The rapid adoption of AI and ML technologies in modern smart grids has significantly improved forecasting accuracy, operational efficiency, fault diagnosis, and demand-response management. However, the increasing complexity of DL and ensemble learning models has introduced a major challenge: the lack of interpretability. While advanced models often achieve superior predictive performance, their black-box nature makes it difficult for operators and decision-makers to understand the rationale behind generated predictions and recommendations. This limitation is particularly critical in smart-grid environments where operational decisions directly affect reliability, economic performance, and energy security. Recent reviews have identified explainability as one of the most important requirements for trustworthy AI deployment in energy systems [4].

XAI has emerged as a promising solution to address these concerns. XAI techniques aim to provide transparent explanations regarding how ML models generate predictions and which variables influence decision outcomes. Among existing methods, SHAP and LIME are the most frequently adopted approaches in energy-related applications. SHAP is particularly attractive because it provides both global and local interpretability while maintaining a strong theoretical foundation derived from cooperative game theory. Recent studies have demonstrated the effectiveness of SHAP in electricity demand forecasting, renewable energy prediction, fault detection, and demand-response optimization. These explanations enable operators to validate model outputs and identify influential variables such as weather conditions, electricity prices, and consumer behavior patterns [4].

Furthermore, explainability has become increasingly important because regulatory agencies and energy stakeholders demand transparency in AI-assisted decision-making processes. Interpretable models improve stakeholder trust, facilitate human oversight, and support compliance with emerging AI governance frameworks. Consequently, XAI is now considered a critical component of next-generation smart-grid intelligence systems rather than merely an optional analytical tool [4].

Energy forecasting represents one of the most important applications of AI in smart-grid environments. Accurate prediction of electricity demand, renewable energy generation, and market prices enables grid operators to optimize resource allocation and improve system stability. Recent studies have shown that ML models such as XGBoost, Random Forest, Artificial Neural Networks, and LSTM networks outperform many traditional statistical forecasting methods in terms of prediction accuracy and adaptability to nonlinear energy consumption patterns [11].

Hybrid intelligent systems have gained particular attention because they combine the predictive capabilities of ML with the interpretability of fuzzy logic. For example, Wang and Nishter [11] proposed a neuro-fuzzy framework for real-time load forecasting and adaptive control in smart grids. Their results demonstrated that integrating neural networks with fuzzy reasoning improved forecasting performance while maintaining interpretability and adaptability under dynamic operating conditions. Such findings indicate that hybrid approaches may provide a suitable foundation for developing intelligent decision-support systems in modern energy infrastructures.

Despite these advances, many forecasting models still focus exclusively on predictive accuracy while neglecting interpretability and uncertainty management. As a result, there remains a need for integrated frameworks capable of simultaneously addressing prediction accuracy, explainability, and decision uncertainty.

Uncertainty is an inherent characteristic of smart-grid environments due to fluctuating renewable energy production, changing consumer behavior, weather variability, and market dynamics. Conventional optimization methods often struggle to manage these uncertainties effectively. Fuzzy logic has therefore become a widely adopted methodology for energy management because it enables the representation of vague, uncertain, and linguistic information through fuzzy sets and inference rules.

Traditional Type-1 Fuzzy Logic Systems (T1FLS) have demonstrated effectiveness in energy management applications; however, they assume precise membership functions and may be insufficient when uncertainty levels are high. To overcome this limitation, T2FLS were developed. Type-2 fuzzy sets incorporate uncertainty directly into membership functions through the FOU, enabling more robust reasoning under ambiguous conditions [12].

Recent research has demonstrated the advantages of Type-2 fuzzy systems in various smart-energy applications. Aljohani [12] proposed an Intelligent Type-2 Fuzzy Logic Controller for hybrid microgrid energy management with electric vehicle integration. The results showed improved energy utilization, enhanced system stability, and reduced operational costs under uncertain renewable generation conditions. Similarly, robust Type-2 fuzzy controllers have been successfully applied to photovoltaic microgrids and battery energy storage systems, outperforming traditional Type-1 fuzzy approaches in power-quality improvement and uncertainty management.

