Abstract:<p>Multi-variate time series prediction models use the historical information of multiple exogenous series to predict the future values of the target series. At present, attention-based deep networks can obtain the spatial correlations between target series and multiple exogenous series, but it is difficult to capture temporal correlations across multiple time steps, which play a role in improving prediction accuracy. To that end, we propose a correlational graph attention-based Long Short-Term Memory network (CGA-LSTM), a nested network that nests the correlational attention mechanism in the graph attention mechanism to strengthen the spatio-temporal correlations. We construct the time series as a graph structure, where nodes represent time steps in exogenous series. To obtain sufficient expressive power, we propose a nonlinear transformation, correlational attention-based LSTM, instead of the original linear transformation to transform the exogenous series into higher-level features. The original linear transformations cannot obtain spatial correlations. The correlational attention mechanism can adaptively select the relevant exogenous series to obtain the spatial correlations. Then calculating the weight coefficients between the node and its neighbors to capture the temporal correlations. The performance of the proposed algorithm was tested on 4 datasets and compared with state-of-the-art methods. The experimental results show that our model is effective, can provide higher prediction accuracy.</p>

What problem does this paper attempt to address?

The paper attempts to address two main issues in multivariate time series forecasting: 1. **Spatial Correlation**: How to capture highly dynamic, time-varying spatial correlations between the target sequence and multiple exogenous sequences. 2. **Temporal Correlation**: How to capture temporal correlations at different time steps. Specifically, multivariate time series data include the values of the target sequence and multiple exogenous sequences, which are widely present in scenarios such as stock data in financial markets, photovoltaic power data from power companies, and air quality and water quality data from urban sensors. The forecasting task is very challenging, mainly influenced by the following two factors: - **Spatial Correlation**: The spatial correlation between the target sequence and multiple exogenous sequences is highly dynamic and changes over time. - **Temporal Correlation**: Temporal correlations at different time steps play an important role in improving forecasting accuracy. For example, solar radiation, humidity, and temperature affect photovoltaic power generation. At night, solar radiation is zero, and no power is generated; as the sun rises, photovoltaic power first increases and then decreases, but sometimes the trend of photovoltaic power differs from solar radiation due to the influence of temperature and humidity. The importance of each exogenous sequence to the target sequence is also different, and most time series data are collected by sensors, which are easily affected by the surrounding environment, so detecting useful exogenous sequences is crucial for forecasting. Traditional Autoregressive Integrated Moving Average (ARIMA) models and Vector Autoregression (VAR) models have limitations in handling multivariate time series data and cannot effectively model spatiotemporal correlations. Therefore, the authors propose a Correlation Graph Attention-based Long Short-Term Memory network (CGA-LSTM), aiming to enhance the ability to capture spatiotemporal correlations through a nested correlation attention mechanism.

Correlational graph attention-based Long Short-Term Memory network for multivariate time series prediction

A Joint Network of Non-Linear Graph Attention and Temporal Attraction Force for Geo-Sensory Time Series Prediction

Temporal Self-Attention-based Conv-LSTM Network for Multivariate Time Series Prediction

TVGCN: Time-varying graph convolutional networks for multivariate and multifeature spatiotemporal series prediction

Multivariate sequence prediction for graph convolutional networks based on ESMD and transfer entropy

AFSTGCN: Prediction for multivariate time series using an adaptive fused spatial-temporal graph convolutional network

A dual‐stage attention‐based Conv‐LSTM network for spatio‐temporal correlation and multivariate time series prediction

Multistage attention network for multivariate time series prediction

Spatiotemporal Self-Attention-Based LSTNet for Multivariate Time Series Prediction

A Temporal Window Attention-Based Window-Dependent Long Short-Term Memory Network for Multivariate Time Series Prediction.

TDG4MSF: A temporal decomposition enhanced graph neural network for multivariate time series forecasting

MGCP: A Multi-Grained Correlation based Prediction Network for Multivariate Time Series

Capturing Spatial-Temporal Correlations with Attention Based Graph Convolutional Networks for Network Traffic Prediction

Graph Attention Recurrent Neural Networks for Correlated Time Series Forecasting

CTS-LSTM: LSTM-based Neural Networks for Correlatedtime Series Prediction

Graph Attention Recurrent Neural Networks for Correlated Time Series Forecasting - Full version

Multivariate long sequence time-series forecasting using dynamic graph learning

Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks

Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting

Learning Time-aware Graph Structures for Spatially Correlated Time Series Forecasting

A Recurrent Spatio-Temporal Graph Neural Network Based on Latent Time Graph for Multi-Channel Time Series Forecasting