1 Introduction
Energy is not only the necessary basis for a country's development, but also an important index to show its comprehensive national power. Traditional energy is mainly minerals such as coal and oil, which is an important support for economic development. However, these are non renewable energy. And the energy crisis and environmental pollution caused by them have attracted more and more attention. These traditional energy sources have been unable to meet or adapt to the needs of future social development. Therefore, the development of new energy has become an important direction of national strategic development. The development and utilization of new and renewable resources of energy has become an important layout of the development strategy. Solar energy resources are inexhaustible. It is a huge amount of renewable resources without pollution. Compared with coal and oil, it is more environmentally friendly and clean. Moreover, solar energy is not affected by region, and these available resources can be broad, efficient and popular. Therefore, photovoltaic industry has become one of the important emerging industries in the world.
So far, China has been the world's largest emitter of carbon dioxide. But China promises to be carbon neutral by 2060. The biggest challenge is to find ways to reduce or offset greenhouse gas emissions. China needs to start generating electricity with net zero carbon emission technologies, such as photovoltaic technology
[1]. After more than ten years of rapid development, China has formed a relatively complete photovoltaic manufacturing industry system, and the industrial scale and application engineering have ranked first in the world. However, there is still a gap between China and the world's developed countries. Facing the current opportunities and challenges, the release and change of industrial policies can effectively macro-control the operation and development of photovoltaic enterprises. Photovoltaic industry is an important strategic emerging industry for the development and utilization of new energy in China and even the world. Therefore, it is important for the government to make policy to ensure the stable and orderly development of photovoltaic enterprises to accelerate the industrial structure transition in China. It is an important strategic task of the Chinese government to maintain the long-term, stable and sustainable utilization of energy resources. On this basis, China will vigorously promote the development and utilization of new energy such as photovoltaic power generation, guide and support the stable and orderly development of photovoltaic enterprises. Therefore, there are a series of policies issued for the development of new energy. These policy helped China transform itself from a contract factory of photovoltaic cells into a solar photovoltaic power generation powerful nation.
Therefore, it is very important to study the impact of industrial policy on photovoltaic enterprise risk and predict its volatility. Previously, few studies used the text information in industrial policy to analyze the risk of photovoltaic enterprises in addition to the traditional financial indicators. In this paper, we will first use LDA to extract the text features hidden in the industrial policy text, then train the data in depth neural network, and add the text features to predict the risk of photovoltaic enterprises. Taking China's photovoltaic industry policy as an example, this paper uses text mining to find out the key contents of the policy, and uses deep neural network model to empirically analyze the industrial policy in the risk fluctuation of photovoltaic enterprises. The main problem focused in this paper is to quantitatively analyze whether industrial policy will have an impact on photovolatile enterprise risk. Moreover, on this basis, whether industrial policy has a lag impact on photovolatile enterprise risk.
2 Literature Review
Energy is an important basis for economic growth and social progress
[2]. With the progress of economy and the increasing demand of human society for energy, renewable energy, with its renewability and pollution-free availability, will grow into an effective and practical choice to ensure the future development of the world
[3]. Developing renewable energy is an inevitable choice for China's sustained economic growth, harmonious coexistence between human and environment and sustainable development. The constant characteristics of photovoltaic power generation make it an invention source of green energy with low operation cost, low maintenance, high availability and reliability
[4, 5]. These advantages have led to the rapid development of photovoltaic production and the improvement of manufacturing methods in the solar industry, becoming one of the most promising technologies in the field of renewable energy and sustainability
[6, 7]. Therefore, China will vigorously develop photovoltaic technology and issues a series of policies to ensure the steady development of photovoltaic enterprises.
The impact of industrial policy orientation on the operation and development of photovoltaic enterprises has been widely concerned by relevant national departments and all sectors of society. Many scholars at home and abroad have conducted researches on this issue, which mainly focus on whether industrial policy has a significant impact on the photovoltaic industry, if so, to what extent, and whether it is positive or negative. And they provide reliable suggestions for the country in policy-making.
