Predicting a Time-Dependent Quantity Using Recursive Generative Query Network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Time series prediction is crucial for many applications.
• Existing models may not fully leverage auxiliary information.
• Graph Quantum Networks (GQN) offer a framework for sequence modeling.

Purpose of the Study:

• To propose a novel neural architecture, the Recurrent Graph Quantum Network (RGQN), for efficient time series prediction.
• To adapt GQN concepts for sequence data with associated meta-information.
• To demonstrate the practical utility of RGQN in real-world scenarios.

Main Methods:

• Developed RGQN, integrating meta-information with dynamic components using Recurrent Neural Networks (RNNs).
• Utilized other related time series as predictors (independent variables).
• Inference phase uses meta-information as a query for expected properties.

Main Results:

• RGQN enables efficient time series prediction by incorporating meta-information.
• Demonstrated application in predicting pressure signals in an industrial pipeline.
• Analysis of prediction residuals under non-routine conditions (e.g., leakage) for anomaly detection.

Conclusions:

• RGQN is a powerful tool for time series forecasting, especially when sequence data has associated properties.
• The model's residual analysis provides a foundation for data-driven Leak Detection and Location frameworks.
• The RGQN approach is versatile and applicable to various sequence-based problems with meta-descriptors.