MCNR: Multiscale feature-based latent data component extraction linear regression model

Abstract

Time series forecasting is of great significance in various fields and is widely applied in industries such as finance and energy management. However, time series data often contains rich periodic features, and there is a high correlation between these features, with the core trend often being implicit in certain features. Therefore, effectively separating and extracting core information from complex multidimensional data, while avoiding noise interference, has become an urgent problem to be solved. To address this, we propose a multi-scale latent feature extraction model, MCNR, for specific periods. It assigns weighted labels to periodic data points through a backward-weighted periodic module, thereby giving more weight to the periodic points and allowing the model to focus on key periodic features. Another core innovation of the MCNR model is the division of the retrospective window into different scales to capture long-term, mid-term, and short-term time features. For data of different scales, the model uses the Regularized Latent Component Regression (RLC) module for latent component extraction and regularization. By focusing on the correlation between each dimension and the predicted value, it uses principal component analysis to extract linear combinations of multivariate features, thus effectively separating the core regions of the data. This process significantly improves the model’s adaptability to different time series structures. Additionally, MCNR introduces a Multilevel Data Normalization (MDN) module. Through the reversibility of MDN, the model can adapt to the distribution differences of the data, normalizing the features based on the mean and standard deviation of the data, thereby removing the trend and seasonal components from the data and further enhancing the model’s stability, robustness, and generalization ability. Compared to the latest mainstream models, MCNR achieves the same experimental results with only approximately 10k parameters, and in experiments on multiple datasets, the model improved the Mean Squared Error (MSE) by 6.58 %.