Towards Sustainable Industry 5.0: An LLM-Based Co-Pilot for Energy-Efficient Factory Scheduling

Abstract

Industry 5.0 promotes sustainable, resilient, and human-centric manufacturing. Many factories struggle to produce energy-aware schedules that balance throughput, energy, and time-of-use (TOU) tariffs. Classical methods (heuristics and optimization) help but lack transparency and adaptability, limiting operator-in-the-loop use. Generative AI, particularly Large Language Models (LLMs), offers reasoning, adaptation, and interaction, yet integration with production scheduling is nascent. We introduce a hybrid framework that combines classical heuristics with GPT-4 reasoning to create an Industry 5.0-compatible Co-Pilot for energy-aware factory scheduling. The Co-Pilot evaluates and adapts machine operation schedules to avoid peak windows and explains trade-offs in natural language. We evaluate on three datasets (CTU synthetic, Kaggle manufacturing, Zenodo benchmark). Results show the heuristic Co-Pilot consistently reduces peak load share versus classical baselines at similar cost; on Zenodo, GPT-4 saves 4–7% in cost and energy, while its performance is less stable on synthetic data. These findings highlight the promise of LLM-powered scheduling and the value of hybrid human-AI strategies in Industry 5.0.

1. Introduction

One of the primary objectives of the fourth industrial revolution, Industry 4.0 [1,2,3], was to develop advanced and high-end automated systems that enhance overall production system performance. In Industry 5.0 [4,5,6], industrial production systems are redefined to generate smooth human-machine collaboration, sustainability, and resilience. Goals now extend beyond throughput to emphasize energy reduction and sustainability. One way for factories to achieve a balanced system between energy consumption, throughput, and quality is to explore the optimization of components such as energy-aware production scheduling [7,8]. Energy-aware production scheduling requires dynamically allocating jobs to machines while hovering production targets and energy costs under time-of-use (TOU) tariffs [9].

For several years, classical scheduling methods [10], such as Shortest Processing Time (SPT), Earliest Due Date First (EDF), or First Come First Served (FCFS), have yielded acceptable energy-aware scheduling outcomes in academia and industry [11,12]. Some more advanced energy-aware scheduling technologies include constraint programming, metaheuristics, and reinforcement learning. They offer more complex interactions between production and energy states [13]. A very notable gap of these scheduling methods is often the lack of flexibility, explainability, and adaptability when integrated with heterogeneous industrial contexts and unstable constraints. In the era of Industry 5.0, it becomes imperative to address this gap to build sustainable energy-aware scheduling methods.

With the advent of Industry 5.0, the industrial sector is witnessing an unprecedented explosion of Large Language Model (LLM) [14] applications, such as GPT-4 [15,16]. Thanks to the integration of LLMs in industrial systems, decision-making tasks can be improved through their ability to interpret structured and semi-structured information, generate alternative solutions, and provide natural language explanations to shop floor operators. When correctly integrated and explored, LLM applications are capable of producing remarkable capabilities in reasoning, planning, and contextual adaptation. In functions such as predictive maintenance and supply chain optimization, LLM applications are steadily expanding their scope. However, their integration in shop floor manufacturing tasks such as energy-aware scheduling for Industry 5.0 applications remains underexplored [17,18].

In our research, we address this gap by suggesting the design of an LLM-based Co-Pilot framework. The proposed framework integrates the GPT-4 LLM with classical scheduling to produce energy-aware reasoning. The system integrates heuristic scheduling with GPT-4 to generate and adapt machine operation schedules under TOU tariffs. Our primary research contributions are as follows:

• We present a system architecture for an LLM Co-Pilot that integrates energy-aware heuristics with GPT-4 reasoning for production scheduling.
• We develop and evaluate the Co-Pilot on three diverse datasets (synthetic, industrial, and benchmark), enabling cross-domain comparison against classical scheduling baselines.
• We provide a comprehensive experimental study demonstrating when LLMs provide tangible benefits, when heuristics remain sufficient, and how hybrid LLM-heuristic approaches align with the principles of Industry 5.0.

This research provides practical cues and guidance for factories seeking to implement AI-assisted decision support systems compatible with Industry 5.0 requirements in manufacturing, specifically in the area of production scheduling. The proposed Co-Pilot contributes toward the design of intelligent, sustainable, and human-centric production ecosystems by introducing explainable, LLM-driven scheduling, capable of adapting to energy tariffs and operational variability.

The remainder of this paper is structured as follows: Section 2 briefly reviews the literature on energy-aware scheduling and LLM applications in industrial contexts. Section 3 introduces the proposed system architecture. Section 4 describes the methodology and experimental design. Section 5 presents and discusses the results. Finally, Section 6 concludes the paper and outlines directions for future research.

2. Brief Literature Review

2.1. Energy-Aware Production Scheduling

Energy-aware production scheduling is a process that involves managing energy demand, environmental aspects (emissions), and cost to effectively control various jobs with different processing times and meet specific energy demand criteria [19,20]. Traditional manufacturing research has addressed a portion of energy-aware production scheduling processes to balance machine utilization, optimize throughput, and minimize energy consumption. Bruzzone, A. et al., Fang, K.T. and Lin, B.M. [21,22] highlight that this process was introduced to manage complex energy flows and production for factories, changing job scheduling into optimisation models solved with associated algorithms. To optimize throughput tardiness with little focus on energy, several classical heuristic methods like Shortest Processing Time (SPT), Earliest Due Date First (EDF), and First Come First Served (FCFS) achieve acceptable results [23,24].

