A Deep Neural Network for Anomaly Detection and Forecasting for Multivariate Time Series in Smart City

In the progress of constructing a smart city, large amounts of univariate and multivariate times series data is generated by complex real-world systems and internet of things(IoT) with sensors such as wearable devices. Abnormal status in univariate and multivariate time series are necessary to be identified by abnormal detection methods. Time series forecasting in univariate and multivariate which refers to detect the different patterns in the input time series is significant for managers. However, building a system for anomaly detection and forecasting is challenging. On the one hand, the temporal dependency is required to capture in time series data, on the other hand, inter-correlations in different pairs of time series data are so important for the system that the system needs to encode inter-correlations. In this work, we propose an Attention based Convolutional Recurrent Encoder-Decoder (ACRED), which is effective to address anomaly detection and forecasting problems in time series. The studies based on a Secure Water Treatment tested (SWaT) dataset suggest that ACRED can outperform popular deep recurrent neural network methods.