Demand Response Potential Forecasting: A Systematic Review of Methods, Challenges, and Future Directions

Abstract

Demand response (DR) is increasingly recognized as a critical flexibility resource for modernizing power systems, enabling the large-scale integration of renewable energy and enhancing grid stability. While the field of general electricity load forecasting is supported by numerous systematic reviews, the specific subfield of DR potential forecasting has received comparatively less synthesized attention. This gap leaves a fragmented understanding of modeling techniques, practical implementation challenges, and future research problems for a function that is essential for market participation. To address this, this paper presents a PRISMA-2020-compliant systematic review of 172 studies to comprehensively analyze the state-of-the-art in DR potential estimation. We categorize and evaluate the evolution of forecasting methodologies, from foundational statistical models to advanced AI architectures. Furthermore, the study identifies key technological enablers and systematically maps the persistent technical, regulatory, and behavioral barriers that impede widespread DR deployment. Our analysis demonstrates a clear trend towards hybrid and ensemble models, which outperform standalone approaches by integrating the strengths of diverse techniques to capture complex, nonlinear consumer dynamics. The findings underscore that while technologies like Advanced Metering Infrastructure (AMI) and the Internet of Things (IoT) are critical enablers, the gap between theoretical potential and realized flexibility is primarily dictated by non-technical factors, including inaccurate baseline methodologies, restrictive market designs, and low consumer engagement. This synthesis brings much-needed structure to a fragmented research area, evaluating the current state of forecasting methods and identifying the critical research directions required to improve the operational effectiveness of DR programs.

1. Introduction

Modernizing power systems involves the growth of renewable energy alongside a move toward electric heating and transportation systems [1]. Implementing renewable power systems with electrical sector decarbonization adds operational difficulties to grid stability [2]. It creates problems with peak system demands while generating power unpredictably. Under current power system conditions, demand response (DR) operates as a fundamental solution to introduce flexibility through electricity demand modulation triggered by price signals, incentive structures, and grid reliability occurrences. The energy consumption becomes more efficient and adaptive through peak shaving and load shifting and targeted load reduction measures of DR programs [3,4]. Figure 1 gives a high-level view of a DR ecosystem. The figure illustrates a community where many stakeholders provide flexibility. Smart homes contribute through rooftop solar, electric vehicles, and behind-the-meter batteries, while factories and small and medium enterprises add controllable loads and on-site resources. An aggregator pools this distributed flexibility and offers it into the electricity market as a dependable product. The utility then uses the awarded capacity to plan operations for the coming days and plan peak-shaving and valley-filling events during high and low demand periods, respectively.

Historically, power systems largely treated retail demand as inelastic, assuming fixed consumption over predefined periods [5]. The rollout of advanced metering infrastructure and the growth of real-time analytics within smart grid platforms now allow demand-side participants to play an active role in operations. In this setting, DR may benefit substantially from distributed energy resources such as residential photovoltaics and behind-the-meter storage, since these technologies enable localized optimization of flexible demand and can unlock building-level adjustments that align with system needs [6].

Electricity markets commonly classify demand–response programs as price-based or incentive-based [7]. In an incentive-based program, participants receive payments when they reduce consumption below an agreed baseline. The Customer Baseline Load (CBL) defines that reference level for comparable non-event periods [8]. Figure 2 illustrates how the CBL and DR potential anchors settlement during a typical DR event. The difference between CBL and reduced load during DR event constitute DR potential and this is the key quantity which needs to be determined as it accurately informs the grid operators of the flexible reserves possess by consumer participating in energy markets [9,10].

Determining and forecasting DR potential remains a substantial difficulty despite delivering numerous advantages. Modeling accuracy depends on modeling various consumer behavioral patterns, demand elasticity variations and power system restrictions. The unpredictable dynamics of user participation in power grids create multiple uncertainties that make traditional modeling strategies inadequate. Modern research focuses on building resilient forecasting models that use machine learning (ML) and optimization algorithms combined with probabilistic modeling techniques according to [11]. New technology developments strengthen the performance and swift response features of DR systems. Adopting IoT devices with AI technology and blockchain systems produces better demand predictions, automatic usage modifications and self-operated energy market involvement [12]. The developments in DR technologies make it possible to use these systems as essential components of future smart grid deployments. DR solutions require addressing multiple persistent challenges because user engagement needs improvement, strengthened regulatory guidelines and protection of data security and privacy [4].