Dusonchet made a comparative analysis of the photovoltaic industry policies of eastern European countries and found that not all the photovoltaic industry promotion policies will have a positive effect, and if the timing is not right, they will have side effects
[8]. Burns, et al. found that industrial policies can encourage innovation and provide deployment support. Although the industrial policy has produced certain positive effects, the manufacturing industry supported by investment has not received enough innovation incentive mechanism, and the R & D support is relatively weak, which has not improved the competitive advantage of the industry to a great extent
[9]. Yu, et al. used a multi perspective approach to study the policy system framework of various countries to seek the best PV policy
[10]. Shuai used Shapley game theory to theoretically analyze the cost sharing among central government, local government and users, and establishes an investment cost sharing model
[11]. Chen believes that the role of supporting the solar photovoltaic industry has changed from subsidized producers to subsidized demanders
[12]. Suh believed that increasing industrial policy support depends on the development of photovoltaic industry, and this can not be generalized
[13].
There are also many scholars using empirical research. Xiong, et al. studied the subsidy policy of large photovoltaic power stations, and finds that the subsidy policy strongly promoted the development of the photovoltaic industry, but also caused a serious oversupply of photovoltaic industry and the lack of competitiveness of photovoltaic enterprises
[14].
On the whole, the existing research mainly focuses on the qualitative description and analysis of the current situation and effectiveness of the policy, but rarely uses the method of quantitative analysis to analyze, summarize and sort out China's government subsidy policy. And there is no research that can quantitatively analyze whether industrial policy actually affects the risk and volatility of photovoltaic enterprises. This paper will analyze the industrial policy documents with text mining technique, and verifies the effectiveness of industrial policies on the photovoltaic enterprise risk prediction based on the deep learning model.
3 The Proposed Framework
As data mining models can help to avoid the judgment error caused by the subjective views of experts and scholars on industrial policies, this paper uses LDA topic model
[15, 16, 17, 18] to process the policy text to extract the topic in the policy text
[19, 20, 21]. Further, we use deep neural network models
[22, 23, 24, 25] to verify the impact of China's industrial policy on the photovoltaic enterprise risk. The combination of these two methods can better explain how industrial policy affects the photovoltaic enterprise risk. Based on these ideas,
Figure 1 shows the research framework of this paper.
Figure 1 The proposed framework |
Full size|PPT slide
3.1 Data Preprocess
Each photovoltaic enterprise consists of 11 attributes and 1 risk value, as shown in Table 1. These 11 attributes can be divided into 10 traditional factor indicators and 1 industrial policy factor indicator. The numerical data can directly obtain the corresponding features, and the text data can be used to extract the subject features in the industrial policy. The extraction method and process will be discussed in detail later.
Table 1 The attributes of data |
| No. | Attributes | No. | Attributes |
| 1 | Age | 7 | Turnover of fixed assets |
| 2 | Asset-liability Ratio | 8 | Growth |
| 3 | ROE | 9 | The top ten majority shareholding ratio |
| 4 | Operating margins | 10 | Total monthly market value |
| 5 | Debt financing cost | 11 | Policy |
| 6 | Equity financing rate | 12 | Z_score |
3.2 Text Mining
The content exists in the text description of the policy document, which can not be directly understood by the computer. Therefore, we need to introduce text mining to help us extract high-value information hidden in text description. In this paper, we use Chinese word segmentation and LDA to solve the incompatibility between text data and data mining algorithms. We use Chinese word segmentation to preprocess the text data, and use the topic model LDA to extract the topics and keywords of important policies in the processed text data.
The text description of policy often appears in the form of long sentences with different structures. However, the key information described is distributed in several words of a sentence. Therefore, we should focus on several key words rather than the whole sentence. Chinese word segmentation refers to the process of dividing a Chinese character sequence into strings and recombining word sequences according to certain standards. Unlike English and other languages, Chinese has no space as a natural separator. And in Chinese, different segmentation may lead to different meanings. During the stage of Chinese word segmentation, we divide the policy text in the form of Chinese sentences into Chinese word sequences, which is convenient for the next processing steps according to the specific meaning of the text. At the same time, stop words can be abandoned during this stage to prevent impact on the analysis. After Chinese word segmentation and removal of stop words, the remaining words are related to our research.
