关键词: 产能过剩, 预警模型, 煤电行业, 预期损失, 可解释性

Abstract: Reliable early warning mechanism of coal power overcapacity is the necessary premise and key to ensure its power supply security in the short term and carbon-neutrality goal in the long term. The “double carbon” strategy has become one of the important national strategies. Under this established strategy, as the largest “contributor” to carbon emissions, coal power overcapacity is an unchangeable development trend and its phase-out is imperative. However, China's economic development stage and coal-based energy resources endowment, coupled with the volatility of renewable energy output and the immaturity of energy storage technology require coal power to be the “ballast” of safe and stable power supply for a long time in the future. Therefore, the exit of coal power overcapacity must be planned in advance, and its foundation lies in accurate early warning of coal power overcapacity.
However, the existing research on early warning of overcapacity has suffered some limitations. First, the existing research on the construction of early warning models does not fully consider the matching relationship between data characteristics and model characteristics, which results in a non-inferior model rather than an optimal model. Second, scholars focus on accuracy when evaluating the models. However, the early warning of overcapacity risk is closely related to capacity regulation. Therefore, it is essentially a cost-sensitive decision-making problem and the potential loss caused by prediction error needs more attention. Third, existing research often pursues prediction performance and builds complex models, ignoring the opacity caused by the complexity of the model while management decision scenarios need not only relevance, but also causality.
Therefore, first, in view of the high-dimension of coal power overcapacity warning indicators and sample's sparseness, we construct a SVM model (linear kernel) good at dealing with small sample and high-dimensional data. Second, due to the difference between the economic consequences of capacity shortage and overcapacity, we build the total cost index to reduce the expected loss of the early warning model. Third, given the decision-making demand for “correlation+causality”, the interpretable method is constructed to reveal model reasoning mechanism and the driving mechanism of factors on risk.
The results show: 1)Under the constraint of the highest accuracy, the accuracy, macro recall, and macro precision of the SVM (linear kernel) are better than in other models, but the total cost of the SVM (linear kernel) is higher, which is approximately 1.5 times that of the BP neural network. 2)Under the constraint of the minimum total cost, the total cost, accuracy, macro recall, and macro precision of the SVM (linear kernel) are better than in other models. Due to sacrificing a small amount of accuracy in exchange for a significant decrease in overall cost, it is recommended to choose the SVM (linear kernel) model with the minimum overall cost constraint. Furthermore, revealed by post interpretability techniques, the evolutionary pattern of key characterization indicators for coal power overcapacity risk (low risk→medium risk→high risk) is sensitive indicators→periodic indicators→comprehensive indicators; the corresponding important cause change law is market factors→policy and transmission factors→comprehensive factors.
To summarize, the paper has contributed to the literature in two ways. First, our models improve the modeling logic of overcapacity risk early warning models under high-dimensional data, expand the model evaluation approach from achieving the highest accuracy to minimizing overall cost, and overcome the opacity of machine learning models. It provides comprehensive, quantitative analytical tools for the governance decision-making of overcapacity risk. Second, we have revealed the primary characterization indicators and important causes of overcapacity under different risk levels, and the evolutionary law of the risk state. This provides a solid decision-based foundation for preventing and controlling coal power overcapacity.

Key words: overcapacity, early warning model, coal power industry, expected loss, interpretability