Construction And Empirical Study On Data-characteristic-driven Forecasting And Early Warning Models For Coal Power Overcapacity

In recent years,coal power industry has suffered serious overcapacity,limited development space,and the trend of excess is becoming increasingly fierce.Moreover,coal power industry has complex industrial linkage and plays a significant role in the national economy.Therefore,Chinese government has intensively introduced a series of measures to defuse overcapacity,but due to the lack of necessary foresight and precision,the implementation effect of relevant policy measures is not very ideal,and it is trapped in the Governance Dilemma of “overcapacity?overcapacity defusing?capacity shorage?capacity incentivizing?overcapacity again”.The main reason lies in the incompleteness and asymmetry of information.Thus,to achieve scientific prevention and accurate control of overcapacity,it is imperative to establish more precise and complete forecasting and risk warning models of overcapacity.Therefore,organically integrating data-characteristic-driven modeling idea with multi-modal information ensemble modeling idea,a combination forecasting method and model of coal power overcapacity(CPO)is constructed.Firstly,the nature and pattern characteristics of the time series data of CPO scale are identified.Secondly,the data is decomposed by the matching decomposition method to obtain multiple components.Thirdly,the data characteristics of each component are identified and the forecast models matching the data characteristics are constructed.Finally,the final prediction results of the CPO scale are obtained by integrating the results of each component prediction.Based on the prediction,using data-characteristic-driven integrated modeling method,an early warning model system for CPO risk is proposed for coal power industry scenarios.Firstly,the early warning indicators of system,completeness,and accuracy are selected,and Surrogate Data Method-Multifractal Detrended Fluctuation Analysis(SMF-DFA)is constructed to determine the warning label adaptively,which lays the foundation for early warning models.Then,due to the data's high dimensionality and sparseness,the cost sensitivity of decision problems,and the machine learning models' lower interpretability,we built a forecasting model system that covers “early warning model?model evaluation?model interpretation”.Based on the empirical and analysis of forecasting and early warning,based on the combination prediction and early warning results,the paper puts forward the countermeasures and suggestions for scientific prevention and precise governance of overcapacity in coal power industry from macro,meso and micro level.The results show that:(1)by identifying the nature characteristics and pattern characteristics of the CPO scale time series,it is found that it not only has nonstationary and non-linear characteristics,but also has high complexity and mutation characteristics,which indicates that the CPO is formed by coupling of multiple driving factors,and its development trend is determined by multiple forces;(2)according to the characteristics of time series data of CPO scale,a Variational Mode DecompositionTriple Exponential Smoothing-Least Square Support Vector Machine(VMD-TESLSSVM)model matching with the characteristics is constructed,and the empirical results show that the contructed model is superior to other models in terms of level and directional forecasting accuracies as well as forecasting stability;(3)the VMD-TESLSSVM model is used to predict the CPO scale,and the forecast results show that the scale of CPO will be still at a relatively high level showing a trend of falling first and then rising and the institutional distortion will still be the decisive factor for thermal power overcapacity;(4)In view of the long-range correlation and multifractal structure of CPO scale time series,SMF-DFA method is contrusted to determine the risk threshold of CPO adaptively from the data evolution law.This method effectively overcomes the subjectivity and lack of theoretical basis limitations of traditional statistical and empirical methods to determine the threshold,and provides the necessary warning label information for the construction of early warning model;(5)Based on the idea of data-characteristic-driven integrated modeling,a risk early warning model system for CPO that covers “model construction?model evaluation?model interpretation” is proposed,and this system improves the modeling approach of the risk forecasting model in a sparse,high-dimensional data environment,enriches the model evaluation technique for cost-sensitive decision problems,and overcomes the “black box” issue of the machine learning model,and the accuracy of risk forecasting,expected losses,and reliability are better considered;(6)through empirical analysis,the important early warning indicators of CPO risk,as well as the typical characteristics and vital causes of overcapacity under different risk levels are revealed.In general,the relevant results of this paper provide effective quantitative analysis tools for CPO governance,improve the prospective and accuracy of governance,and provide some reference and inspiration for other industries.This paper has a total of 35 pictures,20 sheets,and 107 references.