关键词: 零碳智慧园区, 负荷预测, 多时间尺度, 卷积神经网络, 长短期记忆神经网络, 自注意力机制

Abstract:

Accurate load prediction is critical for improving the energy efficiency and profitability of zero-carbon smart parks. However， the application of conventional load prediction techniques faces two main challenges which are the difficulty in obtaining hourly numerical weather forecast data and the need for predictions across different time scales. In the absence of weather forecast data， a method using Convolutional Neural Networks（CNN） was proposed to extract the coupled spatial features between multiple loads. The reconstructed features were input into a Long Short-Term Memory（LSTM） network to extract temporal features of the load， followed by the application of a self-attention mechanism to enhance the model's ability to extract feature information. A fully connected network was then employed for load prediction， resulting in a multi-variable-load， multi-time-scale prediction model based on CNN-LSTM-Self attention. A case study of a park was used to predict its cooling， heating， and electrical loads for the next 1 hour， 1 day， and 1 week. Experimental results showed that the CNN-LSTM-Self attention model outperformed the CNN， LSTM， and CNN-LSTM models in terms of prediction accuracy across multiple time scales. Specifically， the CNN-LSTM-Self attention model showed a more significant advantage in predicting the 1-hour load， with the mean absolute percentage error（MAPE） of cooling， heating， and electrical load predictions improved by 16.25%， 19.16%， and 10.24%， respectively， compared to the CNN-LSTM model.

Key words: zero-carbon smart park, load prediction, multi-time scale, convolutional neural network, long short-term memory neural network, self-attention mechanism