2D Convolutional Neural Markov Models for Spatiotemporal Sequence Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep neural networks, particularly recurrent neural networks, are used for complex spatiotemporal forecasting.
• Real-world spatiotemporal data is often noisy and chaotic, necessitating probabilistic models for robustness.
• Existing deep Markov models (DMMs) struggle to maintain spatial characteristics, often converting data to 1D.

Purpose of the Study:

• To propose a novel DMM that preserves spatial characteristics in spatiotemporal forecasting.
• To enhance the robustness of time series forecasting models against noise and large data variance.
• To improve the accuracy of forecasting over longer periods.

Main Methods:

• Developed a DMM incorporating 2D convolutional neural networks to retain spatial information.
• Utilized synthetic data with high variance to test the model's robustness.
• Compared the proposed model against naive forecasting, vanilla DMM, and convolutional LSTM.

Main Results:

• The proposed 2D convolutional DMM demonstrated superior robustness to noisy data compared to baseline methods.
• The model outperformed deep neural network models, including convolutional LSTM, in longer forecast periods.
• The study identified limitations in forecasting real-world precipitation data.

Conclusions:

• The novel 2D convolutional DMM effectively models spatiotemporal sequences while preserving spatial information and handling noise.
• This approach offers a more robust alternative to existing methods for complex time series forecasting.
• Future work should address limitations with real-world data and explore further research potentials.