Water Quality Prediction Method Based on Reinforcement Learning Graph Neural Network

Abstract

The importance of water quality prediction for management and pollution control has gained significant recognition in recent years. However, existing methods face two main challenges: the interaction between water quality variables and the environment is often overlooked, and even when considered, it is not effectively utilized. To address these issues, we propose a reinforcement learning graph neural network-based approach. Our method, an adjacency reinforcement learning, and multi-channel graph convolution autoencoder, predicts water quality by performing reinforcement learning on the adjacency of water quality indicator images. The obtained adjacencies inform the design of a triple-channel adjacency-attention graph convolution network. Ultimately, water quality prediction is achieved through a deep autoencoder clustering method and an auto-regression model. We evaluate this method on a significant dataset we collected, achieving strong experimental results. Additionally, ablation experiments, robustness analysis, and water quality indicator complexity experiments are conducted to validate the method’s effectiveness.

Keywords

• Water quality prediction
• graph neural network
• reinforcement learning
• autoencoder
• deep clustering