In smart-home environments, Type-2 fuzzy controllers have also been employed to coordinate renewable energy sources, battery storage systems, and electric vehicles. Experimental results demonstrated reductions in electricity consumption, peak demand, and operational costs while maintaining user comfort and system reliability. These findings suggest that Type-2 fuzzy reasoning is particularly suitable for sustainable energy management applications where uncertainty is unavoidable.

Although XAI and Type-2 fuzzy systems have individually demonstrated significant benefits, research integrating these two paradigms remains relatively limited. Existing studies generally focus on either explainable forecasting or uncertainty-aware decision-making. Very few frameworks utilize knowledge extracted from explainable ML models to enhance fuzzy reasoning and decision-support processes.

This gap is particularly significant because explainability and fuzzy reasoning possess complementary characteristics. XAI techniques can identify the most influential variables affecting prediction outcomes, while Type-2 fuzzy systems can use this knowledge to generate transparent and uncertainty-aware decisions. The combination of these approaches has the potential to improve trustworthiness, robustness, and interpretability simultaneously. Recent reviews have identified the integration of XAI, uncertainty quantification, and human-centered decision support as one of the most promising research directions in intelligent energy systems [4].

Therefore, the development of a unified framework that combines ML forecasting, SHAP-based explanations, and IT2FIS represents an important research opportunity. Such a framework can support sustainable energy management by providing accurate predictions, transparent explanations, and robust decisions under uncertainty.

Based on the reviewed literature, three major research gaps can be identified:

I. Most smart-grid forecasting studies prioritize prediction accuracy while neglecting explainability.

II. Existing Type-2 fuzzy energy-management systems rarely exploit information extracted from XAI models.

III. Limited research has investigated integrated frameworks that simultaneously address forecasting, interpretability, uncertainty management, and sustainable energy optimization.

To address these limitations, this study proposes an XAI–Driven Type-2 Fuzzy Decision Support Framework that integrates SHAP-based explainability with Interval Type-2 fuzzy reasoning for sustainable smart-grid energy management.

The proposed framework aims to develop a transparent, uncertainty-aware, and intelligent decision-support system for sustainable energy management in smart-grid environments. The framework integrates three complementary technologies: ML-based forecasting, XAI, and IT2FIS. The rationale behind this integration is that accurate predictions alone are insufficient for critical energy infrastructures. Instead, decision-makers require both interpretable explanations and robust mechanisms for handling uncertainty.

As illustrated, the proposed architecture consists of four interconnected layers:

I. Data Acquisition and Preprocessing Layer II. ML Forecasting Layer III. Explainability Layer IV. Type-2 Fuzzy Decision Support Layer The framework continuously collects operational data from smart-grid infrastructures and converts them into actionable energy-management recommendations. This architecture aligns with recent trends in trustworthy AI and explainable smart-grid intelligence, which emphasize transparency, human-centered decision-making, and uncertainty-aware control mechanisms.

The first layer gathers real-time and historical information from multiple smart-grid components. The collected variables include:

 Historical electricity demand  Electricity market prices  Solar irradiance  Wind speed  Ambient temperature  Relative humidity  Battery State of Charge (SOC)  Renewable energy generation output  Electric vehicle charging demand

Because smart-grid data are typically heterogeneous, noisy, and incomplete, a preprocessing stage is implemented before model training.

The preprocessing pipeline consists of:

Data cleaning: missing values are handled using interpolation or K-nearest neighbor imputation techniques. Outliers are detected through interquartile range analysis and z-score methods.

Feature engineering: additional features are extracted from raw data, including:

 Hour of day  Day of week  Peak/off-peak indicator  Seasonal index  Renewable penetration ratio

Data normalization: all variables are normalized using Min-Max scaling:

X_{norm}=\frac{X-X_{min}}{X_{max}-X_{min}}

This normalization improves learning stability and convergence during model training.

