Experience knowledge decomposition – Data generation: Enhanced multi-step short-term cooling load predictions in data centres with data shortage issues

For new data centres or data centres with inconsistent data interface protocols, the shortage of available cooling load data leads to low prediction accuracy in data-driven cooling load prediction (CLP) models. Data generation (DG) aims to enrich existing cooling data and can help address this shortage in data centre. However, as the prediction horizon extends, DG may become less effective, as the generated cooling load may not always be realistic for multiple future steps. To overcome this limitation, this study proposes a CLP strategy (EKD-DG) that combines DG with experience knowledge decomposition (EKD) to generate both dynamic and static cooling load. The original load is first decomposed into dynamic and static components using EKD. A conditional variational autoencoder (CVAE) is employed to process the dynamic load and generate synthetic dynamic cooling load with a similar distribution. The EKD-DG strategy is then trained using both the raw and the generated dynamic cooling load. Compared to DG, EKD-DG improves the quality of the generated data by producing more realistic dynamic cooling load. This study uses the past 24 hours of data to predict cooling load for the next x hours (x = 1, 6, 12, 18, 24). Annual cooling load from a data centre in Guangzhou are used to validate the results. A comparison of the prediction performance of four models (LSTM, LSTM-EKD, LSTM-DG, and LSTM-EKD-DG) shows that for 1-hour predictions, the mean absolute percentage error (MAPE) of LSTM-EKD-DG improves by 0.52 compared to DG and by 0.9 compared to LSTM, indicating better short-term performance. For predictions beyond 12 hours, the MAPE of LSTM-EKD-DG is similar to LSTM-EKD, but it still improves by 0.50 compared to DG and by 0.67 compared to LSTM. • EKD-DG proposed in this study can improve the data shortage problem in data centre cooling load predcition from two perspectives: increasing data volume and improving data quality. • The predictive performance of the improved model was compared in multi-step prediction. • SHAP reveals the feature contribution of different models in prediction. • ADF test explains the improvement of EKD on load data in terms of time smoothness.