Air Conditioning Load Forecasting Based on CEEMDAN and CNN-BiLSTM-Attention

To address the issues of strong non-stationarity in the electrical load time series of commercial complex air conditioning systems and the insufficient prediction accuracy of traditional models, this study proposes a multi-source spatiotemporal load forecasting method based on CEEMDAN and CNN-BiLSTM-Attention. By integrating passenger flow, temperature, humidity, CO2 concentration, and temporal features to construct an energy consumption feature set, the method utilizes CEEMDAN decomposition to reduce the complexity of the time series and combines CNN, BiLSTM, and Attention mechanisms to achieve air conditioning load forecasting. Experimental results demonstrate that this model outperforms comparison models in metrics such as MAPE, RMSE, MAE, and R2, exhibiting high prediction accuracy and stability for air conditioning energy consumption fluctuations in commercial buildings under complex environmental conditions.