Harnessing Emerging Knowledge and Extensibility Techniques for Stationarity and Normalization in Multi-Agent Time Series Forecasting
Dinesh Asanka, Chathura Rajapakse, Masakazu Takahashi · Engineering Technology & Applied Science Research · 2025
This study introduces a novel multi-agent approach designed to optimize time series forecasting through an efficient normalization framework. The proposed architecture leverages the knowledge-emerging and extensibility properties, enabling an adaptive and scalable forecasting performance. The system is evaluated using the Hiroshima human mobility dataset. It consists of two key components: (1) a stationarity verification module employing the Augmented Dickey-Fuller (ADF) test and (2) a dynamic normalization module that selects the optimal technique based on ADF statistics and p-values. Five normalization methods—MaxAbs, MinMax, Log, Z-Score, and Sigmoid are analyzed to determine the most effective approach for different time series characteristics. Additionally, knowledge-emerging techniques, specifically the J48 decision tree algorithm, are integrated to enhance the system's predictive efficiency. Experimental results demonstrate the effectiveness of the proposed multi-agent architecture in improving the accuracy and adaptability of time series forecasting.