An AI Ladder-Based Energy Management System for Sustainability Score Prediction in Supply Chain

2.png

Figure 2. AI Ladder-based energy management system (EMS) framework for Sustainability Score prediction

2.1 Phase 1: Collect

The Collect phase constitutes the first operational rung of the IBM AI Ladder and is defined in this study as the systematic acquisition of structured, traceable, and decision-relevant supply chain records from distributed edge-level sources before central consolidation [4]. Such a design is methodologically necessary because the Sustainability Score targeted in this study is not determined by a single environmental indicator, but by the interaction of material, energy, emissions, logistics, cost conditions that emerge across different operational sites and reporting boundaries [1, 8]. Thus, the Collect stage is framed as the deliberate construction of an auditable evidence base capable of preserving operational granularity while making heterogeneous records sufficiently accessible for downstream integration in an EMS-oriented architecture [6, 14, 18].

2.1.1 Multi-source supply chain data acquisition model

The proposed acquisition model, illustrated in Figure 3, collects raw material usage and material consumption from material consumption reports and raw material usage reports because materials data are necessary to represent the physical input structure from which waste intensity and resource efficiency can be assessed [1, 19]. Energy consumption, electricity consumption, machine-level energy, and manufacturing energy are acquired from energy meter logs, electricity consumption reports, machine energy logs, and production records because energy-related variables are central to both EMS logic and the environmental interpretation of production performance [20]. Renewable energy share is collected from renewable energy records and utility-related documentation because the study requires not only total energy use, but also the extent to which operational energy demand is met through cleaner energy sources that alter the sustainability meaning of otherwise similar production profiles [6, 20]. Waste generation is captured from waste management reports because contemporary reporting standards explicitly link waste-related practices to materials use and value-chain impacts, making waste a necessary operational rather than merely descriptive variable in sustainability assessment [8, 21]. Carbon footprint and CO₂ emissions are obtained from carbon accounting records and emissions reports because corporate sustainability measurement requires activity data that can support organization-level and value-chain Greenhouse Gases (GHG) inventories in a standardized and comparable manner [22, 23].

3.png

Figure 3. Edge-level data acquisition model

Transport distance, logistics activity, delivery time, and cost are collected from transport records, shipment logs, delivery performance records, and cost records because upstream and intra-chain sustainability conditions are materially shaped by freight distance, transport mode, delivery performance, and service expenditure rather than by production data alone [4, 24].

2.1.2 Edge-level data capture and source coverage

Data are captured as close as possible to their point of origin at the edge level, including production units, metering points, warehouses, transport interfaces, emissions-accounting points, waste handling functions, and supplier-facing reporting channels, because proximity to the source preserves operational specificity and reduces the loss of contextual information that often occurs when sustainability data are reconstructed only after aggregation [18]. This edge-level orientation is particularly important for logistics and transport variables, transport-related emissions assessment depends on shipment-level information such as distance, mode, fuel, mass, storage conditions, and provider activity rather than on generalized annual summaries alone [22]. For that reason, the study treats source heterogeneity, because the Sustainability Score is intended to reflect interactions among purchasing, production, energy use, transport, emissions, and waste generation, all of which are recorded in different operational systems and organizational documents [1, 8].

2.1.3 Data transmission and consolidation to the core layer

After acquisition, as depicted in Figure 4, the collected and aggregated records are transmitted from distributed edge sources to a centralized core data environment through digital extraction and synchronized reporting interfaces, because the methodological purpose of the Collect stage is not only to gather records, but also to make them uniformly accessible across source boundaries in accordance with the AI Ladder logic [18].

4.png

Figure 4. AI Ladder-based supply chain data acquisition model

This transmission step is necessary because raw supply chain evidence is ordinarily fragmented across operational reports, metering logs, transport documents, environmental inventories, and supplier disclosures, and therefore cannot support coherent sustainability scoring unless it is first assembled into a common evidence base with preserved provenance [4]. The design also aligns with EMS principles, since energy-related decision support depends on the regular availability of measured consumption data and associated operational context rather than on isolated or ad hoc reporting snapshots [6]. By treating collection, access, and traceable transfer as one integrated methodological activity, the study ensures that the raw records entering the analytical pipeline remain closely tied to the operational conditions from which they were generated [19-21, 23]. These collected multi-source records are then consolidated into the core data environment, specifically a data lake, for the subsequent Organize stage [4].

