Research On Mid-and Short-term Forecast Methods Of Thermal Coal Price

As the most important form of the use of primary energy in China,the price forecast of thermal coal is of great significance to the production and operation of enterprises,as well as the prospect of the market.Recently,thermal coal price forecasting attracted many research interests from both academic and industrial.Existing research mainly based on econometric methods,ARIMA(Autoregressive Integrated Moving Average)-based time series methods,BP(Back Propagation)neural network and LSTM(Long Short-Term Memory)-based deep learning models.Although presenting the state-of-the-art performance for price forecasting,LSTM-based models hold problems of gradient vanish and the loss of long-term temporal dependence information.What is more,most of the existing studies confuse spot market with future market by ignoring the difference of these two markets so that the thermal prices of these two markets cannot be modeled specifically.Therefore,in this thesis two targeted models are studied for thermal coal spot and futures prices specifically.The main contributions can be summed up as below.1)Based on the analysis on both the inherent characteristics of the thermal coal spot price and its connection to the influencing factors,a deep learning forecasting model with attention mechanism is developed.The long-term temporal characteristics are modeled by bidirectional LSTM,a CNN(Convolutional Neural Network)module is introduced to extract the local temporal characteristics of influencing factors,and then the dependence relationship between different influencing factors is learned through lateral attention mechanism to distill information under long-term steps.Sufficient experiments are implemented on real dataset to verify the performance of the proposed model.By comparing the error indicators of models at different time steps and analyzing fitted trend curves,the proposed model is proved to outperform mainstream time series models.2)In order to reliably predict the futures price of thermal coal,its statistical features are analyzed and compared to spot price,and then a GAN(Generative Adversarial Network)type model is designed.On the basis of Wasserstein GAN,the objective function is optimized based on the characteristics of the futures price series.To solve the problem of frequent fluctuation of futures price series,the generator takes into account the mean square error term and the inflection point prediction loss.Also,the evaluation index ASP(Accuracy of Sign Prediction)is defined to evaluate the performance of the trend prediction.Comparing to LSTM and Wasserstein GAN,the proposed model has better performance on ASP,RMSE(Root-mean-square Deviation)and MAPE(Mean Absolute Percentage Error)metrics,and achieves better overall prediction and trend prediction as well.