前言

系列专栏:【深度学习：算法项目实战】✨︎
涉及医疗健康、财经金融、商业零售、食品饮料、运动健身、交通运输、环境科学、社交媒体以及文本和图像处理等诸多领域，讨论了各种复杂的深度神经网络思想，如卷积神经网络、循环神经网络、生成对抗网络、门控循环单元、长短期记忆、自然语言处理、深度强化学习、大型语言模型和迁移学习。

随着全球能源结构的不断转型和可持续发展战略的深入实施，风能作为一种清洁、可再生的能源，其在全球能源供应中的地位日益凸显。风力涡轮机作为风能利用的主要设备，其发电量的准确预测对于电力系统的稳定运行、优化调度以及风电的并网消纳具有重要意义。然而，由于风能的间歇性和不稳定性，风力涡轮机的发电量预测成为了一个极具挑战性的任务。

近年来，深度学习技术在时间序列预测领域展现出了强大的潜力，特别是循环神经网络（RNN）及其变体，如长短期记忆网络（LSTM）和门控循环单元（GRU），在处理序列数据方面表现出色。双向门控循环单元（BiGRU）作为GRU的一种扩展，通过同时考虑序列数据中的过去和未来信息，进一步提高了模型捕捉长期依赖关系的能力。此外，注意力机制（Attention）的引入，使得模型能够动态地调整不同时间步长输入特征的权重，从而更加关注对预测结果影响较大的关键信息。

基于上述背景，本文提出了基于BiGRU+Attention的风力涡轮机发电量时序预测模型，旨在通过结合BiGRU和Attention机制的优势，提高风力涡轮机发电量预测的准确性和效率。该模型不仅能够捕捉风力发电量数据中的时序依赖关系，还能通过注意力机制关注对预测结果更为重要的输入特征，从而在复杂多变的风电环境中实现高精度的预测。

本文首先介绍了研究背景与意义，概述了BiGRU和Attention机制的基本原理及其在风力涡轮机发电量预测中的应用潜力。随后，详细描述了模型的构建与训练过程，包括数据收集与预处理、特征提取、时间序列建模、预测输出以及模型训练与性能评估等关键步骤。通过实际案例分析和实验结果展示，验证了该模型在风力涡轮机发电量预测中的有效性和优越性。

1. 数据集介绍

在风力涡轮机中，Scada 系统测量并保存风速、风向、发电量等数据，时间间隔为 10 分钟。该文件取自土耳其一台正在发电的风力涡轮机的 scada 系统。原数据英文及解释如下：

• Date/Time (for 10 minutes intervals)
• LV ActivePower (kW): The power generated by the turbine for that moment
• Wind Speed (m/s): The wind speed at the hub height of the turbine (the wind speed that turbine use for electricity generation)
• Theoretical_Power_Curve (KWh): The theoretical power values that the turbine generates with that wind speed which is given by the turbine manufacturer
• Wind Direction (°): The wind direction at the hub height of the turbine (wind turbines turn to this direction automaticly)

---

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as snsfrom sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_absolute_error, \mean_absolute_percentage_error, \mean_squared_error, root_mean_squared_error, \r2_scoreimport torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import TensorDataset, DataLoader, Dataset
from torchinfo import summarynp.random.seed(0)

data = pd.read_csv("T1.csv")
data.head(10).style.background_gradient(cmap='bone')

2. 数据预处理

现在，我们使用 pandas.to_datetime 函数，将日期解析为时间数据类型

print(type(data['LV ActivePower (kW)'].iloc[0]),type(data['Date/Time'].iloc[0]))# Let's convert the data type of timestamp column to datatime format
data['Date/Time'] = pd.to_datetime(data['Date/Time'],format='%d %m %Y %H:%M')
print(type(data['LV ActivePower (kW)'].iloc[0]),type(data['Date/Time'].iloc[0]))cond_1 = data['Date/Time'] >= '2018-01-01 00:00:00'
cond_2 = data['Date/Time'] <= '2018-01-07 23:59:59'
data = data[cond_1 & cond_2]
print(data.shape)

<class 'numpy.float64'> <class 'str'>
<class 'numpy.float64'> <class 'pandas._libs.tslibs.timestamps.Timestamp'>
(987, 5)

3. 数据可视化

3.1 数据相关性

使用 data.corr() 计算相关系数矩阵

correlation = data.corr()
print(correlation['LV ActivePower (kW)'].sort_values(ascending=False))

LV ActivePower (kW)              1.000000
Theoretical_Power_Curve (KWh)    0.996930
Wind Speed (m/s)                 0.965048
Wind Direction (°)               0.217183
Date/Time                       -0.028100
Name: LV ActivePower (kW), dtype: float64

plt.figure(figsize=(10, 8))  # 设置图片大小
sns.heatmap(correlation, annot=True, cmap=sns.cubehelix_palette(dark=.20, light=.95, as_cmap=True), linewidths=0.5)
plt.show()