This review conducts an organized evaluation of available methods which measure and predict DR potential. The research discusses essential technologies that enable DR implementation and addresses both existing implementation barriers and research gaps. The study aims to guide the development of complete strategies that strengthen DR programs to reach global implementation and scaling targets for their essential role in sustainable energy transitions.

1.1. Research Objectives

This systematic review aims to assess the existing methods for predicting DR potential. The main objective of this research is to provide an overview of the methods currently used for DR potential estimation and forecasting. This work examines mathematical, statistical, and data-driven models operating within residential, commercial, and industrial consumer sectors. The analysis divides forecasting approaches into three distinctive categories, including time series and machine learning techniques, with a separate mention of hybrid models that implement multiple paradigms. The evaluation determines which predictive model gains effectiveness through optimization-based frameworks, game-theoretic models, and reinforcement learning technology for DR availability prediction.

Our second objective is to identify the main obstacles that make demand–response modeling difficult. Key uncertainties include differences in household comfort tolerance and willingness to act, variation in program incentives across tariffs and programs, and shifts in demand elasticity with season, time of day, and local weather. This review analyzes multiple horizon forecasting techniques to understand their weaknesses when dealing with dynamic operational situations and DERs such as solar PVs coupled with battery energy storage systems. Studies need to demonstrate consistent performance metrics, which should be standardized because inconsistent evaluation practices between studies create difficulties in benchmarking research.

1.2. Contribution of the Paper

While the domain of general electricity load forecasting is supported by numerous systematic reviews, the specific subfield of DR potential forecasting has received comparatively less synthesized attention. This is a critical gap in the literature. DR potential is the key operational quantity that enables load aggregators and other market participants to bid distributed flexibility into day-ahead and ancillary service markets. Unlike general load forecasting, which predicts total consumption, DR potential forecasting must estimate the available flexibility under specific program conditions which makes it a distinct and more complex challenge. The absence of a systematic review on this topic leaves a fragmented understanding of modeling techniques, practical implementation challenges and future research needs for a modern grid operations. Specifically, the contributions of the review are:

• The review develops a clear classification system that categorizes modern forecasting methods into four key groups: statistical, analytical, optimization-based, and AI approaches. This taxonomy provides researchers with a structured framework to assess, compare, and select the most appropriate techniques based on their specific applications, clarifying their respective strengths and limitations.
• The study provides a holistic analysis of how different modeling paradigms must be integrated for accurate DR potential estimation. We demonstrate how the problem requires a combination of economic optimization for resource dispatch, game-theoretic models for strategic participant behavior, and machine learning for data-driven uncertainty quantification, reflecting the true multi-faceted nature of DR.
• This paper provides a crucial analysis of how the practical effectiveness of forecasting models is shaped by non-technical factors. It systematically analyzes how the accuracy and impact of any model are ultimately constrained by non-technical realities, including consumer behavioral patterns, trust, market regulations, and data privacy concerns, providing a more realistic and grounded perspective on the deployment of DR.
• We introduce a structured framework that moves beyond simple categorization. By providing explicit, pre-specified rules for classifying methods, particularly complex hybrids and AI-coupled optimizers, our taxonomy resolves common ambiguities that have fragmented the literature. This gives the research community a clear and consistent lens to compare, select, and develop forecasting models.

2. Review Methodology

2.1. Research Design and Reporting Standard

This work is designed, conducted, and reported as a systematic review in accordance with PRISMA-2020 [13].

2.2. Research Protocol

A review protocol specifying objectives, eligibility criteria, information sources, search strategies, study selection, data extraction, risk-of-bias (ROB) assessment, and synthesis methods was drafted a priori. The subsequent subsection will provide the details of screening conducted throughout the research data collection process from identification to finally inclusion in the systematic review.

2.3. Research Questions

This review addresses two fundamental questions. We first synthesize the current landscape of methods for forecasting DR potential across residential, commercial, and industrial sectors. We then conduct a critical analysis of the primary obstacles that undermine robust modeling, from behavioral uncertainties and program heterogeneity to the complexities of weather-dependent demand elasticity. To guide this analysis, we introduce a framework centered on the key sources of uncertainty in DR forecasting. We distinguish between four types: (1) data uncertainty, arising from meter noise and missing values; (2) model uncertainty, stemming from model specification and parameterization; (3) exogenous uncertainty, driven by weather and price scenarios; and (4) market uncertainty, linked to participant behavior and rebound effects.

