[pt] Com o aumento das questões energéticas e ambientais, os veículos elétricos (EVs) se tornarão um modo de transporte essencial na distribuição logística. Um cenário vital a ser considerado é a dependência do congestionamento do tráfego nos tempos de viagem dos veículos, como é comum nas áreas urbanas hoje. Esse recurso significa que a velocidade de um EV em cada rota pode ser distinta durante diferentes períodos. Como os EVs possuem autonomia limitada, vários trabalhos na literatura propuseram modelos de consumo de energia em função da velocidade e fatores aerodinâmicos. No entanto, sua aplicação permanece limitada e simplificada devido à sua dependência da velocidade e dos tempos de viagem. No caso da velocidade, os modelos da literatura trabalham sob uma velocidade média durante um determinado arco ou introduzem aproximações com métodos de linearização por partes. Em relação aos tempos de viagem, os atuais algoritmos de roteamento de veículos muitas vezes reformulam a rede viária em um gráfico completo onde cada arco representa o caminho mais rápido entre dois locais. Os resultados obtidos por esses métodos divergem da realidade, principalmente para problemas de roteamento de arco envolvendo serviços nos arcos de uma rede rodoviária. Por essas razões, definimos o Problema de Roteamento de Arco Capacitado Elétrico com tempos de viagem dependentes do tempo e taxa de consumo de energia dependente da velocidade. Ao longo de um horizonte de planejamento, cada arco está associado a uma função de velocidade passo a passo. O objetivo é atender um conjunto de arcos que demandam serviços por meio de uma frota de EVs com carga e capacidade de bateria limitadas, minimizando o tempo total de viagem. Além disso, a taxa de consumo de energia por unidade de tempo percorrido é considerada uma função não linear baseada na velocidade. Propomos um algoritmo de pré-processamento de consumo de energia de forma fechada sem aproximações. Nós o incorporamos em uma metaheurística Iterate Local Search e comparamos o impacto no projeto de rotas com os veículos convencionais. / [en] With energy and environmental issues rising, electric vehicles (EVs) will become an essential mode of transportation in logistics distribution. A vital scenario to consider is the dependence of traffic congestion on vehicle travel times, as it is common in urban areas today. This feature means that the speed of an EV on each route may be distinct during different periods. Because EVs have a limited driving range, various works in the literature have proposed energy consumption models as a function of speed and aerodynamic factors. However, their application remains limited and oversimplified due to their dependence on speed and travel times. In the case of speed, the models in the literature work under an average speed during a given arc or introduce approximations with piece-wise linearization methods. Regarding travel times, current vehicle routing algorithms often reformulate the road network into a complete graph where each arc represents the quickest path between two locations. The results obtained by these methods differ from reality, particularly for Arc Routing Problems involving services on the arcs of a road network. For these reasons, we define the Electric Capacitated Arc Routing Problem with Time-dependent Travel times, and Speed-dependent Energy Consumption Rate (E-TDCARP). Over a planning horizon, each arc is associated with a step-wise speed function. Based on this function, a vehicle s speed can change while traveling on a given arc. The objective is to serve a set of arcs that require services through a fleet of electric vehicles with limited load and battery capacity, minimizing the total travel time. Furthermore, the energy consumption rate per unit of time traveled (ECR) is considered a nonlinear function based on speed. We propose a closed-form energy consumption preprocessing algorithm without approximations. We embed it into an Iterate Local Search metaheuristic (ILS) for E-TDCARP and compare the impact on the design of routes between these alternative vehicles and conventional ones.

Cette thèse s’intéresse au problème classique de tournées de véhicules avec contraintes de capacité (CVRP pour Capacitated Vehicle Routing Problem) ainsi qu’une variante beaucoup plus complexe, soit le problème de tournées de véhicules dépendant du temps avec fenêtres de temps et points de transfert défini sur un réseau routier (TDVRPTWTP-RN pour Time-Dependent Vehicle Routing Problem with Time Windows and Transfer Points on a Road Network). Dans le premier article, le TDVRPTWTP-RN est résolu en adaptant une métaheuristique qui représente l’état de l’art pour le CVRP, appelé Slack Induction for String Removals (SISR). Cette métaheuristique fait appel au principe “détruire et reconstruire” en retirant des séquences de clients consécutifs dans les routes de la solution courante et en réinsérant ensuite ces clients de façon à créer une nouvelle solution. Le problème est défini sur un réseau routier où différents chemins alternatifs peuvent être utilisés pour se déplacer d’un client à l’autre. De plus, le temps de parcours sur chacun des arcs du réseau n’est pas fixe, mais dépend du moment où le véhicule quitte le sommet origine. S’inspirant de problèmes rencontrés en logistique urbaine, nous considérons également deux types de véhicules, de petite et grande capacité, où les grands véhicules sont interdits de passage au centre-ville. Ainsi, les clients du centre-ville ne peuvent être servis que suite au transfert de leur demande d’un grand à un petit véhicule à un point de transfert. Comme un point de transfert n’a pas de capacité, une problématique de synchronisation apparaît quand un grand véhicule doit y rencontrer un ou plusieurs petits véhicules pour leur transférer une partie de son contenu. Contrairement aux problèmes stricts de tournées de véhicules à deux échelons, les grands véhicules peuvent aussi servir des clients localisés à l’extérieur du centre-ville. Comme le problème abordé est beaucoup plus complexe que le CVRP, des modifications importantes ont dû être apportées à la métaheuristique SISR originale. Pour évaluer la performance de notre algorithme, un ensemble d’instances tests a été généré à partir d’instances