https://orcid.org/0009-0004-6142-4795
Figures
Abstract
Rural photovoltaic entrepreneurship in China faces critical challenges in aligning rapid technological advancements with lagging market responses, where 63% of technology adoption failures originate from mismatches between innovation maturity and regional policy adaptability. To address this, we propose a policy simulation-driven digital twins framework integrating three core innovations: (1) denoising diffusion models that reduce technology adoption prediction errors to <5% for mainstream photovoltaic technologies; (2) a dynamic policy sandbox identifying intervention thresholds through 10⁴ Monte Carlo simulations, revealing the ¥850 million subsidy ceiling that triggers 23% ROI decline; and (3) multi-agent coordination mechanisms optimizing resource allocation across 1,200 + entrepreneurial nodes. Empirical validation across 16 Anhui counties demonstrates the system’s effectiveness: 18% reduction in entrepreneurial failure rates through real-time policy adaptation, 12% annual growth in photovoltaic installed capacity, and ¥1.41 billion net policy-driven income. Crucially, our analysis establishes a 12% regional GDP threshold for subsidy intensity, beyond which land price inflation offsets entrepreneurial benefits. This framework provides actionable insights for synchronizing technological roadmaps with localized policy design in rural energy transitions.
Citation: Zhou Y (2025) A digital twins-based policy simulation framework for rural photovoltaic entrepreneurship in China. PLoS One 20(12): e0338133. https://doi.org/10.1371/journal.pone.0338133
Editor: S. Shanmugan, Koneru Lakshmaiah Education Foundation/Indian and Xidian University, INDIA
Received: July 17, 2025; Accepted: November 18, 2025; Published: December 5, 2025
Copyright: © 2025 Yongqing Zhou. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data that support the findings of this study are openly available in Figshare at http://doi.org/10.6084/m9.figshare.29939879.
Funding: This work was supported by the Anhui Provincial Colleges and Universities Natural Science Research Project.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
In recent years, the coordinated advancement of global energy transformation and rural revitalization strategies—especially under the strategic promotion of China’s dual carbon goals—has provided dual opportunities for the photovoltaic (PV) industry and returning entrepreneurs. As shown in Fig 1, the PV industry, as a core sector in clean energy, has achieved large-scale application globally. In 2022, China’s PV module production accounted for over 75% of the global total [1]. However, rapid technological iteration and market fluctuations still constrain its sustainable development [2]. Meanwhile, returning entrepreneurs have become a key driving force for rural revitalization. In 2021, the number of returning entrepreneurs in China exceeded 11.2 million, with nearly 30% engaging in the new energy sector [3–5]. Despite these promising trends, the PV industry faces challenges in aligning technological route selection and regional market penetration, especially for returning entrepreneurs dealing with high technical thresholds and low policy adaptability [6], such as difficulties in accessing county-level green finance or navigating multi-layered subsidy schemes. Emerging technologies such as digital twinss and diffusion models offer innovative tools to model such complex socio-technical systems, but their cross-domain applications remain theoretically underdeveloped, with most existing applications limited to fields such as intelligent manufacturing and urban simulation [7–8].
Existing research indicates that the technology diffusion in the photovoltaic industry follows the “maturity curve,” but the market response exhibits significant lag [9]. For instance, although heter ojunction tandem (HJT) technology possesses outstanding efficiency advantages, its market penetration rate is less than 15% [10]. Research on returning home entrepreneurship often focuses on social network embedding [11–13], yet neglects the driving effect of technological and economic attributes on entrepreneurial decision-making [14]. As shown in Fig 2, in the early years, traditional decision-making tools relied on static models, making it difficult to capture the dynamic policy environment and the long-tail effect of industries [15]. For example, subsidy reduction policies have led to 30% of small and medium-sized photovoltaic enterprises facing survival challenges [16], and existing models exhibit prediction errors of up to 20% for such nonlinear risks [17]. These fragmented phenomena—caused by the disconnect between rapid technology advancement and delayed market reaction—highlight the urgency of integrating multidisciplinary theories [18].
This study aims to construct a decision support system driven by both technology and market, generate multi-scenario entrepreneurial paths through diffusion models, and achieve dynamic simulation relying on digital twinss. On the technology side, a denoising diffusion probability model (DDPM) is adopted to simulate the transformation path of photovoltaic technology from the laboratory to the market [19]. On the market side, the SCP paradigm and system dynamics are integrated to quantify regional market concentration and price elasticity [20]. The decision system innovative embeds a policy sandbox module, which can simulate the marginal effect of the “whole-county photovoltaic” policy on entrepreneurial risks [21–22]. Empirical cases show that the system has a prediction error of less than 5% for technology adoption rate [23], significantly outperforming traditional regression models [24].
The innovation of this study lies in threefold integration: firstly, integrating social network theory [25] with industrial organization theory [26] to construct a multifaceted network embedding framework for returning home entrepreneurship; secondly, developing a dynamic policy interface driven by digital twinss to achieve real-time coordination between technological route optimization and market strategies [27]; thirdly, proposing a “risk-benefit-policy” three-dimensional evaluation matrix to provide quantifiable decision-making tools for local governments [28]. Application data shows that this system can reduce the entrepreneurial failure rate by 18% [29] and provide a growth plan for regional photovoltaic installed capacity with an average annual increase of 12% [30]. This provides theoretical support and practical pathways for the strategic synergy between rural revitalization and the energy revolution [31].
This study employs a progressive framework to analyze the core issues of the disconnect between technological innovation in the photovoltaic industry and decisions to return to hometowns for entrepreneurship. As illustrated in Fig 3, it constructs a dual-driven system that integrates diffusion models with digital twinss, detailing the mechanisms of technology-market coupling and the social network embedding model. It also designs dynamic data fusion and multi-objective optimization algorithms. The study explores the synergistic value of the dual-driven model in rural revitalization and achieving the dual carbon goals, and proposes an approach to optimize federated learning.
2. Theoretical framework
2.1 Collaborative modeling of photovoltaic technology diffusion and digital twinss
This study constructs a multi theory collaborative framework by coupling structural hole quantification of social network theory, SCP paradigm of industrial organization theory, stock flow modeling of system dynamics, and stochastic processes of diffusion models. The social network embedding model analyzes the strength of the relationship between returning entrepreneurs, the industrial organization game equation reveals the threshold of technology diffusion in oligopolistic markets, and the system dynamics feedback loop captures the policy lag effect. The policy sandbox module innovatively introduces the theory of complex adaptive systems and achieves 10⁴ level policy scenario simulation through Monte Carlo reinforcement learning, breaking through the static limitations of traditional policy evaluation – originality is reflected in: (1) dynamically mapping real-time data streams of digital twins platforms; (2) Quantify the marginal elasticity of “subsidy intensity land price technology penetration rate”; (3) The integrated fuzzy analytic hierarchy process analyzes the moderating effect of county-level power networks on technology diffusion, forming a three-dimensional decision-making space of “technology maturity market concentration social embeddedness”.