2.4. Eligibility Criteria

• Inclusion Criteria

Peer-reviewed journal or conference studies (2015–2025) in English that (a) model or forecast DR potential/flexibility (including “customer baseline load (CBL)” estimation and DR event response), or (b) analyze enablers/barriers/challenges directly impacting DR-potential estimation; settings may include residential, commercial, and/or industrial sectors. Outcomes must report at least one evaluation metric or methodological contribution relevant to forecasting/modeling.

• Exclusion Criteria

Non-English, non-peer-reviewed or grey literature; studies without full text access; scope limited to purely conceptual policy commentary without methodological detail; publications prior to 2015. Studies exclusively describing DR program design without quantitative modeling/forecasting were excluded unless they provided data or evaluation relevant to DR-potential estimation (e.g., CBL definition, feature sets, uncertainty handling, or operational value).

2.5. Information Sources

Searches were performed in Web of Science Core Collection, Scopus, and IEEE Xplore. The last comprehensive search was executed on 30 July 2025. Grey literature and preprints were not included to ensure consistent peer review standards.

2.6. Search Strategy (Database-Specific and Reproducible)

Full, database-specific Boolean queries, limits, and filters are provided in detail below. Filters common to all databases include publication years 2015–2025 (inclusive), English language; subject area filters applied to energy/power/electrical engineering where available. Exported metadata included titles, abstracts, keywords, DOIs, venues, and years.

• Web of Science (Core Collection).
  • ((“demand response” OR “demand-side” OR “demand side”) NEAR/3 (potential OR flexibility OR “load shifting” OR “load shedding” OR “DR potential”) OR “customer baseline load” OR CBL)
  • AND (forecast* OR predict* OR estimat*); years = 2015–2025; language = English; categories: Energy Fuels; Electrical Eng.; CS–Interdisciplinary.
• Scopus.
  • TITLE-ABS-KEY(((“demand response” OR “demand-side” OR “demand side”) W/3 (potential OR flexibility OR “load shifting” OR “load shedding” OR “DR potential”)) OR “customer baseline load” OR CBL)
  • AND TITLE-ABS-KEY(forecast* OR predict* OR estimat*); years = 2015–2025; language = English; subject areas: Energy; Engineering; Computer Science.
• IEEE Xplore.
  • ((“demand response” OR “demand-side” OR “demand side”) NEAR/3 (potential OR flexibility OR “load shifting” OR “load shedding” OR “DR potential”) OR “customer baseline load” OR CBL)
  • AND (“forecast” OR “prediction” OR “estimation”); years = 2015–2025; language = English; content type: Journals & Conferences.

Figure 3 illustrates the staged keyword development where Level 1 and Level 2 words are necessary to be in the title, abstract or keywords while the literature was made more relevant by applying the keyword combinations at Level 3 and 4.

2.7. Record Management and De-Duplication

Records were exported from each database and merged in refernce manager (Mendeley version 1.19.8). The duplicates were removed via DOI matching, normalized title (lowercased, punctuation stripped), year, and first-author keys. A secondary duplicate pass was run in systematic review software (Rayyan Web Version) using fuzzy matching. Manual spot checks resolved near-duplicates.

2.8. Selection Process

Screening proceeded in two stages (title/abstract, then full text) against the eligibility criteria. Two researchers independently piloted the criteria on a 50-record calibration set and then screened all remaining records in mentioned database. When a conflict arose, the same two independent screeners first re-reviewed the record separately against the eligibility rules, documented their reasons, and then met to seek consensus. If consensus was still not achieved after this independent re-assessment, a third senior reviewer was consulted. All conflict outcomes and rationales were logged in the screening spreadsheet. No automation tools for inclusion decisions were used; translation software was not required due to English-only scope.

2.9. Data Extraction and Coding Schema

We used a structured extraction template capturing bibliographic data, modeling choices, inputs, metrics, and results. To enable statistical/numerical synthesis across methods, we extracted the following: (i) sector (residential/commercial/industrial), region/country; (ii) dataset type and granularity (e.g., AMI/smart meter, SCADA, weather, price, DR event flags), sampling interval, forecast horizon; (iii) model family and architecture; (iv) feature engineering and exogenous inputs; (v) evaluation protocols (train/test split, cross-validation, rolling origin); (vi) point metrics (MAE, RMSE, MAPE) and, when applicable, probabilistic metrics (Pinball loss, CRPS, interval coverage and calibration); (vii) uncertainty handling; (viii) operational or economic outcomes (e.g., cost savings, DR event success rates, reserve requirement changes); (ix) dataset availability (public/private), and any reported smart meter penetration or coverage; (x) presence of industry case studies or pilots. Two extractors worked independently on a 10% sample to harmonize coding; disagreements were resolved by consensus, and the final extraction was single-reviewer with verification.