In natural language processing, LDA is an important generation model, which can be used to identify the hidden topic information in large-scale documents or corpora. It uses the Bag of Words approach, which treats each document as a word-frequency vector, transforming the text data into numeric data that is compatible with data mining algorithms. Each document can be regarded as a representation of the probability distribution of a topic, and each topic can be regarded as a representation of the probability distribution of many words. Moreover, because of the non-correlation between the components of the random vector in the Dirichlet distribution, the candidate topics are independent of each other. To achieve the above objectives, the LDA uses a joint distribution to compute the conditional distribution of the hidden variable under a given observable variable. The observable variable is a set of words, and the latent variable is the topics.
In our research method, the words generated from the policy text can be regarded as LDA documents to extract the topics related to photovoltaic policy, and the topics are represented by the words with high probability distribution among several corresponding topics. After the processing of LDA, each keyword of each topic is associated with the data set.
3.3 Deep Neural Network Model
All categorical, numeric and topical features are transmitted into the input layer of the deep neural networks (DNN) to start the training process. Then, the input layer maps the features to the first hidden layer, and each hidden layer includes a number of nodes for processing the input data of the layer and transporting the result to the next layer. The activation function of each layer can add nonlinear mapping to the mapping process to guarantee the abstraction ability of the DNN more effective. In this paper, we use the deep neural network model for prediction.
4 Empirical Analysis
4.1 Data Description
The data used in this paper comes from the real financial report data of photovoltaic concept listed companies and the photovoltaic related policy documents published on the national official website. It contains financial statements of 254 photovoltaic enterprises, which are numerical data, and 1606 policy documents published on the three official websites of the Government Website of China, the Ministry of Industry and Information Technology, and the National Energy Administration, which are textual data.
4.1.1 Traditional Financial Indicators
The traditional financial indicators data used in this paper is downloaded from the CSMAR database. The sample selected the financial report data of 254 photovoltaic concept listed companies from 2016 to 2021, and extracted the traditional operation index data such as asset liability ratio, return on net assets and operating profit margin.
4.1.2 Policy Documents
In addition to the traditional business index data of enterprises, this paper needs to obtain the policy documents on Photovoltaic Industry issued by the government. The photovoltaic industry is an industry that involves many fields and is governed by different departments of government. Through our investigation and analysis, the three official websites of the Government Website of China, the Ministry of Industry and Information Technology, and the National Energy Administration are the main online platforms for China to release major policies and measures related to the photovoltaic industry. Moreover, the policies issued by these three platforms have a wide range of influence and play an important role in the operation and development of photovoltaic industry.
The Government Website of China is sponsored by The General Office of the State Council and operated and maintained by Chinese Government Website Operation Center. The Government Website of China is also a channel for interaction between the public and the government. It is a comprehensive platform for the State Council, various departments of the State Council and local governments to publish government information and provide online services on the Internet.
The Ministry of Industry and Information Technology is an important constituent department of The State Council. As an industry management department, the Ministry of Industry and Information Technology mainly manages and standardizes industry plans, policies and standards, and guides industry development, but does not actually intervene in the production and operation activities of industry enterprises. The Ministry of Industry and Information Technology is responsible for putting forward strategies and policies for the development of new industrialization, drawing up plans, policies and standards for industry, communication industry and high-tech industry and organizing their implementation, guiding technological innovation and progress in these industries, and coordinating and solving major problems in the process of new industrialization. For the use of new energy such as photovoltaic, the Ministry of Industry and Information Technology participates in the formulation of policies for energy conservation, comprehensive utilization of resources and promotion of clean production, and organizes and coordinates relevant major demonstration projects and the promotion and application of new products, technologies, equipment and materials.
The National Energy Administration is an important department responsible for comprehensive coordination and service guarantee of national energy development strategy. It is mainly responsible for drafting rules and regulations on energy development and its supervision and administration, promoting energy system reform, coordinating major issues in energy development and reform, and organizing the formulation of policies and related standards for coal, oil, natural gas, electric power, new energy and renewable energy.