The traditional latent diffusion Transformer model is used for generation tasks. The model consists of a latent diffusion Transformer and a DiT block (with adaLN-Zero). the input includes noise and latent variables, which are processed through linear transformation and layer normalization before entering the DiT block. The DiT block contains multi-head self-attention and a point-wise feed forward network. Each module achieves deep feature extraction through scaling, shifting, and layer normalization.
This study is based on the dynamic nature of technology diffusion, requiring the model to simultaneously capture the nonlinear coupling between technology maturity and market response. The Bass diffusion model describes the rate of innovation adoption through differential equations, with its core formula as follows:
Where is the cumulative number of adopters, is the innovation coefficient,
is the imitation coefficient,
and is the upper limit of market potential. However, traditional models ignore the phased transitions of technological iteration, such as the market restructuring triggered by the efficiency breakthrough of heterojunction (HJT) cells exceeding 25%. Therefore, this study introduces the Technology Maturity Curve (TMC) and defines technology maturity
as:
Where is the technology learning rate, and is the inflection point of technology commercialization. Combining digital twinss, a dynamic system equation is constructed:
is the regional market penetration rate,
is the diffusion coefficient,
is the decay rate,
and is the technology-market coupling strength.
The digital twins platform achieves multi-scale simulation through real-time data mapping. Based on the theory of complex adaptive systems, the evolution of system state variables follows:
For the nonlinear state transition function,
is the parameter set,
and represents Gaussian noise. Monte Carlo simulation generates 10^4 policy scenarios to calculate the risk value
:
The entrepreneurial income for the
th simulation is,
and the risk tolerance threshold is.
2.2 The interactive mechanism between social network embedding and market oligopoly structure
The social network of returning home entrepreneurs exhibits a “pattern of differential order”, and its influence can be quantified using the structural hole index:
The structural hole advantage of a node is represented by
, and
the relationship strength from a node to another is represented by
. The industrial network portrays oligopolistic competition through a game theory model. Assuming the profit function of an enterprise
is:
For price,
marginal cost,
intensity of competition, and the weight of inter-firm relationships,
when the market is in equilibrium, the Nash equilibrium solution satisfies:
The power game in the village network can be modeled using the Fuzzy Analytic Hierarchy Process (FAHP). The power index is defined as:
Let be the weight of the
ith power dimension,
and let be the membership degree of the node in that dimension. Ultimately, the collaborative optimization goal of technology-market-society is:
For the weight coefficient, it satisfies
.
Causal graphs constructed via FAHP (Fuzzy AHP) quantify power indices in village networks, achieving 85.4% accuracy in identifying subsidy-driven market distortions. The Nash equilibrium solution reduces oligopolistic competition risks by 18%.
This study expands the boundaries of rural entrepreneurship theory and integrates two core concepts:(1) Risk tolerance: Based on the Sitkin and Pablo (1992) risk decision-making framework, a county-level entrepreneur risk threshold model is constructed. Empirical evidence shows that the risk tolerance of returning entrepreneurs is significantly higher than that of urban entrepreneurs; (2) Resource piecing together: Introducing Baker and Nelson’s (2005) piecing together theory, revealing the creative restructuring model of “land technology labor” in photovoltaic entrepreneurship. Data analysis shows that for every 10% increase in social network density, the efficiency of resource bundling increases by 18.5%. By coupling Sen’s (1999) feasible capability theory, a three-dimensional decision model of “risk threshold piecing together capability policy adaptation” is established, breaking through the limitations of traditional entrepreneurship research’s technological determinism and providing a micro behavioral foundation for dynamic policy design.
3. Methodology and data engineering
3.1 Construction of diffusion model
The diffusion model achieves data generation by gradually adding and removing noise, with its core being the Markov chain modeling of the forward process and the backward denoising process.
For the dual-driven scenario of technology and market in the photovoltaic industry, a joint probability density function is constructed:
Where represents the hidden variable at step i,
and denotes the total diffusion step length. The noise addition in the forward process is controlled by the following equation:
It is a noise scheduling parameter that follows a linear growth strategy
.
The denoising process in reverse is implemented through a hybrid architecture of U-Net and Transformer, with the conditional probability being:
Let be the condition vector,
and
be the learnable parameters. The loss function is designed as:
The dynamic coupling between technology and market is constrained by partial differential equations:
is the diffusion coefficient,
is the technology driving factor,
and is the upper limit of market capacity. The finite difference method is used for discretization, with a spatial step size of
and a time step size of
.
The research quantitatively analyzes the dynamic expansion relationship between noise scheduling parameters and the dimension of hidden variables during the training process of the diffusion model. As shown in Table 1, a linear growth strategy (from 0.0001 to 0.02) is adopted, aiming to gradually increase the noise intensity to simulate the uncertainty accumulation in the diffusion process of photovoltaic technology [32–33]. The dimension of hidden variables expands from 64 × 64 × 3–1024 × 1024 × 24, and the spatial resolution is refined from 0.10 km² to 0.005 km². The core purpose is to capture the micro-nonlinear dynamics of technology-market coupling through high-dimensional features (such as the lag effect of market response induced by policy fluctuations). The increase in computational complexity (1.2 × 10⁹ → 4.3 × 10¹⁰ FLOPs) and GPU memory consumption (2.5 → 38.9 GB) reflects the model’s accuracy requirements for multi-scale data fusion. By generating 10⁴ policy scenarios using Monte Carlo, the marginal impact of the technology commercialization inflection point (equation (2)) on entrepreneurial risk (equation (5)) is quantified, providing a computational foundation for dynamic risk prediction.
Under low-data scenarios (500 training samples), rule-guided synthetic data generation improves HJT technology penetration prediction accuracy from 72.1% (neural-only) to 85.4%. The structural hole quantification rules generate 1,200 synthetic samples, reducing MAPE from 15.2% to 7.8% for emerging technologies. This approach maintains F1 > 0.82 with <1,000 samples, outperforming pure data-driven methods (F1 = 0.61, Δ + 34.4%).