Sustainability pressure and TOU tariffs have prompted energy-aware scheduling, integrating power profiles into assignment and sequencing to inform better scheduling decisions. Several studies have been conducted in this regard. One of the first presented by [24] aimed to minimize both total energy consumption and tardiness on single machines. Their work laid the foundation for multi-objective scheduling approaches. A few years later, Fang, K. et al. [25] built upon their predecessors’ idea to produce energy-aware production scheduling for reducing carbon footprint. Their work linked production planning with environmental metrics. This initiative has been later expanded by other recent works [26]. Overall, advanced energy-aware production scheduling research integrating sustainability pressure and TOU tariffs included Metaheuristics (simulated annealing, tabu, genetic algorithms) for target cost and peak-load reduction [25,27,28]; exact models such as mixed-integer programming (MIP) and constraint programming (CP) to solve small/medium instances. Dai, M. et al. [27] proposed a study that integrates the metaheuristic simulated annealing algorithm to design an improved hybrid production scheduling system for flexible job shops. Their work demonstrated the scalability of heuristic search to solve such problems. Although mathematical programming approaches, such as MIP and CP, are powerful in solving various complex energy-aware production scheduling problems by providing exact solutions for small- to medium-scale instances, they are not suitable for industrial-scale systems due to their computational expense. In recent years, there has been considerable work on energy-aware production scheduling, including energy-aware production scheduling using reinforcement learning (RL). They aim to dynamically adapt schedules under uncertainty, particularly in response to disturbances or real-time changes in energy tariffs that necessitate dynamic rescheduling [29].

Despite the tremendous progress and advances in this area, which aim to achieve better scheduling optimization for factories, many energy-aware scheduling approaches remain opaque to operators. They are designed as “black boxes” with limited visibility, explainability, and transparency for non-expert operators who remain accountable for production outcomes on the shop floor. These shortcomings create an incompatibility with the Industry 5.0 requirements and principles for solution design, calling for more human-centric, transparent, sustainable, and explainable solutions for factories. There is therefore a need to design hybrid solutions combining schedule optimization with interpretability.

2.2. Industry 5.0 and Human-Centric Scheduling

The transition from Industry 4.0 to Industry 5.0 is completely transforming design and operation principles within factories. The European Commission [26] expects this new paradigm of Industry 5.0 to carry more than a simple industrial transformation, rather a societal change that promotes the well-being of the environment and the human being. Industry 4.0’s primary focus was to achieve advanced automation, cyber–physical systems, and data-driven optimization within factories. Under Industry 5.0, we are pointed toward new design characteristics that prioritize human-centricity, sustainability, and resilience beyond automation [30,31]. Considering these requirements, researchers and factory design stakeholders can unlock possibilities for more innovative solutions in various manufacturing domains.

There has been production scheduling work under Industry 4.0 with a focus on the integration of artificial intelligence (AI), machine learning (ML), and advanced algorithms to optimize factory throughput [32]. Industry 5.0 prioritizes energy-aware operations that balance efficiency, operator well-being, and flexible decision-making. The primary goal of Industry 5.0 production scheduling systems is to balance efficiency with adaptability and operator well-being. It looks beyond throughput optimization to generate production schedules that consider energy efficiency, fluctuating tariffs, and align with renewable energy availability. Another essential element added in Industry 5.0 production scheduling systems is the role of human operators. Their role evolves from passive recipients of production schedules to active collaborators in decision-making [33,34].

In recent studies, researchers are investigating the integration of operator-in-the-loop functions in production scheduling to show value in interactive adjustments [35,36]. In these innovative production scheduling systems, human decision-makers interact with automated schedulers to revise or override decisions in real time. Although this new research aims to address the human-centric requirements of Industry 5.0, it lacks proper, explainable tools and functions that enable easy interaction with shop floor operators, who are often non-experts in this field. Most production scheduling systems display numerical interfaces without offering transparent explanations of scheduling trade-offs that are useful to human operators. This lack of transparency may reduce trust between human operators and operating equipment, endangering the whole production system due to a scepticism in using industrial practices. We therefore aim to develop an AI co-pilot that integrates energy-aware production scheduling optimization with natural language interaction, closely aligning with Industry 5.0 design requirements and empowering operators to make informed decisions transparently.

2.3. Large Language Models in Industrial Applications

In recent years, there has been a fast-growing wave in the field of Large Language Models (LLMs). LLMs have become popular in various research domains for their ability to interpret structured/semi-structured data, provide explainable rationales, generate context-aware reasoning, and interact through natural language [37]. Currently, one of the most publicly popular LLMs is the GPT family, developed by OpenAI [38]. The GPT family is reputed to accelerate NLP advances and decision support [39,40,41]. LLMs are already actively integrated and produce successful results in areas such as customer service, software development, and legal reasoning. This remarkable success awakened interest in several researchers to incorporate them in industrial manufacturing applications as well. Industrial proofs-of-concept span predictive maintenance, documentation, anomaly explanation, and supply chain planning [37,42,43]. Wang, T. et al. [15] issues a report on ChatGPT’s version 4 potential in sustainable smart manufacturing. They performed a comprehensive survey and highlighted ChatGPT’s feature to improve operators’ communication with explanatory text support.

However, energy-aware production scheduling with LLMs is still rare; based on existing advances in LLM research, production scheduling could benefit from LLMs’ unique advantage in promoting human interaction and generating explainable decision cues. In this regard, when integrating LLMs into production scheduling, factories can obtain systems that utilize LLMs to justify scheduling choices, highlight trade-offs (for example, between energy savings and deadline risks), and accept feedback in natural language. Our work integrates some of these advantages to address the gap by designing a hybrid, explainable scheduler.

2.4. Research Gap

From this brief literature review, we identify two primary gaps that we aim to address in this study: the lack of explainability and adaptability to heterogeneous industrial contexts in current classical or energy-aware scheduling methods, which primarily focus on optimization, and the rare integration of LLM tools and technologies to improve energy-efficient production scheduling. To contribute to the research field, we propose the creation of an LLM Co-Pilot that combines the robustness of heuristics with the adaptive reasoning of GPT-4. Our solution provides not only optimized schedules but also transparent, explainable support for operators in Industry 5.0 environments.