Therefore, we will obtain all policy documents related to the keyword "photovoltaic" from the Government Website of China, the Ministry of Industry and Information Technology and the National Energy Administration for the following research and analysis of this paper. In order to further research whether there is lag effect in the impact of industrial policy, this paper collects all photovoltaic related policy documents from 2016 to 2021. Finally, it includes 31 policy documents on the Government Website of China, 82 policy documents on the Ministry of Industry and Information Technology, and 1493 policy documents on the National Energy Administration.
4.2 Evaluation Criterion
This paper employs three widely used performance measures, including mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE), to evaluate the performance of different data mining models. Given pairs of actual value () and predicted value (), each of these performance measures is defined as follows.
4.3 Topical Feature Extraction
The text data includes 1606 photovoltaic related industrial policy documents issued by China. This report uses LDA to extract the topics hidden in the policy text. LDA can be used to obtain the keyword distribution of each topic and the topic distribution in each policy document.
According to our training of LDA model, the number of topics we selected is 10. Some representative subject words and their probability values are shown in Table 2.
| Type | Description |
| Topic 1 | Poverty alleviation, Work, Photovoltaic, Shake off poverty, Duty |
| Topic 2 | Development, Technology, Energy storage, Construction |
| Topic 3 | Power supply, Limited, Company, The masses, Enterprise |
| Topic 4 | Energy, Photovoltai, Electricity, Power Generation, Nation |
| Topic 5 | High quality, Support, Hope, Coal Power, Understanding |
| Topic 6 | 10 million Watt, The whole country, Amount, Installation, Generator Set |
| Topic 7 | Construction, Kilovolt, Plan, Project, Channel |
| Topic 8 | Rectification, Death, Occur, Personal, Casualty accident |
| Topic 9 | Risk, Cooperation, Mechanism, China, International |
| Topic 10 | Energy, Development, Clean, National, China |
4.4 Experimental Analysis
4.4.1 Hyperparameter Optimization
Hyperparameters are parameters whose values are set before starting the learning process, rather than parameter data obtained through training. Usually, we need to optimize the hyperparameters and select a set of hyperparameters for the learning machine to improve the performance and effect of learning.
First, for the architecture of the DNN, we use a four hidden layers DNN as an example. Grid search was applied to help us change the number of nodes in each layer to make the model perform best. We don't want too many nodes in each layer, so we set the number of nodes in each layer in the range of
[1, 10]. The number of nodes in the first layer is equal to the number of independent variables. The output of the last layer is 1. The number of nodes in other layers is adjusted accordingly according to the number of layers.
Second, to avoid over-fitting and under-fitting problems, we evaluate the number of the number of trees, maximum depth of trees and number of attributes used by each tree in random forests (RF). We set 100 trees with maximum depth as 5 in this paper.
Third, the kernel of our SVM model is a Gaussian Kernel and we use grid search to tune the gamma and penalty factor . We set the range of and gamma as the general range and then narrow the range by a gradually reduced step. We set and gamma=0.1 in this paper.
Meanwhile, -fold cross validation is applied to each model in order to avoid overfitting, which means each model is carried out times with cross-validation. Each time, a model is built on data from -fold training data and is tested by the remaining 1-fold data. Considering that our data set is not very large, we set .
4.4.2 A Subsection
In our proposed method, LDA is first used to extract the text features hiding in the text of the industrial policies, and deep neural networks then are trained on the data, which include the text features and traditional numeric features for predict photovoltaic enterprise risk. The experimental results in this paper are shown in Table 3 and Figure 2.
Table 3 The performances of the LDA-based DNN model |
| Data & Model | Predictive Index |
| MAE | MAPE | RMSE |
| DNN + Traditional financial indicators | 0.332 | 0.570 | 0.312 |
| DNN + Traditional financial indicators + Policy topic indicators of current quarter | 0.163 | 0.400 | 0.221 |
| DNN + Traditional financial indicators + Policy topic indicators of last quarter | 0.275 | 0.521 | 0.293 |
Figure 2 The performances of DNN models |
Full size|PPT slide
It can be seen from Table 3 and Figure 2, the best results have obtained when we use traditional financial indicators and policy topic indicators of current quarter for prediction. Through -fold cross validation, the MAE value on the test set is 0.163; the MAPE value is 0.400; and the RMSE value is 0.221. When using traditional financial indicators alone, MAE value is 0.332, the MAPE value is 0.312. If the DNN model uses traditional financial indicators and policy topic indicators of last quarter, the MAE value is 0.275; the MAPE value is 0.521; the RMSE value is 0.293. The experimental results show that using traditional financial indicators with policy topic indicators of current quarter can strongly improve the performance of DNN while policy topic indicators of last quarter are also helpful.