2.1.4 Dataset source, scope, and reproducibility

The dataset used in this study was derived from real enterprise records related to green supply chain operations. The dataset was structured to represent product-level supply chain observations across five product categories, namely electronics, pharmaceutical, food, apparel, and automotive. Each observation represents one supply chain record containing operational, environmental, logistics, energy, cost, and LCA-derived sustainability information.

The dataset consists of 1,029 product-level supply chain observations across five product categories, namely electronics, pharmaceutical, food, apparel, and automotive. Each observation represents one supply chain record containing operational, environmental, logistics, energy, cost, and LCA-derived sustainability information. The records were collected from enterprise operations located in the Greater Jakarta metropolitan area (Jabodetabek), including Jakarta, Bogor, Depok, Tangerang, and Bekasi, Indonesia. Since the dataset was constructed from real enterprise records, the terms “reports” and “logs” in this study refer to actual enterprise documentation and system-generated records rather than conceptual or simulated source categories.

To support reproducibility while maintaining confidentiality, the study provides the variable dictionary, data structure, target-variable definition, LCA-based Sustainability_Score scoring logic, data preprocessing procedure, Random Forest (RF) configuration, fixed random seed, 10-fold cross-validation scheme, and evaluation metrics reported as mean ± standard deviation. Although the raw enterprise records cannot be fully disclosed due to confidentiality restrictions, the analytical pipeline can be reproduced using the same variable structure, scoring procedure, preprocessing steps, and modelling configuration.

2.2 Phase 2: Organize

The Organize phase constitutes the second stage of the IBM AI Ladder and, in the context of this study, refers to the consolidation, harmonization, validation, and structuring of the records acquired during the Collect phase so that they become analytically coherent and decision-ready within a centralized core environment [18]. In methodological terms, this stage is necessary because the Sustainability Score investigated in this study is inferred from the interaction of material, energy, emissions, logistics, and cost variables. The Organize stage is therefore designed as a governed transformation process through which distributed reports and logs are converted into a traceable, interoperable, and analysis-ready dataset [25].

2.2.1 Data consolidation and schema harmonization

The first organizing activity, shown in Figure 5, is the consolidation of distributed edge-level records into the core layer, where material consumption reports, raw material usage reports, energy meter logs, electricity records, machine energy logs, production records, renewable energy records, waste reports, emissions inventories, transport documents, delivery records, cost records, and supplier disclosures are brought into a unified tabular environment with consistent schema definitions.

5.png

Figure 5. Data transmission and consolidation to the core layer

This step is methodologically necessary because heterogeneous operational sources frequently encode similar phenomena under different field names, units, reporting intervals, and reference identifiers, which would otherwise distort downstream interpretation by making semantically related variables appear incomparable or duplicative [25]. Accordingly, the Organize stage harmonizes variable naming conventions across sources, aligns equivalent fields under common attribute labels, and standardizes units of measurement so that variables are represented in a common analytical vocabulary rather than in source-specific formats [21].

Timestamp fields and reporting-period boundaries are likewise aligned to a common temporal structure because a predictive framework cannot infer valid relationships among operational and environmental variables if those variables refer to inconsistent observation windows across facilities, processes, or shipments [24]. The same logic applies to identifiers: company, facility, process, product, batch, and shipment references are normalized so that records from internal operations and external supplier disclosures can be joined without weakening provenance or creating ambiguous links between operational events and sustainability attributes.

2.2.2 Operational definition of Sustainability Score

In this study, Sustainability_Score is defined as a continuous target variable representing the overall sustainability performance of each supply chain record. The score was derived from a Life Cycle Assessment (LCA)-based evaluation of each record and was prepared prior to the machine learning (ML) modelling stage. Therefore, Sustainability_Score served as the dependent variable to be predicted from structured supply chain indicators.

