Predicting sun spots using a layered perceptron neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Neural network performance is highly dependent on network structure.
• Optimal structure selection for neural networks remains a significant challenge.
• Determining the appropriate number of hidden units is crucial for accurate time series forecasting.

Purpose of the Study:

• To apply Principal Component Analysis (PCA) for determining the optimal structure of multilayered neural networks.
• To specifically identify the ideal number of hidden units for feedforward networks in time series forecasting.
• To enhance the efficiency and accuracy of neural network models through optimized structure selection.

Main Methods:

• Utilizing Principal Component Analysis (PCA) as a method to analyze and determine neural network architecture.
• Implementing a multilayered feedforward neural network for time series forecasting tasks.
• Conducting empirical experiments using sunspot data to validate the proposed PCA-based approach.

Main Results:

• PCA effectively guides the selection of the number of hidden units in neural networks.
• The proposed method demonstrates improved accuracy in time series forecasting compared to traditional approaches.
• Reduced computational time for training and prediction is observed with the optimized network structure.

Conclusions:

• Principal Component Analysis (PCA) offers a robust method for optimizing neural network structures in time series forecasting.
• The study validates the effectiveness of PCA in determining the number of hidden units, leading to more efficient and accurate models.
• This approach provides a valuable tool for neural network design, particularly for complex time series data.