Light-stacking strengthened fusion based building energy consumption prediction framework via variable weight feature selection

Highlights: • A model-strengthened framework is proposed for building energy prediction. • The framework can balance the speed and accuracy of hyperparameter optimization. • A fusion feature selection method with high generalization performance is proposed. • Ablation analysis and sensitive analysis are used to verify feature selection output. • The effectiveness of the model is evaluated through excellent visualization methods. Building energy consumption prediction plays an irreplaceable role in energy resource management and planning. Continuous improvement in the performance of predictive models is the key to ensure energy management and deployment operations. The imbalance between the speed and the accuracy for hyperparameter optimization is an important factor that limits the performance of the model. A Light-Stacking Strengthened Fusion Framework (LSStFu) is proposed to solve this problem. The optimization and fusion of the multi-type hyperparameter model obtained by random search can greatly improve the accuracy of the model prediction. This process can also assure a reduction in time consumption when compared with grid search. Moreover, a feature selection algorithm can only describe a single aspect of a bunch of multiple types of data sets, which limits the generalization performance ability of the algorithm. In order to address the above limitation, a Variable Weight Feature Selection (VWFS) method is proposed to fuse the contribution of three feature selection algorithms based on particle swarm optimization. To evaluate the robustness of the proposed LSStFu algorithm, it is compared with other algorithms through eight evaluation indicators. This evaluation process verifies the reliability and stability of the Light-Stacking framework to provide an efficient, accurate, and stable hyperparameter optimization framework for energy predictive models. From the ablation analysis, it can be observed that with the optimal subset of features obtained through the VWFS, the accuracy of the prediction models is improved. In addition, the model construction process has also sped up at the same time.