【摘要】 随着我国贺兰山东麓地区葡萄园的大量兴建和大规模引种，各个葡萄园内出现了以霜霉病为典型的严重的病虫害问题。针对目前酿酒葡萄霜霉病精确预测手段缺乏的问题，提出一种基于遗传算法改进长期和短期记忆神经网络的预测模型。将遗传算法(GA)加入长短期记忆神经网络(LSTM)预测模型的参数调节环节中，通过优化算法代替人工手动调参在超参数搜索空间中不断迭代得到最优超参数组合最终确定模型。再建立基于霜霉病—气象时序数据的手动调参LSTM模型和BP神经网络模型，将三种模型在测试集上进行对比试验。GA-LSTM模型的预测结果均方根误差、均方误差、平均绝对误差分别为0.410 3、0.168 4、0.245 0,均小于LSTM模型和BP神经网络模型。预测结果表明LSTM在时间序列问题的应用中预测性能优于BP神经网络模型，使用遗传算法对LSTM模型的超参数选择环节进行优化，最终得到的超参数组合优于手动调参的LSTM模型得到的超参数组合。更多还原

【Abstract】 With the large-scale construction and introduction of vineyards in the eastern foothills of Helan Mountain in China, serious pest problems typified by downy mildew have emerged in each vineyard. In view of the lack of accurate prediction methods for downy mildew in wine grapes, a prediction model based on one kind of long-term and short-term memory neural network which is improved by genetic algorithm is proposed. Genetic algorithm(GA) is added to the parameter adjustment link of the long short-term memory neural network(LSTM) prediction model, and the optimal hyperparameter combination is obtained by continuously iterating in the hyperparameter search space through the optimization algorithm instead of manual parameter tuning. Then, the manual parameter LSTM model and BP neural network model based on downy mildew-meteorological time series data were established, and these three models were compared on the test set. The prediction results including root mean square error, mean square error, mean absolute error of GA-LSTM model are 0.410 3, 0.168 4 and 0.245 0 respectively, which are smaller than that of LSTM model and BP neural network model. The prediction results show that the prediction performance of LSTM in the application of time series problems is better than that of BP neural network model, by using genetic algorithm to optimize the hyperparameter selection link of LSTM model, the final hyperparameter combination is better than that obtained by manually tuned LSTM model.更多还原