算法介绍及实现——基于遗传算法改进的BP神经网络算法（附完整Python实现）

目录

一、算法介绍

1.1 遗传算法

1.2 为什么要使用遗传算法进行改进

二、算法原理

三、算法实现

3.1 算子选择

3.2 代码实现

一、算法介绍

1.1 遗传算法

        遗传算法是受启发于自然界中生物对于自然环境 “适者生存”的强大自适应能力，通过对生物演化过程模拟和抽象，构建了以自然界生物演变进化为逻辑基础的遗传算法。遗传算法包括了自然界生物在演变过程中的主要步骤，即选择、（基因）变异和（基因）交叉，对应着遗传算法中的三个运算算子。在具体的优化问题下，遗传算法会产生多个问题的可行解作为种群，然后让种群进行模拟意义上生物进化中的选择、变异、交叉等操作。在种群繁衍（迭代）一定次数之后，通过计算种群的适应度，寻找最终种群中的最优个体，该个体即代表优化问题的近似最优解。上述此即为遗传算法主要思想。其流程图如下：

1.2 为什么要使用遗传算法进行改进

        BP算法原理不多赘述，可见我之前博文BP原理介绍，在BP训练过程中，很容易出现陷入局部最小值的情况，所以引入遗传算法进行优化。遗传作为一种模拟生物进化的全局寻优算法，有着优秀的全局寻优能力，能够以一个种群为基础不断的迭代进化，最后获得问题的最优解或近似最优解。BP算法和遗传算法都是人们广泛使用的算法，而且两算法具有明显的优势互补，故而很多研究者都在探索两个算法的融合方法，以期能提高算法性能、提升算法精度。

二、算法原理

        基于遗传算法改进的BP神经网络算法（GA-BP算法）的主要思想即为：通过遗传算法的全局寻优能力获得最优的BP网络的初始权值和阈值，将寻优算法获得的最优初始权值和阈值作为BP神经网络的初始权值和阈值，然后进行训练以避免陷入局部最小值。遗传算法改进后的BP神经网络权值不是随机产生的，而是遗传算法寻优模块获得的。BP算法中的初始权值和阈值作为遗传算法个体的基因值，个体长度即为BP神经网络中权值和阈值的个数，每个基因即代表一个权值或阈值，基因上的数值就是BP神经网络中连接权值或阈值的真实值，如此便组成了遗传算法中的一个染色体。一定数量的染色体作为遗传算法训练的初始种群，再经过遗传算法的选择运算、交叉运算、变异运算等迭代过程后获得一个最优个体，然后以最优个体作为BP网络的初始参数进行训练，此即为GA-BP算法的原理。流程图如下：

三、算法实现

3.1 算子选择

         对于(e)所述的组织方法，是当影响因子数据和目标数据没有很强的相关性的情况下，用前一时序区间的数据作为该时序数据的影响因子来进行训练。

3.2 代码实现

         实例为基于一段时序监测数据的滑坡位移预测，监测影响因子数据有：温度、降雨、风力、灌溉等，监测的目标数据是坡体的裂缝宽度数据。实验表明影响因子数据和目标数据不具有强相关性，所以选择用目标数据本身作为影响因子数据。

        将整个算法分成如下模块：

chrom_code  # 基因编码模块
chrom_mutate  # 变异算子模块
chrom_cross  # 交叉算子模块
chrom_select  # 选择算子模块
chrom_fitness  # 染色体适应度计算模块
data_prepare  # 数据准备模块
BP_***work  # BPNN模块
chrom_test  # 染色体检测模块

new_GA-BP   # 改进算法主程序

chrom_test.py  检测生成的染色体基因有没有超限。

# 染色体检查
# 检查染色体中有没有超出基因范围的基因

def test(code_list,bound):
    """
    :param code_list: code_list: 染色体个体
    :param bound: 各基因的取值范围
    :return: bool
    """
    for i in range(len(code_list)):
        if code_list[i] < bound[i][0] or code_list[i] > bound[i][1]:
            return False
        else:
            return True

