Multi-Level Knowledge Inheritance Graph Neural Network for Tobacco Pest Time-Series Forecasting

The quality of tobacco raw materials is the decisive factor affecting the quality of cigarette products. The monitoring and forecasting of tobacco pest distribution in the factory is crucial in preventing the invasion of tobacco pests into the raw materials, thereby mitigating potential economic losses. However, traditional time series forecasting methods fail to deal with the multi-level spatial characteristics within the tobacco pest distribution data, therefore leading to poor forecasting performance. In this work, a multi-level knowledge inheritance graph network (MKIGN) is proposed to handle this problem. First, by creating virtual variables, a workshop-level dataset and a district-level dataset were generated from the original point-level dataset. Then, a multi-level graph network structure with an integrated knowledge inheritance strategy is designed, which will be applied to the two newly constructed datasets to capture the multi-level spatial characteristics of the tobacco pest distribution data. The effectiveness and superiority of the proposed method are illustrated on a real tobacco pest distribution dataset.