Patel et al. (2025) QuSrO-MEnDRN: data assimilation with quest search optimization enabled multihead error minimum learning approach for rainfall prediction

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-11-11
• Authors: Nileshkumar Patel, Jitendra Bhatia, Rajesh Kumar Gupta, Sudeep Tanwar
• DOI: 10.1007/s00704-025-05817-0

Research Groups

Department of Computer Science and Engineering, Institute of Technology, Nirma University, Ahmedabad, Gujarat, India.

Short Summary

This study introduces QuSrO-MEnDRN, a novel deep learning model integrating quest search optimization and a modified deep spatial transformer U-Net for data assimilation, to enhance rainfall prediction accuracy and mitigate overfitting. The model achieves superior performance with minimal error rates compared to conventional methods.

Objective

• To develop and evaluate the QuSrO-MEnDRN model, which integrates quest search optimization, multihead error minimum learning, and a modified deep spatial transformer U-Net (MDST-UNet) for data assimilation, to achieve accurate and reliable rainfall prediction while mitigating overfitting and local convergence issues.

Study Configuration

• Spatial Scale: Regional scale, focusing on various subdivisions within India, specifically utilizing the "Rainfall in India 1901-2015" dataset, including analysis for the VIDARBHA region.
• Temporal Scale: Monthly, seasonal, and annual rainfall predictions over a multi-decadal period (1901-2015, with analyses extending to 2020 in some figures).

Methodology and Data

• Models used:
  • QuSrO-MEnDRN (Quest Search Optimization Enabled Multihead Error Minimum Learning based Deep Neural Network and Recurrent Neural Network)
  • MDST-UNet (Modified Deep Spatial Transformer-based U-Net) for data assimilation
  • QuSrO algorithm (Quest Search Optimization) for hyperparameter tuning
  • Recurrent Neural Network (RNN) blocks
  • Multihead attention block
  • Deep Neural Network (DNN) components (Flatten layer, Fully Connected layers)
  • KNN-based missing data imputation
  • Kalman filter (integrated within the data assimilation process)
• Data sources:
  • Rainfall in India 1901-2015 dataset (Kaggle: https://www.kaggle.com/datasets/aravindpcoder/rainfall-in-india-1901-2015)
  • Raw meteorological information (temperature, humidity, wind speed, atmospheric pressure) for data acquisition.
  • Rain gauge observations (primary measurement method for rainfall in millimeters).

Main Results

• The QuSrO-MEnDRN model achieved a minimum Root Mean Squared Error (RMSE) of 0.04 millimeters and a Mean Absolute Error (MAE) of 0.03 millimeters for monthly rainfall prediction (at 80% training, 100 epochs).
• It demonstrated superior prediction performance and generalization ability compared to conventional methods (CNN-CNN-LSTM, BSRG, SARIMAX, LSTM, CT-DMD, CSO-MEnDRN, HHO-MEnDRN) across monthly, seasonal, and annual analyses.
• For monthly rainfall prediction (80% training, 100 epochs), the model recorded a Mean Squared Error (MSE) of 0.0019 (mm)^2 and a Mean Absolute Percentage Error (MAPE) of 1345.86.

Contributions

• Introduction of the novel QuSrO algorithm, which combines seeking and hunting characteristics to actively trace optimal solutions, improving convergence speed and minimizing prediction error by avoiding local optima.
• Development of the QuSrO-MEnDRN model, integrating MDST-UNet for data assimilation, the QuSrO algorithm for hyperparameter tuning, multihead attention, and error minimum learning, to mitigate overfitting and local convergence problems, thereby enhancing interpretability and reliability in rainfall prediction.
• Utilization of a Modified Deep Spatial Transformer-based U-Net (MDST-UNet) for data assimilation, which leverages contracting and expansive paths along with the QuSrO algorithm to refine the intensity and spatial distribution of the model, leading to more accurate rainfall forecasts.

Funding

This research did not receive any specific funding.

Citation

@article{Patel2025QuSrOMEnDRN,
  author = {Patel, Nileshkumar and Bhatia, Jitendra and Gupta, Rajesh Kumar and Tanwar, Sudeep},
  title = {QuSrO-MEnDRN: data assimilation with quest search optimization enabled multihead error minimum learning approach for rainfall prediction},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05817-0},
  url = {https://doi.org/10.1007/s00704-025-05817-0}
}