 chrom_code.py  基因编码。

# 基因编码模块

import random
import numpy as np
import chrom_test

def code(chrom_len,bound):
    """
    :param chrom_len: 染色体的长度，为一个数，采用实数编码即为基因的个数
    :param bound: 取值范围，为一个二维数组，每个基因允许的取值范围
    :return: 对应长度的编码
    """
    code_list = []
    count = 0
    while True:
        pick = random.uniform(0,1)
        if pick == 0:
            continue
        else:
            pick = round(pick,3)
            temp = bound[count][0] + (bound[count][1] - bound[count][0])*pick
            temp = round(temp,3)
            code_list.append(temp)
            count = count + 1
        if count == chrom_len:
            if chrom_test.test(code_list,bound):
                break
            else:
                count = 0
    return code_list

BP_***work.py    完成网络结构的构建。

# BP模块 借助PyTorch实现

import torch

# 引入了遗传算法参数的BP模型
class BP_***(torch.nn.Module):

    def __init__(self, n_feature, n_hidden, n_output, GA_parameter):
        super(BP_***, self).__init__()
        # 构造隐含层和输出层
        self.hidden = torch.nn.Linear(n_feature, n_hidden)
        self.output = torch.nn.Linear(n_hidden, n_output)
        # 给定网络训练的初始权值和偏执等
        self.hidden.weight = torch.nn.Parameter(GA_parameter[0])
        self.hidden.bias = torch.nn.Parameter(GA_parameter[1])
        self.output.weight = torch.nn.Parameter(GA_parameter[2])
        self.output.bias = torch.nn.Parameter(GA_parameter[3])

    def forward(self, x):
        # 前向计算
        hid = torch.tanh(self.hidden(x))
        out = torch.tanh(self.output(hid))
        return out

# 传统的BP模型
class ini_BP_***(torch.nn.Module):

    def __init__(self, n_feature, n_hidden, n_output):
        super(ini_BP_***, self).__init__()
        # 构造隐含层和输出层
        self.hidden = torch.nn.Linear(n_feature, n_hidden)
        self.output = torch.nn.Linear(n_hidden, n_output)

    def forward(self, x):
        # 前向计算
        hid = torch.tanh(self.hidden(x))
        out = torch.tanh(self.output(hid))
        return out

def train(model, epochs, learning_rate, x_train, y_train):
    """
    :param model: 模型
    :param epochs: 最大迭代次数
    :param learning_rate:学习率
    :param x_train:训练数据（输入）
    :param y_train:训练数据（输出）
    :return: 最终的loss值（MSE）
    """
    # path = "log.txt"
    # f = open(path, 'w',encoding='UTF-8')
    # f.write("train log\n------Train Action------\n"
    #         "Time：{}\n".format(time.ctime()))
    loss_fc = torch.nn.MSELoss(reduction="sum")
    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
    loss_list = []
    for i in range(epochs):
        model.train()
        # 前向计算
        data = model(x_train)
        # 计算误差
        loss = loss_fc(data, y_train)
        loss_list.append(loss)
        # 更新梯度
        optimizer.zero_grad()
        # 方向传播
        loss.backward()
        # 更新参数
        optimizer.step()
        # print("This is {} th iteration,MSE is {}。".format(i+1,loss))
    loss_ls = [loss_list[i].detach().numpy() for i in range(len(loss_list))]
    return loss_ls

 chrom_fitness.py   适应度计算

# 适应度计算模块
# 功能；传入一个编码，返回一个适应度值
from torchvision.transforms import transforms
import torch
import BP_***work
import numpy as np