The LCA-based scoring procedure followed the general principles of ISO 14040 [26] and ISO 14044 [27], which frame LCA through goal and scope definition, life cycle inventory analysis, life cycle impact assessment, and interpretation. The scoring process used the available life cycle inventory and sustainability-related records. These indicators were transformed into normalized impact values and then aggregated into a 0-100 sustainability scale. Lower environmental burden indicators were reverse-coded so that lower impact contributed to a higher Sustainability Score, while beneficial indicators such as renewable energy share contributed positively.

The theoretical scale ranges from 0 to 100, where higher values indicate stronger sustainability performance and lower values indicate weaker sustainability performance. In the organized dataset used in this study, the observed Sustainability_Score ranged from 12.80 to 91.53. Scores between 80 and 100 indicate very strong sustainability performance, 60-79 indicate good sustainability performance, 40-59 indicate moderate sustainability performance requiring managerial review, 20-39 indicate weak sustainability performance, and scores below 20 indicate critical sustainability performance requiring urgent intervention.

The scoring was conducted as a structured LCA-based assessment by the research team before predictive modelling. To improve consistency, the same scoring rule, normalization procedure, and weighting logic were applied to all 1,029 observations. The reliability of the scoring process was checked through source-record verification, range plausibility checks, missing-value inspection, duplicate-record screening, consistency checks between LCA indicators and source variables, and sensitivity testing of the scoring weights. These procedures were intended to ensure that Sustainability_Score represented a consistent and auditable sustainability assessment rather than an arbitrary or manually assigned label.

2.2.3 Data quality assurance with preliminary exploratory

The second organizing activity is quality assurance, which is performed to ensure that the unified dataset preserves both internal consistency and source traceability before any analytical use is considered [25]. Thus, a preliminary exploratory profiling step was conducted following the data consolidation and schema harmonization process shown in Figure 6.

6.png

Figure 6. Data consolidation and schema harmonization process

This is essential because fragmented sustainability records often contain duplicate entries, missing values, incompatible reporting scopes, and inconsistent units or category definitions, all of which can bias later estimation of Sustainability Score by altering the apparent relationship between resource use, emissions, logistics performance [22]. For this reason, overlapping records are deduplicated by checking shared identifiers and reporting periods, incomplete entries are evaluated against source context, and inconsistent values are reconciled through source-priority rules that privilege primary operational logs over derivative summaries where both exist. Provenance metadata are preserved throughout this process by retaining source tags, timestamps, facility identifiers, reporting boundaries, and document lineage, because data quality in a sustainability-oriented EMS context depends not only on numerical correctness but also on the ability to audit how each variable entered the analytical pipeline [22].

In this study, normalization of structured tabular formats, alignment of data types, and preservation of variable definitions are treated as prerequisites for valid downstream analytics, since predictive modelling on incoherent enterprise data would risk learning reporting artifacts rather than sustainability-relevant relationships [14, 25].

2.3 Phase 3: Analyze

Since Sustainability Score has been defined as an existing continuous numerical target in the organized dataset, the Analyze phase formulates the task as a regression problem rather than a classification problem. The Analyze phase applies RF regression to estimate Sustainability Score from the structured supply chain dataset. Since Sustainability Score is a continuous numerical target, the task is formulated as a regression problem rather than a classification problem [16]. In this phase, the organized dataset is transformed into predictive sustainability intelligence by identifying relevant predictors, examining data patterns, and evaluating model performance. Before model fitting, Exploratory Data Analysis (EDA) was conducted through correlation analysis, box plots, and distribution analysis. Correlation analysis examined the direction and strength of relationships between predictors and Sustainability Score, box plots helped identify unusual observations, and distribution analysis showed the spread of each variable. These steps provided an initial understanding of the dataset before predictive modelling was performed [16].

Following EDA, variable ranking was performed using Univariate Regression and RReliefF. Univariate Regression identified predictors with strong individual relationships to Sustainability Score, while RReliefF complemented this by assessing predictor relevance through local differences among neighboring observations. This made RReliefF suitable for the regression setting because it can capture local dependencies and potential interaction effects before RF modelling [28]. After variable ranking, the dataset was evaluated using cross-validation. The training data were used to build the RF regression model, while the testing data were used to evaluate performance on unseen observations. This separation helped reduce overfitting and avoid overly optimistic results caused by data leakage [16, 29].

