In the energy management of smart buildings, the building operation environment is complex and variable, and energy consumption signals exhibit characteristics of strong nonlinearity, high temporal dependence, and long cycles. Traditional single models fail to accurately depict their high-dimensional mapping relationships. To improve the accuracy of energy consumption prediction and optimize control strategies, this study proposes an RNN-BPNN hybrid model that combines Recurrent Neural Network (RNN) and Back Propagation Neural Network (BPNN), and constructs a prediction-driven closed-loop system for energy consumption optimization control. RNN is used to extract time-series features, while BPNN is applied for nonlinear mapping and high-dimensional regression, thus achieving collaborative optimization of prediction and control. Experiments are conducted using a public building energy consumption dataset. In scenarios with a well-established energy management system, the hybrid model outperforms single models, with the Gated Recurrent Unit (GRU)-BPNN achieving the best prediction performance: The Mean Absolute Error (MAE) is 311.43 kW, the Root Mean Square Error (RMSE) is 484.29 kW, and the coefficient of determination (R2) is 0.925. Compared with the optimal baseline model, the MAE, RMSE and R2 metrics are improved by approximately 2.21%, 1.90% and 1.17%, respectively. In data-sparse scenarios, the Coupled Input and Forget Gate (CIFG)-BPNN achieves an MAE of 330.33 kW and an R2 of 0.794, demonstrating excellent generalization ability. The results indicate that the RNN-BPNN hybrid structure can effectively integrate temporal dependence and nonlinear features, providing a feasible solution and application potential for energy consumption prediction and optimal control in smart buildings.
Ma et al. (Mon,) studied this question.