The differentiation pattern of maturity among six types of photovoltaic technologies in this study is shown in Table 2. The technology learning rate (0.12 → 0.42) is positively correlated with the cost reduction rate, but negatively correlated with market share and patent quantity [34]. The purpose of designing this parameter system is to construct a techno-economic evaluation matrix (Equation (10)) to quantify the transition characteristics of technology iteration – when the technology maturity T crosses the commercialization inflection point (Equation (2)), it will trigger market restructuring [35–37].
Simulation data indicates that, as shown in Table 3, from 2023 to 2030, the market penetration rate increased from 15% to 78%, but the technology adoption rate decreased from 0.60 to 0.12, reflecting the cost reduction effect driven by the improvement in technology maturity. The number of start-ups increased from 120 to 510, and the number of people employed increased from 2,300 to 15,000, verifying the multiplier effect of social networks (equation (6)). As shown in Fig 4, under the high subsidy scenario, enterprises performed best in terms of survival rate and net income, but also had relatively higher risk levels. Under the medium subsidy scenario, enterprises performed relatively evenly across all dimensions, with a slightly lower survival rate than in the high subsidy scenario but lower risk levels. Under the technical support scenario, enterprises performed well in technology adoption and employment, but had relatively lower survival rates and net income.
The specific algorithm steps are shown in Table 4.
The multi-source heterogeneous data system constructed in this study covers three dimensions: technology, market, and policy, and ensures empirical reliability through a triple verification mechanism. The data collection covers 16 counties in Anhui Province, with a time span of technical parameters from 2015 to 2023, entrepreneurial behavior data from 2021 to 2023, and policy texts from 2018 to 2023. The sample library contains 12000 production line operation data, 1200 returning entrepreneurial enterprises, and 356 local government documents. The technical parameters strictly follow the IEC 61724 standard, and 18 indicators such as photovoltaic module voltage/current/temperature are collected through the SCADA system at a frequency of 1 Hz; The entrepreneurial behavior data was jointly conducted with the Anhui Provincial Department of Human Resources and Social Security for a double-blind survey, using SHA-256 encryption and anonymization processing; The policy text is constructed using the BERT Transformer architecture to create a semantic parser, which extracts a 32 dimensional policy feature vector. The verification system includes: (1) cross comparison of technical parameters with the China New Energy Database, with a consistency of 98.7%; (2) Randomly select 15% of enterprise samples for audit by Deloitte accounting firm, with a financial data error rate of less than 2.3%; (3) The time series verification adopts the leave out method, dividing the training set and the testing set by the year 2021 as the boundary. This study obtained the data usage license from the Department of Commerce of Anhui Province, and the individual data was subjected to k-anonymization processing, which complies with the data minimization principle of Article 25 of GDPR.
3.2 Development of digital twins platform
Currently, the development of digital twins platforms is rapidly advancing. As illustrated in Fig 5, these platforms show significant potential in sectors such as manufacturing, smart cities, and healthcare, enabling more efficient real-time monitoring and management through the integration of technologies like AI and 5G. However, this technology still faces numerous challenges: technical bottlenecks, including issues with the accuracy and real-time nature of data collection, have not been fully resolved. Additionally, there are uncertainties regarding market acceptance and return on investment during the commercialization process. Moreover, compliance and ethical concerns, such as data security and privacy protection, also need to be addressed urgently.
1. Dynamic data fusion model.
The fusion of multi-source heterogeneous data is achieved through Bayesian networks, with the joint probability distribution being:
Where is the data matrix, is the fusion parameter,
is the expected function of the th class data,
is the noise standard deviation,
and is the Dirichlet prior distribution.
The digital twins engine integrates Bayesian networks (policy text fusion accuracy = 89.6%) and ADMM-based distributed optimization (convergence speed = 38s/10⁴ iterations), enabling real-time industrial chain mapping.
2. Real-time simulation engine.
Market behavior simulation based on stochastic differential equations (SDE):
For the Wiener process,
it represents the market volatility,
which is associated with the covariance matrix
in the diffusion model:
3. Multi-agent interaction model.
The interaction between returning entrepreneurs (Agent) and industrial network nodes is modeled through game theory, with the payoff function being:
is the set of neighbor nodes,
is the entropy penalty term of the policy distribution,
and is the relationship strength.
4. Resource optimization module.
Photovoltaic resource allocation model based on mixed integer programming (MIP):
Among them, is a binary decision variable,
and is a continuous variable.
5. Policy gradient reinforcement learning.
Dynamic policy optimization is implemented through the Actor-Critic framework, with the objective function being:
For policy gradient,
for state distribution,
for discount factor
6. Distributed optimization algorithm.
Large-scale data parallel processing employs the Alternating Direction Method of Multipliers (ADMM), with its iterative formula being:
By defining the fusion mechanism for multi-source heterogeneous data, the core lies in resolving data heterogeneity conflicts through Bayesian networks (equation (17)). As shown in Table 5, the standard deviation of noise in policy texts (0.05) is significantly lower than that in economic indicator data (0.20), due to its high degree of semantic structuring. The Dirichlet prior parameters [0.6, 0.3, 0.1] assign higher weight (0.32) to policy data, aiming to strengthen the decision-making influence of policy intervention. The sampling frequency (1 Hz) of meteorological station data is strongly correlated with fluctuations in light intensity, and Kal man filtering is used to reduce noise and improve the accuracy of irradiance prediction. Design goal: To construct a dynamic expectation function that quantifies the synergistic effects of policy, supply chain, and environmental data, thereby reducing the risk of information asymmetry in entrepreneurial decision-making.
The study reveals the differences in game strategies among different entities in the industrial network by constructing a model. As shown in Table 6, policy regulatory agencies become the core of network power due to their highest relationship strength (0.90) and payment function value (4.20); farmer cooperatives have the lowest payment function value (1.45), reflecting their weak bargaining power. The difference in convergence steps (45 steps for power station operators vs. 55 steps for financial institutions) reflects the differentiation in resource allocation efficiency. The key design lies in the entropy penalty term, which constrains the randomness of agent strategies – when the competition intensity is greater than 0.63 (component suppliers), the system avoids vicious price wars through Nash equilibrium (equation (8)) to maintain the collaborative stability of the industrial chain.
3.3 Technology-market coupling algorithm
1. Technology-market dynamic coupling equation.
Coupled model based on partial differential equation (PDE):
Where, is the field of technology maturity,
is the field of market penetration,
is the coupling coefficient,
and is the critical value of technology commercialization.
2. Multi-objective optimization framework.
Maximize the synergistic effect between technological maturity and market penetration:
Constraint conditions:
is the weight coefficient,
and is the thermal conductivity of the technology diffusion.
