Broad-ESN Based on Radical Activation Function for Predicting Time Series With Multiple Variables.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Multidimensional time series (MTS) present unique challenges due to their inherent multidimensionality and multifeature characteristics.
• Selecting an appropriate prediction model is crucial for effectively analyzing and forecasting MTS data.
• Existing prediction models may struggle with gradient vanishing and local optimization issues in complex time series patterns.

Purpose of the Study:

• To propose a novel Broad Echo State Network (Broad-ESN) model for enhanced Multidimensional Time Series (MTS) prediction.
• To introduce a radical activation function to mitigate gradient vanishing and improve handling of complex data.
• To enhance the optimization process for improved prediction accuracy.

Main Methods:

• A radical activation function was developed to address gradient disappearing issues during iterative processes.
• Sliding window techniques were employed for feature extraction from MTS, with reservoir count determined by feature number.
• The Pied Kingfisher Optimizer (PKO) was initialized using Cubic chaotic mapping and optimized with an exponential spiral equation to prevent local optima.

Main Results:

• The proposed Broad-ESN model significantly outperformed existing models in forecasting performance.
• The model achieved high prediction accuracy and demonstrated a low error rate.
• The radical activation function proved effective in handling complex data patterns.

Conclusions:

• The novel Broad-ESN model offers a significant advancement in Multidimensional Time Series (MTS) prediction.
• The integration of a radical activation function and an optimized PKO enhances model robustness and accuracy.
• This approach provides a superior solution for MTS forecasting tasks.