2.3.1 Random Forest model configuration

RF regression was applied because Sustainability_Score is a continuous target variable and the relationships among supply chain indicators may be nonlinear or interaction-based. Model evaluation was conducted using 10-fold cross-validation. In this procedure, the dataset was divided into 10 folds. In each iteration, nine folds were used for model training and the remaining fold was used for testing. This process was repeated until each fold had served once as the testing subset.

The RF model was implemented as a baseline regression model with bootstrap sampling, squared error as the splitting criterion, 100 regression trees, and a fixed random seed to support reproducibility. Model performance was evaluated using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). The final performance values were reported as mean ± standard deviation across the 10 validation folds. This evaluation strategy provides a more stable estimate of the model’s generalization performance. The RF configuration is summarized in Table 1, and the Analyze phase workflow is shown in Figure 7.

Table 1. Random Forest (RF) model configuration

7.png

Figure 7. Analyze phase workflow

RF is selected because it is suitable for structured tabular data and can model nonlinear relationships and interaction effects without requiring restrictive linear assumptions. In the RF procedure, multiple bootstrap-resampled training subsets are generated from the original training data. Each subset is used to train a separate regression tree, and the collection of these trees forms an ensemble of regression trees. At each split within a tree, only a random subset of predictors is considered, which reduces correlation among trees and improves the generalization ability of the model [15, 30]. Once the ensemble model has been developed, the framework proceeds to performance evaluation. Within each cross-validation fold, the RF model was trained on nine folds and evaluated on the remaining fold. The final performance was obtained by averaging the evaluation metrics across all 10 folds and reporting their standard deviations. The evaluation metrics should correspond to the regression nature of the target variable, meaning that model performance is assessed based on the difference between predicted and actual Sustainability Score values [16]. This ensures that the evaluation process reflects the predictive objective of the study rather than relying on classification-based measures that are not appropriate for a continuous outcome.

The final output of the Analyze phase is the result of the RF regression analysis. This result includes the estimated Sustainability Score, the ranked importance of predictors, and the model performance findings. Together, these outputs provide evidence-based sustainability intelligence for the AI Ladder-based EMS framework. The result is not limited to prediction accuracy alone; it also supports interpretation by identifying which supply chain variables contribute most strongly to sustainability performance. This makes the Analyze phase useful for both predictive assessment and managerial decision-making within a sustainability-oriented supply chain context [1, 15].

2.4 Phase 4: Infuse

This section illustrates the development of an AI Ladder-based EMS that embeds Sustainability Score prediction into green supply chain processes. The approach follows the Infuse phase of the IBM AI Ladder, which focuses on embedding AI into workflows to enable smarter operations and enhanced decision-making [17]. Infusing AI into the core EMS workflows enables the system to transform structured supply chain data into predictive sustainability insights, supporting better prioritization, intervention, and greener managerial decisions [6, 7].

2.4.1 Energy Management System process re-engineering and integration design

As illustrated in Figure 8, the process starts with EMS process re-engineering activities, including the review of the current state (As-Is) and the design of the desired future state (To-Be). These steps are crucial to ensure that the existing EMS process is clearly mapped and improved before introducing predictive intelligence [31].

8.png

Figure 8. Energy management system (EMS) process re-engineering and integration design

A Focus Group Discussion (FGD) is used to gather expert input on how managers currently monitor separate indicators such as energy consumption, renewable energy share, emissions, waste generation, transport distance, cost, and delivery time. This input informs the design of a new EMS process in which the predicted Sustainability Score is integrated into the decision flow as a synthesized indicator for greener supply chain management [1, 6].

The redesigned EMS process allows the system to move beyond descriptive monitoring. Instead of displaying isolated indicators only, the infused process places the predicted Sustainability Score alongside the underlying operational and environmental variables, enabling managers to interpret sustainability performance in a more coordinated and decision-oriented manner [17]. In this way, the prediction output is not treated as a substitute for detailed indicators, but as an additional interpretive layer that supports monitoring, review, and action planning within the EMS environment [6, 12].

