A Multi-Seasonal SS-MSTL-DR Approach With Efficient Training Using Deep Learning and LLaMA: A Case Study of 6-Element Air Quality Prediction in a Subtropical City

Large models and deep learning models for real-time services are trained from time to time to refresh the new features and achieve better accuracy. Multivariate time-series decomposition and recombination (MTS-DR) models effectively improve training efficiency. In this research, an MTS method that is more stable and can handle finer decomposed components helps the training process of deep learning and Large Language Model Meta AI (LLaMA), and is crucial for our subsequent study on the capabilities of adaptive features engineering. Thus, this study introduces an innovative method developed using the Multiple Seasonal-Trend Decomposition using LOESS (MSTL) algorithm, termed “Multiseasonal Scalable Sub-model MSTL decomposition and recombining” (Multiseasonal SS-MSTL-DR) model. This method is designed to upgrade the decomposition and recombining of multiseasonal trends within a scalable sub-model framework. Its architecture is specifically tailored for seamless integration into deep learning systems to enhance the efficiency of model training. The objective is to establish a robust foundation for analyzing time-series data exhibiting distinctive and intricate seasonal patterns. Research methods include conducting a State-of-the-Art (SOTA) literature review, data preparation, model building, feasibility tests, comprehensive evaluation, and summary. Through a case of subtropical urban air quality prediction, it is found that the "Multiseasonal SS-MSTL-DR" model can further reduce the complexity of time-series and highlight its inherent characteristics. The multiple seasonality design and the SS training make the characteristics easier to mine in MTS models. Those experimental results of the cold/warm seasons show that the learning speed and accuracy of deep learning and LLaMA models have been improved, especially in the MTS deep learning model. Since we have extended Multiseasonal SS-MSTL-DR to six different air pollutant concentration data sets, the six of them have completely different spatial characteristics of physical and chemical, which is equivalent to proving the commonality of application, i.e. the proposed method is suitable for other multiseasonal time-series data.