Аңдатпа
The paper presents the development of wheat yield forecasting model in North-Kazakhstan region using artificial neural network. In the course of the study the key predictors of yield were identified, including meteorological indicators (the sum of precipitation during the growing season, average daily air temperature), agrochemical characteristics of soil (moisture at a depth of 0,1 m, 0,5 m and 1 m) and phytosanitary conditions. The constructed neural network model based on the back-propagation algorithm with Levenberg–Marquardt optimization demonstrated high forecasting accuracy: on the test subset the root-mean-square error (RMSE) was 3.368, the mean absolute percentage error (MAPE) was 12.02 %, and the coefficient of determination was R² = 0.9831. These values indicate both low absolute and relative prediction error and stable generalization of the model. The developed system can be used by agrarians to estimate expected yields and optimize resource consumption, as well as integrated into agrometeorological services. Prospects for further development include expansion of the input data set through satellite monitoring and adaptation of the model to changing climatic conditions. The scientific novelty of this study lies in the development and experimental validation of a regional wheat yield forecasting model for the North-Kazakhstan region based on an artificial neural network that simultaneously integrates agrochemical, meteorological and phytosanitary factors at the level of 13 districts. An optimized three-layer network architecture trained with the Levenberg–Marquardt algorithm is tailored to a multi-year (2008–2017) regional data set, and the resulting model achieves high predictive accuracy (R² = 0.9831 with low RMSE and MAPE), demonstrating the advantages of the neural-network approach for regional agrometeorological monitoring and yield planning. Keywords: wheat, yield predictors, artificial neural network, MATLAB, yield forecasting