Nonpooling Convolutional Neural Network Forecasting for Seasonal Time Series With Trends

Area of Science:

• Machine Learning
• Artificial Intelligence
• Time Series Analysis

Background:

• Automatic processing of non-preprocessed time series is crucial for machine learning.
• Seasonal time series with trends are common in forecasting applications.
• Convolutional Neural Networks (CNNs) are popular for pattern recognition.

Purpose of the Study:

• To investigate the effectiveness of CNNs for analyzing seasonal time series with trends.
• To compare the performance of fully connected neural networks (FNNs), non-pooling CNNs (NPCNNs), and standard CNNs.
• To identify optimal network architectures and parameters for time series forecasting.

Main Methods:

• Experimental analysis using simulated and real time-series data.
• Implementation and comparison of FNN, NPCNN, and CNN models.
• Evaluation of feature extraction capabilities of convolutional and pooling layers.

Main Results:

• Convolutional layers generally improve performance in extracting seasonal and trend patterns.
• Pooling layers were found to introduce negative effects, hindering performance.
• NPCNNs demonstrated superior performance for seasonal time series with trends compared to FNNs and standard CNNs.

Conclusions:

• Non-pooling CNNs are recommended for processing seasonal time-series data with trends.
• The Adam optimizer and ReLU or linear activation functions are suggested for optimal results.
• This research offers an automated approach for building neural networks tailored to seasonal time series with trends.