Deep Learning-based Probabilistic Forecasting Method For Thermal Coal Price

Among the primary energy in China,coal plays the most important role.The co-ordination of supply and demand and the price of thermal coal are vital to the social and economic development.In recent years,there exists dynamic unbalance of supply and demand in both abroad and domestic thermal coal markets,which results in violent price fluctuation.Accurate price forecasting can greatly benefit coal industry chain,not only the coal supply and demand sides but also energy policy makers.Thus,many research works have been published on coal price forecasting,which mainly are point forecasting models,including classic statistical and econometric models,feature engineering-aided machine learning models,and data-driven deep learning models.There are numerous factors affecting the price of thermal coal,which hold diverse time granularities,existing deep learning models,such as LSTM,GRU,TCN models,cannot fully capture long-range temporal and spatial dependencies.Therefore,this thesis studies the optimization of a point forecasting model for thermal coal prices via long-range dependence and effect from multiple related factors by attention mechanism.Furthermore,the probabilistic forecasting model outperforms point forecasting ones over the capability on formulation of possible fluctuation range and the risks,and is a powerful tool to quantify the uncertainty of forecasting.Based on the deep learning-based point forecasting model,this thesis further studies the probabilistic forecasting method of thermal coal price.Firstly,the price formation mechanism of thermal coal is analyzed,combining with the analysis of external factors that affect thermal coal price,a deep learning-based point forecasting model with spatiotemporal attention mechanism is proposed.Through correlation analysis on the time sequences of external factors,the suitable factors are selected and their optimal combination is designed.According to the deployed performance metrics,an attention mechanism is developed to enable the deep learning model to fully exploit spatiotemporal features and adaptively extract the weight distribution of external influencing factors.Sufficient experiments are carried out on an open thermal coal price dataset to evaluate the proposed model,and the comparative results verify the enhancement on the forecasting accuracy of the proposed model over the baseline models.Furthermore,a probabilistic forecasting model is developed to enhance the representation capability of price forecasting,which is based on the forecasting error whiting via cloud transformation and error calibration.Firstly,the forecasting error sequence of point forecasting model is statistically analyzed to obtain its distribution,then a combinational model of multiple sub-clouds is formulated to fit the error distribution,which is used to calculate a forecasting interval under a certain confidence level.Adopting rooted mean square error(RMSE),mean absolute error(MAE)and coefficient of determination as the performance metrics of forecasting error fitting model and forecasting interval coverage,forecasting interval average width and coverage width-based criterion as the metrics of probabilistic forecasting model,the proposed model is evaluated,and the comparative results verify the advantage of the proposed model over reference models.