RAINER: A Robust Ensemble Learning Grid Search-Tuned Framework for Rainfall Patterns Prediction

Zhenqi Li, Junhao Zhong, Hewei Wang, Jinfeng Xu, Yijie Li, Jinjiang You, Jiayi Zhang, Runzhi Wu, Soumyabrata Dev · 2025

Read RAINER: A Robust Ensemble Learning Grid Search-Tuned Framework for Rainfall Patterns Prediction by Zhenqi Li, Junhao Zhong, Hewei Wang, Jinfeng Xu, Yijie Li, Jinjiang You, Jiayi Zhang, Runzhi Wu, Soumyabrata Dev (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-16900.txt.

Introduction

Zhenqi Li leads a team of nine on this paper from two thousand twenty-five. It presents RAINER, a frame for rainfall prediction that pairs ensemble learning with a full grid search for the best settings.

Rain data is tangled and nonlinear, and past work often picks model settings by hand or by hunch. RAINER first cleans the data, fills gaps, and builds fresh weather traits, such as the gap between temperature and humidity. It then shrinks the trait set by principal components. A wide field of models is tried, from weak learners to Kolmogorov-Arnold Networks, and merged by vote. On real data sets the frame posts strong scores.

Catalog record

CREATORZhenqi Li, Junhao Zhong, Hewei Wang, Jinfeng Xu, Yijie Li, Jinjiang You, Jiayi Zhang, Runzhi Wu, Soumyabrata Dev
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-16900

