A Sustainable Performance Assessment System for Road Freight Transport Based on Artificial Neural Networks

Before creating our sustainable performance measurement model, we conducted a thorough review of existing literature. As stated in our original paper [9], we categorized the models that evaluate sustainable performance in the field of transport. These models take into account different dimensions in their assessment. Out of the 17 models we analyzed [9], we compared them based on their applicability in other contexts to identify the most robust model [10]. We chose this model as a starting point since it considers the four most significant dimensions related to sustainability in its evaluation [10]. We further enhanced the model by expanding it from 4 dimensions (economic, social, environmental, operational) and 16 fields to 5 dimensions and 25 fields, as depicted in Figure 1. The Table 1 and Table 2 below summarize the main findings and highlight the dimensions that each model considers.

1.png

Figure 1. Multidimensional sustainability performance measurement model

Table 1. Sustainable performance assessment models comparison

It is crucial to acknowledge and integrate the needs and concerns of stakeholders who have a direct interest in a company's development strategies and plans [27]. This is because such stakeholders are directly impacted by the efforts made towards sustainable development. Hence, we recognized it as important to include them as an integral dimension in the model we are developing. We have identified five dimensions that are essential in our model, namely economic (ECO), environmental (ENV), social (SOC), operational (OP), and stakeholders (STA).

Table 2. Dimensions and fields identification

Typically, stakeholders can be divided into two groups: "Contractual" stakeholders, who are directly involved with the company through contracts (e.g., customers, suppliers, employees, shareholders), and "diffuse" stakeholders, who are indirectly affected by the actions of the company (e.g., local authorities, public bodies, non-governmental organizations) [28]. Both groups of stakeholders are important to consider when evaluating a company's sustainability performance, as including all stakeholders can increase transparency and collaboration. It is essential to satisfy all stakeholders in order to measure a company's performance effectively [29-31]. Different stakeholders have different views on sustainability, with some arguing that it involves meeting the needs of all stakeholders both now and in the future (Leader of the Area of Social Corporate Responsibility-CHEM) while others believe that it involves being responsible and working together as a collective (Engineer of the Management Systems Directorate-ENG).

In order to measure and evaluate the relative sustainability of internal and external stakeholders, we propose, through this article, 4 fields of evaluation covering different points of common interest of the stakeholders. An added dimension to the model they have chosen as the basis for the development of a new evaluation model, namely the "Multidimensional Sustainability Performance Measurement Model for Road Freight Transport". In addition to this and to further improve the application of this framework, some fields have been eliminated while others have been added to develop a new more efficient model. The fields have been defined on the basis of a large number of documentary resources and judgments of industrial experts.

We considered the Air factor as a dual influence affecting both the environmental aspect through Air Pollution and the social aspect through Air Quality.

In view of the extent of its direct consequences for humans, we consider that the area of ‘noise pollution’, defined as untimely or excessive sound that can have harmful and deleterious effects on humans and human health, fully deserves its place in the social dimension.

The cost of delay has also been included as a field in the economic dimension. It is a means of communicating the impact of time on expected outcomes and is of great importance because it combines urgency and value and generates both indirect and direct impacts.

The purpose of using a neural network is to generalize the model and evaluate its performance and accuracy of calculation. In the future, a scientific paper will be published to elaborate on the supervised machine learning model and its implementation with real-time data. The current article briefly presents the design and identifies the next phase of research. To test the consistency of our performance measurement system and enhance its capabilities, we generated 200 different scenarios using a new machine learning model based on a multilayer perceptron with backpropagation. Artificial Neural Network (ANN) modeling was chosen as it can handle data with high volatility more accurately. ANN is an innovative and beneficial model for problem-solving and machine learning, functioning similarly to the nervous system of the human brain, and overcoming the limitations of classic forecasting methods. Setting up a multilayer Perceptron to solve a problem requires determining the optimal weights applied to each inter-neuronal connection, which is done using a feed-forward propagation algorithm.

7.png

Figure 7. General layout of ML system

Figure 7 above demonstrates a general layout of the ML system. As input to the ML system, labeled or unlabeled training data is obtained from many sources. The learning system's knowledge base determines the usage of an appropriate ML method, taking into account the decisions to be made by the ML predictions produced from the present datasets for quality output, hence increasing the organization's decision-making performance [32].

The general approach of this ML is structured as shown in Figure 8.

8.png

Figure 8. Followed approach during learning

Note that the model parameters are adjusted by gradient descent during backward propagation. In what follows, we will detail the algorithms used during this machine learning.

Training a multilayer perceptron is a process of determining the values for the various weights in order to properly resolve the connection that the network is modeling [33]. At this stage, we will look at a basic multilayer perceptron with only five weights. The network error for a particular pattern may be determined for any weight combination. The goal of training is to determine the weighting combination that produces the minimum error.

To locate the maximum or minimum point of a given function, one technique is to use gradient descent or steepest descent method, which utilizes the steepest gradients. The optimum point obtained may either be local or global [34]. The backpropagation training algorithm employs this method to look for the global minimum of the error surface. Among the different methods used for training multilayer perceptrons, backpropagation is the easiest to compute.

Multilayer perceptron is a widely used model for prediction, function approximation, and model classification tasks [33]. This study focuses on prediction, specifically forecasting future trends based on learning values. The neural network used in this study has five inputs in the input layer, three hidden layers, and a single output layer, as illustrated in Figure 9.

9.png

Figure 9. Structure of the neural network

The network's weights are first set to modest random values. This is equivalent to picking a random location on the error surface. The backpropagation method then computes the error surface's local gradient and adjusts the weights in the direction of the steepest local gradient. Given a fairly smooth error surface, it is hoped that the weights would converge to the error surface's global minimum.

The aim of finding a minimal error reinforces the accuracy of the model: The smaller the error, the better the performance of the model. When the performance of the model is high, the prediction is made in the best condition and the difference between the actual value and the predicted value obtained is minimal (or even zero).

The following Table 5 shows the prediction performance. The results are promising and show a very high accuracy validating the generalization of the multidimensional performance measurement system model presented in this paper.

Table 5. Prediction performances