Models and Technologies for Applying Neuro-Symbolic Intelligence to Multifactor Forecasting of Feed Wheat Yield

Abstract

The paper examines ways to improve the quality of multifactor forecasts of feed wheat yield through the combined use of statistical initial data as well as fuzzy-possibilistic and neural network approaches and models. The problem of multifactor forecasting of temporal changes in feed wheat yield is solved by constructing a multifactor model that describes the dependence of yield on forecasted values of parameters characterizing the state of natural and climatic factors as well as on the expected agrotechnological measures. The use of expert knowledge made it possible to perform preliminary processing of the initial data by removing from consideration factors with weak influence and factors that are costly to monitor, which increased the quality of the yield forecast in the example presented in the article by approximately two times. One-dimensional convolutional neural network models, recurrent neural network models, and a hybrid neural network model based on the sequential connection of blocks of three main types of neural network layers (convolutional, recurrent, and fully connected) were also developed, studied, and tested. The conducted comparative analysis showed that the proposed hybrid model ARIMA + hybrid neural network has a significant advantage in terms of error value and coefficient of determination compared with the other models under study. Approaches to constructing a hybrid architecture that combines fuzzy-possibilistic modeling and neural network methods based on principles close to ontology-oriented neuro-symbolic intelligence are also considered.