In order to optimize the hyperparameters of a multi-layer Recurrent Neural Network (RNN) for crop yield prediction, this research compares three metaheuristic optimization approaches: Genetic Algorithm (GA), Whale Optimization Algorithm (WOA), and Particle Swarm Optimization (PSO). The study is done for wheat yield in Rajasthan, India, using a multi-source data set that includes climatic parameters, soil properties, nutrient content (NPK), and past yield values. Deep models such as RNNs can learn temporal dependencies in such data. However, their performance tends to be very sensitive to architecture and training hyperparameters. To solve this problem, this study uses GA, WOA, and PSO to optimize nine essential RNN hyperparameters automatically. Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Coefficient of Determination (R2) are used to test the best models. Out of the three, the WOA-optimized RNN performed the best with the lowest RMSE (390.03), highest R2 value (0.7340), and highest accuracy in prediction (85.00%). In addition, this work contributes to Sustainable Development Goal 2: Zero Hunger, which aims to eradicate hunger and attain food security via sustainable agriculture. By enhancing the prediction of wheat yield, the research aids farmers in better planning, utilizing resources more effectively, and eventually producing more wheat, giving rise to increased productivity and improved food supply. Proper forecasting also helps governments and policy planners coordinate food supplies and agricultural schemes. In simple terms, the suggested model serves to produce more wheat, which is a useful component in the global fight against hunger.