Scholarlink

<p>Nowadays, the Internet of Things (IoT) sensing technology has been widely applied in fields such as meteorology, agriculture, and industry, enabling real-time monitoring and accumulation of massive temporal data. However, the dynamic data collected by sensors often exhibit strong randomness, significant nonlinearity, and obvious volatility, which greatly affect their application value in decision-making. Moreover, single models have limited prediction accuracy when dealing with such multi-sensor, high-frequency sampled data, making it difficult to meet the computational requirements and decision-making needs of application scenarios. Therefore, this study proposes a forecasting method based on feature selection and a hybrid model, validated using multivariate wind speed sensing data from a complex mountainous area. Firstly, Pearson correlation coefficient and Least Absolute Shrinkage and Selection Operator (LASSO) regression are employed to quantify the linear relationships between features and the target variable. A weighted Recursive Feature Elimination (RFE) algorithm, integrated with multiple nonlinear models, is utilized to analyze nonlinear relationships. The input features of the prediction model are determined through this decomposition-combination strategy. Subsequently, a hybrid model, KF-CNN-LSTM, is proposed. This model integrates the capability of Kalman Filter (KF) to reduce dynamic data errors, the advantage of Convolutional Neural Network (CNN) in efficiently extracting spatial features, and the memory capacity of Long Short-Term Memory (LSTM) network to capture temporal dependencies in sequence data. Meanwhile, the recursive nature of KF eliminates the need for storing large amounts of historical data, and its collaboration with the computational units of CNN and LSTM not only meets the computational requirements of multivariate sensing data but also enhances the extraction of data features. Experimental results demonstrate that the proposed KF-CNN-LSTM hybrid model significantly outperforms other benchmark models in prediction accuracy. Compared with the relatively superior CNN-LSTM model, the KF-CNN-LSTM model achieves reductions of 43.94%, 26.52%, 25.15%, and 19.43% in Mean Square Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), respectively.</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 short-time wind speed forecasting method based on feature selection and KF-CNN-LSTM hybrid model

  • 期刊论文

  • 发表日期：

    2025年10月4日

查看全文
  • Xu He,
  • Jiale Wang

摘要

Nowadays, the Internet of Things (IoT) sensing technology has been widely applied in fields such as meteorology, agriculture, and industry, enabling real-time monitoring and accumulation of massive temporal data. However, the dynamic data collected by sensors often exhibit strong randomness, significant nonlinearity, and obvious volatility, which greatly affect their application value in decision-making. Moreover, single models have limited prediction accuracy when dealing with such multi-sensor, high-frequency sampled data, making it difficult to meet the computational requirements and decision-making needs of application scenarios. Therefore, this study proposes a forecasting method based on feature selection and a hybrid model, validated using multivariate wind speed sensing data from a complex mountainous area. Firstly, Pearson correlation coefficient and Least Absolute Shrinkage and Selection Operator (LASSO) regression are employed to quantify the linear relationships between features and the target variable. A weighted Recursive Feature Elimination (RFE) algorithm, integrated with multiple nonlinear models, is utilized to analyze nonlinear relationships. The input features of the prediction model are determined through this decomposition-combination strategy. Subsequently, a hybrid model, KF-CNN-LSTM, is proposed. This model integrates the capability of Kalman Filter (KF) to reduce dynamic data errors, the advantage of Convolutional Neural Network (CNN) in efficiently extracting spatial features, and the memory capacity of Long Short-Term Memory (LSTM) network to capture temporal dependencies in sequence data. Meanwhile, the recursive nature of KF eliminates the need for storing large amounts of historical data, and its collaboration with the computational units of CNN and LSTM not only meets the computational requirements of multivariate sensing data but also enhances the extraction of data features. Experimental results demonstrate that the proposed KF-CNN-LSTM hybrid model significantly outperforms other benchmark models in prediction accuracy. Compared with the relatively superior CNN-LSTM model, the KF-CNN-LSTM model achieves reductions of 43.94%, 26.52%, 25.15%, and 19.43% in Mean Square Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), respectively.