Long Sequence Multivariate Time-Series Forecasting for Industrial Processes Using SASGNN

Abstract

In process industries, the processes are usually very long, which results in long residence of the material in the process. In addition, in many processes, the quality of the final product cannot be detected online, making it challenging to achieve timely and effective control. Therefore, in many processes, middle-state variables are observed and used for control. In many cases, these middle-state variables cannot be measured online. Long sequence multivariate time series forecasting (MTSF) reveals the changes in the process in advance and allows us to make timely adjustments in predictive control mode. However, many current MTSF methods do not consider the relationships between variables adequately, making it difficult to achieve satisfactory results for long sequence forecasting. To address this problem, a novel sparse attention spectral graph neural network (SASGNN) is proposed. In SASGNN, a graph structure learning module is first developed to establish relationships between different variables by combining prior knowledge with a sparse attention mechanism. Then, a long sequence forecasting module transforms the constructed graph into the spectral domain to extract the latent temporal pattern of each variable and the correlation feature between different variables. Besides, considering the temporal information loss during modeling, SASGNN uses a variant GRU to extract the long-term dependencies from the historical data. Experimental results show the superior and stable performance for long sequence forecasting using the SASGNN.