# 最小二乘思想获得两组数据的误差
def zxec_PC(X, Y):
    X = np.array(X, dtype=np.float).flatten()
    Y = np.array(Y, dtype=np.float).flatten()
    if len(X) != len(Y):
        print("Wrong!")
    n = len(X)
    Wc = 0
    for i in range(n):
        Wc = Wc + (X[i] - Y[i]) * (X[i] - Y[i])
    return Wc

def calculate_fitness(code,n_feature,n_hidden,n_output,epochs
                      ,learning_rate,x_train,y_train):
    """
    :param code: 染色体编码
    :param n_feature: 输入层个数
    :param n_hidden: 隐含层个数
    :param n_output: 输出层个数
    :param epochs: 最多迭代次数
    :param learning_rate: 学习率
    :param x_train: 训练（输入）数据
    :param y_train: 训练（输出）数据
    :return: fitness 适应度值
    """
    Parameter = code[:]
    # 参数提取
    hidden_weight = Parameter[0:n_feature * n_hidden]
    hidden_bias = Parameter[n_feature * n_hidden:
                  n_feature * n_hidden + n_hidden]
    output_weight = Parameter[n_feature * n_hidden + n_hidden:
                  n_feature * n_hidden + n_hidden + n_hidden * n_output]
    output_bias = Parameter[n_feature * n_hidden + n_hidden + n_hidden * n_output:
                  n_feature * n_hidden + n_hidden + n_hidden * n_output + n_output]

    # 类型转换
    tensor_tran = transforms.ToTensor()
    hidden_weight = tensor_tran(np.array(hidden_weight).reshape((n_hidden, n_feature))).to(torch.float32)
    hidden_bias = tensor_tran(np.array(hidden_bias).reshape((1, n_hidden))).to(torch.float32)
    output_weight = tensor_tran(np.array(output_weight).reshape((n_output,n_hidden))).to(torch.float32)
    output_bias = tensor_tran(np.array(output_bias).reshape((1, n_output))).to(torch.float32)
    # 形装转换
    hidden_weight = hidden_weight.reshape((n_hidden,n_feature))
    hidden_bias = hidden_bias.reshape(n_hidden)
    output_weight = output_weight.reshape((n_output,n_hidden))
    output_bias = output_bias.reshape(n_output)
    # 带入模型计算
    GA = [hidden_weight, hidden_bias, output_weight, output_bias]
    BP_model = BP_***work.BP_***(n_feature,n_hidden,n_output,GA)
    loss = BP_***work.train(BP_model,epochs,learning_rate,x_train,y_train)
    # 计算适应度
    prediction = BP_model(x_train)
    fitness = 10 - zxec_PC(prediction.detach().numpy(),y_train.detach().numpy())
    return round(fitness,4)

 chrom_mutate.py  选择算子

# 变异算子
import random

def mutate(chrom_sum, size, p_mutate, chrom_len, bound, maxgen, nowgen):
    """
    :param chrom_sum: 染色体群，即种群，里面为一定数量的染色体  类型为一个二维列表
    :param size: 种群规模，即染色体群里面有多少个染色体  为一个数
    :param p_mutate: 交叉概率 为一个浮点数
    :param chrom_len: 种群长度，即一条染色体的长度，即基因的个数 为一个数
    :param bound: 各基因的取值范围
    :param maxgen:  最大迭代次数
    :param nowgen: 当前迭代次数
    :return: 变异算子后的种群
    """
    count = 0
    # print("\n---这是第{}次遗传迭代...".format(nowgen))
    while True:
        # 随机选择变异染色体
        # print("{}-{}".format(nowgen,count+1))
        seek = random.uniform(0,1)
        while seek == 1:
            seek = random.uniform(0,1)
        index = int(seek * size)
        # print("可能变异的染色体号数为：",index)
        # 判断是否变异
        flag = random.uniform(0,1)
        if p_mutate >= flag:
            # 选择变异位置
            # print("发生变异中...")
            seek1 = random.uniform(0,1)
            while seek1 == 1:
                seek1 = random.uniform(0,1)
            pos = int(seek1 * chrom_len)
            # print("变异的基因号数为：",pos)
            # 开始变异
            seek3 = random.uniform(0,1)
            fg = pow(seek3*(1-nowgen/maxgen),2) # 约到迭代后期，其至越接近0，变异波动就越小
            # print("变异前基因为：",chrom_sum[index][pos])
            if seek3 > 0.5:
                chrom_sum[index][pos] = round(chrom_sum[index][pos] +
                                              (bound[pos][1] - chrom_sum[index][pos])*fg,3)
            else:
                chrom_sum[index][pos] = round(chrom_sum[index][pos] -
                                              (chrom_sum[index][pos] - bound[pos][0])*fg,3)
            # print("变异后基因为：", chrom_sum[index][pos])
            count = count + 1
        else:
            # print("未发生变异。")
            count = count + 1
        if count == size:
            break
    return chrom_sum

