Accurately predicting food production requires modelling the complex nonlinear and multivariate temporal relationships identified in crop yield data. Hence, this paper proposes an enhanced forecasting framework that integrates Temporal Fusion Transformer (TFT) with Particle Swarm Optimisation (PSO) for hyperparameter adjustment. The designed approach preprocesses raw agricultural datasets, including data transformation and cleaning for addressing missing values, discrepancies and outliers. Transformation procedures are then made to ensure uniform representation and temporal alignment. To create lag features, rolling statistics, seasonal indicators and interaction variables, feature engineering based normalization approach is deployed. The Temporal Fusion Transformer, a Deep Learning (DL) architecture created especially for interpretable multi-horizon time-series prediction, is used in the primary prediction model. TFT combines gated residual networks, variable selection networks, recurrent layers, and multi-head attention mechanisms. Particle Swarm Optimisation optimize important hyperparameters to improve prediction performance. From python software, it is proven that with the use of PSO optimized TFT model, minimal error of Mean Square Error (MSE) and 0.002 Root Mean Square Error (RMSE) value compared to the other technical models.