基于GA-Adam优化算法的BP神经网络农业灌水量预测模型

Forecast Model of Agricultural Irrigation Water Volume Based on GA-Adam Optimization Algorithm BP Neural Network

针对传统BP神经网络预测农业灌水量时存在易陷入局部最小值、难以选择合适学习率的问题,提出了一种基于遗传算法和Adam算法并行优化BP神经网络的农业灌水量预测模型。该模型利用遗传算法对BP神经网络进行初始权值和阈值的预筛选,然后采用Adam算法来实现学习率自适应于参数梯度不断更新。收集黄河流域陇中片灌溉分区内7个典型灌区的气象数据以及玉米实测灌水数据对模型进行训练,同时与传统GD法、GA法、Adam法下的网络模型进行对比。结果表明:GA-Adam模型仅在训练次数为67次,训练时长为0.403 s时便达到预设精度;且GA-Adam模型预测值与期望值的RMSE和MAE最小,分别为54.73和47.76,决定系数R^(2)为0.81,总体预测效果最好。

To solve the issue that the traditional BP neural network predicts the amount of agricultural irrigation water,fall into the local minimum and select the appropriate learning rate with difficulty,this paper proposes an agricultural irrigation amount prediction model based on the genetic algorithm and the Adam algorithm to optimize the BP neural network in parallel. This model uses the global search characteristics of genetic algorithm to prescreen the BP neural network with initial weights and thresholds,and then uses Adam algorithm to realize the learning rate adaptation to the constant update of the parameter gradient. The meteorological data of 7 typical irrigation areas in the Longzhong Area of the Yellow River Basin and the measured irrigation data of corn are collected to train the model. At the same time,it is compared with the network model under the traditional GD method,GA method,and Adam method. The results show that the GA-Adam model reaches the preset accuracy only when the training times are 67 times and the training duration is 0.403 s;and the RMSE and MAE of the predicted value and expected value of the GA-Adam model are the smallest:54.73 and 47.76 respectively. The coefficient of determination R^(2) is 0.81. In a word,the overall prediction effect of GA-Adam model is the best.