chrom_cross.py  交叉算子

# 交叉算子
import random
import chrom_test

def cross(chrom_sum, size, p_cross, chrom_len, bound):
    """
    :param chrom_sum:种群集合，为二维列表
    :param size:种群总数，即染色体的个数
    :param p_cross:交叉概率
    :param chrom_len:染色提长度，每个染色体含基因数
    :param bound:每个基因的范围
    :return: 交叉后的种群集合
    """
    count = 0
    while True:
        # 第一步 先选择要交叉的染色体
        seek1 = random.uniform(0,1)
        seek2 = random.uniform(0,1)
        while seek1 == 0 or seek2 == 0 or seek1 == 1 or seek2 == 1:
            seek1 = random.uniform(0, 1)
            seek2 = random.uniform(0, 1)
        # index_1(2)为选中交叉的个体在种群中的索引
        index_1 = int(seek1 * size)
        index_2 = int(seek2 * size)
        if index_1 == index_2:
            if index_2 == size - 1:
                index_2 = index_2 - 1
            else:
                index_2 = index_2 + 1
        # print("可能交叉的两个染色体为：",index_1,index_2)
        # 第二步 判断是否进行交叉
        flag = random.uniform(0,1)
        while flag == 0:
            flag = random.uniform(0,1)
        if p_cross >= flag:
            # 第三步 开始交叉
            # print("开始交叉...")
            p_pos = random.uniform(0, 1)
            while p_pos == 0 or p_pos == 1:
                p_pos = random.uniform(0, 1)
            pos = int(p_pos * chrom_len)
            # print("交叉的极影位置为：",pos)
            var1 = chrom_sum[index_1][pos]
            var2 = chrom_sum[index_2][pos]
            pick = random.uniform(0,1)
            # print("交叉前染色体为：")
            # print(chrom_sum[index_1])
            # print(chrom_sum[index_2])
            chrom_sum[index_1][pos] = round((1-pick) * var1 + pick * var2,3)
            chrom_sum[index_2][pos] = round(pick * var1 + (1-pick) * var2,3)
            # print("交叉后染色体为：")
            # print(chrom_sum[index_1])
            # print(chrom_sum[index_2])
            if chrom_test.test(chrom_sum[index_1],bound) and chrom_test.test(chrom_sum[index_2],bound):
                count = count + 1
            else:
                continue
        else:
            # print("没有发生交叉。")
            count = count + 1
        # print("本次循环结束\n")
        if count == size:
            break
    return chrom_sum

chrom_select.py   选择算子 

# 选择算子
import numpy as np
import random

def select(chrom_sum,fitness_ls):
    """
    :param chrom_sum:种群
    :param fitness_ls: 各染色体的适应度值
    :return: 更新后的种群
    """
    # print("种群适应度分别为：",fitness_ls)
    fitness_ls = np.array(fitness_ls,dtype=np.float64)
    sum_fitness_ls = np.sum(fitness_ls,dtype=np.float64)
    P_inh = []
    M = len(fitness_ls)
    for i in range(M):
        P_inh.append(fitness_ls[i]/sum_fitness_ls)
    # 将概率累加
    for i in range(len(P_inh)-1):
        P_temp = P_inh[i] + P_inh[i+1]
        P_inh[i+1] = round(P_temp, 2)
    P_inh[-1] = 1
    # 轮盘赌算法选择染色体
    a***ount = []
    for i in range(M):
        rand = random.random()
        for j in range(len(P_inh)):
            if rand <= P_inh[j]:
                a***ount.append(j)
                break
            else:
                continue
    # 根据索引号跟新种群
    # print("轮盘赌的结果为：",a***ount)
    new_chrom_sum = []
    for i in a***ount:
        new_chrom_sum.append(chrom_sum[i])
    return new_chrom_sum