4.4.3 Comparative Results
Some widely used data mining modes including linear regression, SVM and RF are used for comparison, and the prediction performances of the models are shown in Table 4 and Figure 3.
Table 4 The comparison of various LDA-based data mining model |
| Model | Predictive Index |
| MAE | MAPE | RMSE |
| DNN | 0.163 | 0.400 | 0.221 |
| Linear Regression | 1.453 | 1.113 | 0.498 |
| SVM (Gaussian Kernel, C = 1, gamma=0.1) | 0.748 | 0.822 | 0.491 |
| RF | 0.304 | 0.704 | 0.340 |
Figure 3 The performances of four models with LDA |
Full size|PPT slide
It can be seen from Table 4 and Figure 3, the accuracy of the results has been effectively improved when the policy subject indicators are added to predict the photovoltaic enterprise risk. It proves that policy do significantly affect the photovoltaic enterprise risk, which can help us better predict the photovoltaic enterprise risk. Moreover, it can be seen from the results in the table, when using the traditional financial indicators and the policy indicator of the last quarter for risk prediction, the results are better than when using the traditional financial indicators alone. However, the results are poorer than when using traditional financial indicators and policy indicator of the current quarter. This shows that the policies of last quarter have a weak impact on the photovoltaic enterprise risk in this quarter. Therefore, industrial policy has an impact on the risk of photovoltaic enterprises, and there is a weak lag impact.
In order to better confirm the results, a -test is conducted on 4 results of 5-fold cross-validation, and the results shows that the MAE, MAPE and MAPE of model 1 is significantly lower than that of SVM, LR and RF (all , ), which is considered to be statistically significant. Therefore, we believe our proposed LDA-based deep learning model outperformed some widely used data mining models such as RF, SVM, and LR.
5 Conclusions and Future Work
In this paper, we research the impact of industrial policies on the photovoltaic enterprise risk based on topic model LDA and deep neural network technology. This paper extracts the ten topic features of China's photovoltaic policies, and on this basis, constructs a deep neural network model for photovoltaic enterprise risk prediction using enterprise traditional financial data and policy textual data. And we test the effectiveness of the policy topic features for the photovoltaic enterprise risk prediction through ablation experiments. The method proposed in this paper avoids the subjective factors of some experts. Moreover, we look at the impact of industrial policy on the photovoltaic enterprise risk prediction from a quantitative perspective, which enhances the credibility and persuasion.
According to the experimental results, the industrial policy can enhance the prediction results of the photovoltaic enterprise risk on the basis of traditional financial indicators. This shows that the photovoltaic enterprise risk is indeed affected by industrial policy. Through further research, this paper finds that there is a certain lag in the impact of policies on the photovoltaic enterprise risk. The industrial policy of last quarter still has a weak impact on the photovoltaic enterprise risk in this quarter. This shows that industrial policy generally plays a very important role in photovoltaic enterprises. Moreover, the new energy policies of the state can get good feedback and implementation, which can play a due role in regulating photovoltaic enterprises in time. Therefore, we can effectively predict the photovoltaic enterprise risk through the relevant industrial policy of the last quarter and the current quarter, and know the possible fluctuation trend of the photovoltaic enterprise risk. The above conclusions will bring important enlightenment to the subsequent influencing factors and prediction analysis of the photovoltaic enterprise risk.
In the future research, on the one hand, we can further study the impact of a specific industrial policy on the photovoltaic enterprise risk and conduct case analysis. On the other hand, we can study the impact of production costs, network public opinion and other factors on the photovoltaic enterprise risk, so as to provide effective analysis tools and decision support for the sustainable and healthy development of photovoltaic enterprises.
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
PDF(209 KB)
Figure 1 The proposed framework
Table 1 The attributes of data
