Frequency-domain analysis for cooling load prediction: A Multi-Scale FourierGNN Crossformer approach

With the rise in energy consumption in buildings, particularly for air conditioning systems, intelligent air conditioning control has become more important. Accurate prediction of central air conditioning system cooling load is crucial to optimizing energy efficiency. This paper proposes a Multi-Scale FourierGNN Crossformer (MFGformer) model based on Crossformer, which integrates Fourier Graph Neural Networks (FourierGNN) and Multi-scale Cross-axis Attention (MCA) mechanisms for multivariate cooling load time series forecasting of central air conditioning systems. The complexity of dynamic characteristics of multivariable time series data is often misunderstood by current forecasting models, which this model takes into account. The FourierGNN module maps time series data into the frequency domain through the Discrete Fourier Transform, effectively capturing periodic and trend features of the data. The MCA mechanism captures multi-scale characteristics and local detail information in time series data through dual cross-attention computation, enhancing the ability to capture dependencies between different variables. Validation on two real case datasets shows that the MFGformer model performs well in predicting the cooling loads of central air conditioning systems, especially when outputting 24-hour predictions at a 96-hour input time step, the model’s Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Arctangent Absolute Percentage Error (MAAPE), and Coefficient of Variation of Root Mean Square Error (CV-RMSE) are 130.7765 kW, 322.2957 kW, 25.2611% and 56.6846%. Compared with eight other typical models, the MAE, RMSE, MAAPE and CV-RMSE of the MFGformer model were reduced by 22.8869-156.1751 kW, 59.6766-336.4538 kW, 0.9422-51.109% and 8.6393-32.6965%, respectively, for different output step sizes. These results demonstrate the model’s ability to provide accurate predictions when dealing with data characterized by volatility and nonlinearity, while maintaining the ability to identify long-range dependencies.