3. System Architecture

We design our Co-Pilot system to support energy-efficient, human-centric production scheduling compatible with Industry 5.0 principles. Figure 1 illustrates the overall system architecture with the key components: a structured data ingestion, hybrid scheduling strategies, and an operator-facing interface that facilitates explainable decision support.

3.1. Inputs and Pre-Processing

Our system receives four input categories that need to be processed: jobs and machines, time-of-use (TOU) electricity tariffs, renewable availability, and factory schedule. All these inputs together define the scheduling problem. Job identifiers, machine assignments, and processing times are essential elements in jobs and machines input to provide the fundamental structure of the scheduling environment. The time-of-use (TOU) electricity tariffs input ensures that the scheduling system integrates tariffs, allowing the system to evaluate not only production throughput but also energy costs. In short, it consists of the cost of energy consumption during peak, off-peak, and intermediate time windows. The renewable availability input includes renewable energy components such as on-site solar or wind generation. These elements are very advantageous for factories and can be integrated into scheduling to prioritize green energy use requirements. The factory schedule input represents an existing production plan that records actual or generated job start and finish times on each machine within a realistic factory environment. It could be exported from a factory manufacturing execution system (MES) or other scheduling software when available. It is worth noting that this input is not available in all datasets. In our experiment, we extracted the factory schedule input from the Zenodo optimization dataset.

Table 1. Summary of datasets used for evaluation and preprocessing details.

Dataset	Jobs	Machines	Power (kW)	Preprocessing Summary
CTU Synthetic	∼1200	1 (aggregated)	3–6	Aggregated into a single virtual machine; rescaled to an 8-h horizon; uniform due times.
Kaggle Manufacturing	∼2500	5	2–10	Cleaned missing values; normalized IDs; standardized time units (minutes).
Zenodo Optimization	∼1000	3	3–9	Extracted start/end times; mapped to TOU tariff windows; normalized to unified job schema.

presents an example of the factory schedule data structure. The system provides the factory schedule through the scheduled_start and scheduled_end fields that define the planned execution of each job within the factory. From the table, each record sample includes job and machine identifiers, processing time, power consumption, and corresponding scheduled start and finish timestamps. This baseline schedule serves as a practical benchmark to evaluate the heuristic and GPT-4 Co-Pilot schedules. It is a useful baseline that reflects practical shop-floor constraints.

We standardized all inputs into a unified representation with fields such as job_id, machine_id, proc_time_min, release_time, due_time, power_kw, and optional scheduled_start/scheduled_end. Pre-processing’s role is also to encode critical constraints (deadlines, machine capacity, tariff structures, and renewable priorities).

3.2. Scheduling Core

The system’s “brain” lies in the LLM–Co-Pilot Core which combines classical and AI-driven strategies. We establish deterministic performance benchmarks by implementing three baseline scheduling methods: SPT, EDF, FCFS, and two AI-based metaheuristic approaches: Simulated Annealing (SA) and Randomized Restart Greedy (RRG). These last two methods represent stochastic and population-based search paradigms commonly used in production scheduling optimization. They are helpful as valuable benchmarks to evaluate the efficiency and adaptability of the proposed LLM Co-Pilot framework. As mentioned in the input section, for data containing ‘factory schedules’ information, we also utilize it as a baseline. Our hybrid system combines a heuristic and LLM Co-Pilot. The heuristic Co-Pilot consists of a lightweight energy-aware adjustment mechanism, which can be seen as a variant of EDF that shifts tasks out of peak periods when slack is available. The heuristic Co-Pilot ensures a balance between small energy savings, feasibility, and simplicity. To prevent excessive disruptions to production flow, we capped delays at 60 min (capped delay). The LLM Co-Pilot integrates GPT-4’s natural language reasoning capabilities to model a comprehensive, energy-aware production scheduling system. We encode different jobs from the experimental data in a compact JSON-like structure before feeding them into the LLM Co-Pilot. We also designed two prompts to lead GPT-4 into our desired objectives: “minimize energy cost under TOU tariffs” and “avoid overlaps and meeting due times”. As an outcome, GPT-4 generates candidate schedules that will be parsed, validated, and refined before integration. In short, in the background of the LLM Co-Pilot system, GPT-4 is prompted with a compact job representation and instructed to minimize cost under TOU tariffs while respecting release and due dates.

The primary difference between the two Co-Pilots is that the heuristic one is purely rule-driven, while the GPT-4 one can adapt dynamically to contextual constraints. This feature offers a more flexible but less predictable approach.

3.3. Validation and Outputs

We integrate a deterministic feasibility checker module that prevents AI-generated outputs from violating industrial feasibility requirements. At the core of the feasibility checker’s operation is an evaluation of all generated schedules to ensure they meet these requirements. The feasibility checker eliminates machine overlaps by repacking jobs sequentially where necessary. It also enforces deadlines by shifting within Slack and applies an EDF (Earliest Deadline First) fallback for any missing jobs. A significant advantage of the feasibility checker is that it ensures every schedule output is executable in practice, even if the LLM overgeneralizes or under-specifies certain cases.

The system outputs a final schedule and KPIs: total energy (kWh), cost ($) under TOU, and peak-load share (%). When renewable data are provided, we also report renewable-aligned and off-peak minutes.

3.4. Operator Interaction and Explainability

The Co-Pilot includes an operator-facing interface built with Streamlit [44] to visualize Gantt charts, performance metrics, and natural-language rationales. Through this GUI, operators can review and approve schedules or request revisions as needed. A feedback loop (see Figure 1) routes operator preferences back into the LLM Co-Pilot, enabling adaptive scheduling that remains transparent, explainable, and aligned with Industry 5.0 principles.

