| Coal plays a crucial role in ensuring China’s energy security,and maintaining a stable coal price is of paramount importance in safeguarding the nation’s economic and social development.While deep learning methods have been widely employed in thermal coal futures price forecasting,there is still room for improvement,particularly in cases of price bubbles.To address this issue,this thesis leverages insights from the formation mechanism of thermal coal price,supply-demand theory,price bubbles theory,and other related fields to design an in-depth learning model.By combining knowledge and data,this model establishes a prediction framework for thermal coal futures price,thereby enhancing the reliability of price forecasting.The primary focus of this thesis encompasses the following key areas:1)A price bubble recognition model is designed and its causes is analyzed for thermal coal futures prices.To achieve this,we employ the Generalized sup ADF test(GSADF)and Backward sup ADF test(BSADF)methods based on sequence stationarity and cointegration.Our analysis reveals the existence of four long-lasting price bubbles in China’s thermal coal futures prices from January 17,2019 to May 11,2022.To design an interpretable analysis model with high reliability,we leverage Stacking strategy,knowledge distillation and Layer-wise Relevance Propagation(LRP)algorithm.We then conduct statistical description and cause analysis of the four bubbles in the sample test period to provide mechanism guidance for future futures price prediction models.Additionally,we provide suggestions on policy regulation by combining the cause of bubbles with the actual situation of the thermal coal market.2)A thermal coal futures price forecasting model based on Dual-stage Two-phase Attention-based Recurrent Neural Network and Wasserstein Generative Adversarial Nets(DSTP-RNN+WGAN)is designed.Based on the thorough understanding of the underlying price fluctuation laws,a novel price classification forecast model is proposed,which is designed to capture the differences between the normal fluctuation period and the bubbles period.To model the long-term dynamic space-time relationship between the price impact index series and the price series during the normal fluctuation period,we employ a Dual-stage Two-phase Attention-based Recurrent Neural Network DSTP-RNN model.This model is capable of accurately predicting price trends and capturing the partial laws of bubbles prices under small samples.To further improve the accuracy of trend prediction,we define a new accuracy index AT and improve the objective function of the prediction task.Additionally,we design a price forecasting calibration model based on WGAN to model the forecasting error of the DSTP-RNN model during the bubbles period.Our experimental results on real futures price data demonstrate that our proposed model outperforms mainstream benchmark models such as ARIMAX,SVR,LSTM,and DSTP-RNN in terms of price prediction and trend prediction accuracy.3)A thermal coal futures price forecasting model based on Dual-stage Two-phase Attention-based Recurrent Neural Network and Denoising Diffusion Probabilistic Model(DSTP-RNN+DDPM)is designed.With the powerful generation ability of DDPM,a two-stage denoising process with different guidance conditions is designed to enhance the model’s understanding of sequences and achieve high-precision price prediction.The price foam test results are taken as the leading conditions for noise reduction,providing a stronger knowledge driven role for price law modeling;And based on the results of the analysis of the causes of price foam,feature selection is carried out to achieve efficient feature extraction of price laws.In addition,the noise estimation network of the proposed model adopts the DSTP-RNN model,which effectively extracts the longterm dynamic spatial correlation of the input sequence;The loss of trend similarity is introduced into the loss function to ensure the accuracy of trend prediction of the model.The experimental results show that the prediction performance of the proposed DSTP-RNN+DDPM model is significantly improved compared to the DSTP-RNN+WGAN model. |
