Multi-Task Prediction of Hydrochar Properties from High-Moisture Waste with Machine Learning Methods

Hydrothermal carbonization (HTC) is a promising way to treat sludge, food waste, and manure with hydrochar production due to its low energy intake and high efficiency. Hydrochar with excellent fuel property is expected to be a promising fuel to replace fossil fuels for heat and electricity generation. Meanwhile, it can be used as fertilizer, conditioner or remediation agent in soils as a carbon-rich porous material. The char with stable carbon could provide a mean for carbon sequestration in soils. The traditional way to understand these important characteristics of hydrochar is to conduct HTC experiments and then detect the characteristics individually, which is quite expensive, labor-intensive and time-consuming. Machine learning (ML), as a data-driven approach, can perform prediction tasks after training with HTC dataset and facilitate understanding of the relative importance of input features. In this work, three state-of-the-art ML models including supporting vector machine (SVM), random forest (RF) and deep neural network (DNN) were used to predict the characteristics of hydrochar. The average R² of single and multi-task prediction of the optimized DNN model were 0.87 and 0.91, which indicated that the multi-task prediction performance of DNN was the best compared to other two ML models. This prediction can also help us make some evaluations about the hydrothermal technology and the application of hydrochar before taking action, which is beneficial to the labor, time, energy and resources saving compared with the experimental process.