The second layer is responsible for predicting future energy conditions. Recent studies indicate that advanced ML models significantly outperform traditional statistical forecasting approaches in smart-grid applications. Hybrid forecasting systems have demonstrated superior performance in capturing nonlinear energy consumption patterns and renewable generation fluctuations.

In the proposed framework, XGBoost is selected as the primary forecasting model because of its:

 High predictive accuracy  Computational efficiency  Robustness against overfitting  Compatibility with SHAP explanations

The forecasting model receives the preprocessed feature vector:

X_t = {D_t, T_t, H_t, P_t, S_t, W_t}

 (D_t) = historical demand  (T_t) = temperature  (H_t) = humidity  (P_t) = electricity price  (S_t) = solar irradiance  (W_t) = wind speed

The forecasting function is expressed as:

\hat{Y}_{t+1}=f(X_t) \hat{Y}_{t+1}

represents predicted electricity demand or renewable generation for the next time interval.

The forecasting output serves as the primary input to the explainability and decision-support modules.

The third layer addresses one of the major limitations of modern ML systems: lack of interpretability. Recent reviews emphasize that explainability is becoming an essential requirement for trustworthy smart-grid applications because operators must understand why a prediction was generated before taking corrective actions. SHAP has emerged as one of the most reliable explanation techniques in energy forecasting due to its theoretical consistency and ability to provide both local and global interpretations.

The proposed framework adopts SHAP to quantify feature contributions.

For a prediction (f(x)), the SHAP explanation is represented as:

f(x)=\phi_0+\sum_{i=1}^{M}\phi_i

 (\phi_0) denotes the average prediction,  (\phi_i) represents the contribution of feature (i),  (M) is the number of input variables.

The explainability layer generates:

Global explanations: identification of the most influential variables affecting forecasting performance across the entire dataset.

Local explanations: interpretation of individual predictions.

Temporal explanations: analysis of feature importance changes over time.

The extracted knowledge is subsequently transferred to the fuzzy decision-support module, enabling the incorporation of explainable information into the decision-making process.

Although explainability improves transparency, uncertainty remains unavoidable in smart-grid operations. Renewable generation, consumer behavior, and electricity prices are inherently uncertain and difficult to model precisely.

To address this challenge, the proposed framework employs an IT2FIS.

Recent studies have shown that Type-2 fuzzy systems outperform conventional Type-1 approaches in energy-management applications because they explicitly represent uncertainty through the FOU. This capability leads to more robust decision-making under dynamic operating conditions.

The fuzzy decision engine receives four inputs:

I. Forecasted Demand II. Renewable Energy Availability III. Electricity Price IV. Battery SOC The output variable is:

Energy Management Action Level (EMAL): with five linguistic states:

I. Very Low II. Low III. Medium IV. High V. Critical Fuzzy rule base: example rules include:

Rule 1.

IF Demand is High AND Renewable Generation is Low AND Electricity Price is High THEN Energy Saving Action is Critical Rule 2.

IF Demand is Medium AND Renewable Generation is High THEN Energy Saving Action is Low Rule 3.

IF Battery SOC is High AND Demand is High THEN Battery Discharge Action is High The fuzzy reasoning process consists of:

 Fuzzification  Rule Evaluation  Type Reduction  Defuzzification

The final crisp output determines the optimal energy-management strategy.

The complete decision-support process operates as follows:

Step 1. Collect smart-grid operational data.

Step 2. Perform preprocessing and feature extraction.

Step 3. Predict future demand and renewable generation using XGBoost.

Step 4. Generate SHAP explanations.

Step 5. Extract influential factors.

Step 6. Feed explainable information into the IT2FIS engine.

Step 7. Generate energy-management recommendations.

Step 8. Execute operational actions and update the system continuously.