2.10. Definitions and Taxonomy Harmonization

To remove category overlap and inconsistencies, we adopted the following a priori taxonomy with disambiguation rules:

• Statistical/analytical: AR/ARIMA/ARIMAX, linear/GLM/GAM, state-space, survival/ time-to-event where used for DR potential.
• Optimization-based: stochastic/robust optimization, bilevel/game-theoretic, economic dispatch models when the optimization itself produces the forecast/estimate.
• AI/ML: tree ensembles, SVM, kernel methods, and deep learning (RNN/LSTM/GRU, CNN, transformers). Deep learning remains under ML unless explicitly coupled with a separate optimizer/logic module.
• Hybrid: explicit coupling of two or more paradigms (e.g., optimization + ML; fuzzy logic + ML; physics-informed + ML), or stacked/ensemble models combining fundamentally different learners.

We harmonize key terms: Customer Baseline Load (CBL) denotes reference consumption absent DR; DR potential/flexibility is the achievable shed/shift under specified conditions; shed = net load reduction; shift = temporal redistribution preserving energy. Moreover, all acronyms are defined at first use.

2.11. Risk of Bias/Study Quality Assessment

To ensure the credibility of our synthesis, we conducted a formal study-level appraisal to assess the risk of bias in each included article. For this, we employed the Risk of Bias in Systematic Reviews (ROBIS) tool, a structured instrument tailored for evaluating evidence syntheses. The appraisal process was performed by two independent raters. To ensure consistency, the raters first calibrated their judgments on a 15% random subset of the studies, resolving any disagreements on the level of concern (low, some, high, or unclear) through consensus. Following this calibration phase, the established rating criteria were applied consistently across the full set of included works.

2.12. Outcomes and Measures

Primary outcomes for synthesis were methodological: model families/architectures, input types, validation designs, and reported performance metrics. We prioritized standardized point metrics (MAE, RMSE, MAPE) and, where available, probabilistic metrics (Pinball, CRPS, interval coverage/calibration) to quantify uncertainty. We additionally extracted operational/economic value proxies (e.g., cost reduction, DR event success/default rates, reserve requirement changes, regret/cost–loss analysis) to link forecasting quality to decision value.

2.13. Synthesis Methods

Given heterogeneity in targets, horizons, data, and metrics, we conducted a structured narrative synthesis with quantitative descriptive components rather than meta-analysis. We will provide (i) frequency distributions across method families, sectors, data types, regions, and years; (ii) a comparative evidence table summarizing each study’s contributions, data, metrics, and limitations; and (iii) a metric comparability table mapping which accuracy/uncertainty metrics were used by which studies. An evidence map (heat/bubble plot) will depict method × sector × data-type × region/year coverage. Where studies reported paired comparisons versus baselines (e.g., coupled/hybrid vs. standard), we extracted relative performance gains/skill scores.

2.14. Reporting, Reproducibility, and Figure Standards

All figures (including PRISMA) are supplied as vector PDFs or ≥300 dpi rasters with clear labeling and legible fonts. The tables include complete headers/units. The PRISMA-2020 checklist, full search strings, ROBIS materials, and code to recreate tables/figures will be provided as per requirement. Abbreviations are unified across captions and legends.

2.15. PRISMA-2020 Flow Chart

Figure 4 presents the PRISMA-2020 flow diagram.

Table 1. PRISMA 2020 counts across identification, screening, eligibility, and inclusion (revised totals).

Stage	n	Notes/Reason
Records identified (WoS, Scopus, IEEE)	10,250	2015–2025; English filters applied at query/export.
Duplicates removed	882	DOI/title/year/author key plus fuzzy duplicate pass.
Records after de-duplication	9368	Proceeded to title/abstract screening.
Excluded at title/abstract screening	8012	Pre-2015, off-topic, or insufficient methodological detail.
Reports sought for retrieval	1356	Full text requested.
Reports not retrieved	17	Access restricted or unavailable.
Full-text assessed for eligibility	1339	Screened against criteria.
Excluded at full text	1167	241 non-English; 926 irrelevant/insufficient data.
Studies included in review	172	Final set for extraction and synthesis.

lists each decision point demonstrating the complete process of identification of search results from database, screening process, and final selection of articles in systematic review process.

