Python实现ALO-LSTM蚁狮算法优化长短期记忆神经网络时间序列预测（完整源码和数据)

#!/usr/bin/env python # coding: utf-8 # In[1]: import math # 导入math模块进行数学运算 import random # 导入random模块用于生成随机数 import tensorflow as tf # 导入TensorFlow库 import matplotlib.pyplot as plt # 导入matplotlib库的pyplot模块进行绘图 import numpy as np # 导入numpy库进行数值操作 import pandas as pd # 导入pandas库进行数据处理 from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score # 从scikit-learn导入评估指标 from sklearn.preprocessing import MinMaxScaler # 导入MinMaxScaler进行数据缩放 from tensorflow.python.keras.callbacks import EarlyStopping # 导入EarlyStopping回调函数 from tensorflow.python.keras.layers import Dense, Dropout, LSTM # 导入Dense、Dropout和LSTM层 from tensorflow.python.keras.layers.core import Activation # 导入Activation激活函数 from tensorflow.python.keras.models import Sequential # 导入Sequential模型 import numpy # 导入numpy库 import copy # 导入copy模块 import warnings # 导入warnings模块进行警告控制 warnings.filterwarnings("ignore") # 忽略警告信息 # #更多模型搜索机器学习之心，支持模型定制 # In[2]: plt.rcParams['font.sans-serif'] = ['Microsoft YaHei'] # 用来正常显示中文标签 plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号 # In[3]: def ALO(N, Max_iter, lb, ub, dim, fobj): # 初始化位置的函数 def initialization(N, dim, ub, lb): return np.random.uniform(lb, ub, (N, dim)) # 轮盘赌选择的函数 def roulette_wheel_selection(weights): selection = np.random.rand() * sum(weights) cumulative_sum = np.cumsum(weights) selected_index = np.where(selection <= cumulative_sum)[0][0] return selected_index # 围绕蚁狮进行随机漫步的函数 def random_walk_around_antlion(dim, Max_iter, lb, ub, antlion_position, current_iter): walk = np.zeros((Max_iter, dim)) for i in range(dim): walk[:, i] = antlion_position[i] + np.random.uniform(-1, 1, Max_iter) * (ub[i] - lb[i]) # 边界检查 walk[:, i] = np.maximum(np.minimum(walk[:, i], ub[i]), lb[i]) return walk # 初始化蚁狮和蚂蚁的位置 antlion_position = initialization(N, dim, ub, lb) ant_position = initialization(N, dim, ub, lb) sorted_antlions = np.zeros((N, dim), dtype=float) elite_antlion_position = np.zeros(dim, dtype=float) elite_antlion_fitness = np.inf convergence_curve = np.zeros(Max_iter) antlions_fitness = np.zeros(N) ants_fitness = np.zeros(N) # 计算初始蚁狮的适应度并排序 for i in range(antlion_position.shape[0]): antlions_fitness[i] = fobj(antlion_position[i, :]) sorted_indexes = np.argsort(antlions_fitness) sorted_antlions = antlion_position[sorted_indexes, :] elite_antlion_position = sorted_antlions[0, :] elite_antlion_fitness = antlions_fitness[0] current_iter = 0 while current_iter < Max_iter: # 模拟随机漫步 for i in range(ant_position.shape[0]): # 根据适应度选择蚁狮（适应度越好，选择的概率越高） rolette_index = roulette_wheel_selection(1. / antlions_fitness) if rolette_index == -1: rolette_index = 1 # RA是围绕选定的蚁狮进行的随机漫步 RA = random_walk_around_antlion(dim, Max_iter, lb, ub, sorted_antlions[rolette_index, :], current_iter) # RE是围绕精英（迄今为止最好的蚁狮）进行的随机漫步 RE = random_walk_around_antlion(dim, Max_iter, lb, ub, elite_antlion_position, current_iter) # 