Аннотация
The aim of the study was to optimise the parameters of a neural network model for forecasting plant pest population dynamics in agrocenoses with variable environmental characteristics. The methodology included weekly field data collection over five months at 18 sites in northern Kazakhstan, taking into account six pest species, environmental parameters, crop characteristics, and pest control schemes. Data cleaning, linear interpolation, normalisation, and one-hot encoding were applied. A multilayer perceptron (MLP) architecture with two hidden layers (32 and 16 neurons) and Rectified Linear Unit (ReLU) activation, trained using various configurations, was used as the baseline. The results showed that the highest values of the determination coefficient (R²>0.82) were achieved when modelling Helicoverpa armigera and Frankliniella occidentalis, while the poorest performance was observed for Loxostege sticticalis (R²≈0.64). In plots with tomatoes and sunflowers, forecast accuracy was higher, which was associated with stable irrigation conditions and protective measures. The average Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values were 4.36 and 3.52, respectively. The optimal configuration was identified as a two-layer model (64 and 32 neurons) with ReLU activation and the Adaptive Moment Estimation (ADAM) optimiser, yielding MAE=2.9 and R²=0.83. Deeper models led to overfitting, whereas simplified ones resulted in insufficient accuracy. Statistical tests confirmed the significance of differences between configurations (p<0.001). Cross-validation and sliding window analysis demonstrated the model’s high robustness to spatiotemporal variability, with minimal accuracy degradation when transferred across regions and seasons. The results demonstrate the high practical value of the developed approach for integration into pest monitoring and forecasting systems in agricultural practice.