Machine Learning-Driven Adaptive Charging Scheduling for Electric Truck Fleets under Dynamic Tariffs and Public Charging Constraints

The electrification of heavy-duty transport is accelerating, and companies like Einride are leading the transition. However, operating large electric truck fleets presents new challenges in energy management. Truck charging requires massive amounts of power, and if unmanaged, it can lead to:

• Very high electricity costs when charging occurs during peak tariff hours.
• Increased stress on the local grid infrastructure.
• Trucks not ready on time for their next shipment
• High charging infrastructure costs due to overcapacity

The key challenge is how to transition to a zero-emission truck fleet and meet transportation demand while avoiding costly power tariff penalties and over installing charging infrastructure and securing unnecessarily large grid connection. This means ensuring that charging the trucks get charged in time, does not exceed the site’s available power capacity or trigger higher peak demand charges, which could otherwise increase the total energy cost for the entire company. This challenge becomes even more complex when considering not just depot charging but also public charging station availability, which can influence cost and scheduling decisions.

Traditional approaches, such as rule-based charging or static optimization, fail to adapt to uncertainty in tariffs, spot price, truck schedules, and public station congestion.

• Rule-based methods typically rely on fixed charging rules, such as initiating charging immediately upon truck arrival or during predefined off-peak hours. While simple to implement, these approaches disregard variability in electricity prices, renewable energy availability, and truck arrival patterns, often leading to suboptimal costs.
• Static optimization models (for example linear or mixed-integer programming) determine a charging plan based on known or assumed inputs, such as fixed tariffs and scheduled arrivals. Although more sophisticated than rule-based approaches, they operate under the assumption of perfect data is available and therefore perform poorly when confronted with unexpected price fluctuations, late truck arrivals, or sudden public station congestion.

Machine Learning (ML), by contrast, can predict these uncertainties and dynamically adjust schedules in real-time, offering a scalable and cost-effective solution for the power node. 

Thesis goal:

The main goal of this thesis is to develop a machine learning-driven framework that forecasts electricity tariffs and public charging station usage, and utilizes these forecasts to adapt truck charging schedules in real time. The objective is to minimize total energy costs while ensuring that all operational constraints, such as truck departure times, limited charging dock availability, and site power limits are meet.

The study will explore how different ML approaches can be used to forecast dynamic variables like tariffs and charging station congestion, and how these forecasts can be integrated into an adaptive scheduling framework. The thesis will also evaluate how such a system influences overall depot efficiency, charger utilization, and operational robustness compared to static and rule-based scheduling methods.

Methodology

1.     Data Collection & Preprocessing: Gather historical data from Einride (truck schedules, energy demand), collect spot price data from relevant energy markets and gather public charging station usage data.

2.     Forecasting with ML: Train ML models for public charging usage and tariff forecasting.

3.     Adaptive Scheduling Algorithm: Modify existing charging schedules based on predicted tariffs and station usage. There are different approaches that can be used (Optimization-based, Reinforcement Learning or a Hybrid approach)

4.     Evaluation: Based on a real-life use case from one of the Einride depot  run the ML together with the scheduling algorithm. Then compare static optimization and ML-driven scheduling approaches based on cost reduction, robustness, scalability.

Expected Contributions

The expected contributions of the thesis would include:

·       A novel ML-driven framework for fleet-scale adaptive charging.

·       Application of ML-based scheduling to real-world depot operations of Einride with the possibility of reduced occasions of exceeding power use. 

The expected contributions of Einride thesis would include:  

·       Organizing a formal start-up meeting at Einride’s facilities.

·       Bi-weekly follow-up meetings with the students to provide feedback and guidance.

·       Supplying relevant data and a representative depot use case for the research.

·       Appointing a dedicated contact person for the thesis project.

·       Participating in the final thesis presentation, to be held either at Einride or at Knightec in Västerås.

If interested please apply via the link or visit our website. For any questions please contact Aldin Berisa and Edin Jelacic.

Preferably we would like a pair of students with good knowledge in Embedded Systems, AI, ML and good programming skills.

Aldin Berisa (aldin.berisa@knightec.se), Knightec Group AB and Einride AB (The thesis is financially supported)