无线电工程

气温变化与人类的生产生活密切相关，对人类出行计划、农林生产以及军事作战等方面都有着重要影响，因此对于大气温度更加精准的预测具有一定现实意义。针对传统预测模型对气温预测精度不佳的问题，提出了一种融合极端梯度提升树(Extreme Gradient Boosting, XGBoost)和改进长短期时序网络(Long and Short-Term Temporal Patterns with Deep Neural Network, LSTNet)的气温预测模型。利用XGBoost进行特征筛选，降低数据维度；利用LSTNet进行改进，在其卷积层嵌入通道注意力(Channel Attention, CA)机制，强化显著特征；把循环神经网络层中的循环门单元(Gate Recurrent Unit, GRU)改为双向长短时记忆网络(Bidirectional Long Short-Term Memory, BiLSTM),并加入了时序注意力(Temporal Attention, TA)机制，使模型拥有同时提取正反向信息的能力且突出了重要时间步的信息；用建立好的模型进行预测实验和对比实验。实验结果表明，提出的改进模型在气温单步预测时，平均绝对误差(Mean Absolute Error, MAE)为0.303℃,相较于长短期记忆(Long Short-Term Memory, LSTM)神经网络等对比模型，MAE最多降低3.429℃,最少降低0.225℃;多步预测时，MAE随时间步增加，最多降低3.827℃,最少降低0.288℃,说明所提模型在单步预测和多步预测上预测精度都更高，优于同类模型。

Temperature change is closely related to human production and life.It has an important impact on human travel planning, agricultural and forestry production and military operations.Therefore, it is of certain practical significance to make more accurate prediction of atmospheric temperature.To address the low accuracy of traditional temperature prediction models, a temperature prediction model integrated with Extreme Gradient Boosting(XGBoost) and improved Long and Short-term Temporal Patterns with Deep Neural Network(LSTNet) is proposed.Firstly, XGBoost is used for feature filtering to reduce the data dimension.Then, LSTNet is improved by embedding the Channel Attention(CA) mechanism in its convolution layer to strengthen the salient features, the Gate Recurrent Unit(GRU) in the cyclic neural network layer is replaced by BiLSTM,and the Temporal Attention(TA) mechanism is added to make the model be able to extract the forward and reverse information at the same time and highlight the information of important time steps.Finally, prediction experiment and comparison experiment are carried out with the established model.The experimental results show that the Mean Absolute Error(MAE) of the proposed model is 0.303 ℃.As compared with comparison models such as LSTM neural network, the MAE decreases by 3.429 ℃ at the maximum and 0.225 ℃ at the minimum.In multi-step prediction, the MAE decreases by 3.827 ℃ at the maximum and 0.288 ℃ at the minimum with the increase of time step, which indicates that the proposed improved model is better than similar models with a higher prediction accuracy in single-step and multi-step prediction.