很明显，LV ActivePower (kW) 与 Theoretical_Power_Curve (KWh)、Wind Speed (m/s) 存在强相关性

4. 特征工程

4.1 特征缩放（归一化）

StandardScaler()函数将数据的特征值转换为符合正态分布的形式，它将数据缩放到均值为0，‌标准差为1的区间‌。在机器学习中，StandardScaler()函数常用于不同尺度特征数据的标准化，以提高模型的泛化能力。

# 使用选定的特征来训练模型
features = data.drop('Date/Time', axis=1)
target = data['LV ActivePower (kW)'].values.reshape(-1, 1)

# 创建 StandardScaler实例，对特征进行拟合和变换，生成NumPy数组
scaler = StandardScaler()
features_scaled = scaler.fit_transform(features)
target_scaled = scaler.fit_transform(target)
print(features_scaled.shape, target_scaled.shape)

4.2 构建时间序列数据

我们创建一个时间序列数据，时间步 time_steps 假设设置为10

time_steps = 10
X_list = []
y_list = []for i in range(len(features_scaled) - time_steps):X_list.append(features_scaled[i:i+time_steps])y_list.append(target_scaled[i+time_steps])X = np.array(X_list) # [samples, time_steps, num_features]
y = np.array(y_list) # [target]

samples, time_steps, num_features = X.shape

4.3 数据集划分

train_test_split 函数将数组或矩阵随机分成训练子集和测试子集。

X_train, X_valid,\y_train, y_valid = train_test_split(X, y, test_size=0.2, random_state=45,shuffle=False)
print(X_train.shape, X_valid.shape, y_train.shape, y_valid.shape)

以上代码中 random_state=45 设置了随机种子，以确保每次运行代码时分割结果的一致性。shuffle=False 表示在分割数据时不进行随机打乱。如果设置为True（默认值），则会在分割之前对数据进行随机打乱，这样可以增加数据的随机性，但时间序列数据具有连续性，所以设置为False。

4.4 数据集张量

# 将 NumPy数组转换为 tensor张量
X_train_tensor = torch.from_numpy(X_train).type(torch.Tensor)
X_valid_tensor = torch.from_numpy(X_valid).type(torch.Tensor)
y_train_tensor = torch.from_numpy(y_train).type(torch.Tensor).view(-1, 1)
y_valid_tensor = torch.from_numpy(y_valid).type(torch.Tensor).view(-1, 1)print(X_train_tensor.shape, X_valid_tensor.shape, y_train_tensor.shape, y_valid_tensor.shape)

以上代码通过 train_test_split 划分得到的训练集和验证集中的特征数据 X_train、X_valid 以及标签数据 y_train、y_valid 从 numpy 数组转换为 PyTorch 的张量（tensor）类型。.type(torch.Tensor) 确保张量的数据类型为标准的 torch.Tensor 类型，.view(-1, 1) 对张量进行维度调整

class DataHandler(Dataset):def __init__(self, X_train_tensor, y_train_tensor, X_valid_tensor, y_valid_tensor):self.X_train_tensor = X_train_tensorself.y_train_tensor = y_train_tensorself.X_valid_tensor = X_valid_tensorself.y_valid_tensor = y_valid_tensordef __len__(self):return len(self.X_train_tensor)def __getitem__(self, idx):sample = self.X_train_tensor[idx]labels = self.y_train_tensor[idx]return sample, labelsdef train_loader(self):train_dataset = TensorDataset(self.X_train_tensor, self.y_train_tensor)return DataLoader(train_dataset, batch_size=32, shuffle=True)def valid_loader(self):valid_dataset = TensorDataset(self.X_valid_tensor, self.y_valid_tensor)return DataLoader(valid_dataset, batch_size=32, shuffle=False)

在上述代码中，定义了一个名为 TSCDataset 的类，它继承自 torch.utils.data.Dataset
__init__ 方法用于接收数据和标签。__len__ 方法返回数据集的长度。__getitem__ 方法根据给定的索引 idx 返回相应的数据样本和标签。

data_handler = DataHandler(X_train_tensor, y_train_tensor, X_valid_tensor, y_valid_tensor)
train_loader = data_handler.train_loader()
valid_loader = data_handler.valid_loader()