data_prepare.py  数据准备

# 数据准备
import numpy as np
import pandas as pd

def Data_loader():
    # 文件路径
    ENU_measure_path = "18-10-25至19-3-25三方向位移数据.xlsx"
    t_path = "天气数据.xls"
    M_path = "data.csv"
    # 三方向数据
    df_1 = pd.read_excel(ENU_measure_path)
    ENU_df = pd.DataFrame(df_1)
    ENU_E = ENU_df["E/m"]
    ENU_E = np.array(ENU_E)
    ENU_N = ENU_df["N/m"]
    ENU_N = np.array(ENU_N)
    ENU_U = ENU_df["U/m"]
    ENU_U = np.array(ENU_U)
    ENU_R = ENU_df['R/m']
    ENU_R = np.array(ENU_R)
    
    df_2 = pd.read_excel(t_path)
    t_df = pd.DataFrame(df_2)
    # 最大温度数据
    max_tem = t_df["bWendu"]
    max_tem_ls = []
    for i in range(len(max_tem)):
        temp = str(max_tem[i])
        temp = temp.replace("℃","")
        max_tem_ls.append(eval(temp))
    max_tem = np.array(max_tem_ls)
    # 最低温度数据
    min_tem = t_df["yWendu"]
    min_tem_ls = []
    for i in range(len(min_tem)):
        temp = str(min_tem[i])
        temp = temp.replace("℃","")
        min_tem_ls.append(eval(temp))
    min_tem =np.array(min_tem_ls)
    # 天气数据
    tianqi = t_df["Tian_Qi"]
    tianqi = np.array(tianqi)
    # 风力数据
    Feng = t_df["Feng"]
    Feng = np.array(Feng)
    # 降雨数据
    rain = t_df["rainfall"]
    rain = np.array(rain)
    # 灌溉数据
    guangai = t_df["guangai"]
    guangai = np.array(guangai)
    # 获取时间数据
    namels = t_df["ymd"]
    name_ls = []
    for i in range(len(namels)):
        temp = str(namels[i])
        temp = temp.replace(" 00:00:00","")
        name_ls.append(str(temp))
    # 读取另一文件数据，该数据为位移计和GNSS监测数据
    df_3 = pd.read_csv(M_path)
    M_df = pd.DataFrame(df_3)
    M_data = M_df["Measurerel"]
    R_data = M_df["R"]
    M_data = np.array(M_data)
    R_data = np.array(R_data)

    return [ENU_R, M_data, R_data, ENU_U, ENU_E, ENU_N,max_tem,min_tem,name_ls]

主程序！！！！

# 改进算法主程序
import sys
import chrom_code  # 基因编码模块
import chrom_mutate  # 变异算子模块
import chrom_cross  # 交叉算子模块
import chrom_select  # 选择算子模块
import chrom_fitness  # 染色体适应度计算模块
import data_prepare  # 数据准备模块
import BP_***work  # BPNN模块
import torch
import torch.nn.functional as F
from torchvision.transforms import transforms
import numpy as np
import matplotlib.pyplot as plt
import time

plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

# -----参数设置-----
epochs = 300  # 神经网络最大迭代次数
learning_rate = 0.01  # 学习率
n_feature = 6  # 输入层个数
n_hidden = 9  # 隐含层个数
n_output = 1  # 输出层个数

chrom_len = n_feature * n_hidden + n_hidden + n_hidden * n_output + n_output  # 染色体长度
size = 15  # 种群规模
bound = np.ones((chrom_len, 2))
sz = np.array([[-1, 0], [0, 1]])
bound = np.dot(bound, sz)  # 各基因取值范围
p_cross = 0.4  # 交叉概率
p_mutate = 0.01  # 变异概率
maxgen = 30  # 遗传最大迭代次数

# 数据准备
# ========================================= #
data_set = data_prepare.Data_loader()
displace = data_set[1]
name_ls = data_set[-1]
in_train_data = []
in_test_data = []

