【Transformer-Adaboost回归预测】Transformer-Adaboost多变量回归预测

import numpy as np
from sklearn.ensemble import AdaBoostRegressor
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.tree import DecisionTreeRegressor
from tensorflow import keras
from tensorflow.keras import layers

数据准备

假设有两个输入变量 X1 和 X2，以及一个目标变量 Y

假设数据已经存储在 X1、X2 和 Y 中，每个变量为二维数组，形状为 (样本数, 时间步长)

参数设置

inputWindowSize = 10 # 输入窗口大小
outputWindowSize = 1 # 输出窗口大小
numFeatures = 2 # 输入变量的数量
numBoostingIterations = 10 # Adaboost 迭代次数
numTransformerLayers = 2 # Transformer 模型的层数
numAttentionHeads = 2 # Transformer 模型的注意力头数
hiddenUnits = 32 # Transformer 模型中的隐藏单元数量
numEpochs = 50 # 训练迭代次数

数据预处理

inputFeatures = np.concatenate((X1, X2), axis=1) # 将输入变量连接起来
outputTarget = Y

数据划分为训练集和测试集

trainInput, testInput, trainTarget, testTarget = train_test_split(inputFeatures, outputTarget, test_size=0.2)

特征标准化

scaler = StandardScaler()
trainInput = scaler.fit_transform(trainInput)
testInput = scaler.transform(testInput)

使用 Transformer 模型进行特征提取

inputShape = (inputWindowSize, numFeatures)

model = keras.Sequential()
model.add(layers.Transformer(
num_layers=numTransformerLayers,
num_heads=numAttentionHeads,
d_model=hiddenUnits,
output_shape=inputShape,
))
model.add(layers.GlobalAveragePooling1D())
model.add(layers.Dense(hiddenUnits, activation=‘relu’))
model.add(layers.Dense(outputWindowSize))

model.compile(optimizer=‘adam’, loss=‘mse’)

模型训练

model.fit(trainInput, trainTarget, epochs=numEpochs, batch_size=32, verbose=0)

使用 Adaboost 进行预测

boostedModels = []
for i in range(numBoostingIterations):
# 用决策树作为基础模型
baseModel = DecisionTreeRegressor(max_depth=3)
baseModel.fit(trainInput, trainTarget)

# 预测残差
residuals = trainTarget - baseModel.predict(trainInput)

# 添加到集成模型中
boostedModels.append((baseModel, 1.0))

# 更新训练目标
trainTarget = residuals

模型预测

predictions = np.zeros_like(testTarget)
for model, weight in boostedModels:
predictions += weight * model.predict(testInput)

计算均方误差

mse = mean_squared_error(testTarget, predictions)
print(“测试集均方误差：”, mse)