Volume 65

Multi-Objective Prediction of Building Loads Based on Similar Days Considering Window Inconsistency Hongtao Wei, Zaiqing Ren, Baorui Zhang, Chongyi Tian

Abstract

To address the limitations that traditional time series methods and regression analysis struggle to characterize the nonlinearity, dynamics, and uncertainty of building energy consumption, and that existing deep learning prediction models suffer from the issue of window inconsistency, this study proposes a multi-objective prediction method for building loads based on similar days that accounts for window inconsistency. The primary goal of this method is to achieve accurate prediction of both the total building load and chiller energy consumption.The implementation process of the proposed method is as follows: First, combined with the meteorological features and date attributes of the day to be predicted, 100 similar days are selected to construct a matched training dataset; Second, three parallel convolutional channels (for meteorological data, temporal information, and historical load) are designed to extract features, followed by cross-dimensional fusion of these features; Finally, the fused features are input into the X-LSTM-Transformer model: the X-LSTM component captures long-term dependencies, while the Transformer component focuses on key time nodes. Experiments were conducted using one year of measured data from an office building. The results indicate that,For the total building load prediction: MAE = 6.138, RMSE = 7.623, R² = 0.979;For the chiller energy consumption prediction: MAE = 4.097, RMSE = 6.358, R² = 0.956.The predicted values show high consistency with the actual measured values. Compared with the X-LSTM-Transformer model without the similar-day strategy, as well as benchmark models including CNN-Attention-LSTM, Bi-LSTM, and LSTM, the proposed method demonstrates significantly superior prediction accuracy.

Keywords uilding load; X-LSTM-Transformer; Similar days