existantes pour le TDVRPTW-RN. Les réseaux omt été divisés en trois régions : centre-ville, frontière et extérieur. Le centre-ville et l’extérieur sont respectivemnt les royaumes des petits et grands véhicules, tandis que la frontière (où l’on retrouve les points de transfert) peut être visité par les deux types de véhicules. Les résultats numériques montrent que la métaheuristique proposée exploite les opportunités d’optimiser une solution en déplaçant autant que possible les clients neutres, soit ceux qui peuvent être servis indifféremment par un petit ou un grand véhicule, des routes des petits véhicules vers les routes des grands véhicules, réduisant ainsi les coûteuses visites aux points de transfert. Les deuxième et troisième article s’intéressent à des concepts plus fondamentaux et font appel au problème plus simple du CVRP pour les évaluer. Dans le second article, un étude expérimentale est conçue afin d’examiner l’impact de données (distances) imprécises sur la performance de différents types d’heuristiques, ainsi qu’une méthode exacte, pour le CVRP. À cette fin, différents niveaux d’imprécision ont été introduits dans des instances tests classiques pour le CVRP avec 100 à 1 000 clients. Nous avons observé que les meilleures métaheuristiques demeurent les meilleures, même en présence de hauts niveaux d’imprécision, et qu’elles ne sont pas affectées autant par les imprécisions qu’une heuristique simple. Des expériences avec des instances réelles ont mené aux mêmes conclusions. Le troisième article s’intéresse à l’intégration de l’apprentissage automatique dans la métaheuristique SISR qui représente l’état de l’art pour le CVRP. Dans ce travail, le principe “détruire et reconstruire” au coeur de SISR est hybridé avec une méthode d’apprentissage par renforcement qui s’inspire des systèmes de colonies de fourmis. L’ap- prentissage automatique a pour but d’identifier les arêtes les plus intéressantes, soit celles qui se retrouvent le plus fréquemment dans les solutions de grande qualité précédemment rencontrées au cours de la recherche. L’inclusion de telles arêtes est alors favorisé lors de la réinsertion des clients ayant été retirés de la solution par le mécanisme de destruction. Les instances utilisées pour tester notre approche hybride sont les mêmes que celles du second article. Nous avons observé que notre algorithme ne peut produire que des solutions lé- gèrement meilleures que la métaheuristique SISR originale, celle-ci étant déjà quasi-optimale. / This thesis is concerned both with the classical Capacitated Vehicle Routing Problem (CVRP) and a much more complex variant called the Time-Dependent Vehicle Routing Problem with Time Windows and Transfer Points on a Road Network (TDVRPTWTP-RN ). In the first paper, the TDVRPTWTP RN is solved by adapting a state-of-the-art metaheuris- tic for the CVRP, called Slack Induction for String Removals (SISR). This metaheuristic is based on the ruin and recreate principle and removes strings of consecutive customers in the routes of the current solution and then reinserts the removed customers to create a new solution. The problem is formulated in a full road network where different alternative paths can be used to go from one customer to the next. Also, the travel time on each arc of the road network is not fixed, but depends on the departure time from the origin node. Motivated from city logistics applications, we also consider two types of vehicles, large and small, with large vehicles being forbidden from the downtown area. Thus, downtown customers can only be served through a transfer of their goods from large to small vehicles at designated transfer points. Since transfer points have no capacity, synchronization issues arise when a large vehicle must meet one or more small vehicles to transfer goods. As opposed to strict two-echelon VRPs, large vehicles can also directly serve customers that are outside of the downtown area. Given that the TDVRPTWTP-RN is much more complex than the CVRP, important modifications to the original SISR metaheuristic were required. To evaluate the performance of our algorithm, we generated a set of test instances by extending existing instances of the TDVRPTW-RN . The road networks are divided into three regions: downtown, boundary and outside. The downtown and outside areas are the realm of small and large vehicles, respectively, while the boundary area that contains the transfer points can be visited by both small and large vehicles. The results show that the proposed metaheuristic exploits optimization opportunities by moving as much as possible neutral customers (which can be served by either small or large vehicles) from the routes of small vehicles to those of large vehicles, thus avoiding costly visits to transfer points. The second and third papers examine more fundamental issues, using the classical CVRP as a testbed. In the second paper, an experimental study is designed to examine the impact of inaccurate data (distances) on the performance of different types of heuristics, as well as one exact method, for the CVRP. For this purpose, different levels of distance inaccuracies were introduced into well-known benchmark instances for the CVRP with 100 to 1,000 customers. We observed that the best state-of-the-art metaheuristics remain the best, even in the presence of high inaccuracy levels, and that they are not as much affected by inaccuracies when compared to a simple heuristic. Some experiments performed on real-world instances led to the same conclusions. The third paper focuses on the integration of learning into the state-of-the-art SISR for the CVRP. In this work, the ruin and recreate mechanism at the core of SISR is enhanced by a reinforcement learning technique inspired from ant colony systems. The learning component is aimed at identifying promising edges, namely those that are often found in previously encountered high-quality solutions. The inclusion of these promising edges is then favored during the reinsertion of removed customers. The benchmark instances of the second paper were also used here to test the new hybrid algorithm. We observed that the latter can produce only slightly better solutions than the original SISR, due to the quasi-optimality of the original solutions.