Our Co-Pilot framework aligns with Industry 5.0’s vision of human-centric manufacturing by offering a scheduling tool that augments transparency and explainable decision support. Through the Streamlit-based operator interface, users can visualize schedules, inspect energy and cost performance, and interact with the Co-Pilot through simple controls such as tariff sliders or job-priority adjustments. Thanks to the GPT-4 LLM, the system can produce natural-language explanations that justify its scheduling decisions. For instance, it may indicate that a specific task was delayed to avoid peak-tariff hours or to synchronize with renewable availability. With this explainable feedback, operators can understand the scheduling changes and also the rationale behind them. The system helps reduce the cognitive load on operators to interpret results by offering interpretable reasoning and easy visualization. It allows human supervisors to focus on higher-level objectives (balancing production priorities, assessing maintenance needs, etc.)as required by Industry 5.0 core values.

3.5. Scheduling Workflow

Figure 2 presents the experimental workflow that ensures reproducibility and comparability across datasets and methods.

The workflow proceeds as follows:

• Datasets: We evaluate our system using three open-source datasets namely CTU synthetic benchmarks [45], Kaggle factory data [46], and Zenodo optimization schedules [47]. Each dataset has a different scale, structure, and realism. We use these differences to achieve a comprehensive evaluation of both the heuristic and LLM Co-Pilot under heterogeneous conditions. CTU Synthetic Dataset: This dataset comes from academic benchmarking of energy-aware scheduling algorithms. In the original dataset there are approximately 1200 synthetic jobs distributed across 10 machines with simplified power profiles and constant processing rates. Each job has release and due times uniformly sampled within an 8-h horizon. The CTU data is noise-free but somewhat lacks realism. They are ideal for controlled stress-testing of heuristic and GPT-4 reasoning stability. Kaggle Manufacturing Dataset: The Kaggle dataset provides realistic production records from an industrial manufacturing process. The Kaggle dataset has about 2500 job entries with five machines (M01–M05). Each job includes processing time, due date, and power consumption ranging from 2–10 kW. During the preprocessing phase we removed incomplete records, standardized timestamps to minutes, and normalized job IDs and machine IDs to a consistent schema due to the dataset’s heterogeneity. Zenodo Optimization Dataset: The Zenodo dataset has been generated from a prior optimization study where approximately 1000 tasks were mapped to three machines (MAQ118–MAQ120). In the Zenodo dataset we have attributes such as complete factory schedules with start and finish times, energy cost coefficients, and TOU tariff windows. This dataset served as the most realistic benchmark for evaluating the Co-Pilot under near-industrial conditions. To ensure comparability across datasets we standardized all variables to a unified schema with attributes job_id, machine_id, proc_time_min, release_time, due_time, and power_kw. We also normalized power and time attributes to kW and minutes, respectively, and job release and due times were rescaled to a 0–1440 min daily window. We present in

  Table 2. Summary of datasets used for evaluation and preprocessing details.

  Dataset	Jobs	Machines	Power (kW)	Preprocessing Summary
  CTU Synthetic	∼1200	1 (aggregated)	3–6	Aggregated into a single virtual machine; rescaled to an 8 h horizon; uniform due times.
  Kaggle Manufacturing	∼2500	5	2–10	Cleaned missing values; normalized IDs; standardized time units (minutes).
  Zenodo Optimization	∼1000	3	3–9	Extracted start/end times; mapped to TOU tariff windows; normalized to unified job schema.

  a summary of datasets used for evaluation and important preprocessing details.
• Standardization: As previously mentioned, the standardization process consists of mapping all datasets to a unified format containing job, machine, power, release, and due attributes.
• Baseline Scheduling: generation of factory schedules (when available), rule-based baselines (SPT, EDF, FCFS), and AI-based metaheuristic approaches (Simulated Annealing (SA) and Randomized Restart Greedy (RRG)).
• LLM Co-Pilot Scheduling: GPT-4 and the heuristic Co-Pilot produce energy-aware schedules.
• Evaluation: all schedules are benchmarked on energy, cost, and peak-load metrics.

4. Methodology and Experimental Design

In this section, we describe the scheduling strategies, Co-Pilot designs, evaluation metrics, and implementation of the proposed LLM Co-Pilot framework. We combine rule-based heuristics, a lightweight energy-aware Co-Pilot, and a GPT-4-powered Co-Pilot into a single evaluation workflow (Figure 2). We evaluate the performance of our framework using three datasets representing synthetic, real-world, and optimization-derived scheduling problems. The datasets used are: CTU synthetic benchmarks for stress-testing, Kaggle factory records with real job/due dates, and Zenodo optimization-derived schedules as industrial baselines.

4.1. Scheduling Strategies

We compare eight scheduling strategies implemented across all datasets to assess trade-offs between computational heuristics, metaheuristic search, and intelligent LLM-based reasoning:

• Factory Schedule: When available (for example in the Zenodo dataset), the reported schedule serves as a baseline reflecting practical constraints.
• Shortest Processing Time (SPT): Jobs are ordered by ascending processing time, prioritizing short tasks to minimize average flow time.
• Earliest Due First (EDF): Jobs are sorted by due time, ensuring that time-critical tasks are prioritized.
• First Come First Served (FCFS): Jobs are scheduled according to release order, which is a common factory practice.
• Heuristic Co-Pilot: A modification of EDF that shifts tasks with slack out of peak tariff windows, subject to a capped delay (≤60 min). This provides a deterministic baseline for energy-aware scheduling.
• LLM Co-Pilot (GPT-4): Integrates natural language reasoning to directly generate start-finish times while optimizing for TOU tariffs and deadline compliance.
• Simulated Annealing (SA, Metaheuristic): A stochastic optimization approach that iteratively refines job sequences by probabilistically accepting swaps that increase cost during early iterations, enabling exploration of the search space and gradual convergence to lower-cost energy schedules.
• Randomized Restart Greedy (RRG, Metaheuristic): A population-based heuristic that generates multiple randomized job orders and retains the best-performing sequence according to total cost and energy criteria. This strategy balances exploration and exploitation through multiple greedy restarts.