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

RAINER: A Robust Ensemble Learning Grid Search-Tuned Framework for Rainfall Patterns Prediction Zhenqi Li, Junhao Zhong, Hewei Wang, Jinfeng Xu, Yijie Li, Jinjiang You, Jiayi Zhang, Runzhi Wu, Soumyabrata Dev Abstract Rainfall prediction remains a persistent challenge due to the highly nonlinear and complex nature of meteorological data. Existing approaches lack systematic utilization of grid search for optimal hyperparameter tuning, relying instead on heuristic or manual selection, frequently resulting in sub-optimal results. Additionally, these methods rarely incorporate newly constructed meteorological features such as differences between temperature and humidity, to capture critical weather dynamics. Furthermore, there is a lack of systematic evaluation of ensemble learning techniques and limited exploration of diverse advanced models introduced in the past one or two years. To address these limitations, we propose a R obust ensemble Le A rning gr I d search-tu N ed fram E wo R k (RAINER) for rainfall prediction. RAINER incorporates a comprehensive feature engineering pipeline, including outlier removal, imputation of missing values, feature reconstruction, and dimensionality reduction via Principal Component Analysis (PCA). The framework integrates novel meteorological features to capture dynamic weather patterns and systematically evaluates non-learning mathematical-based methods and a variety of machine learning models, from weak classifiers (e.g., KNN, LASSO, and Random Forest) to advanced neural networks such as Kolmogorov-Arnold Networks (KAN). By leveraging grid search for hyperparameter tuning and ensemble voting techniques, RAINER achieves state-of-the-art results within real-world dataset. Extensive experiments demonstrate the framework’s superior performance across diverse metrics, such as Accuracy, Precision, Recall, F1-score, and AUC, highlighting the pivotal role of feature engineering, grid search, and ensemble learning in rainfall prediction tasks. 1 Introduction Geoscience has profound implications for tackling global challenges, spanning meteorology, climate resilience, disaster management, and sustainable urban planning. In recent years, data mining, machine learning, and deep learning have emerged as transformative tools in predicting patterns, enabling the extraction of valuable patterns from complex datasets and fostering advancements in various forecasting domains including healthcare prediction [ 1 , 2 , 3 , 4 ] , where models assist in disease diagnosis and patient outcome prediction; entertainment trend prediction [ 5 , 6 , 7 , 8 , 9 ] , focusing on content popularity and audience preference analysis, enabling tailored recommendation systems; and geoscience forecasting [ 10 , 11 , 12 , 13 ] , which addresses critical environmental and geological challenges. Among the various geoscience prediction domains, multiple specific tasks in geoscience and remote sensing have greatly benefited from machine learning and deep learning-based methods, which have been applied to tasks such as land cover classification and vegetation mapping [ 14 , 15 , 16 , 17 , 18 , 19 ] , cloud segmentation [ 20 , 21 , 22 ] , seismic activity prediction [ 23 , 12 , 24 ] , soil moisture prediction [ 25 , 26 , 10 ] , drought assessment [ 27 , 28 , 11 ] , glacier dynamics modeling [ 29 , 30 , 13 ] , air temperature prediction [ 31 , 32 , 33 ] , agriculture monitoring [ 34 ] , and rainfall prediction [ 35 , 36 , 37 , 38 ] . Despite these advancements, each of these domains continues to present unique challenges. Remote sensing often involves the integration of heterogeneous data sources and the need for high-resolution imagery, which can be computationally intensive to process. Cloud segmentation requires robust models capable of handling noise and occlusions in satellite imagery. Rainfall prediction remains particularly challenging due to the inherent variability and nonlinearity of meteorological data, highlighting the importance of effective feature extraction and model generalization. These challenges underscore the need for continued research and innovation in the application of machine learning to geoscience problems. Among these domains, rainfall prediction stands out due to its intricate dependence on multiple meteorological variables such as humidity, temperature, and atmospheric pressure. Feature extraction and dimensionality reduction techniques, including PCA, have been employed to simplify high-dimensional data while preserving key relationships [ 39 , 40 , 41 ] . Additionally, using machine learning classifiers such as Random Forests, Gradient Boosting, Support Vector Machines (SVM), and XGBoost for rainfall prediction tasks has demonstrated superior performance compared to traditional statistical methods. These ML classifiers excel at capturing nonlinear relationships and handling high-dimensional datasets, enabling robust predictions and better generalization [ 42 , 43 , 44 ] . However, existing approaches for rain prediction face several challenges below, and our motivation is to solve these issues. 1. Some relevant work fails to fully leverage grid search for optimal hyperparameter tuning, only relying on heuristic selection. Furthermore, some studies that utilize grid search do not perform preliminary exploration to observe parameter trends, instead relying on arbitrary or heuristic parameter ranges, thereby increasing the likelihood of obtaining sub-optimal output. 2. Many existing models fail to leverage newly constructed features, such as the differences between temperature and humidity, which are critical for capturing specific meteorological dynamics. 3. Ensemble methods, including voting and advanced integration techniques, have been insufficiently explored, particularly for assessing the combined effects of multiple models. 4. Advanced models such as KAN [ 45 ] , have rarely been evaluated in comparison to methods like MLP, Transformer, LSTM, and weaker classifiers in this domain, limiting insights into their effectiveness in rainfall prediction. To ameliorate the aforementioned problems, we propose a hybrid approach that integrates systematic data analysis with advanced modeling techniques to enhance rainfall prediction accuracy. Our approach includes a robust data preprocessing pipeline (e.g., imputing missing values, removing outliers, and dropping highly correlated features), feature construction, correlation analysis, and dimensionality reduction using PCA. Together, these steps form a comprehensive data processing system. Additionally, we conduct extensive experiments to explore parameter tuning and dataset partitioning strategies for each classifier, employing grid search to identify the relatively optimal parameters for a variety of weak and strong classifiers. We also evaluate the impact of ensemble methods such as voting on improving prediction performance. The main contributions of our RAINER are threefold: 1. We develop a robust and systematic data analysis and processing pipeline including feature construction (e.g., temperature/humidity max differences) and PCA on real-world data, which ensures high-quality inputs for predictive modeling. 2. We perform extensive exploration on diverse non-learning mathematical and weak-to-advanced learning-based approaches, investigating optimal parameter settings through grid-search and examining the effects of ensemble voting methods. 3. We conduct comprehensive experiments with detailed performance evaluation across multiple metrics (e.g., Accuracy, Precision, Recall, ROC Curve, AUC, and F1-score), providing quantitative comparisons of various methods and four distinct feature engineering strategies in rainfall prediction tasks, highlighting their respective strengths and limitations. The structure of this paper is as follows. Section 2 reviews related work. Section 3 introduces the dataset and outlines the assumptions made for the experiments. Section 4 explains the data preprocessing and feature engineering process. Section 5 presents the methods, detailing non-learning mathematical approaches and weak/advanced learning-based models. Section 6 conducts experiments with quantitative analysis and ROC comparisons, and Section 7 concludes the paper with discussions on future work. 2 Related Work 2.1 Non-learning Approaches (e.g., mathematical, physical, statistical, and hybrid) for Rainfall Prediction Traditional approaches to rainfall prediction have predominantly relied on mathematical and statistical models. These models aim to describe atmospheric processes

Read the complete text