This integration enables simultaneous achievement of:

 High forecasting accuracy,  Explainability,  Uncertainty management,  Sustainable energy utilization.

Compared with existing smart-grid decision-support systems, the proposed framework offers several advantages:

 Improved forecasting accuracy through ML.

 Transparent prediction interpretation through SHAP.

 Robust uncertainty management using Interval Type-2 Fuzzy Logic.

 Increased operator trust and acceptance.

 Enhanced REU.

 Reduced operational costs and carbon emissions.

Consequently, the framework provides a comprehensive solution for trustworthy and sustainable smart-grid energy management.

The proposed XAI–Driven Type-2 Fuzzy Decision Support Framework integrates three mathematical components: 1) ML-based forecasting, 2) SHAP-based explainability analysis, and 3) IT2FIS-based decision-making. The objective is to transform raw smart-grid data into interpretable and uncertainty-aware energy management recommendations. Recent studies emphasize that future smart-grid intelligence systems should combine predictive accuracy, explainability, and uncertainty quantification within a unified framework to support trustworthy operational decision-making.

Let the smart-grid input vector at time (t) be defined as:

X_t = [D_t, T_t, H_t, S_t, W_t, P_t, SOC_t]

 (D_t): electricity demand,  (T_t): ambient temperature,  (H_t): relative humidity,  (S_t): solar irradiance,  (W_t): wind speed,  (P_t): electricity price,  (SOC_t): battery state of charge.

The objective of the forecasting model is to estimate future demand and renewable generation:

\hat{Y}_{t+1}=f(X_t)

where (f(\cdot)) denotes the ML forecasting function.

In the proposed framework, XGBoost is employed as the forecasting engine because of its robustness, scalability, and compatibility with explainability techniques. XGBoost constructs an ensemble of regression trees and generates predictions by minimizing a regularized objective function.

The prediction model is expressed as:

\hat{Y}{i}=\sum{k=1}^{K} f_k(X_i)

 (K) represents the number of decision trees,  (f_k) denotes the prediction generated by tree (k).

The corresponding optimization objective is:

Obj=\sum_{i=1}^{n}l(y_i,\hat{y}i)+\sum{k=1}^{K}\Omega(f_k)

 (l(\cdot)) denotes the loss function,  (\Omega(\cdot)) represents the regularization term used to control model complexity.

The forecasting model generates:

 Predicted electricity demand.

 Predicted renewable energy generation.

 Predicted energy imbalance indicators.

Accurate forecasting provides the foundation for subsequent explainability and decision-support processes. Recent smart-grid forecasting studies demonstrate that ML models significantly improve prediction accuracy when compared with conventional statistical approaches.

To improve transparency and trustworthiness, the framework employs SHAP.

SHAP quantifies the contribution of each feature to a prediction based on Shapley values derived from cooperative game theory. Recent reviews identify SHAP as one of the most reliable and widely adopted explainability techniques in energy forecasting and smart-grid intelligence.

For an instance (x), the prediction is represented as:

f(x)=\phi_0+\sum_{i=1}^{M}\phi_i

 (\phi_0) is the average model prediction,  (\phi_i) is the contribution of feature (i),  (M) is the number of input features.

The Shapley value is calculated as:

\phi_i=\sum_{S\subseteq N\setminus {i}} \frac{|S|!(M-|S|-1)!}{M!} [f(S\cup {i})-f(S)]

 (S) represents a subset of input features,  (N) is the complete feature set.

The resulting SHAP values provide:

Global importance:

GI_i=\frac{1}{n}\sum_{j=1}^{n}|\phi_{ij}|

which measures the overall influence of feature (i) across all observations.

Local importance:

LI_i=\phi_i

which explains a specific prediction.

The obtained SHAP information is transferred to the fuzzy decision-support layer, enabling explainability-aware decision-making.

Although explainability provides transparency, smart-grid environments remain highly uncertain due to renewable intermittency, weather variability, and consumer behavior fluctuations.