更新蚂蚁的位置，根据论文中的公式（2.13） ant_position[i, :] = (RA[current_iter, :] + RE[current_iter, :]) / 2 for i in range(ant_position.shape[0]): # 边界检查（将蚂蚁的位置限制在搜索空间内） flag4ub = ant_position[i, :] > ub flag4lb = ant_position[i, :] < lb ant_position[i, :] = (ant_position[i, :] * ~(flag4ub + flag4lb)) + ub * flag4ub + lb * flag4lb ants_fitness[i] = fobj(ant_position[i, :]) # 更新蚁狮的位置和适应度，基于蚂蚁（如果蚂蚁的适应度高于蚁狮，则假设蚂蚁被蚁狮捕获，蚁狮更新到蚂蚁的位置以建立陷阱） double_population = np.vstack((sorted_antlions, ant_position)) double_fitness = np.concatenate((antlions_fitness, ants_fitness)) sorted_indexes = np.argsort(double_fitness) double_sorted_population = double_population[sorted_indexes, :] antlions_fitness = double_fitness[sorted_indexes][:N] sorted_antlions = double_sorted_population[:N, :] # 如果任何蚁狮的适应度高于精英，更新精英的位置 if antlions_fitness[0] < elite_antlion_fitness: elite_antlion_position = sorted_antlions[0, :] elite_antlion_fitness = antlions_fitness[0] # 保持精英在种群中 sorted_antlions[0, :] = elite_antlion_position antlions_fitness[0] = elite_antlion_fitness # 更新收敛曲线 convergence_curve[current_iter] = elite_antlion_fitness # 每隔50次迭代显示迭代次数和迄今为止获得的最佳最优解 if current_iter % 10 == 0: print(f"迭代 {current_iter + 1}次时的最佳适应度为 {elite_antlion_fitness}") current_iter += 1 return elite_antlion_fitness, elite_antlion_position, convergence_curve # In[4]: ''' 定义一个函数用于创建数据集，输入参数包括: dataset: 数据集 look_back: 回溯长度，即用多少个时间步作为输入预测下一个时间步 ''' def creat_dataset(dataset, look_back): dataX, dataY = [], [] for i in range(len(dataset) - look_back - 1):# 循环遍历数据集 a = dataset[i: (i + look_back)]# 提取回溯长度的数据作为输入 dataX.append(a)# 将该输入数据添加到输入数据集 dataY.append(dataset[i + look_back])# 将下一个时间步的数据添加到输出数据集 return np.array(dataX), np.array(dataY)# 将输入输出数据集转换为numpy数组并返回 # In[5]: dataframe = pd.read_csv('焦作.csv',header=0, parse_dates=[0],index_col=0, usecols=[0, 1])# 读取CSV文件，并设置第一列为日期，作为索引 dataset = dataframe.values# 获取数据集的值 dataframe.head(10)# 查看前10行数据 # In[6]: scaler = MinMaxScaler(feature_range=(0, 1))# 使用 MinMaxScaler 进行数据归一化 dataset = scaler.fit_transform(dataset.reshape(-1, 1))# 将数据变为二维数组并进行归一化 # In[7]: train_size = int(len(dataset) * 0.8) # 计算训练集大小，将数据集长度的 80% 转换为整数 test_size = len(dataset) - train_size # 计算测试集大小，将剩余的数据作为测试集 train, test = dataset[0: train_size], dataset[train_size: len(dataset)] # 将数据集划分为训练集和测试集，使用切片操作实现 # In[8]: def build_model(lk, neurons1, neurons2, neurons3, dropout): look_back = lk # 设置滑动窗口的大小为look_back trainX, trainY = creat_dataset(train, look_back) # 根据训练集数据和滑动窗口大小创建训练集的输入序列和输出值 testX, testY = creat_dataset(test, look_back) # 根据测试集数据和滑动窗口大小创建测试集的输入序列和输出值 X_train, y_train = trainX, trainY # 将训练集的输入序列和输出值赋值给X_train和y_train X_test, y_test = testX, testY # 将测试集的输入序列和输出值赋值给X_test和y_test model = Sequential() # 创建Sequential模型 model.add(LSTM(units=neurons1, return_sequences=True, input_shape=(look_back, 1))) # 添加第一个LSTM层，设置神经元个数为neurons1，输入