Simultaneous prediction for multiple source–loads based sliding time window and convolutional neural network

• Main Authors: Li, Q. (Author), Ouyang, H. (Author), Yang, T. (Author), Zhang, G. (Author), Zhang, L. (Author), Zhen, L. (Author)
• Format: Article
• Language: English
• Published: Elsevier Ltd 2022
• Subjects: Convolution; Convolutional neural network; Convolutional neural networks; Energy industry; Energy source; Extraction; Forecasting; Information filtering; Large scale data; Median filters; Multiple source; Multiple source–load; Multiple source–loads; Parallel strategies; Prediction; Sampling; Sliding time window; Sliding time windows; Time sequences; Time series analysis; Training sample
• Online Access: View Fulltext in Publisher

 Practical applications of perception, communication, computing, etc. in modern energy industry continually generate large scale data from diversified monitoring terminals. Such massive information has the characteristics like heterogeneity, time sequence, low value density, etc. To promptly obtain high quality data served for safe operation of power network, this paper developed a simultaneous prediction algorithm for multiple source–loads combing sliding time window with convolutional neural network (CNN). The contributions lie in such aspects as extraction of high-value training samples, construction of CNN catered for multiple source–loads and prediction efficiency. Firstly, after the incomplete and abnormal information of raw samples are processed through median filtering and interpolation, the correlation analysis of time series and sliding time window are employed to extract high-value training samples served for CNN. Also, a simultaneous prediction model suitable for multiple source–loads is constructed by modifying CNN. In addition, integrated with the characteristics of the proposed CNN architecture, the parallel strategies based on task parallel and data parallel are designed to achieve rapidly quality prediction for massive and heterogeneous energy sources and loads. Extensive experimental results demonstrate that the proposed algorithm can obtain the higher predicting accuracy simultaneously satisfying requirements of diversified energy source and loads. Loads of experimental comparisons in multicore chips show that the proposed parallel strategies can offer increased speedup under massive prediction. © 2022 The Authors