To model uncertainty explicitly, the framework employs Interval Type-2 Fuzzy Sets (IT2FS). Recent energy-management studies demonstrate that IT2FLS provides superior robustness compared with Type-1 fuzzy systems under uncertain operating conditions.

An IT2FS (\tilde{A}) is defined as:

\tilde{A}= \left{ (x,u),\mu_{\tilde{A}}(x,u) \right}

u\in J_x\subseteq[0,1]

The uncertainty is represented through the FOU:

FOU(\tilde{A})= \bigcup_{x\in X}J_x

The FOU is bounded by:

Upper Membership Function (UMF):

\overline{\mu}_{A}(x)

Lower Membership Function (LMF):

\underline{\mu}_{A}(x)

The interval between UMF and LMF captures uncertainty associated with linguistic variables.

The fuzzy decision-support system utilizes four inputs:

[D_f,R_g,P_e,SOC]

 (D_f): forecasted demand,  (R_g): renewable generation,  (P_e): electricity price,  (SOC): battery state of charge.

Each variable is represented by five linguistic terms:

VL,L,M,H,VH

corresponding to:

 Very Low,  Low,  Medium,  High,  Very High.

Gaussian interval Type-2 membership functions are employed:

\mu(x)= \exp \left( -\frac{(x-c)^2}{2\sigma^2} \right)

 (c) is the center,  (\sigma) is the standard deviation.

The fuzzy knowledge base consists of expert-defined IF–THEN rules.

A typical rule is:

R_i:

IF\ D_f\ is\ High \ AND R_g\ is\ Low \ AND P_e\ is\ High \ THEN EMAL\ is\ Critical

The firing strength of Rule 1 is computed as:

w_i= \mu_{D_f} \cdot \mu_{R_g} \cdot \mu_{P_e} \cdot \mu_{SOC}

For interval Type-2 systems:

w_i= [\underline{w_i},\overline{w_i}]

 (\underline{w_i}) denotes lower firing strength,  (\overline{w_i}) denotes upper firing strength.

A unique characteristic of Type-2 fuzzy systems is the type-reduction stage.

The framework employs the Karnik–Mendel (KM) algorithm, which remains the most widely adopted approach for centroid type reduction in interval Type-2 fuzzy systems. Recent studies continue to improve computational efficiency of KM-based procedures while preserving accuracy.

The left and right centroids are computed as:

Y_L= \frac{\sum_{i=1}^{N}x_iw_i^{L}} {\sum_{i=1}^{N}w_i^{L}}

Y_R= \frac{\sum_{i=1}^{N}x_iw_i^{R}} {\sum_{i=1}^{N}w_i^{R}}

The type-reduced interval is:

Y_{TR}=[Y_L,Y_R]

The final crisp decision output is obtained through centroid averaging:

\frac{Y_L+Y_R}{2}

The resulting value corresponds to the EMAL.

To evaluate the effectiveness of the proposed framework, a sustainability index is introduced:

SI= w_1REU+ w_2PCR+ w_3CER+ w_4CSR

 REU = Renewable Energy Utilization,  PCR = Peak Consumption Reduction,  CER = Carbon Emission Reduction,  CSR = Cost Saving Ratio.

Subject to:

\sum_{i=1}^{4}w_i=1

This index provides a unified measure of sustainability performance and enables comparison with benchmark energy-management approaches.

The complete mathematical workflow can be summarized as follows:

 Acquire smart-grid data.

 Forecast future demand and renewable generation using XGBoost.

 Generate SHAP explanations and feature importance values.

 Transform prediction outputs into fuzzy linguistic variables.

 Execute Interval Type-2 fuzzy inference.

 Apply KM type reduction.

 Perform defuzzification.

 Generate final energy-management recommendations.

This formulation establishes a mathematically rigorous and explainable framework capable of supporting sustainable smart-grid energy management under uncertainty.