Broad Echo State Network with Reservoir Pruning for Nonstationary Time Series Prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Nonstationary time series prediction is crucial for system control and management.
• Existing neural networks struggle with nonstationary and nonlinear data features.
• Echo State Networks (ESN) and Broad Learning Systems (BLS) offer potential but have limitations.

Purpose of the Study:

• To develop a novel network for accurate nonstationary time series prediction.
• To enhance the efficiency and performance of time series modeling.
• To address the limitations of current neural network approaches in handling complex data.

Main Methods:

• Integration of Echo State Network (ESN) into Broad Learning System (BLS) to create a Broad Echo State Network (BESN).
• Implementation of an incremental learning algorithm to improve training efficiency by eliminating error backpropagation.
• Development of a neuron pruning optimization algorithm to reduce redundant information based on contribution degree.

Main Results:

• The proposed Broad Echo State Network (BESN) demonstrates superior performance in nonstationary time series prediction.
• The network effectively models complex nonstationary and nonlinear data features.
• Comparative tests show improved accuracy over traditional methods like ESN, BLS, and Recurrent Neural Networks.

Conclusions:

• The Broad Echo State Network (BESN) offers a significant advancement in nonstationary time series prediction.
• The dynamic optimization and incremental learning contribute to enhanced model efficiency and accuracy.
• This method provides a robust solution for predicting complex time series in various applications.