Cooling Load Prediction for Data Center Based on Re-LSTM

For the load prediction for the data center, the traditional methods cover the thermal modeling process and data-driven models. The thermal modeling process takes a long time to generate prediction results. However, some historical data of the data center is not available due to the data encryption for the sake of data security, making the traditional data-driven models challenging. This paper proposed the Re-LSTM model based on the LSTM block and shortcut connection with the selected features as prior knowledge for the untrained model. The proposed model could handle the insufficient operation data and decrease the training cost to improve construction efficiency. As for the training results of different models, the proposed Re-LSTM demonstrated a significantly remarkable improvement compared with the data-driven model. The training time of Re-LSTM is around three times less than other models. By means of the Re-LSTM, the cooling load could be predicted accurately and efficiently, and it is practical to predict the cooling load when the operation data is not accessible.