✅博主简介：擅长数据搜集与处理、建模仿真、程序设计、仿真代码、论文写作与指导，毕业论文、期刊论文经验交流。
✅成品或者定制，扫描文章底部微信二维码。
(1) 基于机器学习的风电非平稳态时间序列数据清洗方法
风电功率数据具有强烈的非平稳特性,原始数据中包含大量异常值和缺失值,直接影响预测模型的训练效果和预测精度。本研究提出了一套完整的风电数据清洗流程,首先将风电历史功率数据、历史气象数据和风机运行状态数据进行多元融合,构建包含风速、风向、温度、湿度、气压等气象特征以及风机转速、桨距角、发电机温度等设备状态特征的综合数据集。针对异常值检测问题,采用孤立森林算法进行无监督异常检测,该算法通过随机选择特征和分割点构建隔离树,异常样本由于与正常样本的分布差异较大,在隔离树中更容易被分离出来,因此具有较短的平均路径长度。孤立森林算法不依赖于数据的分布假设,能够有效识别风电数据中由于传感器故障、数据传输错误等原因造成的异常点,并将其标记为缺失值待后续处理。对于缺失值填补,本研究采用基于门控循环单元插值网络构建的对抗生成网络模型,生成器网络学习数据的时序分布特征生成填补值,判别器网络判断填补后的数据序列是否符合真实数据的分布规律,两个网络通过对抗训练不断优化,最终生成与真实数据分布一致的填补值。实验表明该方法相比传统的均值填充和前后值填充方法能够更好地保留数据的时序特征和统计特性,为后续深度学习模型的训练提供高质量的数据基础。
(2) 基于时间滑动窗口的深度学习输入数据集构造算法
风电输出功率具有明显的时间周期性和自相关性,如何有效提取这些时序特征并构造适合深度学习模型的输入数据集是提高预测精度的关键。本研究提出了基于时间滑动窗口的数据集构造算法,该算法以固定长度的时间窗口在历史数据序列上滑动,每次滑动生成一个训练样本,样本的输入为窗口内的历史数据序列,输出为窗口后续时刻的功率值。时间窗口的长度根据风电功率的自相关分析结果确定,既要包含足够的历史信息以捕获功率变化的规律,又要避免引入过多的冗余信息增加模型的计算负担。滑动窗口的步长决定了相邻样本之间的重叠程度,较小的步长能够生成更多的训练样本,提高数据利用率,但也会增加样本之间的相关性。本研究通过交叉验证实验确定了最优的窗口长度和滑动步长,在扩展原始数据集规模的同时有效提取了风电功率的周期性特征。构造的数据集包含多个时间尺度的特征,短期特征反映功率的快速波动,长期特征反映功率的日周期和季节性变化规律。数据集经过归一化处理后送入深度神经网络进行训练,使网络能够学习到风电功率与气象因素、设备状态之间的复杂非线性映射关系,为实现高精度的功率预测奠定了数据基础。
(3) 基于长短时记忆网络与卷积神经网络融合的风电功率预测模型
本研究构建了两种深度学习预测模型用于风电短期功率预测。第一种模型为TLW-LSTM模型,采用时间滑动窗口构造的数据集作为输入,网络结构包含两层全连接层分别作为输入层和输出层,中间三层多节点长短时记忆层作为隐藏层。长短时记忆网络通过门控机制有效解决了传统循环神经网络在长序列训练中的梯度消失问题,能够捕获风电功率序列中的长期依赖关系。模型训练采用Nadam优化器,该优化器结合了动量加速和自适应学习率的优点,能够加快收敛速度并提高训练稳定性。为防止过拟合,网络中加入了Dropout正则化层,在训练过程中随机丢弃部分神经元,增强模型的泛化能力。实验结果表明TLW-LSTM模型的预测准确率达到百分之九十二点七,显著优于决策树、随机森林和支持向量机等传统机器学习方法。
import numpy as np import pandas as pd import torch import torch.nn as nn import torch.optim as optim from sklearn.ensemble import IsolationForest from sklearn.preprocessing import MinMaxScaler from torch.utils.data import DataLoader, TensorDataset class IsolationForestDetector: def __init__(self, contamination=0.1): self.model = IsolationForest(contamination=contamination, random_state=42) def fit_predict(self, data): predictions = self.model.fit_predict(data) anomaly_mask = predictions == -1 return anomaly_mask class GRUICell(nn.Module): def __init__(self, input_dim, hidden_dim): super(GRUICell, self).__init__() self.hidden_dim = hidden_dim self.update_gate = nn.Linear(input_dim + hidden_dim + 1, hidden_dim) self.reset_gate = nn.Linear(input_dim + hidden_dim + 1, hidden_dim) self.candidate = nn.Linear(input_dim + hidden_dim + 1, hidden_dim) self.decay = nn.Linear(1, hidden_dim) def forward(self, x, h, mask, delta): gamma = torch.exp(-torch.relu(self.decay(delta))) h = h * gamma combined = torch.cat([x * mask, h, delta], dim=1) z = torch.sigmoid(self.update_gate(combined)) r = torch.sigmoid(self.reset_gate(combined)) h_tilde = torch.tanh(self.candidate(torch.cat([x * mask, r * h, delta], dim=1))) h_new = (1 - z) * h + z * h_tilde return h_new class WGAN_Generator(nn.Module): def __init__(self, input_dim, hidden_dim, output_dim): super(WGAN_Generator, self).__init__() self.grui = GRUICell(input_dim, hidden_dim) self.fc = nn.Linear(hidden_dim, output_dim) def forward(self, x, mask, delta): batch_size, seq_len, _ = x.shape h = torch.zeros(batch_size, self.grui.hidden_dim) outputs = [] for t in range(seq_len): h = self.grui(x[:, t, :], h, mask[:, t, :], delta[:, t, :]) out = self.fc(h) outputs.append(out) return torch.stack(outputs, dim=1) class SlidingWindowDataset: def __init__(self, data, window_size, prediction_horizon): self.data = data self.window_size = window_size self.prediction_horizon = prediction_horizon def create_sequences(self): X, y = [], [] for i in range(len(self.data) - self.window_size - self.prediction_horizon + 1): X.append(self.data[i:i + self.window_size]) y.append(self.data[i + self.window_size:i + self.window_size + self.prediction_horizon, 0]) return np.array(X), np.array(y) class TLW_LSTM(nn.Module): def __init__(self, input_dim, hidden_dim, num_layers, output_dim, dropout=0.2): super(TLW_LSTM, self).