These strategies together enable a comprehensive comparison between deterministic heuristics, stochastic metaheuristics, and intelligent language-based scheduling approaches for Industry 5.0 energy-aware manufacturing systems.

4.2. Heuristic Co-Pilot Design

The heuristic Co-Pilot is designed to ensure reliability while modestly improving load profiles without inflating costs. It follows three main steps: first, generate an EDF schedule as a baseline, then for each job, check if its execution overlaps with peak tariff windows, and finally, if slack is available, delay the job by a fraction of its slack (default 50%), up to 60 min, to move it out of peak. More details are available in the pseudo code of Algorithm 1 below. Without the integration of energy-related parameters such as time-of-use (TOU) tariffs or explicit energy-awareness, the heuristic will become a conventional Earliest Due Date (EDF) scheduling strategy. The distinguishing energy-aware behavior is triggered only when the tariff structure contains peak and off-peak intervals. When we include energy-aware components, the heuristic algorithm performs conditional job shifting within available slack and aims to relocate jobs from peak tariff periods to off-peak windows. This shifting logic in Algorithm 1 is bounded by a maximum delay of 60 min. It is the key element introducing energy optimization into the scheduling process. Without TOU information, no shifts occur, and the output schedule remains equivalent to the EDF baseline.

Algorithm 1 Heuristic Co-Pilot.

1: Input: Jobs with $\{release\_time,due\_time,proc\_time,power\}$ 2: Generate EDF baseline schedule 3: for each job j in EDF schedule do 4:    Check overlap of j with peak tariff windows 5:    if slack(j) > 0 then 6:      Delay j by $min(0.5\times slack(j),60)$ 7:    end if 8: end for 9: Repack jobs sequentially per machine 10: Output: Adjusted energy-aware schedule

4.3. LLM Co-Pilot Design

We guarantee flexibility and ensure bounded outputs in our schedules by integrating a structured prompting workflow into the GPT-4 Co-Pilot design, which includes hybrid prompting and fallback mechanisms. Each job is represented as a JSON-like entry with fields{job_id, machine_id, proc_time_min, release_time, due_time, power_kw}. One of the prompt’s tasks is to specify schedule objectives (“Minimize energy cost under TOU tariffs while meeting due times”) and constraints (no overlaps, deadlines). Through the LLM reasoning process, GPT-4 generates candidate schedules as a JSON list of start/finish times. Finally, outputs are parsed, infeasible jobs are repaired, and missing jobs are filled with EDF fallback. Algorithm 2 presents in detail the operation of the LLM Co-Pilot. In this algorithm, we incorporate energy cost minimization characteristics within the prompt as an optimization objective. The GPT-4 Co-Pilot uses job timing and machine allocation to reduce the total cost computed under TOU tariffs. If we exclude tariff and energy-intensity data, the LLM will only consider due-time and feasibility objectives, effectively becoming a standard non-energy-aware EDF policy. The inclusion of TOU tariffs and power consumption attributes, therefore, activates the sustainability-driven reasoning layer that distinguishes the LLM Co-Pilot from a purely throughput-oriented scheduler.

To ensure reproducibility, we provide an example of the prompt template used to instruct GPT-4 during the design of the LLM Co-Pilot. Each job is encoded as a structured entry in JSON format as follows:

• [{"job_id": "J001", "machine_id": "M1", "proc_time_min": 45,
   "release_time": "08:00", "due_time": "12:00", "power_kw": 5.0}]

The corresponding GPT-4 prompt specifies the objective and constraints as shown below:

• You are an energy-aware factory scheduler. Using the list of jobs below,
  generate a feasible production schedule that:
  (1) Minimizes total energy cost under the given time-of-use (TOU) tariffs,
  (2) Avoids overlapping jobs on the same machine,
  (3) Ensures all jobs finish before their due_time,
  (4) Uses the following tariff periods:
     peak = 09:00--17:00 ($0.25/kWh),
     off-peak = 17:00--09:00 ($0.12/kWh).
  Output a JSON list of jobs with their proposed start and end times.

GPT-4 then produces a structured response in the following format:

• [{"job_id": "J001", "machine_id": "M1",
   "start_time": "07:15", "end_time": "08:00"},
   {"job_id": "J002", "machine_id": "M2",
   "start_time": "08:10", "end_time": "09:50"}]

The system parses this response, checks for infeasible or missing jobs, and applies an EDF fallback for those cases. This workflow ensures deterministic schedule completion while maintaining transparency of LLM reasoning and reproducibility of experimental results.

Algorithm 2 LLM Co-Pilot (GPT-4 Scheduler).

1: Input: Jobs with $\{job\_id,machine\_id,proc\_time\_min,release\_time,due\_time,power\_kw\}$ 2: Construct prompt with objectives and constraints 3: Send prompt → GPT-4 4: Receive candidate schedule as JSON 5: Parse schedule and validate: Repair infeasible jobs Apply EDF fallback for missing jobs 6: Output: Energy-aware LLM-generated schedule

4.4. Evaluation Metrics

We use three primary key performance indicators (KPIs) to evaluate all schedules:

• Total Energy (kWh): Sum of energy consumed by all jobs.
• Cost ($): Computed under TOU tariffs, with higher prices in peak hours.
• Peak Load Share (%): Percentage of scheduled minutes falling within peak tariff windows.

