A hybrid model based on multivariate fast iterative filtering and long short-term memory for ultra-short-term cooling load prediction

Abstract

The current ultra-short-term cooling load forecasting models have not given due attention to the data pre-processing stage. In this paper, multivariate signal decomposition methods MEMD and MvFIF are used in the preprocessing phase to replace the complex signal with simpler subcomponents. The resulting increase in the number of features is tackled through a dimensionality reduction technique, PCA. Finally, prediction is done using two rigorous machine learning algorithms – LSTM and XGBoost. By combining these algorithms at different stages, four hybrid algorithms are formed - MEMD-PCA-LSTM, MEMD-PCA-XGBoost and MvFIF-PCA-LSTM, and MvFIF-PCA-XGBoost. Following a thorough performance comparison, this paper proposes MvFIF-PCA-LSTM for the prediction of ultra-short-term cooling loads. Additionally, experiments are performed to compare the running time of the proposed model, to endorse the importance of using PCA in the proposed model, and to evaluate the choice of parameters that undergo feature reduction. Compared to the base LSTM model assayed on the same datasets, the proposed model offered an improvement of 24.94%, 33.65%, and 23.82% in R2 values for SIT@Dover, SIT@NYP, and simulated datasets, respectively. MAPE achieved by the proposed model is exceptionally low, measuring at 1.13% for the SIT@Dover dataset, 1.42% for the SIT@NYP dataset, and a mere 0.36% for the simulated dataset. The best values of performance metrics computed for the proposed model demonstrate its accuracy in ultra-short-term cooling load prediction.