# 数目分配
train_num = 120
test_num = len(displace) - train_num - n_feature

for i in range(len(displace)):
    temp = []
    if i <= train_num-1:  # 用于控制训练数据和预测数据的分配
        temp = [round(displace[i + j], 5) for j in range(n_feature)]
        in_train_data.append(temp)
    else:
        temp = [round(displace[i + j], 5) for j in range(n_feature)]
        in_test_data.append(temp)
    if i == len(displace)-n_feature-1:
        break

# 格式转化
in_train_data = np.array(in_train_data)
in_test_data = np.array(in_test_data)
# 数据分割，用于建模和预测
out_train_data = displace[n_feature:train_num+n_feature]
out_test_data = displace[train_num+n_feature:len(displace)]

# 测试输出
# print(in_train_data)
# print(out_train_data)
# print(in_test_data)
# print(out_test_data)
# print(train_num)
# print(test_num)

# 数据格式转换及数据归一化
tensor_tran = transforms.ToTensor()
# 训练过程中的输入层数据
in_train_data = tensor_tran(in_train_data).to(torch.float)
in_train_data = F.normalize(in_train_data)
in_train_data = in_train_data.reshape(train_num, n_feature)
# 预测过程中的输入层数据
in_test_data = tensor_tran(in_test_data).to(torch.float)
in_test_data = F.normalize(in_test_data)
in_test_data = in_test_data.reshape(test_num, n_feature)
# 训练过程中的输出层数据
out_train_data = out_train_data.reshape(len(out_train_data), 1)
out_train_data = tensor_tran(out_train_data).to(torch.float)
un_norm1 = out_train_data[0][0]
out_train_data = F.normalize(out_train_data)
norm1 = out_train_data[0][0]
out_train_data = out_train_data.reshape(train_num, n_output)
fanshu_train = round(float(un_norm1 / norm1), 4)  # 建模时，训练数据中输出数据的范数
# 预测中用于检验的输出层数据
out_test_data = out_test_data.reshape(len(out_test_data), 1)
out_test_data = tensor_tran(out_test_data).to(torch.float)
un_norm = out_test_data[0][0]   # 归一化前
out_test_data = F.normalize(out_test_data)
norm = out_test_data[0][0]  # 归一化后
out_test_data = out_test_data.reshape(test_num, n_output)
fanshu = round(float(un_norm / norm), 4)  # 预测时，测试数据中输出数据的范数

# 建模训练数据
x_train = in_train_data
y_train = out_train_data
x_test = in_test_data
y_label = out_test_data
# ========================================== #

chrom_sum = []  # 种群，染色体集合
for i in range(size):
    chrom_sum.append(chrom_code.code(chrom_len, bound))
a***ount = 0  # 遗传迭代次数计数器
best_fitness_ls = []  # 每代最优适应度
ave_fitness_ls = []  # 每代平均适应度
best_code = []  # 迭代完成适应度最高的编码值

# 适应度计算
fitness_ls = []
for i in range(size):
    fitness = chrom_fitness.calculate_fitness(chrom_sum[i], n_feature, n_hidden, n_output,
                                              epochs, learning_rate, x_train, y_train)
    fitness_ls.append(fitness)
# 收集每次迭代的最优适应值和平均适应值
fitness_array = np.array(fitness_ls).flatten()
fitness_array_sort = fitness_array.copy()
fitness_array_sort.sort()
best_fitness = fitness_array_sort[-1]
best_fitness_ls.append(best_fitness)
ave_fitness_ls.append(fitness_array.sum() / size)