__init__() self.input_fc = nn.Linear(input_dim, hidden_dim) self.lstm1 = nn.LSTM(hidden_dim, hidden_dim, batch_first=True, dropout=dropout) self.lstm2 = nn.LSTM(hidden_dim, hidden_dim, batch_first=True, dropout=dropout) self.lstm3 = nn.LSTM(hidden_dim, hidden_dim, batch_first=True, dropout=dropout) self.output_fc = nn.Linear(hidden_dim, output_dim) self.dropout = nn.Dropout(dropout) def forward(self, x): x = torch.relu(self.input_fc(x)) x, _ = self.lstm1(x) x = self.dropout(x) x, _ = self.lstm2(x) x = self.dropout(x) x, _ = self.lstm3(x) out = self.output_fc(x[:, -1, :]) return out class LW_CLSTM(nn.Module): def __init__(self, input_dim, hidden_dim, num_filters, kernel_size, output_dim, dropout=0.2): super(LW_CLSTM, self).__init__() self.conv1 = nn.Conv1d(input_dim, num_filters, kernel_size, padding=kernel_size//2) self.conv2 = nn.Conv1d(num_filters, num_filters, kernel_size, padding=kernel_size//2) self.pool = nn.MaxPool1d(2) self.lstm = nn.LSTM(num_filters, hidden_dim, num_layers=2, batch_first=True, dropout=dropout) self.fc = nn.Linear(hidden_dim, output_dim) self.dropout = nn.Dropout(dropout) def forward(self, x): x = x.permute(0, 2, 1) x = torch.relu(self.conv1(x)) x = torch.relu(self.conv2(x)) x = self.pool(x) x = x.permute(0, 2, 1) x, _ = self.lstm(x) x = self.dropout(x[:, -1, :]) out = self.fc(x) return out def train_model(model, train_loader, val_loader, epochs, learning_rate): optimizer = optim.NAdam(model.parameters(), lr=learning_rate) criterion = nn.MSELoss() best_val_loss = float('inf') for epoch in range(epochs): model.train() train_loss = 0 for batch_x, batch_y in train_loader: optimizer.zero_grad() predictions = model(batch_x) loss = criterion(predictions, batch_y) loss.backward() optimizer.step() train_loss += loss.item() model.eval() val_loss = 0 with torch.no_grad(): for batch_x, batch_y in val_loader: predictions = model(batch_x) loss = criterion(predictions, batch_y) val_loss += loss.item() if val_loss < best_val_loss: best_val_loss = val_loss torch.save(model.state_dict(), 'best_model.pth') return model def calculate_metrics(predictions, targets): mae = np.mean(np.abs(predictions - targets)) rmse = np.sqrt(np.mean((predictions - targets) ** 2)) mape = np.mean(np.abs((predictions - targets) / (targets + 1e-8))) * 100 accuracy = 1 - mae / np.mean(np.abs(targets)) return {'MAE': mae, 'RMSE': rmse, 'MAPE': mape, 'Accuracy': accuracy} if __name__ == "__main__": np.random.seed(42) raw_data = np.random.randn(10000, 10) detector = IsolationForestDetector(contamination=0.05) anomalies = detector.fit_predict(raw_data) scaler = MinMaxScaler() cleaned_data = scaler.fit_transform(raw_data) dataset = SlidingWindowDataset(cleaned_data, window_size=24, prediction_horizon=6) X, y = dataset.create_sequences() train_size = int(len(X) * 0.8) y_train, y_val = y[:train_size], y[train_size:] train_dataset = TensorDataset(torch.FloatTensor(X_train), torch.FloatTensor(y_train)) val_dataset = TensorDataset(torch.FloatTensor(X_val), torch.FloatTensor(y_val)) train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True) val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False) tlw_lstm = TLW_LSTM(input_dim=10, hidden_dim=128, num_layers=3, output_dim=6) lw_clstm = LW_CLSTM(input_dim=10, hidden_dim=128, num_filters=64, kernel_size=3, output_dim=6)如有问题，可以直接沟通
👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇👇
)
)
)
