Abstract
The paper conducted an in-depth research of urban distribution system's resource integration process; It built the evaluation index system of this issue at the beginning. Later, it studied the normalization methods to these corresponding indicators. Then, it constructed the evaluation model of urban distribution system's resource integration based on generalized regression neural network. At the end of this paper, it conducted a numerical experiment example and made certain conclusions. The innovation of this paper can be described as that the evaluation model constructed in this paper provides an effective and convenient way to evaluate the urban distribution system's resource integration process.
Introduction
Urban Distribution Systems (UDS) are the critical infrastructure for city logistics, aiming to enhance enterprise competitiveness through integrated resource management. However, quantifying the effectiveness of such integration is a complex task. Traditional methods like the Delphi method or Analytic Hierarchy Process (AHP) are often hampered by subjective bias. While artificial neural networks offer a more objective alternative, standard backpropagation models can suffer from slow convergence or local optima. This research proposes a Generalized Regression Neural Network (GRNN) approach to create a comprehensive and computationally efficient evaluation model.
Methodology
Evaluation Index System and Normalization
The researchers established a robust evaluation index system comprising 22 specific indicators categorized into six primary dimensions: Economic, Technical, Resource, Policy, Social, and Environmental. Given the diverse nature of these indicators—ranging from qualitative policy assessments to quantitative energy consumption metrics—four distinct normalization formulas were applied to map all data points onto a standard [0, 1] scale, ensuring compatibility for neural network processing.
The GRNN Model Architecture
The core of this research is the construction of a Generalized Regression Neural Network. As a variant of the Radial Basis Function (RBF) network, the GRNN architecture is designed for rapid, one-way training without the iterative looping required by standard backpropagation.
The model consists of three main layers:
- Linear Input Layer: Receives the normalized indicator values.
- RBF Hidden Layer: Utilizes Gaussian functions to determine the distance between input patterns and training samples.
- Linear Output Layer: Aggregates the weighted signals to produce the final evaluation score.
This architecture prevents common issues like slow convergence and ensures that weights are uniquely determined by the training data, enhancing objectivity.
Experimental Results and Discussion
To validate the model, a numerical experiment was conducted using six UDS samples. Samples 1–4 were utilized for training, while samples 5 and 6 served as the test set. Benchmark "actual" scores were established using expert consensus to measure the network's predictive accuracy. The model was implemented using the MATLAB R2010a GRNN toolbox.
The Impact of the Smooth Factor
A critical finding of the study is the sensitivity of the GRNN to the "smooth factor" (σ). The performance of the network in approximating the target function is heavily dependent on this parameter.
As shown in the experimental results, a smaller smooth factor—specifically in the range of 0.1 to 0.2—yielded the lowest error rates for both the training and testing datasets. This precision allows the model to closely mirror expert evaluations while maintaining computational efficiency.
Conclusion
The GRNN-based model provides a powerful tool for the objective evaluation of UDS resource integration. By incorporating modern logistics metrics—including environmental sustainability and social aspirations—the framework offers a holistic view of urban logistics performance. While the current study relies on a small sample size, it demonstrates a clear path toward more automated, real-time evaluation systems in the context of smart city development. Future research should focus on scaling the model with larger, purely objective datasets to further validate its robustness.
Summary updated