5. 构建时序模型（TSF）

5.1 构建BiGRU_Attention模型

class BiGRU_Attention(nn.Module):def __init__(self, input_dim, hidden_dim, num_layers, output_dim, num_heads):super(BiGRU_Attention, self).__init__()self.hidden_dim = hidden_dimself.num_layers = num_layersself.bigru = nn.GRU(input_dim, hidden_dim, num_layers=num_layers, batch_first=True, bidirectional=True)self.attention = nn.MultiheadAttention(embed_dim=hidden_dim*2, num_heads=num_heads, batch_first=True)self.dropout = nn.Dropout(p=0.5)self.fc = nn.Linear(hidden_dim*2, output_dim)def forward(self, x):# x: [batch_size, seq_len, input_dim]# h0: [num_layers * num_directions, batch_size, hidden_dim]h0 = torch.zeros(self.num_layers*2, x.size(0), self.hidden_dim).to(x.device)out, _ = self.bigru(x, h0)drop_gruout = self.dropout(out)# 注意力层attn_output, _ = self.attention(drop_gruout, drop_gruout, drop_gruout)drop_output = self.dropout(attn_output[:, -1, :])# 全连接层out = self.fc(drop_output)return out

5.2 定义模型、损失函数与优化器

model = BiGRU_Attention(input_dim = num_features,hidden_dim = 8,num_layers = 2, output_dim = 1, num_heads = 16)
criterion_mse = nn.MSELoss()  # 定义均方误差损失函数
criterion_mae = nn.L1Loss()  # 定义平均绝对误差损失
optimizer = torch.optim.Adam(model.parameters(), lr=1e-04) # 定义优化器

nn.L1Loss() 是 PyTorch 中的一个损失函数，用于计算模型预测值和真实值之间的平均绝对误差（Mean Absolute Error, MAE）

5.3 模型概要

# batch_size, seq_len(time_steps), input_size(in_channels)
summary(model, (32, time_steps, num_features)) 

==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
BiGRU_Attention                          [32, 1]                   --
├─GRU: 1-1                               [32, 10, 16]              1,920
├─Dropout: 1-2                           [32, 10, 16]              --
├─MultiheadAttention: 1-3                [32, 10, 16]              1,088
├─Dropout: 1-4                           [32, 16]                  --
├─Linear: 1-5                            [32, 1]                   17
==========================================================================================
Total params: 3,025
Trainable params: 3,025
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 0.61
==========================================================================================
Input size (MB): 0.01
Forward/backward pass size (MB): 0.04
Params size (MB): 0.01
Estimated Total Size (MB): 0.05
==========================================================================================

6. 模型训练与可视化

6.1 定义训练与评估函数

def train(model, iterator, optimizer):epoch_loss_mse = 0epoch_loss_mae = 0model.train()  # 确保模型处于训练模式for batch in iterator:optimizer.zero_grad()  # 清空梯度inputs, targets = batch  # 获取输入和目标值outputs = model(inputs)  # 前向传播loss_mse = criterion_mse(outputs, targets)  # 计算损失loss_mae = criterion_mae(outputs, targets)combined_loss = loss_mse + loss_mae  # 可以根据需要调整两者的权重combined_loss.backward()optimizer.step()epoch_loss_mse += loss_mse.item()  # 累计损失epoch_loss_mae += loss_mae.item()average_loss_mse = epoch_loss_mse / len(iterator)  # 计算平均损失average_loss_mae = epoch_loss_mae / len(iterator)return average_loss_mse, average_loss_mae

上述代码定义了一个名为 train 的函数，用于训练给定的模型。它接收模型、数据迭代器、优化器作为参数，并返回训练过程中的平均损失。

def evaluate(model, iterator):epoch_loss_mse = 0epoch_loss_mae = 0model.eval()  # 将模型设置为评估模式，例如关闭 Dropout 等with torch.no_grad():  # 不需要计算梯度for batch in iterator:inputs, targets = batchoutputs = model(inputs)  # 前向传播loss_mse = criterion_mse(outputs, targets)  # 计算损失loss_mae = criterion_mae(outputs, targets)epoch_loss_mse += loss_mse.item()  # 累计损失epoch_loss_mae += loss_mae.item()return epoch_loss_mse / len(iterator), epoch_loss_mae / len(iterator)

上述代码定义了一个名为 evaluate 的函数，用于评估给定模型在给定数据迭代器上的性能。它接收模型、数据迭代器作为参数，并返回评估过程中的平均损失。这个函数通常在模型训练的过程中定期被调用，以监控模型在验证集或测试集上的性能。通过评估模型的性能，可以了解模型的泛化能力和训练的进展情况。

epoch = 500
train_mselosses = []
valid_mselosses = []
train_maelosses = []
valid_maelosses = []for epoch in range(epoch):train_loss_mse, train_loss_mae = train(model, train_loader, optimizer)valid_loss_mse, valid_loss_mae = evaluate(model, valid_loader)train_mselosses.append(train_loss_mse)valid_mselosses.append(valid_loss_mse)train_maelosses.append(train_loss_mae)valid_maelosses.append(valid_loss_mae)print(f'Epoch: {epoch+1:02}, Train MSELoss: {train_loss_mse:.5f}, Train MAELoss: {train_loss_mae:.3f}, Val. MSELoss: {valid_loss_mse:.5f}, Val. MAELoss: {valid_loss_mae:.3f}')

