A hybrid prediction model of improved bidirectional long short-term memory network for cooling load based on PCANet and attention mechanism

Accurate and reliable cooling load forecasting is a prerequisite for air-conditioning system control and the basis for building-side energy management. Therefore, a hybrid prediction model of an improved bidirectional long short-term memory (BiLSTM) network based on principal component analysis network (PCANet) and attention mechanism (CNN-IBiLSTM-Attention) is proposed to predict the cooling load of large commercial buildings. First of all, the PCANet algorithm is used to analyze the sensitivity of the influencing factors. Then, the hybrid strategy improved whale optimization algorithm (HSIWOA) is used to optimize the hyperparameter of BiLSTM. At last, the performance of the proposed algorithm is verified by using the actual data of two commercial buildings in Xi'an. The results show that using the PCANet algorithm for sensitivity analysis avoids feature redundancy. HSIWOA is suitable for hyperparameter optimization of BiLSTM. Compared with the other three prediction models, CNN-IBiLSTM-Attention reduced the mean absolute percentage error (MAPE) of Building 1 and 2 test sets by 31.55 %, 55.59 %, and 60.58 % and 56.49 %, 60.3 %, and 67.37 %, respectively. The proposed prediction model has superior hyperparameter optimization ability, better model complexity, and stronger generalization ability. Therefore, the proposed prediction model becomes a reliable tool for predicting the cooling load of large commercial buildings. •Comprehensive classification of factors affecting commercial building cooling load.•Feature extraction method based on the PCANet algorithm.•Combining PCANet, BiLSTM, and attention mechanism for cooling load forecasting.•Using the HSIWOA to optimize the hyperparameters of BiLSTM.•Verifying the validity of the CNN-IBiLSTM-Attention model using actual data.