Memory-Augmented Short Time Series Forecasting

Time series forecasting is widely applied in many applications such as finance, healthcare, and transportation. Despite its broad use, real-world scenarios often present the challenge of cold start problems. For example, newly built wind farms face challenges in forecasting wind speed due to insufficient historical observation records. Cold start problems make it challenging for the current time series forecasting solutions to perform effectively. Therefore, addressing the issue and achieving high-quality time series predictions with limited data (i.e., short time series data) is crucial. In response to these challenges, we propose a novel Memory-Augmented Short Time Series Forecasting model (MEMSTSF). Specifically, we design a novel Transformer architecture with a memory module to not only capture temporal dependencies within the sequence, but also leverage auxiliary information from multiple data sources as memory to assist short time series in learning intricate temporal patterns. Furthermore, the novel Transformer architecture features a built-in time series decomposition module, which captures the global properties of the target time series, enhancing the comprehensiveness and overall predictive performance of MEMSTSF. Extensive experiments on three datasets demonstrate the superiority of our model in forecasting short time series with insufficient historical data.