while True:
    # 选择算子
    # print("\n这是第{}次遗传迭代。".format(a***ount+1))
    # print("平均适应度为：",fitness_array.sum()/size)
    chrom_sum = chrom_select.select(chrom_sum, fitness_ls)
    # 交叉算子
    chrom_sum = chrom_cross.cross(chrom_sum, size, p_cross, chrom_len, bound)
    # 变异算子
    chrom_sum = chrom_mutate.mutate(chrom_sum, size, p_mutate, chrom_len, bound, maxgen, a***ount + 1)
    # 适应度计算
    fitness_ls = []
    for i in range(size):
        fitness = chrom_fitness.calculate_fitness(chrom_sum[i], n_feature, n_hidden, n_output,
                                                  epochs, learning_rate, x_train, y_train)
        fitness_ls.append(fitness)
    # 收集每次迭代的最优适应值和平均适应值
    fitness_array = np.array(fitness_ls).flatten()
    fitness_array_sort = fitness_array.copy()
    fitness_array_sort.sort()
    best_fitness = fitness_array_sort[-1]  # 获取最优适应度值
    best_fitness_ls.append(best_fitness)
    ave_fitness_ls.append(fitness_array.sum() / size)
    # 计数器加一
    a***ount = a***ount + 1
    if a***ount == maxgen:
        index = fitness_ls.index(max(fitness_ls))  # 返回最大值的索引
        best_code = chrom_sum[index]  # 通过索引获得对于染色体
        break

# 参数提取
hidden_weight = best_code[0:n_feature * n_hidden]
hidden_bias = best_code[n_feature * n_hidden:
                        n_feature * n_hidden + n_hidden]
output_weight = best_code[n_feature * n_hidden + n_hidden:
                          n_feature * n_hidden + n_hidden + n_hidden * n_output]
output_bias = best_code[n_feature * n_hidden + n_hidden + n_hidden * n_output:
                        n_feature * n_hidden + n_hidden + n_hidden * n_output + n_output]
# 类型转换
tensor_tran = transforms.ToTensor()
hidden_weight = tensor_tran(np.array(hidden_weight).reshape((n_hidden, n_feature))).to(torch.float32)
hidden_bias = tensor_tran(np.array(hidden_bias).reshape((1, n_hidden))).to(torch.float32)
output_weight = tensor_tran(np.array(output_weight).reshape((n_output, n_hidden))).to(torch.float32)
output_bias = tensor_tran(np.array(output_bias).reshape((1, n_output))).to(torch.float32)
# 形装转换
hidden_weight = hidden_weight.reshape((n_hidden, n_feature))
hidden_bias = hidden_bias.reshape(n_hidden)
output_weight = output_weight.reshape((n_output, n_hidden))
output_bias = output_bias.reshape(n_output)
GA = [hidden_weight, hidden_bias, output_weight, output_bias]

# 带入模型计算
BP_model = BP_***work.BP_***(n_feature, n_hidden, n_output, GA)
ini_BP_model = BP_***work.ini_BP_***(n_feature, n_hidden, n_output)
# 网络训练
loss = BP_***work.train(BP_model, epochs, learning_rate, x_train, y_train)
ini_loss = BP_***work.train(ini_BP_model, epochs, learning_rate, x_train, y_train)
# 建模效果
model_x = BP_model(x_train)
ini_model_x = ini_BP_model(x_train)
# 网络预测
prediction = BP_model(x_test)
ini_prediction = ini_BP_model(x_test)

# 建模数据反归一化(都换算到厘米级)
y_train = y_train.detach().numpy() * fanshu_train
model_x = model_x.detach().numpy() * fanshu_train
ini_model_x = ini_model_x.detach().numpy() * fanshu_train
# 建模绘图
train_name_ls = name_ls[6:126]
xlabel = [i for i in range(0, 120, 14)]
plt.plot(y_train, markersize=4, marker='.', label="真值", c='r')
plt.plot(model_x, markersize=4, marker='.', label="GA-BP预测值", c='b')
plt.title("GA-BP算法建模情况")
plt.ylabel("累计裂缝宽度（mm）")
plt.xticks(xlabel, [train_name_ls[i] for i in xlabel], rotation=25)
plt.grid(linestyle='-.')  # 设置虚线
plt.legend()

f2 = plt.figure()
plt.plot(y_train, markersize=4, marker='.', label="真值", c='r')
plt.plot(ini_model_x, markersize=4, marker='.', label="BP预测值", c='g')
plt.title("BP算法建模情况")
plt.ylabel("累计裂缝宽度（mm）")
plt.xticks(xlabel, [train_name_ls[i] for i in xlabel], rotation=25)
plt.grid(linestyle='-.')
plt.legend()