3. Results

3.1. Comparative Evidence (Main Studies, 2015–2025)

Table 2. Main studies on DR-potential estimation/forecasting (methods, data, metrics). Abbreviation legend can be found below the table.

Study (Year)	Sec.	Data	Hor.	Family	Unc.	Metrics	Notes
Li et al. [11] (2024)	Res	AMI + WX	DA	ML (TL)	Prob	MAE, RMSE, Pinball	Online transfer across heterogeneous users; portability.
Muqtadir et al. [14] (2025)	Com	AMI + WX	DA	ML	Pt	MAE, RMSE	Building heterogeneity; event-oriented eval.
Ruiz-Abellón et al. [15] (2024)	Multi	AMI + Phys	Multi	ML	Prob	Pinball, CRPS, Cov.	Prob. DR with explicit calibration.
Xu et al. [16] (2023)	Res	AMI + Prc	ID	Hybrid (RL + Opt)	Pt	MAPE	Cost/comfort trade-off; control focus.
Chan et al. [17] (2024)	Multi	AMI	Multi	ML	Pt	MAE, RMSE	Attention captures long deps.
Spencer et al. [18] (2025)	Bldg	AMI + WX	DA	ML	Pt	MAE, RMSE	Cross-building transfer; limited DR events.
Ochoa et al. [19] (2025)	Sys	AMI + Scen	ST	ML	Prob	Pinball, CRPS	Scenario utility; stress calibration open.
Sridhar et al. [20] (2025)	Res	AMI + Prc	ID	Hybrid (Opt + ML)	Pt	MAPE	Bidding or reserves linkage; participation bias.
CBL baselines [21] (2024)	Res	AMI	ID	ML (DL)	Pt	MAPE	DR settlement; limited prob. checks.
Bashyal et al. [22] (2025)	Ind	AMI + Prc	ID	Hybrid (MARL)	Pt	RMSE	Production-aware; safety not benchmarked.
Nygård et al. [23] (2025)	Sys	AMI + WX	ST	ML (LSTM)	Pt	RMSE	Custom loss; profit link implicit.
Bashyal et al. [22] (2025)	Ind	AMI + Prc	ID	Hybrid (MARL)	Pt	RMSE	Production-aware; safety not benchmarked.
Nygård et al. [23] (2025)	Sys	AMI + WX	ST	ML (LSTM)	Pt	RMSE	Custom loss; profit link implicit.
Balakrishnan et al. [24] (2025)	Multi	AMI	DA	Hybrid (Stacked)	Pt	MAE, RMSE	Fuzzy/stacked ensemble.
Hong & Fan [25] (2016)	Multi	AMI + WX	Multi	Stat	Prob	Pinball, PI	PLF foundations for DR baselines.
Alipour et al. [26] (2017)	Hub	Meter	Multi	Opt (MINLP)	Scen	Cost/Emis	2m + 1/PEM uncertainty in hubs.
Chow et al. [27] (2021)	Multi	AMI + WX	ST	Stat	Pt	MAE, RMSE	SARIMA/STL baselines for DR screens.
Moslemi et al. [28] (2024)	Multi	AMI + WX	ST	Stat	Pt	MAE, RMSE	State-space/ES; regime shifts.
Kim et al. [29] (2023)	Multi	AMI + WX	ST	Stat	Pt	MAE	Adaptive state-space.
Neshat et al. [30] (2018)	Res/Com	AMI + WX	ST	Stat	Pt	RMSE	Nonlinear seasonality via STL.
Ruiz et al. [31] (2020)	Sys	AMI + PV	DA	Stat	Prob	PI, Cov.	Coordinated load + PV for DR timing.
Macedo et al. [32] (2015)	Res	AMI + WX	ST	ML (ANN)	Pt	MAPE	Early ANN with DR-relevant features.
Pallonetto et al. [33] (2019)	Com	BEMS + AMI	DA	ML (SVR/Ens)	Pt	MAE, RMSE	Occupancy, weather importance.
Rahman et al. [34] (2018)	Multi	AMI + WX	Multi	Hybrid (SVR + RNN)	Pt	MAE, RMSE	Residual learning for nonlinearity.
Guo et al. [35] (2021)	Res	AMI	ST	ML (CNN/RNN)	Pt	RMSE	Deep seq for short windows.
Zhang et al. [36] (2020)	Res	AMI + Prc	ID	Hybrid (DRL)	Pt	Reward, RMSE	Policy learning for shifting.
Chen et al. [37] (2019)	Multi	AMI	ST	ML (Attention)	Pt	MAE	Attention improves STLF.
Zhao et al. [38] (2021)	Multi	AMI	ST	ML (Transformer)	Pt	RMSE	Transformer for STLF.
Xiao et al. [39] (2018)	Multi	AMI	DA	Hybrid (ARIMA + ANN)	Pt	MAE	Additive residual hybridization.
Raju et al. [40] (2022)	Com	AMI + BEMS	DA	Hybrid (Ens)	Pt	MAE, RMSE	Stacking or voting robustness.
Jnr et al. [41] (2021)	Multi	AMI	ST	Hybrid (DL + Stats)	Pt	MAPE	Deep residual hybrids.
Li et al. [42] (2019)	Multi	AMI	DA	Hybrid (ARIMA + ML)	Pt	RMSE	Trend  +  ML residuals.
Roh et al. [43] (2015)	Res	AMI	DA	Opt (MILP)	Scen	Cost	DR as dispatchable resource.
Eshraghi et al. [44] (2019)	Sys	AMI	Multi	Opt (UC)	Scen	Cost, Res.	DR in UC with constraints.
Veras et al. [45] (2018)	Multi	AMI	DA	Opt (NSGA-II)	Scen	Pareto	Cost/comfort trade-offs.
Khezri et al. [46] (2022)	Multi	AMI + WX	DA	Opt (MOO)	Scen	Cost, Emis	Multi-objective planning.