In practical industrial deployments, Cost is the most important KPI that directly impacts factories’ profitability. It represents factories’ energy expenditure under TOU tariffs. Peak Load Share is seen as the secondary KPI because when it is exceeded, peak windows often lead to demand surcharges and higher carbon intensity. The last KPI, Total Energy Consumption, acts as a supporting sustainability metric that reflects overall system efficiency. It does not have an immediate financial impact. In practice, the final production schedule is chosen by balancing Cost, peak reduction, and throughput. Although Cost has priority, peak management ensures long-term stability and regulatory compliance.

4.5. Implementation Details

The Co-Pilot framework was implemented in Python 3.12 (with Google Colab and Ngrok for deployment) using Streamlit as the web-based operator interface with file upload, tariff parameter sliders, and Gantt chart visualization. We integrate the GPT-4 LLM for the LLM Co-Pilot reasoning via the OpenAI API. It is worth noting that the LLM Co-Pilot integration introduces additional computational and financial considerations compared to rule-based or heuristic methods. Each GPT-4 scheduling of approximately 30–50 jobs required an average inference latency of 15–25 s and an estimated API cost between $0.03 and $0.08, based on OpenAI’s public pricing as of 2025. Classical algorithms, including SPT, EDF, FCFS, SA, and RRG, as well as the Heuristic Co-Pilot, executed almost instantaneously (without the least latency) on the same CPU hardware and did not require any external API dependence. For real-time or large-scale industrial deployment, we recommend a hybrid approach where GPT-4 acts as a supervisory reasoning layer guiding fast local heuristics or constraint solvers with its ability to explain scheduling trade-offs, interact naturally with human operators, and adapt to changing tariff constraints. It would bring an excellent balance between computational efficiency and explainable, human-centered decision support in line with the goals of Industry 5.0.

5. Results and Discussion

We demonstrate the performance of our hybrid Co-Pilot framework by evaluating its operation in detail using the CTU synthetic, Kaggle manufacturing, and Zenodo optimization datasets. We compute classical scheduling strategies, namely the Shortest Processing Time (SPT), Earliest Due First (EDF), and First Come First Served (FCFS), as well as factory-reported schedules, when available, as the initial benchmark on the three datasets. Then, we compare the hybrid Co-Pilots, the heuristic, and the GPT-4, generated schedules against the classical baselines. To enhance results comparisons of the baseline heuristics (SPT, EDF, FCFS) and the LLM-based Co-Pilot, we also implemented two additional AI metaheuristic approaches, namely Simulated Annealing (SA) and Randomized Restart Greedy (RRG). On one side, the SA algorithm refines job sequences through iterative probabilistic improvements. On the other side, the RRG method explores multiple randomized orderings to avoid local optima.

In

Table 3. Consolidated scheduling results (Energy = kWh, Cost = $, Peak = %).

Dataset	Strategy	Energy	Cost	Peak
CTU Synthetic	SPT/EDF/FCFS	106.5	12.36	30.0
	Heuristic Co-Pilot	106.5	12.40	27.3
	GPT-4 Co-Pilot	286.8	31.60	25.2
	SA (Metaheuristic)	1062.00	127.20	33.1
	RRG (Metaheuristic)	1062.00	127.20	33.1
Kaggle Manufacturing	SPT	85.0	10.0	30.5
	EDF	85.2	10.1	29.8
	FCFS	86.0	10.3	31.2
	Heuristic Co-Pilot	85.5	10.2	28.7
	GPT-4 Co-Pilot	86.1	10.4	28.5
	SA (Metaheuristic)	8521.34	994.20	33.1
	RRG (Metaheuristic)	8521.34	1002.89	32.6
Zenodo Optimization	Factory	240.0	29.8	32.0
	SPT	225.0	27.5	31.0
	EDF	230.0	28.1	30.8
	FCFS	228.0	27.9	31.5
	Heuristic Co-Pilot	220.0	24.6	28.0
	GPT-4 Co-Pilot	190.0	23.2	24.5
	SA (Metaheuristic)	87,837.60	9155.78	20.2
	RRG (Metaheuristic)	87,837.60	9879.15	27.1

, we present a summary of the consolidated performance across all datasets in terms of energy consumption (kWh), total cost ($), and peak load share (% of time in peak windows). We present in Figure 3, Figure 4 and Figure 5, the scheduling outputs (Gantt chart views) for all three datasets generated by both the heuristic and GPT-4 Co-Pilots. Figure 6, Figure 7 and Figure 8 illustrate the SA and RRG AI approaches ’scheduling patterns in Gantt charts, running on the three datasets. They show visual insight into how metaheuristic optimization compares with heuristic and LLM-driven scheduling in terms of energy use, cost, and peak load performance. In these figures, the x-axis represents the scheduling time, and each color bar represents a job’s scheduled interval. In other words, we visualize the scheduled jobs over time (x-axis in minutes). We illustrate energy consumption (kWh), total cost ($), and peak load share (% of time in peak windows) results for each dataset individually in Figure 9 (CTU synthetic dataset), Figure 10 (Kaggle Manufacturing dataset), and Figure 11 (Zenodo Optimization dataset).

Overall, the heuristic Co-Pilot produces compact and rule-based job sequences. The GPT-4 Co-Pilot leverages contextual reasoning to shift jobs away from peak tariff windows. The differences in spacing and task distribution visually reflect the trade-offs between the KPIs energy consumption, cost, and peak load share.