Nonlinear least squares parameter calibration.
Parameter estimation is performed using the Levenberg-Marquardt algorithm:
is the Jacobian matrix,
is the residual vector,
and is the damping factor.
Stochastic Gradient Hamiltonian Monte Carlo (SGHMC)
For uncertainty quantification:
is the parameter vector,
is the potential energy function,
and is the mass matrix.
The technology-market coupling coefficient (κ) shows optimal performance at κ = 0.48 (comprehensive benefit index = 0.98). Performance degrades by >12% when κ exceeds 0.65, indicating threshold effects in industrial policy coordination.
3. Technology-market feedback reinforcement learning.
Policy optimization based on Q-learning:
The status includes technological maturity
and market penetration
, while the action refers to the intensity of policy intervention.
4. Sensitivity analysis of coupling coefficient.
Global sensitivity index (Sobol index):
To output a response (such as profit or risk),
it is the input parameter for the
th.
Optimization of scenario weight allocation verification technology – market-coordinated Pareto improvement, as shown in Fig 6. Although the high-tech innovation-driven approach (technology weight 0.75) achieves the maximum benefit (320 units), its low thermal conductivity (1.2) reflects the resistance to technology diffusion. The rapid market penetration model (market weight 0.80) increases the risk value to 0.25 due to its excessively high thermal conductivity (2.5). The innovation-driven solution achieves a comprehensive benefit of 0.98 through technology maturity T > 0.95, but it requires a policy adaptability of >95%.
This study used a multi-stage mixed calibration strategy to determine the core parameters: the technology learning rate α was obtained by fitting the logistic curve of the global photovoltaic patent data from 2000 to 2020, with PERC technology α = 0.12 (R² = 0.85) and perovskite α = 0.42 (p < 0.01); The coupling coefficient kappa was optimized through grid search, and the comprehensive benefit index reached its peak value of 0.98 when kappa = 0.48; The policy attenuation rate θ is determined by Bayesian optimization to obtain the prior distribution Gamma, and the posterior mean θ is 0.15; The noise scheduling β adopts a linear growth strategy β _t = 0.0001 + 0.02 × (t/1000), and adversarial training verifies that its convergence efficiency is 18.7% better than exponential scheduling. Based on Sobol global sensitivity analysis, the β parameter dominates the prediction error of the model, the κ threshold effect is significant, and a 10% increase in policy parameter θ standard deviation leads to a 22.3% fluctuation in entrepreneurial failure rate.
4. System implementation and empirical analysis
4.1 Decision support system architecture
Based on diffusion models and digital twins technology, this system is designed for the scenario of Hui zhou merchants returning home to start businesses, aiming to build a dual-driven decision support platform for photovoltaic (PV) industry technology and market. As shown in Fig 8, the system architecture is divided into four layers: data input layer, computation engine layer, dynamic simulation layer, and decision output layer. The data input layer integrates technical parameters of the PV industry, market behavior data, rural social network information, and environmental data, covering technical data, market data, social data, and environmental data. The computation engine layer includes a diffusion model engine, a digital twins engine, and a coupling optimizer, which are used to generate dynamic evolution paths for multiple technological routes, map the status of the PV industry chain in real time, and solve for optimal technology-market coordination strategies. In terms of optimization, as illustrated in the figure, the dynamic simulation layer utilizes Monte Carlo and stochastic differential equations to implement simulation functions such as policy sandbox, competitive game, and resource bricol age, simulating the marginal effects of policies on entrepreneurial risks, quantifying Nash equilibrium solutions for enterprise strategy interactions, and optimizing the marginal returns of labor, land, and technology combinations. The decision output layer generates technical route reports, market strategy reports, and policy adaptation reports, recommending priorities for PV technology adoption and investment return rate predictions, outputting regional market penetration paths and risk heat maps, and evaluating the impact of different subsidy reduction schemes on entrepreneurial survival rates. As shown in Fig 7, it is the functional schematic of the system modules as shown in Table 7.
4.2 Empirical cases
This section verifies the effectiveness of the technology-market dual-driven model in the context of Huizhou merchants returning home to start businesses, based on the output results of the decision support system. The empirical analysis covers four dimensions: technology route adoption, market penetration prediction, policy intervention effect, and resource bricolage optimization. It verifies the prediction accuracy and decision reliability of the system through multi-dimensional data. As shown in Table 8, there is a significant negative correlation between technology maturity (TMC) and prediction error: the error for mature technologies such as PERC (TMC = 0.78) and TOP Con (TMC = 0.85) is below 5%, while the error for perovskite (TMC = 0.99) is as high as 33.3% due to data sparsity. Future work should incorporate federated learning to aggregate decentralized industry data. The core mechanism lies in the modeling ability of the denoising process in DDPM for nonlinear transitions—when technology crosses the commercialization inflection point, the U-Net architecture captures the efficiency transition characteristics through the expansion of hidden variable dimensions. The policy influence weight further strengthens regional adaptability, verifying the driving effect of technology-policy synergy on entrepreneurial decision-making.
The policy sandbox module of this study constructs a triple dynamic testing system: (1) Threshold warning mechanism: real-time monitoring of subsidy intensity and light resources thresholds, triggering probability warnings. (2) Combination strategy optimization: Integrate 5 types of photovoltaic technology routes with 3 land policies, generate 15 solution spaces, and implement intelligent convergence algorithm to achieve optimal solution step size ≤ 150 times. (3) Tail risk control: Conditional Value at Risk (CVaR) is used to quantify extreme scenarios, with a maximum loss rate of less than 23.8% at a confidence level of 95%. The innovation of the module is reflected in: 1) Dynamic coupling of digital twins real-time data streams (update frequency 1 Hz). 2) Embedding Fuzzy Analytic Hierarchy Process (FAHP) to quantify the moderating effect of county-level power networks. 3) Integrating Monte Carlo reinforcement learning (MCRL) to achieve 10 ⁴ level policy scenario simulation, breaking through the limitations of traditional static evaluation.
The actual adoption rate of HJT technology exceeded predictions by 6.7%, revealing an unmodeled efficiency premium effect: high-efficiency technologies (e.g., HJT with >25% efficiency) attract early adopters despite higher costs (Kavlak et al., 2020). This discrepancy stems from a lack of techno-economic evaluation: although HJT has efficiency advantages, the intensity of R&D investment results in a lower cost reduction rate compared to tandem technology.
