Multi-scale time series prediction model based on deep learning and its application.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional time series prediction models like LSTM and CNN struggle with complex nonlinear dependencies in data such as traffic flow.
• Existing methods often lack adaptive feature focus due to reliance on manually designed attention mechanisms.

Purpose of the Study:

• To propose and evaluate a novel Multiscale Convolutional Attention Long Short-Term Memory (MSCALSTM) model for enhanced time series prediction.
• To address limitations in capturing complex nonlinearities and adaptively focusing on critical features in time series data.

Main Methods:

• The proposed MSCALSTM model integrates a Multiscale Convolutional Neural Network (MSCNN) for capturing dynamic patterns.
• Incorporates a Multiscale Convolutional Block Attention Module (MSCBAM) for adaptive feature selection.
• Leverages Long Short-Term Memory (LSTM) networks for modeling complex temporal dependencies.

Main Results:

• The MSCALSTM model demonstrated superior performance over state-of-the-art methods on the California Performance Measurement System (PEMS) traffic flow dataset.
• Achieved significant improvements in accuracy and robustness for time series prediction tasks.
• Experiments in the Energy domain confirmed the model's strong generalization capabilities.

Conclusions:

• The MSCALSTM model effectively combines multiscale convolutional networks, attention mechanisms, and LSTM for superior time series forecasting.
• The proposed approach offers a robust and accurate solution for complex time series prediction challenges.
• The model shows promise for various forecasting applications beyond traffic flow, including energy data.