of India and performance of these techniques are evaluated against observations. All the aforesaid techniques have performed reasonably well and in comparison, ELM technique has shown better performance with minimal mean absolute percentage error scores for summer monsoon (3.075) and post-monsoon (3.149) respectively than KNN and ANN techniques. The prediction accuracy is highly influenced by the number of hidden nodes in the hidden layer and more accurate results are provided by the ELM architecture (8-15-1). This study reveals that the proposed artificial intelligence approaches have the potential of predicting both summer monsoon and post-monsoon of the Kerala state of India with minimal prediction error scores.
1. Introduction
Accurate prediction of rainfall is highly desirable for states like Kerala where economy of the state and livelihood of people are highly sensitive to rainfall. Kerala receives approximately 2.5 times higher annual mean rainfall than the average of all India rainfall, nevertheless the state needs to resolve water scarcity issues in the upcoming years as the majority of the rainwater flows into the Arabian Sea within 48 to 72 hours of rainfall [1]. Summer monsoon and post-monsoon are the two rainfall seasons occur in the state. Summer monsoon occurs from June to September (JJAS) and is the primary rainy season of the state. Owing to wind reversal, the state has also received rainfall during the post-monsoon period which occurs from October to December (OND) [2].
4. Conclusion
Performances of three artificial intelligence techniques such as K-nearest neighbor (KNN), artificial neural network (ANN), and extreme learning machine (ELM) were evaluated for the prediction of summer monsoon (JJAS) and post-monsoon (OND) rainfall for the Kerala state of India. The performance of the aforementioned approaches has been gauzed by different statistical tests. ELM approach is found to be more accurate than KNN and ANN approaches for the prediction task. There is substantial impact of hidden nodes on the prediction accuracy. ELM structure (8-15-1) provides more accurate results for both JJAS and OND seasons than KNN and ANN techniques in the currently used rainfall dataset. In future, we will further apply different machine learning algorithms for prediction of the chaotic monsoon of southern India. The impacts of climate variability and extreme weather events will also be studied by taking different environmental parameters into consideration.
