Deep Learning for Commercial Building Load Forecasting: Hyperparameter Fine-Tuning Convolution Neural Network-Multivariate Multilayered Long Short-Term Memory Time-Series Model

The existing multivariate multilayered long short-term memory model (M-LSTM) of the Swinburne Energy Laboratory requires improvement in efficiency and performs robustness load forecasting for commercial buildings. Accordingly, this research proposes a hyperparameter fine-tuning (HFT) convolution neural network (CNN) M-LSTM (HFT CNN-M-LSTM). By cascading CNN to M-LSTM, CNN can learn hierarchical features and produce dense representations that M-LSTM leverages to capture long-term dependencies for accurate predictions. In addition, HFT is added to CNN and M-LSTM support to adjust the hyperparameters on the basis of feedback during training. The HFT uses a Bayesian Optimization algorithm for hyperparameter searching. Bayesian Optimization is beneficial in reducing computation cost.The performance of HFT CNN-M-LSTM is evaluated by comparing it with those of M-LSTM and CNN-M-LSTM by using the energy consumption datasets from Swinburne University of Technology, Hawthorn Campus, i.e., ATC Building (Dataset S1) and AMDC Building (Dataset S2). The metrics applied to the evaluation are normalized root-mean-square error (NRMSE), R2 score, and mean absolute percentage error (MAPE). For Dataset S1, HFT CNN-M-LSTM exhibits outstanding performance when compared with M-LSTM and CNN-M-LSTM, with NRMSE = 0.0358, R2 score = 0.8825, and MAPE = 0.1112. For Dataset S2, HFT CNN-M-LSTM exhibits a slight degradation compared with M-LSTM, with NRMSE = 0.0353, R2 score = 0.6323, and MAPE = 0.1133. However, the research result emphasizes that HFT CNN-M-LSTM is an excellent candidate for long-term commercial building load forecasting because the proposed model can capture complex patterns and extract residual errors.