Spatial-Temporal Correlation Neural Network for Long Short-Term Demand Forecasting During COVID-19

Demand forecasting is an important method for dealing with the supply-demand relationship in social resource management. The demands of daily life discussed in this study are mainly about hotels, restaurants, gas stations, drugstores, shopping malls, etc. Accurate demand forecasting can help enterprises meet people’s needs properly, especially since COVID-19 has spread worldwide and significantly changed people’s lives. Previous studies have proven the effectiveness of statistical and deep learning models in demand forecasting, including the Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), and their variants. These methods focus only on one-step prediction of one demand and do not perform well in long-term or multi-step predictions. To address this issue, we propose a framework called the Deep Spatial-Temporal Neural Network (DSTNN) for multi-demand forecasting, which performs well in long short-term forecasting. The DSTNN adopts a transformer encoder-based structure and establishes a direct correlation between the demand distribution and spatiotemporal features. The transformer encoder is designed to learn high-order feature interactions. An auxiliary task of pandemic trend forecasting was used to help the main task learn about pandemic trends. As regional mobility affects the spread of COVID-19, a population mobility graph was designed to achieve a better spatial feature representation. The DSTNN eliminates the dependency on historical sequential data, which helps the model to perform well in long-term forecasting, and its performance does not obviously decline as the prediction period increases. The experimental results show that the DSTNN performs much better than conventional models, particularly in terms of long-term forecasting.