Sugarcane Yield and Quality Forecasting Models : Adaptive ES vs. Deep Learning

This paper presents three forecasting models (based on a backpropagation neural network (BPNN), (mu+lambda) adaptive evolution strategies (A-ES) [2], and a deep neural network (DNN)) for predicting sugarcane quality levels (called commercial cane sugar, CCS) and yield. The performance analysis of the three models is also discussed. Sugarcane is an important economic plant in many countries as the sugar industry is also related to many other manufacturing sectors. However, the sugarcane yield and quality levels have often been volatile leading to poor resource management and economic loss. Our forecasting models would further be developed to help sugar mills avoid such situations. We collected the data from sugarcane farmers residing in Thailand's 24 provinces (during 2010-2014). Comparatively analyzing the accuracies of forecasting the sugarcane CCS and yield obtained from the three models, we found that 1) the DNN-based model is promising in some cases where its errors are less than the other models, 2) A-ES- and DNN- based models have comparable predicting performance on average, and 3) the DNN-based model's prediction accuracy is sensitive to its initial values and the network structure (i.e, the train and testing error ranges are 10.0-11.6 and 12.25-13.76, respectively, while varying network structures and random seeds).