Scholarlink

<p>The combination of multiple machine learning algorithms can improve the performance of time series forecasting data in diversity and complexity. Based on conditional generative adversarial network (CGAN), one-dimensional convolutional neural network (Conv1D), and long short-term memory (LSTM), a hybrid forecasting model LSTM-Conv1D-CGAN and a novel performance evaluation system are proposed in this paper. Firstly, to capture the internal features, LSTM is used to improve the GAN generator. To extract the nonlinear features and dynamic behaviors, Conv1D is used to enhance the GAN discriminator. Meanwhile, Dropout, L2 regularization, and various activation function are used to optimize neural network performance. Secondly, a conditional label matrix is constructed, incorporating weather, electricity price, and historical data. Fourier feature transform is used to extract historical information, and a CGAN model is formed. Thirdly, performance of LSTM-Conv1D-CGAN is evaluated, focusing on prediction accuracy, forecasting stability, model generalization ability, and robustness. Experimental results show that RMSE, MAPE, MFE, <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="202_2025_3301_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>, and QL of LSTM-Conv1D-CGAN are 294.2773, 4.4078, −7.5069, 0.9486, and 117.5991, respectively. LSTM-Conv1D-CGAN outperforms LSTM, CNN, LSTM-GAN, CNN-GAN, and XGBoost. Furthermore, to evaluate the local prediction performance on a single time step, a two-dimensional K-means clustering approach based on dynamic programming is proposed, which builds an energy quantization system by labeling absolute percentage error and absolute error at multiple levels to quantify load energy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid time series forecasting CGAN model based on LSTM and Conv1D and its application for load forecasting

  • 期刊论文

  • 发表日期：

    2025年8月12日

查看全文
  • Jingyu Liu,
  • Yonghong Wu

摘要

The combination of multiple machine learning algorithms can improve the performance of time series forecasting data in diversity and complexity. Based on conditional generative adversarial network (CGAN), one-dimensional convolutional neural network (Conv1D), and long short-term memory (LSTM), a hybrid forecasting model LSTM-Conv1D-CGAN and a novel performance evaluation system are proposed in this paper. Firstly, to capture the internal features, LSTM is used to improve the GAN generator. To extract the nonlinear features and dynamic behaviors, Conv1D is used to enhance the GAN discriminator. Meanwhile, Dropout, L2 regularization, and various activation function are used to optimize neural network performance. Secondly, a conditional label matrix is constructed, incorporating weather, electricity price, and historical data. Fourier feature transform is used to extract historical information, and a CGAN model is formed. Thirdly, performance of LSTM-Conv1D-CGAN is evaluated, focusing on prediction accuracy, forecasting stability, model generalization ability, and robustness. Experimental results show that RMSE, MAPE, MFE, \(R^{2}\) R 2 , and QL of LSTM-Conv1D-CGAN are 294.2773, 4.4078, −7.5069, 0.9486, and 117.5991, respectively. LSTM-Conv1D-CGAN outperforms LSTM, CNN, LSTM-GAN, CNN-GAN, and XGBoost. Furthermore, to evaluate the local prediction performance on a single time step, a two-dimensional K-means clustering approach based on dynamic programming is proposed, which builds an energy quantization system by labeling absolute percentage error and absolute error at multiple levels to quantify load energy.