KUALA LUMPUR, Jul 7 – Universiti Teknologi Malaysia (UTM) researchers have received the Best Paper Association for Computing Machinery (ACM) Special Interest Group on Artificial Intelligence (SIGAI) Award at the 39th International Conference on Industrial, Engineering & Other Applications of Applied Intelligent Systems (IEA/AIE 2026), marking another significant achievement for UTM in artificial intelligence (AI), environmental data analytics and disaster risk reduction.

The award-winning paper, entitled Integrating WGAN-GP Imputation with Broad Learning System for Time-series Rainfall Prediction, was authored by Dato’ Mohd Shahar Abdullah, Prof. Ts. Dr. Ali Selamat, Nguyen Quang Do and Ir. Ts. Dr. Mohd Azlan Abu. The research proposes a hybrid AI model known as Generative Adversarial Network-Broad Learning System (GAN-BLS), integrating Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and Broad Learning System (BLS) to improve rainfall prediction when hydrological data contains missing values.

The study addresses a critical challenge in flood forecasting systems, especially in tropical and monsoon-affected regions such as Malaysia. During extreme weather events, hydrological sensors may fail or produce incomplete readings. These missing values can reduce the accuracy of rainfall prediction and weaken the effectiveness of flood early warning systems at the very moment when reliable data is most needed.

Using rainfall data from the Kuantan River Basin in Pahang, Malaysia, the research analysed records from six telemetry stations over 15 years from January 2009 to December 2023. The dataset contained 33,600 daily observations, including 790 missing values, equivalent to 2.35% of the historical rainfall records. This makes the study highly relevant to Malaysia, where monsoon seasons and recurring flood events continue to affect communities, infrastructure and local economies.

Through the proposed model, WGAN-GP is used to generate realistic values for missing rainfall data while preserving spatial and temporal rainfall patterns. The BLS component then enables fast rainfall prediction and rapid incremental learning without requiring complete retraining whenever new data is introduced.

This capability is particularly important for operational flood monitoring systems, where new telemetry data arrives continuously, and prediction models must adapt quickly. By improving both the reliability of data and the speed of model updates, the proposed approach can support more timely flood warnings, better emergency planning and stronger disaster preparedness.

The paper reports that GAN-BLS outperformed conventional imputation methods, machine learning approaches and selected deep learning models. It achieved the best imputation performance with Mean Squared Error (MSE) = 0.1245, Root Mean Squared Error (RMSE) = 0.0353, Mean Absolute Error (MAE) = 0.0261 and Coefficient of Determination (R2) = 0.942.

In terms of computational efficiency, GAN-BLS recorded 0.1 seconds for initial training and 0.8 seconds for incremental update. This was substantially faster than Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN)-LSTM, Temporal Convolutional Network (TCN) and GAN-TCN models evaluated in the study, highlighting the advantage of BLS for real-time and resource-efficient deployment.

The research has strong practical relevance for Malaysia. Flood events remain one of the country’s recurring natural hazards, particularly during the Northeast monsoon season. A system that can recover missing rainfall data and update predictions rapidly can help strengthen early warning capabilities across vulnerable river basins.

Beyond technical performance, the work demonstrates how AI can be used to address societal challenges. It contributes to climate resilience, smart infrastructure, sustainability and AI-driven public safety, while supporting national efforts to improve disaster risk reduction and environmental intelligence.

The paper also outlines future enhancements, including the integration of additional meteorological variables such as temperature, humidity and wind speed. The authors also plan to include attention mechanisms and further ablation analysis to strengthen model accuracy and explainability.

With these improvements, the GAN-BLS framework could be expanded into a more comprehensive flood early warning platform and tested across multiple river basins in Malaysia.