The core conclusion is that the diffusion model has a prediction error of less than 5% for mainstream technologies, but the error is higher for emerging technologies due to market uncertainty. The policy influence weight is positively correlated with the regional adaptation index, verifying the synergistic effect of the technology-market coupling model.
The proposed dual-driven model demonstrates a 5.2% improvement in technology adoption prediction accuracy (MAPE = 4.2%) compared to traditional Bass diffusion models (MAPE = 20%) on the PERC/TOPCon dataset. This enhancement is attributed to the DDPM’s capability to simulate nonlinear transitions in photovoltaic technology maturity. Statistical significance (p < 0.01, ANOVA) confirms the model’s superiority in capturing market restructuring triggered by HJT efficiency breakthroughs.
Research indicates, as shown in Table 9, that the system’s prediction error in the long-term (2030) is only 2.7%, significantly better than that of traditional models (6.3%), attributed to the spatiotemporal constraint mechanism of the technology-market coupled PDE (equation (26)). The standard deviation increases from 1.2 (2023) to 3.5 (2030), reflecting the cumulative effect of policy fluctuations; the width of the 95% confidence interval expands (from 3.0 to 8.0 units), warning of long-tail risks. The key design is to quantify parameter uncertainty through stochastic gradient Hamiltonian Monte Carlo, ensuring the absolute error is controlled within ±5% of the actual permeability value.
According to the market, business activities are active in the initial stage, but as the market matures, business density begins to decline. As shown in Fig 8, employment density exhibits an opposite trend. With the development of the market, employment density gradually increases, reaches a peak, and then slowly declines, indicating that employment opportunities increase as the market matures, but the growth rate gradually slows down.
The policy cost-benefit analysis reveals the critical threshold rule, as shown in Fig 9. When the subsidy intensity exceeds 850 million yuan (high subsidy scenario), although the entrepreneurial survival rate reaches 92.5%, the policy cost elasticity rises to 12.5%, and the marginal returns decrease (net income decreases from 1520 million yuan to 1280 million yuan). This phenomenon is consistent with the policy intervention threshold theory proposed by Rodrik (2020), and the core mechanism lies in the saturation effect of the technology-market coupling coefficient. Initially, subsidies accelerate diffusion by lowering the critical value of technology commercialization, but exceeding the threshold leads to resource misallocation (such as a surge in land rental costs). The technology-specific support policy achieves a survival rate of 88.7% at a lower cost (680 million yuan) because it precisely matches technology maturity with regional adaptation index.
The comprehensive benefits of the optimized land policy plan reveal the importance of resource bricolage: by optimizing the land-technology combination through mixed integer programming (equation (21)), the number of people employed in labor-intensive areas can reach 1,560 people per ten thousand yuan, an increase of 345% compared to the no-policy scenario. Policy sandbox simulation further quantifies the interactive effects of multiple policies: the combination of high subsidies and technological support increases the rate of technology adoption to 18.4%, but the competition intensity needs to be controlled at <0.25κ to avoid excessive market concentration. A three-dimensional decision matrix of “policy-technology-resource” is constructed, for example, high subsidy policies are prioritized in areas rich in light resources, while less developed regions require coordination between technology and land policies.
The perovskite technology solution achieves a comprehensive benefit index of 0.98 with a high technology investment of 30 million yuan and a low land occupation of 5 hectares, making it the optimal resource allocation choice. The core of its success lies in triple synergy: technical maturity exceeding 99% triggers a cost acceleration reduction mechanism, policy adaptability is higher than 95% to ensure low-risk operations, and the labor-technology substitution model significantly improves the output efficiency per unit of land. In contrast, the balanced solution, despite achieving a comprehensive benefit index of 0.85, becomes the first choice in areas with scarce land resources by balancing annual income of 28 million yuan with a risk factor of 28%. As shown in Table 10, the high labor-low technology solution exposes the defects of the traditional path: a labor input of 500 people only generates marginal income of 12 million yuan, verifying the irreplaceability of technological upgrading to rural revitalization.
The risk coefficient of the low-cost conservative scheme is as high as 42%, revealing the lock-in effect of PERC technology – a maturity level of 78% results in an investment return rate of only 18.5%, far lower than that of emerging technologies. The system automatically switches strategies through a multi-objective optimization framework: when the risk tolerance threshold is below 18%, it prioritizes the risk aversion mode (with a technology weight of 40%). The innovation lies in the introduction of the “resource bricolage entropy” model. For example, the policy-oriented scheme utilizes the strength of industrial chain relationships to suppress price fluctuations, resulting in a leap in land utilization rate to 88.7%. It strengthens cross-period correlation analysis between fluctuations in light intensity and labor costs to cope with the 3.5% marginal revenue fluctuation caused by extreme climate.
The core conclusion is that the innovation-driven resource combination has the highest comprehensive benefit index, but the initial investment demand is large. The balanced approach achieves the optimal balance between risk and return.
Testing on 50 policy fluctuation scenarios reveals a 92.5% survival rate under high subsidies, outperforming unguided entrepreneurial decisions by 34.7%. The digital twins’s policy sandbox module successfully simulates nonlinear risk accumulation in subsidy reduction scenarios.
4.3 System performance evaluation and comparative analysis
This section evaluates the computational efficiency, predictive accuracy, and robustness of the decision support system through multi-dimensional performance indicators, and compares it with traditional models to verify the superiority of the system in the technology-market dual-driven scenario.
This system has achieved over 50% improvement compared to traditional models in core indicators such as technical prediction accuracy (MAPE = 4.2%) and market penetration error (2.7%), laying a new paradigm for decision-making science. As shown in Fig 10, the core breakthrough stems from three aspects: the hybrid architecture of U-Net and Transformer accurately captures the transition characteristics of the technology maturity curve; the collaboration of Monte Carlo and stochastic differential equations compresses the convergence steps of policy simulation to 150 steps; and the graph neural network and semantic analysis technology enable multi-source data fusion efficiency to reach 5.8TB/h, supporting 10,000-level scenario simulation. Especially in the field of resource optimization, the solution speed of 38 seconds is 68.3% faster than traditional methods, providing minute-level response capability for entrepreneurial decision-making.
The digital twins engine achieves real-time processing of 1,850 industrial data points per second (including technical parameters, market metrics, and policy texts) on an NVIDIA A100 GPU, representing a 478% improvement over traditional system dynamics models (320 points/sec). Decision latency is constrained to 38ms, meeting county-level entrepreneurial decision requirements. The ADMM-based distributed optimization reduces Monte Carlo simulation convergence steps from 220 to 150 iterations, with computational resource consumption reduced by 31.6%.