Epoch: 01, Train MSELoss: 0.86945, Train MAELoss: 0.852, Val. MSELoss: 1.08921, Val. MAELoss: 0.997
Epoch: 02, Train MSELoss: 0.82845, Train MAELoss: 0.827, Val. MSELoss: 1.07914, Val. MAELoss: 0.992
Epoch: 03, Train MSELoss: 0.77970, Train MAELoss: 0.800, Val. MSELoss: 1.06364, Val. MAELoss: 0.984
Epoch: 04, Train MSELoss: 0.72964, Train MAELoss: 0.770, Val. MSELoss: 1.03752, Val. MAELoss: 0.971
Epoch: 05, Train MSELoss: 0.67883, Train MAELoss: 0.740, Val. MSELoss: 1.00662, Val. MAELoss: 0.954
******
Epoch: 496, Train MSELoss: 0.08508, Train MAELoss: 0.223, Val. MSELoss: 0.05806, Val. MAELoss: 0.203
Epoch: 497, Train MSELoss: 0.08262, Train MAELoss: 0.218, Val. MSELoss: 0.05554, Val. MAELoss: 0.195
Epoch: 498, Train MSELoss: 0.08714, Train MAELoss: 0.220, Val. MSELoss: 0.04519, Val. MAELoss: 0.168
Epoch: 499, Train MSELoss: 0.09767, Train MAELoss: 0.239, Val. MSELoss: 0.05352, Val. MAELoss: 0.191
Epoch: 500, Train MSELoss: 0.08566, Train MAELoss: 0.220, Val. MSELoss: 0.05005, Val. MAELoss: 0.183

6.2 绘制训练与验证损失曲线

# 绘制 MSE损失图
plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
plt.plot(train_mselosses, label='Train MSELoss')
plt.plot(valid_mselosses, label='Validation MSELoss')
plt.xlabel('Epoch')
plt.ylabel('MSELoss')
plt.title('Train and Validation MSELoss')
plt.legend()
plt.grid(True)# 绘制 MAE损失图
plt.subplot(1, 2, 2)
plt.plot(train_maelosses, label='Train MAELoss')
plt.plot(valid_maelosses, label='Validation MAELoss')
plt.xlabel('Epoch')
plt.ylabel('MAELoss')
plt.title('Train and Validation MAELoss')
plt.legend()
plt.grid(True)plt.show()

7. 模型评估与可视化

7.1 构建预测函数

定义预测函数prediction 方便调用

# 定义 prediction函数
def prediction(model, iterator): all_targets = []all_predictions = []model.eval()with torch.no_grad():for batch in iterator:inputs, targets = batchpredictions = model(inputs)all_targets.extend(targets.numpy())all_predictions.extend(predictions.numpy())return all_targets, all_predictions

7.2 验证集预测

# 模型预测
targets, predictions = prediction(model, valid_loader)

denormalized_targets = scaler.inverse_transform(targets)
denormalized_predictions = scaler.inverse_transform(predictions)

targets 是经过标准化处理后的目标值数组，predictions 是经过标准化处理后的预测值数组。scaler 是StandardScaler() 标准化类的实例，inverse_transform 方法会将标准化后的数组还原为原始数据的尺度，即对预测值进行反标准化操作。

# Visualize the data
plt.figure(figsize=(12,6))
plt.style.use('_mpl-gallery')
plt.title('Comparison of validation set prediction results')
plt.plot(denormalized_targets, color='blue',label='Actual Value')
plt.plot(denormalized_predictions, color='orange', label='Valid Value')
plt.legend()
plt.show()

7.3 回归拟合图

plt.figure(figsize=(5, 5), dpi=100)
sns.regplot(x=denormalized_targets, y=denormalized_predictions, scatter=True, marker="x", color='blue',line_kws={'color': 'red'})
plt.show()

7.4 评估指标

这里我们将通过调用 sklearn.metrics 模块中的 mean_absolute_error mean_absolute_percentage_error mean_squared_error root_mean_squared_error r2_score 函数来评估模型性能

mae = mean_absolute_error(targets, predictions)
print(f"MAE: {mae:.4f}")mape = mean_absolute_percentage_error(targets, predictions)
print(f"MAPE: {mape * 100:.4f}%")mse = mean_squared_error(targets, predictions)
print(f"MSE: {mse:.4f}")rmse = root_mean_squared_error(targets, predictions)
print(f"RMSE: {rmse:.4f}")r2 = r2_score(targets, predictions)
print(f"R²: {r2:.4f}")

MAE: 0.1893
MAPE: 58.1806%
MSE: 0.0519
RMSE: 0.2278
R²: 0.9173