2.4.2 System embedding, configuration, and validation

In practical terms, integration between the RF model and the EMS is carried out through an analytical integration layer. This layer receives prediction outputs from the model and forwards them to EMS modules for visualization and decision support. This pattern ensures that the prediction result is connected to the relevant unit of analysis, such as facility, process, supplier, shipment, or reporting period, while preserving access to the underlying sustainability-related variables [1, 15].

The EMS is configured to display the predicted Sustainability Score together with the operational indicators from which it is inferred. In this way, managers can understand not only the predicted condition, but also the operational pattern behind it. Once embedded into the system, the score can support decisions related to energy-efficiency adjustment, process redesign, logistics optimization, emissions reduction, waste minimization, and renewable-energy prioritization [6].

To ensure that the developed EMS workflow functions properly, the infused system is subsequently validated through functional verification, consistency checks between displayed scores and source records, scenario-based review of alert logic, and managerial usability assessment of dashboards and reporting outputs. This validation confirms that the prediction output is correctly rendered, interpretable, and usable within the EMS environment [31]. As a result, the Infuse phase completes the AI Ladder cycle by turning Sustainability Score prediction into a reusable decision-support capability for greener and more coherent supply chain management [6, 17].

The results of this study demonstrate that Sustainability Score can be effectively predicted from structured green supply chain indicators when operational, environmental, logistics, energy, and cost variables are organized into a coherent analytical framework. Based on 10-fold cross-validation, the RF regression model achieved an R² value of 0.889 ± 0.007, with an MSE of 22.439 ± 0.948, RMSE of 4.737 ± 0.178, MAE of 3.711 ± 0.125, and MAPE of 8.588% ± 0.365%. These values indicate that the selected predictors explain a substantial proportion of the variation in Sustainability_Score and produce relatively low prediction error across validation folds. The small standard deviations across folds further suggest that the model performance is stable [1, 4].

The strongest empirical finding concerns the dominant role of renewable energy share in shaping Sustainability Score. In the variable ranking result, Renewable_Energy_% obtained the highest Univariate Regression score and the highest RReliefF score, while the correlation analysis also showed the strongest positive relationship between renewable energy share and Sustainability Score. This indicates that the quality of energy sourcing is a central determinant of sustainability performance in the observed dataset. Energy management systems can function as structured mechanisms for improving energy performance [12], and the result of this study extends that logic by showing that renewable energy share can function as a key predictive signal in an EMS-oriented sustainability framework. In practical terms, two supply chain activities may consume similar amounts of energy, but their sustainability implications can differ substantially when one relies more heavily on renewable energy. This result also helps explain why renewable energy share appears more influential than total energy consumption in the observed findings. Total energy consumption captures the quantity of energy used, whereas renewable energy share captures the quality of energy sourcing. Therefore, sustainability-oriented EMS should not only monitor how much energy is consumed, but also whether the energy profile supports cleaner and lower-carbon operations. Recent studies on AI-enabled energy management also emphasize that digital and ML approaches become more valuable when they help organizations identify operational patterns that influence energy efficiency and sustainability outcomes [6, 9].

The second and third strongest predictors, CO₂ emissions and waste generation, further validate the environmental logic of the Sustainability Score. CO₂ emissions showed a moderate-to-strong negative relationship with the target variable, while waste generation also showed a negative relationship. These results indicate that higher environmental burden is associated with lower sustainability performance. This interpretation is consistent with the Greenhouse Gas Protocol, which treats emissions measurement as a foundation for organizational greenhouse gas accounting and reporting [22]. Similarly, Global Reporting Initiative (GRI) Standards recognize emissions, energy, materials, and waste as important environmental dimensions in sustainability disclosure. The negative association between waste generation and Sustainability Score is also theoretically coherent. Waste reflects material loss, process inefficiency, disposal burden. In GRI 306: Waste 2020, waste is treated as a material sustainability topic when it reflects significant organizational and value-chain impacts [21]. In the context of this study, the empirical result supports that logic: as waste generation increases, the predicted sustainability condition tends to decline. Waste should therefore be understood not merely as a reporting attribute, but as a core predictor of green supply chain performance.