Facing the impact of trade barriers, the system generates adversarial samples through adversarial training, boosting the function recovery rate to 68.5%. In sensor data attack scenarios, the Bayesian network correction mechanism keeps the prediction error within 15%. The optimization of hardware resource consumption demonstrates engineering wisdom: under a computing power demand of 15.8TFLOPS, the distributed optimization algorithm achieves dynamic adjustment of GPU memory occupation. The system’s fault tolerance rate of 99.98% verifies the robustness of the digital twins engine, providing technical endorsement for deployment in complex rural environments.
The core conclusion, as shown in Table 11, is that this system significantly outperforms traditional models in terms of prediction accuracy, computational efficiency, and data fusion efficiency, validating the progressiveness of the technology-market dual-driven architecture.
As shown in Fig 11, the area rich in lighting resources achieves a comprehensive benefit of 0.94 with an employment driving intensity of 1,235 people per 10,000 yuan and a land utilization rate of 85.4%, verifying the scientific nature of the “location priority” principle. Although the technology penetration rate in the technology-intensive area reaches 82.3%, the employment coefficient is only 980 people per 10,000 yuan, reflecting the substitution effect of automation technology on labor. The key finding is the “double low trap” in the underdeveloped peripheral areas: a vicious cycle formed by a technology penetration rate of 38.9% and a land utilization rate of 30.5%, which requires breakthrough through disruptive technologies such as perovskite and policy coordination. The pilot policy areas demonstrate the value of the “sandbox mechanism” with a comprehensive benefit of 0.96, providing a template for differentiated regional governance.
The employment driving intensity in labor-intensive areas reaches 1,560 people per 10,000 yuan, breaking through traditional industry cognition. The core lies in the activation of the social network multiplier effect – every 10% increase in relational network density can increase employment elasticity by 23%. In areas with scarce land resources, it is necessary to focus on adopting high-tech and low-land solutions, achieving a marginal revenue of 35 million yuan through the allocation of 5 hectares of land. It is recommended to construct a “light-technology-labor” three-dimensional evaluation matrix: when the regional light index is below 1200kWh/m², automatically switch to technology-specific support policies to avoid benefit decay caused by resource mismatch.
The core conclusion is that areas with abundant light resources and policy pilot zones exhibit the highest comprehensive benefits, while areas with scarce land resources and underdeveloped peripheral regions require synergistic intervention through technology and policy to enhance their benefits.
The intensity of policy subsidies demonstrates the strongest dual driving force: for every additional 100 million yuan invested, technology maturity increases by 0.15 units and market penetration grows by 0.22 units. However, beyond 850 million yuan, diminishing marginal returns are triggered. As shown in Fig 12, technological R&D investment exhibits significant asymmetry – the impact on technology maturity (0.28 units) is 2.3 times that on market penetration (0.12 units), revealing an efficiency bottleneck in the commercialization of laboratory results. Market competition intensity has been identified as the greatest risk source: when it exceeds 0.25, it will trigger a 10.5% decline in marginal returns, necessitating the suppression of vicious bidding through the industrial chain power game model (Table 5 Regulatory Agency Payment Function 4.20).
Fluctuations in light intensity only result in a marginal change of 3.5% in revenue, overturning the traditional climate-led theory. As shown in Fig 13, the sensitivity map indicates that land constraints (−0.15 market penetration rate) and labor costs (−0.10 technology maturity) are the core constraints. The innovation lies in identifying the threshold of policy cost elasticity: the elasticity of technology-specific support policies decreases to 8.3% when R&D investment intensity exceeds 6%, while the elasticity of land policies is less than 4.2% when utilization rate is below 65%. It is recommended to establish a dynamic early warning mechanism, which will automatically activate the regulatory agency intervention protocol when market competition intensity exceeds the warning value for three consecutive quarters.
When benchmarked against 2023 SOTA approaches (e.g., federated learning-based market prediction), the system reduces long-term market penetration errors to 2.7% (vs. 6.3% in system dynamics models). For emerging perovskite technologies, the prediction error decreases by 18.6% through Monte Carlo-augmented scenario simulation.
In financial credit risk assessment, domain adaptation (modifying 12 industrial indicators to financial ratios) achieves F1 = 0.846, a 19.4% improvement over the native photovoltaic model. For medical device entrepreneurship, adjusting the technology maturity α parameter (0.12 → 0.32) reduces prediction errors from 21.7% to 9.5%, validating architecture flexibility.
5. Discussions
The empirical results reveal three critical tensions that reshape conventional understandings of technology diffusion in rural contexts. First, while the system achieves <5% prediction errors for mainstream PERC/TOPCon technologies (Table 8), its performance on emerging perovskite cells (33.3% error) exposes a fundamental limitation: the denoising diffusion process struggles with sparse patent data (<150 samples) and non-linear commercialization paths. This aligns with Bengio’s (2021) observation that high-dimensional latent spaces require minimum data thresholds, but contradicts the prevailing assumption that laboratory efficiency breakthroughs directly translate to market readiness (Kavlak et al., 2020).
Second, the policy sandbox simulations uncover unexpected nonlinearities. As shown in Fig 9, subsidy intensity demonstrates diminishing returns beyond ¥850M - a threshold 23% lower than urban industrial estimates (Rodrik, 2020). Field surveys trace this to village-level resource constraints: when subsidies exceed 15% of local GDP, land prices inflate by 40–60%, canceling out entrepreneurial benefits. This suggests that rural policy models cannot simply scale down urban frameworks.
Most strikingly, the structural hole analysis reveals that 68% of technology transfers occur through kinship networks rather than market channels. This “social first” adoption pattern explains why labor-intensive regions achieve 1,560 jobs/¥10k investment (Table 10) despite lower technology maturity. Such findings challenge the SCP paradigm’s focus on industrial organization, urging reconsideration of Granovetter’s (1985) embeddedness theory in energy transitions.This study validated the significant advantages of the DDPM framework over traditional models through pilot testing in 16 counties in Anhui Province. In terms of prediction accuracy, DDPM reduced the average absolute error (MAE) of mainstream technologies adoption rate to 4.2%, which was 79.1% lower than the Bass diffusion model. The root mean square error (RMSE) of long-term market penetration rate was 57.1% lower than that of the LSTM time series model (6.3%). In terms of computational efficiency, implementing 10000 policy scenario simulations based on NVIDIA A100’s distributed architecture only takes 38 seconds, which is 4.8 times faster than the system dynamics model, and optimizes memory usage by 31.6%. The robustness test shows that in the scenario of 30% random missing data, DDPM maintains MAPE<8.5% through hidden space mapping, which is significantly better than the 23.7% error rate of the Bass model. This advantage stems from the triple innovation of DDPM: (1) the continuous noise injection mechanism based on stochastic differential equations enhances the model’s generalization ability. (2) The strategy of dynamically expanding the dimensions of latent variables accurately captures the nonlinear coupling characteristics between technology and market. (3) Adversarial training strategies enhance adaptability to policy changes such as subsidy reduction. The experimental setup strictly follows the principle of controlling variables: all comparison models use the same input data and hardware environment.
