Comparative Analysis of Short-Term Grain Temperature Forecasting Models in Grain Storages

Short-term forecasting of grain temperature is an important component of intelligent monitoring of grain storage facilities, since it enables the timely detection of dangerous temperature trends, assessment of the risk of local self-heating of the grain mass, and improvement of decision-making efficiency during storage. The temperature state of grain is formed under the influence of the thermal inertia of the grain mass, external meteorological conditions, moisture-related processes, crop type, sensor depth, and spatial heterogeneity of the temperature field. Therefore, it is relevant to study forecasting models that can take into account both the temporal dynamics of the process and nonlinear relationships between meteorological, technological, spatial, and regime-related parameters. This paper investigates multiple linear regression, LSTM, Compact ANFIS, and hybrid LSTM-ANFIS architectures for short-term forecasting of grain temperature at prediction horizons of 1, 3, and 6 hours ahead. The study was carried out using a multi-series time dataset that includes two grain crops, four bins, 24 sensor suspensions, and 72 temperature sensors. The input data contained hourly measurements of grain temperature, external meteorological parameters, sensor location characteristics, and additional engineered features. During data preprocessing, lag, statistical, cyclic, spatial, and mode -related features were generated, as well as target variables for each prediction horizon. To prevent information leakage, the dataset was divided into training, validation, and test subsets according to the chronological principle. The correlation analysis showed that the most informative features for grain temperature forecasting are the current grain temperature, its lagged values, rolling statistics, the difference between ambient air temperature and grain temperature, estimated grain moisture, crop type, sensor depth, and spatial-regime characteristics. It was shown that model performance depends significantly on the prediction horizon. For the 1-hour horizon, the best result was achieved by the multiple linear regression model, which provided MAE = 0.10266, RMSE = 0.16233, and R² = 0.99228. For the 3-hour horizon, the highest accuracy was demonstrated by the hybrid LSTMEncoder + Compact ANFIS model, with MAE = 0.29073, RMSE = 0.39682, and R² = 0.95375. For the 6-hour horizon, the best performance was achieved by the LSTMEncoder + Compact ANFIS Residual + Regime Features model, which reached MAE = 0.53045, RMSE = 0.73544, and R² = 0.84081. The obtained results confirm that, for ultra-short-term grain temperature forecasting, tabular models with properly generated lag, statistical, and engineered features may be sufficiently effective. At the same time, as the prediction horizon increases, the use of hybrid architectures becomes more appropriate. Such models combine the ability of LSTM to capture temporal dependencies with the capability of Compact ANFIS to perform nonlinear forecast refinement based on physically meaningful and spatial-regime features. The practical significance of the proposed approach lies in its potential application in grain storage monitoring systems, early warning of dangerous temperature changes, and decision support for microclimate control in stored grain.