Several other variables, including transport distance, energy consumption, manufacturing energy, raw material usage, delivery time, cost, and carbon footprint, showed weaker individual linear relationships with Sustainability Score. However, weak correlation does not necessarily imply irrelevance. In predictive modelling, relationships among variables may be nonlinear, indirect, or interaction-based, especially when the outcome is shaped by multiple operational dimensions [16]. For example, transport distance may not independently predict Sustainability Score strongly, but it can become meaningful when interpreted together with emissions, delivery time, logistics cost, and energy use. The same logic applies to manufacturing energy, which may have different sustainability implications depending on renewable energy share and process efficiency. This supports the selection of RF regression as the main predictive model. Breiman [15] introduced RF as an ensemble method that combines multiple decision trees and can capture nonlinear relationships and interaction effects. In this study, that capability is important because Sustainability Score is shaped by the combined influence of renewable energy share, emissions, waste, logistics conditions, material use, cost, delivery performance, and carbon-related indicators. A purely linear model may underrepresent these complex relationships, whereas RF can learn predictive patterns from multiple variables without assuming that each predictor contributes independently or linearly.

The methodological validity of RF is further strengthened by the structure of the prediction task. Since Sustainability Score is a continuous numerical target, the study appropriately formulates the task as a regression problem. The use of MSE, RMSE, MAE, MAPE, and R² is therefore suitable because these metrics evaluate the difference between predicted and actual numerical values [16]. The R² value of 0.889 ± 0.007 indicates strong predictive capacity, while the MAPE value of 8.588% ± 0.365% suggests an average prediction error of approximately 8.6% across validation folds. These results should be interpreted as evidence of predictive usefulness rather than causal proof. The model shows that observed supply chain indicators can estimate Sustainability Score with reasonable accuracy, but it does not prove that changing one predictor will automatically cause a proportional change in the score.

From a system perspective, the most important implication lies in the Infuse phase. Many AI and predictive analytics studies stop after model evaluation, leaving the prediction output as an isolated analytical result. This study extends the analytical output into an EMS workflow by embedding the predicted Sustainability Score into dashboard visualization, risk indicators, alert notifications, recommendation panels, reporting, and audit logs. This design is aligned with the role of EMS as a structured framework for managing and improving energy performance [33]. It also supports the argument that AI-based energy management becomes more valuable when prediction and optimization outputs are connected to operational decision-making [6]. The As-Is and To-Be process comparison strengthens the managerial contribution of the study. In the As-Is condition, managers review separate indicators manually and depend on descriptive monitoring. In the To-Be condition, the organized dataset is processed by RF regression, transferred into EMS through an analytical integration layer, displayed through dashboards, and connected to alerts, recommendations, and reports. This transformation is consistent with the direction of AI-based logistics optimization, where analytics becomes more useful when sustainability criteria are embedded into operational decision processes rather than treated as separate reports [4]. In this way, the proposed EMS does not replace managerial judgment, but strengthens it by providing a structured, predictive, and traceable decision-support signal.

The UAT results further support the practical feasibility of the proposed EMS integration. Dashboard elements and EMS modules were reported as passed, indicating that the system can display predicted Sustainability Score, show supporting variables, filter records, generate alerts, provide recommendations, track historical logs, and export reports. These results validate the functional readiness of the proposed EMS design. However, the UAT should be interpreted as functional validation rather than long-term operational validation. Further deployment testing would still be needed to examine whether the EMS improves sustainability outcomes over multiple reporting periods, across different organizations, and under real-time operational conditions.

Overall, the findings validate the central proposition of this study: the value of predictive analytics in green supply chain management depends not only on model accuracy, but also on whether prediction outputs are embedded into managerial systems. The AI Ladder provides a coherent implementation pathway by linking data collection, organization, analysis, and system infusion. By combining RF regression with EMS-based workflows, the proposed framework helps organizations convert fragmented supply chain data into predictive sustainability intelligence that is visible, interpretable, traceable, and actionable. This makes the study theoretically relevant for sustainability measurement, methodologically relevant for predictive modelling, and practically relevant for energy-aware supply chain decision support.