Developing a Python neural network model for potato yield forecasting with limited data

This research focuses on developing a neural network model in the Python programming language to solve the technological problem of forecasting potato yields with limited agroclimatic data. The developed program is considered as an element of the digital transformation of technological processes in crop production, aimed at optimizing the use of agricultural machinery and material resources in the agro-industrial complex. The study used data for 2022–2024 on climatic conditions, soil types, and actual potato yields at agricultural enterprises in the Ulyanovsk region. To ensure the quality of the initial information, comprehensive data preprocessing was performed, including the elimination of missing values, filtering of statistical outliers, and normalization of numerical parameters. A neural network model was created using the TensorFlow and Keras libraries with an architecture including an input layer, two hidden layers of 64 neurons each, and an output layer. A distinctive feature of the study is that it works with a limited amount of data (15 observations over 3 years). Therefore, regularization (L2, Dropout) and data augmentation methods were used to prevent overfitting. The architecture was optimized using hyperparameter selection, and a cross-validation split of the data was used to evaluate the reliability of the model. Validation of the model on a test set showed that the developed neural network provides a mean absolute error (MAE) of 0.32 t/ha and a determination coefficient (R²) of 0.87. The model outperformed multiple linear regression (MAE = 0.45 t/ha, R² = 0.75) and random forest (MAE = 0.38 t/ha, R² = 0.81) in forecast quality. The obtained results demonstrate the potential of the developed software for improving the efficiency of planning agricultural operations. It is recommended to implement the model as a software module for decision support systems in the management of technological processes in potato growing.