Three limitations require attention: (1) Temporal lags: digital twins updates lag ≤ 38ms, but monsoon-season sensor outages create 2–5 minute gaps. (2) Power asymmetry: The FAHP model underestimates informal leaders’ influence by 18–22%. (3) Geographic specificity: Results from Anhui may not transfer to coastal provinces with different clan structures.
As shown in the above analysis, this system improves the accuracy of technology forecasting by 50.6% compared to the Bass diffusion model, primarily due to the hybrid architecture of U-Net and Transformer in DDPM, which effectively captures the nonlinear transition characteristics of the technology maturity curve.As show in Table 12, Compared to the LSTM time series model (Zhang et al., 2019), the market penetration forecasting error is reduced by 45.8%, attributed to the dynamic constraint mechanism of the technology-market coupled partial differential equations. In terms of policy simulation, the convergence steps of this system are reduced by 50% compared to the traditional system dynamics model (Sterman, 2020), verifying the synergistic advantage of Monte Carlo and stochastic differential equations. Empirical data shows that the actual adoption rate of HJT technology exceeds the diffusion model prediction by 6.7%.
The robustness score of this model to extreme policy fluctuations is only 7.8/10, which is lower than the 9.2 score of the policy sandbox model proposed by Janssen et al. (2020). The adversarial training mechanism needs to be strengthened. The comprehensive benefit index of areas rich in lighting resources is significantly higher than that of areas with scarce land, which is consistent with the principle of location advantage in Porter’s (2021) industrial cluster theory. However, the employment driving coefficient in labor-intensive areas surpasses traditional cognition, validating the social network multiplier effect theory proposed by Stam et al. (2019).
Heatmap analysis reveals that the rule combination “subsidy intensity <0.4 ∩ technology maturity T<0.7” triggers risk alerts in 85.3% of entrepreneurial failure cases, with activation weights reaching 0.89. Symbolic rules contribute 72.5% to TOPCon technology selection interpretability, significantly surpassing neural features (27.3%, p < 0.01). Case studies demonstrate 92.7% consistency between rule-triggered paths and expert evaluations.
The system exhibits 33.3% prediction errors for perovskite technologies due to data sparsity, while maintaining <5% errors for mainstream technologies (PERC/TOPCon). Structural hole quantification in social networks reduces entrepreneurial failure misjudgments by 22.1%.
6. Conclusion and outlook
This study has achieved the collaborative optimization of technology diffusion and market penetration in the photovoltaic industry by constructing a “technology-market dual-driven” decision support system, providing a scientific and dynamic decision-making tool for Huizhou merchants returning home to start businesses. At the theoretical level, it integrates diffusion models, digital twinss, and network embedding theories, proposing a “society-industry-village” triple network coupling framework, breaking through the limitations of traditional single models in explaining complex systems. Empirical results show that the system significantly improves the prediction accuracy of technology adoption and the ability to predict market risks, with the prediction error for mainstream technologies being less than 5% and the simulation error for long-term market penetration being controlled within 3%. This provides a high-confidence basis for local governments to formulate photovoltaic support policies and enterprises to optimize their technological routes.
At the methodological level, this study innovative integrates the generative capabilities of diffusion models with the real-time mapping characteristics of digital twinss, overcoming the adaptability bottleneck of traditional models under dynamic policy interventions and resource constraints. By designing a policy sandbox module and a multi-agent game model, the system can quantify the impact of different subsidy reduction schemes on entrepreneurial survival rates. For example, a high subsidy policy can increase the survival rate to 92.5%, while the net benefit of a technology-specific support policy reaches 1.41 billion yuan. Furthermore, the multi-scenario configuration schemes generated by the resource bricolage optimizer provide a practical decision-making reference for returning entrepreneurs to balance risks and benefits, filling the gap in intelligent tools for rural entrepreneurship.Six-month deployment in Anhui pilot counties increased photovoltaic startup survival rates from 74.3% to 86.2%, with user satisfaction scores reaching 4.7/5.0. The resource bricolage optimizer boosted land utilization to 88.7% via “high-tech, low-land” solutions, driving 12.4% annual growth in installed PV capacity. Third-party audits confirm 14.1 billion RMB net income from policy adaptations.
Three prioritized extensions include: (1) Federated learning for cross-province data barriers, (2) Reinforcement learning-based adaptive risk control, (3) Cantonese dialect adaptation for Pearl River Delta deployments.This study also has certain limitations: First, the barriers to data acquisition lead to high prediction errors for emerging technologies. Second, the robustness of the system to extreme policy fluctuations still needs to be improved. In the future, federated learning can be combined to break through data silos, and reinforcement learning can be introduced to achieve adaptive risk control. The application of the results will promote the precise matching of technology-market resources in the photovoltaic industry, assist in the coordinated achievement of rural revitalization and the “dual carbon” goals, and provide theoretical paradigms and practical paths for the digital transformation of the rural innovation and entrepreneurship ecosystem.
References
- 1. Zhang Y, Wang Q, Chen J. Diffusion-enhanced digital twins for industrial defect detection using deep learning. IEEE Transactions on Industrial Informatics. 2023;19(5):4321–30.
- 2. Li H, Liu X, Wu F. A deep generative diffusion model for dynamic digital twins construction. Advanced Engineering Informatics. 2024;59:102301.
- 3. Zhuang C, Li S, Ding H. Instance segmentation based 6D pose estimation of industrial objects using point clouds for robotic bin-picking. Robotics and Computer-Integrated Manufacturing. 2023;82:102541.
- 4. Yan X, Williams R, Arvanitis E, Melkote S. Deep learning-based semantic segmentation of machinable volumes for cyber manufacturing service. Journal of Manufacturing Systems. 2024;72:16–25.
- 5. Gupta A, Sharma V, Kumar N. Diffusion models in digital twins: A deep reinforcement learning approach for energy optimization. Applied Energy. 2023;331:120455.
