基于时空序列的CNN-GRU气温预测模型

Temperature prediction model based on CNN-GRU with spatio-temporal sequence

针对传统气温预测方法在面对多维度样本时模型收敛速率低、模型拟合程度差等问题,提出一种基于卷积神经网络(CNN)和门控循环单元(GRU)的组合预测模型。首先,使用CNN对原始样本数据进行降维,提取样本中各特征向量之间的隐含关系;其次,通过GRU网络学习降维后的样本数据的特征动态变化趋势与规律来实现气温预测。通过德国耶拿马克斯普朗克生物地球化学研究所气象站的气象观测数据验证所提模型。实验结果表明,与未采用卷积操作的Prophet-LSTM和PCA-GRU结构神经网络相比,所提模型在处理高维度样本数据时能在保证模型收敛速率的同时,分别将平均绝对误差(MAE)降低24.0%和22.8%,均方根误差(RMSE)降低42.2%和39.4%,验证了它在实际环境中预测温度趋势的有效性。

A combination prediction model based on Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU)was proposed to address the low convergence rate and poor fitting degree of traditional temperature prediction methods when dealing with multidimensional samples.Firstly,the original sample data was dimensionally reduced and the hidden relationships between feature vectors in the sample were extracted by CNN.Secondly,the dynamic trends and regularities of the reduced sample data features were learned through the GRU network to achieve temperature prediction.The proposed model was validated using meteorological observation data from the Max Planck Institute for Biogeochemistry meteorological station in Jena,Germany.The experimental results indicate that compared to the Prophet-LSTM and PCAGRU neural network structures without convolution operations,the proposed model,when dealing with high-dimensional sample data,was able to reduce the Mean Absolute Error(MAE)by 24.0% and 22.8%,respectively,and the Root Mean Square Error(RMSE)by 42.2% and 39.4%,while maintaining the convergence rate of the model.This validates its effectiveness in predicting temperature trends in practical environments.