For the heuristic and GPT-4 co-pilots, the output results of the CTU synthetic data are presented in Figure 3. The machine loads are aggregated in a single line (‘0’ from the y-axis). The heuristic Co-Pilot keeps jobs densely packed with minimal idle periods. The GPT-4 Co-Pilot spreads tasks more widely, introducing gaps that correspond to avoided peak tariff windows. Results from these figures match the result summary in Table 3, where GPT-4 reduced peak share (25.2%) but increased energy/cost due to over-slack scheduling. We present the Kaggle dataset output results in Figure 4. The GPT-4 Co-pilot introduces minor temporal spacing, which slightly reduces peak load share. The Kaggle dataset involves multiple machines (M01–M05). For the Zenodo datasets, the output results in Figure 5 show that the heuristic Co-Pilot maintains a consistent load, albeit with several overlapping dense segments. The GPT-4 Co-Pilot redistributes job starts more intelligently. The visible gaps correspond to avoided peak windows, which visually illustrate the best performance (190 kWh, $23.2, 24.5% peak) achieved by the GPT-4 Co-pilot. The Zenodo dataset involves multiple machines (MAQ118–120).

5.1. CTU Synthetic Dataset-SPT, EDF, FCFS and Co-Pilots

As its name suggests, the CTU synthetic dataset is an artificial dataset created to serve as a stress-test environment. It contains simplified job–machine–energy profiles. The evaluation of the CTU synthetic dataset on the classical baselines (SPT, EDF, and FCFS) yielded a similar outcome, with approximately 106.5 kWh of energy consumption for $12.36, and 30% of operations scheduled in peak tariff windows. By implementing the heuristic Co-Pilot, we achieved a modest improvement because the Co-Pilot shifts selected tasks out of peak hours. Energy and cost remained essentially unchanged compared to the baselines, but the peak load share slightly decreased from 30% to 27.3%. These results indicate that in synthetic data, rule-based adjustments can effectively reduce the load. The GPT-4 Co-Pilot reduced the peak load share even more to 25.2% by performing a more aggressive reallocation. However, it came at the cost of inflated energy consumption (286.8 kWh) and a significantly higher operating cost ($31.60). The LLM Co-Pilot overgeneralized slack constraints and failed to interpret the simplicity of the synthetic dataset. It shows that heuristics are more reliable for synthetic stress tests, whereas GPT-4 requires tighter constraint encoding. Figure 9 clearly illustrates these results.

5.2. Kaggle Manufacturing Dataset-SPT, EDF, FCFS and Co-Pilots

In contrast to the CTU synthetic dataset, the Kaggle dataset is more realistic. It reflects medium-scale industrial conditions with heterogeneous jobs and realistic due dates. In this dataset, the three classical baseline strategies also performed similarly. They consumed approximately 85–86 kWh of energy, and the costs were around $10–10.3, with peak load shares of close to 30%. The heuristic Co-Pilot slightly reduced the peak load share to 28.7%. But resulted in similar outcomes to the classical baselines in terms of energy and cost. The LLM Co-Pilot achieved identical results to the heuristic Co-Pilot, with an energy consumption of 86.1 kWh for $10.40 and a peak share of 28.5%. Unlike in the CTU dataset, the GPT-4 Co-Pilot produced more stable results, comparable to those of the heuristic Co-Pilot, on the Kaggle dataset. This outcome shows that in realistic industrial settings, GPT-4 can provide scheduling solutions that are not only comparable to heuristic baselines but also offer enhanced explainability for operator-in-the-loop applications. Figure 10 provides a clear illustration of the Kaggle dataset results.

5.3. Zenodo Optimization Dataset-SPT, EDF, FCFS, and Co-Pilots

The Zenodo dataset is another very realistic industrial-type dataset resulting from previous optimization studies. Its structure closely resembles industrial-grade scheduling. From this dataset, the classical baselines computed higher consumption and cost than in the earlier datasets. The results range from 225 to 240 kWh of energy, with a cost of $27–30, and peak load shares exceeding 30%. The heuristic Co-Pilot achieves improved results by lowering energy consumption to 220 kWh, reducing costs to $24.60, and decreasing peak usage to 28%. In this dataset, the GPT-4 Co-Pilot achieved the most significant improvements overall. It reduced energy consumption to 190 kWh, the cost to $23.20, and the peak load share to 24.5%. In short, the GPT-4 Co-Pilot achieved an improvement of 4–7% over classical baselines and the heuristic Co-Pilot. It achieves this outcome by effectively exploiting slack. The GPT-4 Co-Pilot clearly outperformed heuristics and baselines under realistic constraints. According to results from the Zenodo dataset, we can conclude that the GPT-4 Co-Pilot has the potential to be successfully integrated into industrial decision-support systems. It can deliver tangible benefits in complex and realistic scheduling scenarios. Figure 11 illustrates the Zenodo optimization dataset scheduling results.

5.4. Comparative Discussion

When analysing all the classical baselines, heuristic Co-Pilot, and GPT-4 Co-Pilot results obtained through the three datasets, we conclude that the heuristic Co-Pilot is reliable across all datasets. It slightly but consistently reduces peak usage without cost increases. It is a safe choice for constrained or safety-critical scenarios. The GPT-4 Co-Pilot’s performance depends on the realism of the dataset. On the synthetic dataset, GPT-4 tends to overgeneralize constraints and inflate energy and cost. On real and optimization-based data, GPT-4 adapted well, achieving measurable improvements in energy and cost while maintaining feasibility. In summary, on the CTU dataset, GPT-4 inflates energy/cost and reduces peak share. The Co-Pilot maintains baseline efficiency. On the Kaggle dataset, GPT-4 and the heuristic closely match baselines, with slight improvements in peak usage. On the Zenodo dataset, GPT-4 clearly outperforms both baselines and the heuristic, lowering both energy and cost by 4–7%.

