Predicting the factors influencing construction enterprises’ adoption of green development behaviors using artificial neural network

The paper addresses the environmental challenges posed by China’s construction industry, which contributes substantially to national output and carbon emissions. It seeks to identify and predict the key factors influencing construction enterprises’ adoption of green development behaviors. Grounded in the Technology-Organization-Environment (TOE) framework, the study builds an index system of influencing factors, then uses data from 2000–2020 to develop an artificial neural network (ANN) prediction model. The research questions include: (1) What factors influence construction enterprises’ adoption of green development behaviors? (2) Which factor is most critical for enterprises and government attention? (3) What is the future evolution trend of such adoption? The study’s importance lies in extending green development behavior research to construction enterprises, informing enterprise strategy and government policy for accelerating green transformation.

The literature review frames enterprise green development behavior as closely tied to environmental protection and sustainable economic development. Prior studies have examined green behaviors such as green purchasing, procurement, consumption, and management across multiple industries, with relatively fewer focusing on construction enterprises. The TOE framework (Tornatzky & Fleischer, 1990) categorizes determinants of technology adoption into technological, organizational, and environmental factors and has been widely applied in IT adoption research. Drawing from this, the paper posits that construction enterprises’ green development behavior is similarly influenced by: (a) technological factors (e.g., technological level, innovation capabilities), (b) organizational factors (e.g., enterprise size, resources), and (c) environmental factors (e.g., market conditions, enterprise demand/competitiveness, government behavior). The authors synthesize indicators into a multi-level index system: 3 first-level, 5 second-level, and 12 third-level indicators (covering equipment, R&D, learning rates; size, employees, assets; urbanization, real estate value, construction output, building area, completed area, and number of policies). This structure operationalizes TOE for quantitative modeling in the construction sector.

Data sources: The study compiles annual data (2000–2020) on China’s construction enterprises from the National Bureau of Statistics (NBS), CNKI for government policy counts, and CNIPA for patents. The target variable (P) is the annual number of patents related to green construction (search terms include green building, building energy saving, low carbon building, environmentally friendly building, sustainable building, ecological building). Environmental variables include urbanization level, total output value of real estate, total output value of the construction industry, building construction area, completed housing area, and number of policies. Organizational variables include number of enterprises, number of employees, and total assets. Technological variables include total power of construction machinery and equipment, R&D expenditure, and (initially) technological learning rates. Measurement of technological learning rates: Following Yang et al. (2012) and Zhang et al. (2020), a dynamic two-factor measurement model is used. Dependent variable: cost per unit construction area (=(fixed asset investment + main business cost − fixed asset depreciation)/total construction area), deflated using the GDP deflator. Independent variables: cumulative construction area and processed cumulative R&D expenditure (with 3-year lag and 20% R&D depreciation). EViews 10 regression estimates inform calculation of experience-based and research-based technological learning rates. Modeling approach: IBM SPSS Modeler 18.0 is used to build an ANN prediction model using a multilayer perceptron (MLP) algorithm (chosen over RBF for stronger predictive capability). Prior to modeling, Pearson correlation analysis (IBM SPSS Statistics 27) evaluates associations between inputs and the target. Based on low/insignificant correlations with the target, five variables are excluded: total number of machinery and equipment, technical equipment rate, power equipment rate, technological learning rate (experience), and technological learning rate (research). The final 12 input variables retained are: total power of equipment; R&D expenditure; number of enterprises; number of employees; total assets; total profits and taxes; urbanization level; total output value of real estate; total output value of the construction industry; building construction area; completed housing area; and number of policies. The target is number of patents. Training/testing partitioning: Data are randomly split using nine ratios (10:90 through 90:10). Model performance is assessed for each, and the 60:40 split (model 6) achieves the highest accuracy (99.6%), and is selected for analysis. ANN structure: The best model comprises an input layer (10 predictive variables indicated in figure, plus bias), a hidden layer with 7 neurons (plus bias), and an output layer with 1 target variable. Variable importance is derived from learned connection weights to assess factor influence.

- Variable selection: Pearson analysis led to exclusion of five variables (total number of equipment, technical equipment rate, power equipment rate, and both technological learning rate measures). Twelve inputs remained for ANN modeling. - Model performance: Across nine train/test splits, the best accuracy (99.6%) was obtained with a 60:40 split. For the target variable (number of patents), the linear correlation between actual and predicted values was 0.903 on the testing set (0.998 on training). Testing set errors: minimum −152.688, maximum 456.234, mean 38.786, mean absolute error 132.201, standard deviation 201.808. - Importance of predictors (to number of patents): total output value of construction industry (0.18), R&D expenditure (0.17), number of construction enterprises (0.16), total output value of real estate (0.13), total power of machinery and equipment (0.12), number of employees (0.08), building construction area (0.04), urbanization level (0.03), number of policies (0.02), total profits and taxes (0.02). These indicate environmental, technological, and organizational drivers. - Trend: The number of green-related patents increased over 2000–2020, indicating an upward trend in construction enterprises’ adoption of green development behaviors. - Core conclusion: Market share (proxied by total output value of construction industry) is the most influential factor; technological innovation investment (R&D) and enterprise size (number of enterprises) also play major roles.

The findings directly address the research questions. First, the study identifies and quantifies the key technological, organizational, and environmental factors influencing green development behavior adoption, aligning with the TOE framework. Second, among all factors, market share (environmental factor; measured by total output value of the construction industry) emerges as the most important driver, suggesting that competitive positioning and market dynamics strongly incentivize green innovation. Third, the observed increase in green-related patents over time evidences a growing adoption trend, and the ANN’s reliable predictive performance (testing correlation 0.903) supports the model’s utility for forecasting. The results corroborate extant literature that environmental pressures and opportunities (market share, policies), internal technological capabilities (R&D), and organizational attributes (enterprise size, workforce) shape enterprise environmental/green behaviors. Validating TOE within the construction context broadens the framework’s applicability beyond traditional IT adoption, highlighting policy leverage points (e.g., stimulating market demand and supportive regulations) and enterprise strategies (e.g., boosting R&D and scaling capabilities) to accelerate green transitions.

The study constructs a TOE-based index system and an ANN prediction model to analyze determinants of green development behavior adoption by China’s construction enterprises (2000–2020). Key contributions include: (1) Empirical evidence that adoption shows a rising trend, as proxied by increasing green-related patents; (2) Identification of market share (total output value of construction industry) as the most influential factor, followed by R&D expenditure and enterprise size; (3) Validation that technological, organizational, and environmental dimensions collectively explain adoption, consistent with the TOE framework. Theoretically, the paper extends green development behavior research to the construction sector and demonstrates ANN’s applicability for predicting factor importance in this domain. Practically, it suggests that enterprises should enhance green technological innovation and optimize organizational scale/resources, while governments should foster enabling market and policy environments to encourage green adoption.

- Data and imputation: Despite high overall accuracy, large prediction errors exist in the testing set (max error 456.234). Mean imputation for missing time-series data likely introduced bias; future work should obtain complete series or use time-series-appropriate imputation. - Scope of indicators: The study focuses on quantifiable indicators; qualitative organizational factors (e.g., employee attitudes) were not included. Incorporating mixed methods could yield richer insights. - Generalizability: The analysis is limited to China’s construction enterprises; extending to multiple countries would enhance external validity. - Methods: Only ANN was employed; future research could triangulate with Bayesian networks, decision trees, or hybrid models to improve robustness and accuracy.