Deep Domain Adaption Transferable Model for Various Time Series Forecasting

Abstract

Time series forecasting (TSF) is one of the most important and frequent tasks in the process of building a smart factory regarding intelligent energy supply and scheduling, which has attracted massive attention. Recent boosting deep learning technology has been proposed and obtained great success in the domain of TSF. However, most of them under the assumption of having sufficient training data. Collecting all kinds of time series is very costly and even not available sometimes, which derives deep transfer learning (DTL). There are some challenging factors that highly limit the performance of TSF using DTL: (1) Different distribution patterns between source and target domains; (2) Time series’ complexity, randomness, and multi-dimension. Besides, most of the existing methods only focused on one or two types of forecasts, which cannot satisfy the actual needs of decision-making. To overcome the above issues, this thesis proposed a novel deep transferable model (DTM) for multiple TSFs, including short-term, medium-term, and long-term forecasts. In the proposed model, we adopted a convolutional neural network (CNN) to extract the rich hidden features from raw time series and frequency-domain features transformed by wavelet transformation (WT). CNN extracted features are fed into long short-term memory (LSTM) to mine the hidden features within a long-time dependency. Thus, the extracted features are the fusion of time-, frequency-domain features, and the features have a long-time dependency with low noise. Besides, correlation alignment (CORAL) is used to minimize the distribution discrepancy between source and target domains to achieve “transfer”. We designed, trained, and tested the proposed DTM on some public time-series data sets. The results indicated its state-of-the-art performance for multiple TSFs only using few training samples. With the proposed method, multiple future time series values could be obtained only with tiny training samples in self-domain. Even one is possible, which provides robust evidence to support decision-makers in measuring proper policy and actions to avoid unexpected things and implement smart factory.