【摘要】 温度数据具有明显的反向、时序相关性及多尺度特征，提升温度预测精度的关键在于能否有效提取温度数据的上述特征.为提取这些特征，该文提出一种多通道卷积双向长短期记忆网络(convolutional neural network-bidirection long short-term memory, CNN-BiLSTM)的短时温度预测模型.该模型首先利用双向长短期记忆网络(BiLSTM)提取数据的反向特征、时序相关性特征；再利用多通道且不同尺寸、不同膨胀率的卷积神经网络(CNN)提取数据的多尺度特征，组成在学习多尺度特征后的数据，将其和原始数据作为BiLSTM层的多通道输入，输出的数据经过全连接层，形成最终的预测结果.实验结果表明：多通道CNN-BiLSTM的短时温度预测模型能有效地提取数据的时序相关性、反向及多尺度特征，可有效地提升温度预测精度，是一种行之有效的短时温度预测模型.更多还原

【Abstract】 Temperature data have obvious reverse, temporal correlation and multi-scale features.The key to improve the accuracy of temperature prediction is to extract the above features from temperature data effectively.In order to extract these features, a short-time temperature prediction model is propsed for Convolutional Neural Network-Bidirection Long Short-Term Memory(CNN-BiLSTM).BiLSTM is used to extract reverse feature and temporal correlation feature from data.Multi-channel CNN with different sizes and expansion rates is used to extract multi-scale feature from the data and composed the data after learning multi-scale feature.The data and the original data are used as multi-channel input of BiLSTM layer, and the output data passes through the full connection layer to form the final prediction result.The experimental results show that the short-time temperature prediction model for multi-channel CNN-BiLSTM can effectively extract the reverse, temporal correlation and multi-scale features of the data, and can effectively improve the accuracy of temperature prediction.Therefore, it is an effective short-time temperature prediction model.更多还原