GAP-LSTM: Graph-Based Autocorrelation Preserving Networks for Geo-Distributed Forecasting.

Area of Science:

• Computer Science
• Data Science
• Sensor Networks

Background:

• Forecasting is crucial for decision support in geo-distributed sensor networks.
• Challenges include multivariate data, multiple nodes, and spatio-temporal autocorrelation, limiting current methods.
• Existing methods often fail to address these complexities simultaneously, impacting forecast accuracy.

Purpose of the Study:

• To propose a novel forecasting method, GAP-LSTM, designed for geo-distributed sensor networks.
• To effectively exploit spatio-temporal autocorrelation in multivariate data from multiple nodes.
• To improve the accuracy and modeling capabilities of forecasting in complex sensor network environments.

Main Methods:

• Developed GAP-LSTM, integrating graph convolution, attention-based long short-term memory (LSTM), and 2-D convolution.
• Utilized latent memory states to capture complex spatio-temporal dependencies.
• Employed a combination of techniques to synergistically exploit data characteristics.

Main Results:

• GAP-LSTM demonstrated superior performance compared to state-of-the-art methods on real-world traffic, energy, and pollution datasets.
• An ablation study confirmed the significant contribution of each component of the GAP-LSTM method.
• The method provides interpretable visualizations to aid domain experts.

Conclusions:

• GAP-LSTM offers enhanced forecasting capabilities for geo-distributed sensor networks by effectively handling multivariate spatio-temporal data.
• The proposed method represents a significant advancement in accurate and insightful forecasting for complex sensor network applications.
• The integration of graph convolution and attention-based LSTM provides a powerful framework for future research in this domain.