Research Of Meta-learning-based Few-shot Time Series Forecasting Approach

Time-series forecasting is a crucial research subject,which can widely be applicable to finance field,health-care,meteorology,transportation,power system and etc.Therefore,the study and exploration for time-series forecasting is always a concerned topic by researchers.To begin with quantities of works is about the statistics-based time-series forecasting approaches,and to the early of 21th century,the classical machinelearning-based approaches gradually were used to conduct time-series analysis,and later,with the development of deep learning technique and the advance of computing power,using DNNs-based models to cope with time-series forecasting has become a popular research topic.However,when it comes to few-shot time-series forecasting scenarios,such as the auxiliary diagnosis of medical records of rare diseases and electric power forecasting of new customers,etc.,due to sufficient training data may be unavailable in such scenarios,which causes that DNNs-based models suffer from over-fitting problem during training phase and further meet with serious performance degradation during predicting phase.In light of aforementioned issues,the meta-learning-based approach is proposed to handle with few-shot time-series forecasting,in which the most key idea is to generalize meta-knowledge and improve model's sensitivity for new forecasting task by fast cross-tasks training,and further to alleviate over-fitting problem of DNNs-based models under few-shot scenarios and improve models' predict accuracy during the forecasting phase.The experimental data consists of one electric boiler with heat reservoir's power dataset from Jilin province and 13 public time-series datasets from UCR Time Series Archive,and in this work,MLP(MultiLayer Perceptron),CNN(Convolutional Neural Networks),LSTM(Long Short-Term Memory networks),CCL(CNN concatenating LSTM)are selected as baseline model to conduct experiments about few-shot time series forecasting.According to experiment settings,we also design two groups of comparison models.Extensive experimental results indicate that meta-learning-based time-series forecasting approach outperforms the baseline models from forecasting performance and convergence speed,and effectively alleviates over-fitting problem of DNNs-based models facing few-shot time-series forecasting scenarios.In addition,by statistical analysis and sensitivity analysis,the models trained by meta-learningbased few-shot time-series forecasting approach have better generalization performance from the statistics perspective,and the proposed approach also has strong robustness for forecasting horizons and data scales.