Creating an artificial neural network for predicting the dynamics of retrospective yield series

Abstract

The article considers the specifics of constructing an artificial neural network (ANN) for predicting the dynamics of time series levels (TS) of grain yield in arid conditions on the example of the Volgograd region. In order to justify the choice of the architecture and hyperparameters of ANN, a preliminary statistical analysis of the simulated TS was performed. The autocorrelation functions of the retrospective yield series constructed in this case can be taken into account when choosing ANN hyperparameters for predicting grain yield. A number of ANN architectures based on recurrent layers, including LSTM, were analyzed. The best results of neural network modeling are obtained for cascading groups of layers of a seriesparallel architecture. The proposed neural network technology for predicting TS yield levels using a pre-forecast autocorrelation analysis of retrospective levels reduces the error of short-term forecasting of grain yield in the arid natural and climatic conditions of the Lower Volga region. Taking into account the results of autocorrelation analysis allows you to choose the values of ANN hyperparameters more reasonably. The achieved accuracy of the regression problem was 82..87%, which can be considered acceptable for planning agricultural production for 1-2 years. The ways of improving the accuracy of the neural network solution of the problem of predicting productivity in the arid conditions of the Lower Volga region using retrospective TS of productivity are formulated.