Carbon Emission Prediction System for Comprehensive Utilization of Straw Based on Multi-Source Data

Abstract

In order to overcome the drawbacks of the conventional methods of carbon emission prediction, which include low precision and the inability to adapt effectively since they rely on single sources of data and are unable to generalize predictions to different items, this paper develops a prediction system that yields on multi-source data. To identify the spatiotemporally aligned feature engineering framework, a composite of heterogeneous data, which are remote sensing, meteorology, and statistically based, is established. Then a hybrid prediction model that integrates mechanistic constraints and adaptive graph neural networks is suggested, which directs the learning of the data through the laws of physics and describes the spatial relationship of the carbon emissions dynamically. Lastly, a prediction engine that is based on multi-scenario simulation is created. It has been demonstrated that the system minimizes RMSE by 22.5 percent compared to the XGBoost (eXtreme Gradient Boosting) model, increases R² to 0.923, and the generalization stability in untrained regions improves a lot. This system can efficiently address the issue of precise prediction and cross-region adaption, which presents a stable analytical instrument of managing the carbon emission of agriculture.