统计与数据科学学院知识库

       原油是当今社会工业发展和日常生活中的重要能源资源，对全球经济和世界金融市场有着举重若轻的作用，国际原油价格的波动也受到多种因素的影响，准确预测国际原油价格波动已成为如今的研究热点。

       本文基于对多模态分解和混合深度神经网络模型的研究，提出了一种多尺度分解-集成框架下的原油价格预测模型。具体而言，首先使用经验模态分解模型（EMD）分解原油价格序列，并通过多模态分量特征重构出三类序列（低频分量、高频分量和趋势分量），其次通过基于随机森林构建的递归特征选择法（RF-RFE）分别对三类序列选择最优特征子集，然后使用双向长短期记忆神经网络（BiLSTM）与卷积神经网络（CNN）融合的混合模型（BiLSTM-CNN）拟合三类序列变化趋势，最后集成获得原油价格序列预测结果。

       为了验证模型的性能，采用世界原油市场上的三大基准价格之一的美国西德克萨斯的中质原油（WTI）每日现货价格验证模型性能，通过从评价指标（MAE、RMSE、MAPE 和𝑅2 Sore ）和预测曲线两个角度，与四类对比模型（LSTM、LSTM-CNN、LSTM-DNN、BiLSTM-CNN）比较结果显示，所提出的模型能够有效提升预测精度。

       为了拓展模型的泛用性，提升模型对突发事件冲击下的原油价格预测精度，使用特征迁移学习对模型再训练，使用突发公共卫生安全事件——新冠疫情对原油价格影响为例，验证了迁移模型的拟合精度和有效性；为了拓展模型的应用范围，将模型对原油价格序列的预测结果作为特征加入不同预测任务，以原油期货价格和新能源市场指数预测为例，验证了合理性和预测精度提升。

       本文提出一种集成框架下的混合深度学习国际原油价格预测模型，通过融合 EMD 分解集成框架和混合神经网络模型，提高国际油价预测模型精度；同时，提出迁移学习提高模型处理突发事件泛用性和时效性，旨在降低预测成本，为国内市场提供预测参考。

       Crude oil is an important energy resource in industrial development and daily life in today's society. It plays a crucial role in the global economy and financial markets, and the fluctuation of international crude oil prices is also influenced by various factors. Accurately predicting international crude oil price fluctuations has become a research hotspot today.

       This thesis proposes a crude oil price prediction model based on the study of multimodal decomposition and mixed deep neural network models under a multi-scale decomposition integration framework. Specifically, the empirical mode decomposition (EMD) model is first used to decompose the crude oil price series, and three types of sequences (low-frequency, high-frequency, and trend components) are reconstructed through multimodal component features. Then, the recursive feature selection method based on random forest construction (RF-RFE) is used to select the optimal feature subsets for each of the three types of sequences. Then, a hybrid model (BiLSTM-CNN) fused with bidirectional long short-term memory neural network (BiLSTM) and convolutional neural network (CNN) is used to fit the changing trends of the three types of sequences. Finally, the prediction results of the crude oil price series are integrated and obtained.

       To verify the performance of the model, the daily spot price of WTI crude oil from West Texas, one of the three benchmark prices in the world crude oil market, was used to verify the performance of the model. From the perspectives of evaluation indicators (MAE, RMSE, MAPE, and Score) and prediction curves, the proposed model was compared with four types of comparative models (LSTM, LSTM-CNN, LSTM-DNN, BiLSTM-CNN), and the results showed that it can effectively improve prediction accuracy.

       In order to expand the universality of the model and improve the prediction accuracy of the model for crude oil prices under the impact of emergencies, the model was retrained using feature transfer learning, and the impact of public health and security emergencies - COVID-19 epidemic on crude oil prices was used as an example to verify the fitting accuracy and effectiveness of the migration model; In order to expand the application scope of the model, the prediction results of the crude oil price series were added as features to different prediction tasks. Taking crude oil futures prices and new energy market index prediction as examples, the rationality and prediction accuracy were verified.

       This thesis proposes a hybrid deep learning international crude oil price prediction model under an integrated framework, which improves the accuracy of the international oil price prediction model by integrating the EMD decomposition integration framework and the hybrid neural network model; At the same time, transfer learning is proposed to improve the universality and timeliness of the model in handling unexpected events, aiming to reduce prediction costs and provide prediction references for the domestic market.