The AI metaheuristics (SA and RRG) results provide an essential comparative reference for our framework. They do not operate on any language model reasoning but demonstrate the potential of stochastic search to reach competitive or near-optimal configurations. Both SA and RRG reached stable schedules on the CTU Synthetic dataset. But they demonstrated higher energy and cost due to the synthetic dataset’s aggregated single-machine representation. This aspect of the dataset limits its flexibility in load balancing. It also explains why the Co-Pilot approaches, the GPT-4 Co-Pilot, outperform SA and RRG in peak reduction despite their broader search capability. Both SA and RRG achieve total energy values comparable to other methods on the Kaggle dataset. But they slightly increased the peak load. On the Zenodo dataset, SA achieves the lowest peak value (20.2%), which suggests an intense exploration of the solution space under time-of-use constraints. However, the total cost and energy magnitudes for the SA and RRG runs are significantly higher due to differences in dataset scaling and unit normalization. It highlights the sensitivity of metaheuristic performance to data pre-processing and objective formulation.

Overall, we conclude that, through this experiment, the LLM-based Co-Pilot achieves more consistent and interpretable energy-aware scheduling results than both rule-based heuristics and stochastic metaheuristics. The GPT-4 Co-Pilot in particular demonstrates a superior ability to trade off cost and peak minimization. It also respects operational constraints, which indicates that natural language reasoning can complement and even outperform optimization-based techniques in dynamic, human-in-the-loop manufacturing environments.

Across datasets, we observe that cost is the first KPI to influence industrial decision-making. The cost is linked to electricity tariffs that are directly related to operating expenses. However, we also noticed that the reduction of peak load share provides an additional long-term advantage as it minimizes demand charges and supports grid-friendly operations, aligning with sustainability goals. Total energy consumption is an indicator of overall process efficiency. In our framework, we integrate and assess the three KPIs together to help operators select a balanced schedule that best fits the factory’s economic and environmental priorities.

We can deduce that the LLM Co-Pilot has high potential to be safely utilized as an efficient scheduling tool when deployed under realistic industrial conditions, especially when they are integrated with fallback heuristics that safeguard feasibility. Our proposed hybrid human–AI scheduling framework is a solution well-aligned with the principles of Industry 5.0, which emphasizes sustainability, resilience, and transparency in operator-in-the-loop decision-making.

It is also worth mentioning that when omitting energy-awareness, both the heuristic and LLM Co-Pilots reduce to classical scheduling policies dominated by due-time adherence and throughput optimization. The heuristic Co-Pilot becomes identical to an EDF, and the GPT-4 Co-Pilot converges toward similar start-finish sequences. The improvement in sustainability metrics (energy cost and peak load), therefore, develops entirely from the TOU-driven rescheduling logic embedded in both algorithms.

The explainable reasoning generated by the GPT-4 Co-Pilot increases operator understanding of scheduling trade-offs with clear natural-language rationales for adjustments (job shifting or prioritization under time-of-use tariffs). This capability enables production staff to interpret complex optimization outcomes without requiring expert knowledge in programming or modeling. Therefore, we conclude that beyond the measurable improvements in energy and cost efficiency, the Co-Pilot framework provides tangible support for human-centered manufacturing. Practically, the framework helps reduce mental workload and improve trust between human operators and AI systems. These features reflect the collaborative intelligence philosophy of Industry 5.0, where technology amplifies human decision-making instead of replacing it. Our Co-Pilot framework contributes to sustainable energy management, as well as the creation of resilient, human-centric factories of the future.

6. Conclusions

In this research, we presented the design of an LLM-based Co-Pilot for energy-aware production scheduling in the context of Industry 5.0. It aims to balance throughput, energy consumption, and cost while providing explainable, human-in-the-loop decision support for operators. We created two co-pilots: a heuristic co-pilot and a GPT-4 one, which integrates GPT-4 LLM reasoning features with classical heuristic production scheduling strategies to achieve the desired outcomes.

We evaluated our system design using three primary open-source datasets: the CTU synthetic dataset, the Kaggle manufacturing dataset, and the Zenodo optimization dataset, each reflecting different production contexts. We observed that the heuristic Co-Pilot results consistently match classical baselines (SPT, EDF, FCFS). They maintained energy, cost, and lower peak load exposure in most cases. The GPT-4 Co-Pilot delivered mixed results: it slightly underperforms on synthetic CTU data containing simpler and synthetic scheduling patterns, on the Kaggle dataset it produced balanced results similar to the heuristic Co-Pilot confirming that GPT-4 adapts effectively to manufacturing data without excessive cost penalties, and it achieved superior results on the more industrial realistic and complex Zenodo dataset, achieving 4–7% reductions in energy and cost and a noticeable reduction in peak-share.

To further benchmark the LLM Co-Pilot performance, we incorporated two AI metaheuristic baselines (SA and RRG) to provide valuable comparative insights into stochastic optimization behavior under time-of-use energy constraints and compared their results. Both AI metaheuristic approaches achieved valid feasible schedules, but their cost and peak efficiency were generally lower than the Co-Pilot’s, especially in the CTU and Kaggle datasets, revealing limited adaptability. These outcomes highlight the complementary value of integrating deterministic heuristics, LLM reasoning, and metaheuristic search within a unified scheduling framework, as stochastic exploration primarily benefits from higher job and machine diversity. The proposed LLM Co-Pilot demonstrates a promising balance between transparency, flexibility, and sustainability, which are the core principles of Industry 5.0’s human-centric and environmentally conscious objectives.

7. Future Works

Our future research will focus on enhancing the LLM Co-Pilot’s system performance, robustness, and scalability. We will reduce performance inconsistency across datasets by using prompt optimization and fine-tuning. We will implement internal hybrid architectures that combine GPT-4 LLM reasoning with constraint solvers or reinforcement learning to improve system robustness and scalability. We aim to extend the Co-Pilot to facilitate real-time operator interaction, enabling adaptive and transparent scheduling in sustainable factory environments. Additionally, we also plan to compare the LLM Co-Pilot’s performance with other advanced AI optimization approaches, such as genetic algorithms and deep reinforcement learning (deep Q-learning and evolutionary optimization), to further position the Co-Pilot’s performance within the broader landscape of intelligent manufacturing optimization.