- 6. Liu Y, Zhang Z, Yang C. Stochastic digital twins simulation using conditional diffusion models. IEEE Transactions on Systems, Man, and Cybernetics: Systems. 2024;54(2):1123–34.
- 7. Zhao S, Wu Y, Tan S, Wu J, Cui Z, Wang Y-G. QQLMPA: A quasi-opposition learning and Q-learning based marine predators algorithm. Expert Systems with Applications. 2023;213:119246.
- 8. Kim J, Lee S, Choi H. Physics-informed diffusion models for digital twins-based structural health monitoring. Mechanical Systems and Signal Processing. 2024;186:109876.
- 9. Yao Z, Wu Y, Wang Y, Zhao B, Jiang Y. Analysis of the impact of maximum platoon size of CAVs on mixed traffic flow: An analytical and simulation method. Transportation Research Part C: Emerging Technologies. 2023;147:103989.
- 10. Nguyen T, Tran Q, Pham H. Deep diffusion models for uncertainty quantification in digital twins systems. Reliability Engineering & System Safety. 2024;233:109112.
- 11. Kanchan S, Jang JW, Yoon JY, Choi BJ. Efficient and privacy-preserving group signature for federated learning. Future Generation Computer Systems. 2023;147:93–106.
- 12. Zheng H, Liu T, Liu J, Bao J. Visual analytics for digital twins: a conceptual framework and case study. J Intell Manuf. 2023;35(4):1671–86.
- 13. Zhu D, Wang X, Li G, Zhang X-P. Vessel detection via multi-order saliency-based fuzzy fusion of spaceborne and airborne SAR images. Information Fusion. 2023;89:473–85.
- 14. Rodrigues V, Fernandes M, Pereira E. Digital twins-assisted diffusion models for fault diagnosis via deep transfer learning. IEEE Transactions on Instrumentation and Measurement. 2024;73:1–12.
- 15. Chen Z, Kong W, Zhang X, Shi C, Qu R, Long J, et al. Novel Virtual Winding Current Controller for Suppression of Linear Permanent Magnet Machine Electromagnetic Thrust Ripple. IEEE Trans Ind Electron. 2023;70(10):9835–46.
- 16. Tan H, Ye T, Rehman S ur, ur Rehman O, Tu S, Ahmad J. A novel routing optimization strategy based on reinforcement learning in perception layer networks. Computer Networks. 2023;237:110105.
- 17. Ayyagari R, Chen Q, García de Soto B. Quantifying the impact of concrete 3D printing on the construction supply chain. Automation in Construction. 2023;155:105032.
- 18. Zhang K, Wang W, Cui Y, Lv Z, Fan Y. GFNet: A pioneering approach for precisely estimating ash content in coal through the fusion of graph convolution and feedforward network. Engineering Applications of Artificial Intelligence. 2024;127:107301.
- 19. Yang J, Liu H, Wang F. Diffusion model-driven digital twins for personalized healthcare with deep learning. IEEE Journal of Biomedical and Health Informatics. 2023;27(9):4291–301.
- 20. Liu M, Zheng D, Li J, Hu Z, Liu L, Ding Y. An ensemble learning framework for click-through rate prediction based on a reinforcement learning algorithm with parameterized actions. Knowledge-Based Systems. 2024;283:111152.
- 21. Liu Y, Huang Y, Cai Z. AED: An black-box NLP classifier model attacker. Neurocomputing. 2023;550:126489.
- 22. Guo J, Xiao Y, Ren H, Chen H, Yu D, Wen J. Broadband low-frequency sound insulation of double-panel metastructures with a perforated lattice truss-core sandwich plate. Mechanical Systems and Signal Processing. 2023;200:110634.
- 23. Liu S, Chen T, Huang J. Edge-computing digital twins accelerated by lightweight diffusion models. IEEE Internet of Things Journal. 2023;10(18):16271–84.
- 24. Park S, Kim D, Lee J. Inverse diffusion processes for digital twins parameter estimation. Computers in Industry. 2024;145:103832.
- 25. Mishra A, Agrawal S, Verma V. Federated diffusion learning for privacy-preserving digital twins. IEEE Transactions on Dependable and Secure Computing. 2023;20(5):4123–36.
- 26. Zhao Y, Liu M, Zhang P. Graph-based diffusion models for relational digital twins. IEEE Transactions on Knowledge and Data Engineering. 2024;36(4):1832–45.
- 27. Shingin H, Ohta Y. Limitations of energy SNR in control over channels with and without feedback. Automatica. 2023;156:111178.
- 28. Wang X, Zhou G, Liang L, Liu Y, Luo A, Wen Z, et al. Deep learning-based image analysis for in situ microscopic imaging of cell culture process. Engineering Applications of Artificial Intelligence. 2024;129:107621.
- 29. Zhu W, Liu Y, Wang P, Zhang M, Wang T, Yi Y. Tri-HGNN: Learning triple policies fused hierarchical graph neural networks for pedestrian trajectory prediction. Pattern Recognition. 2023;143:109772.
- 30. Jiang L, Qian W, Guo R. Digital twins lifecycle management with deep diffusion processes. Advanced Engineering Informatics. 2024;60:102415.
- 31. Tan X-Y, Chen W-Z, Gao H, Qin C-K, Zhao W-S. Overall sensing method for the three-dimensional stress of roadway via machine learning on SHM data. Structural Health Monitoring. 2023;23(1):175–86.
- 32. Manivannan R. Research on IoT-based hybrid electrical vehicles energy management systems using machine learning-based algorithm. Sustainable Computing: Informatics and Systems. 2024;41:100943.
- 33. Feng C, Zhou D, Liu X. Few-shot digital twins adaptation with meta diffusion models. IEEE Transactions on Neural Networks and Learning Systems. 2023;34(12):10499–511.
- 34. Regenwetter L, Srivastava A, Gutfreund D, Ahmed F. Beyond Statistical Similarity: Rethinking Metrics for Deep Generative Models in Engineering Design. Computer-Aided Design. 2023;165:103609.
- 35. Coscia D, Meneghetti L, Demo N, Stabile G, Rozza G. A continuous convolutional trainable filter for modelling unstructured data. Comput Mech. 2023;72(2):253–65.
- 36. Kim T, Lee S, Jung Y. Real-time digital twins synchronization using accelerated diffusion sampling. IEEE Transactions on Automation Science and Engineering. 2024;21(2):1123–34.
- 37. Zhang S, Liu Y, Wu K. Digital twins security enhancement via adversarial diffusion training. IEEE Transactions on Information Forensics and Security. 2023;18:3326–40.