Threshold single multiplicative neuron artificial neural networks for non-linear time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Single multiplicative neuron artificial neural networks (SMNN-ANNs) offer advantages in simplicity and parameter efficiency.
• Traditional SMNN-ANNs often assume a single data generation process for time series, limiting their applicability.
• Many real-world time series exhibit multiple data generation processes, necessitating more sophisticated modeling approaches.

Purpose of the Study:

• To propose a novel artificial neural network architecture, the threshold single multiplicative neuron artificial neural network (TSMN-ANN).
• To address the limitation of single data generation process assumption in SMNN-ANNs by incorporating a threshold mechanism.
• To enhance time series forecasting accuracy for data with multiple underlying generation processes.

Main Methods:

• Introduction of the threshold single multiplicative neuron artificial neural network (TSMN-ANN) architecture.
• Development of training algorithms for TSMN-ANN utilizing harmony search (HS) and particle swarm optimization (PSO).
• Evaluation of the proposed TSMN-ANN using diverse time series datasets and comparison with established forecasting methods.

Main Results:

• The TSMN-ANN effectively models time series with two distinct data generation processes.
• Training algorithms based on HS and PSO demonstrate successful optimization of the TSMN-ANN parameters.
• Comparative analysis indicates superior forecasting performance of the proposed TSMN-ANN over existing methods based on various error metrics.

Conclusions:

• The threshold single multiplicative neuron artificial neural network (TSMN-ANN) provides a robust framework for time series forecasting when multiple data generation processes are present.
• The integration of harmony search and particle swarm optimization offers efficient training strategies for the TSMN-ANN.
• The TSMN-ANN represents a significant advancement in artificial neural network applications for complex time series analysis.