为了给农业及其相关部门制定粮食策略提供理论依据，提出一种基于改进鲸鱼优化算法（improved whale optimization algorithm, IWOA）的反向传播（back propagation, BP）神经网络混合算法（IWOA-BP）。该混合算法先通过引入改进收敛因子、非线性惯性权重和最优邻域扰动策略改进鲸鱼优化算法，再将其最优解赋值给BP神经网络的权值和阈值，最终提高IWOA-BP的收敛速度和收敛精度。选取全国近45年粮食总产量和7种影响因素（有效灌溉面积、化肥施用量、农村用电量、农业机械总动力、粮食作物播种面积、受灾面积和农村人均消费支出）作为数据集，构建基于改进鲸鱼算法的反向传播神经网络粮食产量预测模型。多次实验表明，IWOA-BP模型在测试集上的表现均优于其他预测模型，包括长短期记忆网络（long short-term memory network, LSTM）预测模型、极限学习机（extreme learning machine, ELM）预测模型、基于鲸鱼优化算法的BP神经网络（WOA-BP）预测模型以及基于粒子群算法的BP神经网络（PSO-BP）预测模型。IWOA-BP模型和ELM模型相比，前者的均方根误差（root mean square error, RMSE）、平均绝对百分比误差（mean absolute percentage error, MAPE）分别降低了77.12%、88.18%；和LSTM模型相比，前者的RMSE、MAPE分别降低了69.11%、47.36%；和WOA-BP模型相比，前者的平均绝对误差（mean absolute error, MAE）、RMSE和MAPE分别降低了43.78%、43.22%、45.96%。和PSO-BP模型相比，前者的MAE、RMSE、MAPE分别降低了89.67%、90.61%、90.82%。因此IWOA-BP预测模型的决定系数更高、预测误差更小且收敛速度更快，可有效地预测粮食产量，对于农业部门和相关政策制定者来说具有重要的技术参考价值。

A hybrid algorithm (IWOA-BP) combining the improved whale optimization algorithm (IWOA) and backpropagation neural network (BP) was proposed to offer theoretical support for the formulation of grain strategies in the agriculture sector and its related industries. By introducing an improved convergence factor, nonlinear inertia weight, and optimal neighborhood disturbance strategy into the modified whale optimization algorithm, the optimal solution of the algorithm was obtained. This solution was then utilized as the initial weights and thresholds of the BP neural network, thereby enhancing the convergence speed and accuracy of the IWOA-BP hybrid algorithm. Subsequently, a grain yield prediction model based on the improved whale optimization algorithm was established using data from China's grain yield over 45 years and seven influencing factors including effective irrigation area, chemical fertilizer application, rural electricity consumption, total power of agricultural machinery, sowing area of grain crops, disaster-affected area, and per capita consumption expenditure in rural areas. Through extensive experiments on a test set, it was found that the IWOA-BP model consistently outperformed other prediction models such as long short-term memory (LSTM), extreme learning machine (ELM), BP neural network with whale optimization algorithm (WOA-BP), and BP neural network with particle swarm optimization (PSO-BP). Compared to the ELM model, the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the IWOA-BP model were reduced by 77.12% and 88.18% respectively. When compared to the LSTM model, the RMSE and MAPE of the IWOA-BP model were reduced by 69.11% and 47.36% respectively. Furthermore, in comparison to the WOA-BP model, the mean absolute error (MAE), RMSE, and MAPE of the IWOA-BP model were reduced by 43.78%, 43.22% and 45.96% respectively. Additionally, when compared to the PSO-BP model, the MAE, RMSE, and MAPE of the IWOA-BP model were reduced by 89.67%, 90.61% and 90.82% respectively. Therefore, the proposed IWOA-BP prediction model can be effectively used to predict grain yield due to its higher coefficient of determination, smaller prediction error, and faster convergence speed. It has important technical reference value for agricultural departments and relevant policymakers.