Dual Memory LSTM with Dual Attention Neural Network for Spatiotemporal Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Geophysics

Background:

• Spatiotemporal prediction faces challenges in efficient representation extraction and capturing contextual dependencies.
• Existing methods struggle with long-term spatiotemporal relationships.

Purpose of the Study:

• To propose a novel dual memory LSTM with dual attention neural network (DMANet) for enhanced spatiotemporal prediction.
• To improve the modeling of short-term dynamics and long-term contextual representations.

Main Methods:

• Introduced a dual memory LSTM (DMLSTM) unit utilizing differencing operations and a dual memory transition mechanism.
• Developed a dual attention mechanism to capture long-term spatiotemporal dependencies across temporal and spatial dimensions.
• Integrated dual attention into DMLSTM to create DMANet.
• Utilized B-spline interpolation to enhance the apparent resistivity map (AR Map) dataset.

Main Results:

• The DMANet demonstrated superior performance in spatiotemporal prediction tasks.
• The dual attention mechanism effectively captured long-term spatiotemporal correlations.
• The enhanced AR Map dataset facilitated continuous derivative analysis.

Conclusions:

• The proposed DMANet offers a powerful approach for spatiotemporal prediction, outperforming state-of-the-art methods.
• The DMANet's architecture is effective for modeling complex spatiotemporal dynamics and contextual information.