# 预测数据格式转换（厘米级）
GABP_prediction = prediction.detach().numpy()
BP_prediction = ini_prediction.detach().numpy()
y_label = y_label.detach().numpy()
# 预测数据反归一化（厘米级）
GABP_prediction = GABP_prediction * fanshu
BP_prediction = BP_prediction * fanshu
y_label = y_label * fanshu

# 计算预测结果的SSE误差
def get_MSE(argu1, argu2):
    if len(argu1) != len(argu2):
        return 0
    error = 0
    for i in range(len(argu1)):
        error = error + pow((argu1[i] - argu2[i]), 2)
    error = float(error[0])
    return round(error, 5)


error_BP = get_MSE(y_label, BP_prediction)
error_GA_BP = get_MSE(y_label, GABP_prediction)
print("BP算法预测MSE误差为：", error_BP)
print("GA-BP算法预测MSE误差为：", error_GA_BP)

# 将巡行情况和运行结果写入日志
f = open("log.txt",'a',encoding='UTF-8')     # 追加写打开文件
f.write("运行时间：" + str(time.ctime()) + '\n')
f.write("训练数据长度为：" + str(train_num) + '\n'
        + "测试数据长度为：" + str(test_num) + '\n')
f.write("网络结构层数为：{}、{}、{}\n".format(n_feature,n_hidden,n_output))
f.write("遗传迭代所获得的最优权值为：" + str(best_code) + "\n")
f.write("======预测结果如下======\n真值数据为：" + str(y_label.flatten()) + '\n')
f.write("BP预测结果为：" + str(BP_prediction.flatten()) + "\n"
        + "GA-BP预测结果为：" + str(GABP_prediction.flatten()) + '\n')
f.write("-->>BP预测MSE误差为：" + str(error_BP) + '平方厘米\n'
        + "-->>GA-BP预测MSE误差为：" + str(error_GA_BP) + '平方厘米\n\n')
f.close()

# 预测绘图
test_name_ls = name_ls[126:152]
xlabel2 = [i for i in range(0, 26, 4)]

f3 = plt.figure()
plt.plot(y_label, markersize=4, marker='.', label="真值", c='r')
plt.plot(GABP_prediction, markersize=4, marker='*', label="GA-BP预测值", c='b')
plt.plot(BP_prediction, markersize=4, marker='^', label="BP预测值", c='g')
plt.title("算法预测情况对比")
plt.ylabel("累计裂缝宽度（mm）")
plt.xticks(xlabel2, [test_name_ls[i] for i in xlabel2], rotation=20)
plt.legend()
plt.grid(linestyle='-.')

f4 = plt.figure()
plt.plot(y_label, markersize=4, marker='.', label="真值", c='r')
plt.plot(BP_prediction, markersize=4, marker='^', label="BP预测值", c='g')
plt.title("BP算法预测情况")
plt.ylabel("累计裂缝宽度（mm）")
plt.xticks(xlabel2, [test_name_ls[i] for i in xlabel2], rotation=20)
plt.legend()
plt.grid(linestyle='-.')

f5 = plt.figure()
plt.plot(y_label, markersize=4, marker='.', label="真值", c='r')
plt.plot(GABP_prediction, markersize=4, marker='*', label="GA-BP预测值", c='b')
plt.title("GA-BP算法预测情况")
plt.ylabel("累计裂缝宽度（mm）")
plt.xticks(xlabel2, [test_name_ls[i] for i in xlabel2], rotation=20)
plt.legend()
plt.grid(linestyle='-.')

plt.show()

对比结果确实有提升：

 资源获取：

链接：https://pan.baidu.***/s/1ZiqgN98bhnyEdoQxuDB3SQ?pwd=ervf 
提取码：ervf 
--来自百度网盘超级会员V4的分享

才疏学浅，水平有限。敬请批评指正！

共勉！