A multi-strategy hybrid model for air-conditioning cooling load forecasting

Accurate forecasting of air-conditioning cooling load is important for building energy management, but cooling load series are non-stationary and multiscale. This study examines a hybrid framework combining Multivariate Variational Mode Decomposition (MVMD), a Temporal Convolutional Network (TCN), a Particle Swarm Optimization-based Echo State Network (PSO-ESN), and Multi-Head Attention (MHA) for cooling-load modeling. Using DeST-generated summer data for an office-building model in Beijing, the framework achieved an MAE of 0.0124 and an RMSE of 0.0207 on a chronologically held-out 30% test period. The test-period cooling-load observations were not used for normalization, feature selection, MVMD input, component-model training, or MHA fusion training; they were used only for final evaluation.