IMPROVING ROBUSTNESS IN AI FLOOD FORECASTING VIA BLOCKCHAIN-INSPIRED CONSENSUS MODELS

Abstract

Accurate streamflow forecasting is critical for effective water resource management and flood mitigation. While ensemble forecasting improves robustness, its potential is often limited by static aggregation techniques that fail to leverage the dynamic reliability of individual models, often dampening the signal from the most accurate forecast. This study addresses this gap by presenting and evaluating a model-agnostic ensemble framework inspired by decentralized consensus mechanisms in blockchain technology, designed to enhance forecast accuracy and robustness. The framework integrates daily predictions from four distinct Long Short Term Memory models using three dynamic aggregation strategies: Quorum-based Median Agreement, Skill-Weighted Voting, and Adaptive Leader Selection. For comparison, we also evaluate established adaptive ensemble methods, namely Online Super Learner and Dynamic Model Averaging. The blockchain-inspired strategies treat each model as an independent node, reaching a collective agreement based on dynamic performance metrics. To further improve operational reliability, an online, no-leakage debiasing module was applied as a post-processing step to correct for systematic forecast errors. Experimental results show that the consensus strategies outperform both the individual models, the traditional ensemble average, and the additional adaptive baselines. After debiasing, the Skill-Weighted Voting approach achieved the highest overall accuracy with a Kling-Gupta Efficiency of 0.965 and a Nash-Sutcliffe Efficiency of 0.933, while the Adaptive Leader Selection strategy proved most robust, attaining the lowest 90th percentile absolute error, thus reducing the magnitude of large forecast errors. These findings demonstrate that combining a blockchain-inspired consensus approach with real-time error correction provides a practical and effective pathway for developing more resilient forecasting models.