consolidates the principal studies included in our PRISMA corpus (2015–2025) and standardizes their comparators to enable like-for-like reading. For each work, we report sector, region, data sources, forecast horizon, method family, treatment of uncertainty, accuracy metrics, any decision-value outcomes, and salient limitations. This condensed view foregrounds patterns and trade-offs across approaches and anchors the subsequent quantitative synthesis and discussion of uncertainty, deployment evidence, and taxonomy edge cases for comparability.

• Legend (abbreviations).
  • Sec.—Sector: Res (residential), Com (commercial), Ind (industrial), Sys (system/ISO), Bldg (buildings), Multi (mixed).
  • Data—AMI (smart meter), WX (weather), Prc (price/tariff), Phys (physics/first-principles), Scen (scenario set), PV (photovoltaics), BEMS (building EMS).
  • Hor.—ST (short-term), ID (intra-day), DA (day-ahead), Multi (multi-horizon).
  • Family—Stat (statistical), Opt (optimization), ML (machine learning), Hybrid (explicit coupling/ensembles).
  • Unc.—Pt (point), Prob (probabilistic), Scen (scenario), Cov. (coverage).
  • Metrics—MAE, RMSE, MAPE, CRPS, Pinball, PI.

3.2. Uncertainty & Probabilistic Forecasting with Numerical Synthesis

Across the 34-study subset as numerically synthesized in

Table 3. Uncertainty and numerical synthesis derived from Table 2 (n = 34).

Aspect	Count	Share	Notes
Method family: ML/Hybrid/Statistical/Optimization	14/10/6/4	41%/29%/18%/12%	ML dominates; Hybrids frequent when coupling control/scheduling or residual learning.
Probabilistic forecasts reported (Pinball, CRPS, PI)	6	18%	Mostly Pinball/CRPS; a few report PIs; calibration seldom assessed.
Calibration/reliability checked (PIT/coverage test)	3	9%	Coverage/PIT explicitly shown in a minority.
Decision-linked outcomes (cost, reserve, DR success)	8	24%	Often via dispatch/bidding proxies or reserve impact.
Typical horizons: Intra-day/Day-ahead/Multi	12/16/6	35%/47%/18%	DA most common; ID for control/DRL.
Common inputs: AMI + WX; Price/Tariff; DR-event tags	30; 18; 9	88%; 53%; 26%	AMI + WX near-universal; explicit event flags less frequent.
Validation: simple split/rolling origin/k-fold	18/9/7	53%/26%/21%	Rolling origin used to avoid temporal leakage.
Reported point metrics (MAE/RMSE/MAPE)	31	91%	RMSE, MAE dominant; MAPE common in building-level works.

, ML and Hybrid approaches dominate (41% and 29%), with Statistical and Optimization families comprising 18% and 12%. Only 18% report probabilistic forecasts (Pinball, CRPS, or prediction intervals), and just 9% include calibration or coverage checks (e.g., PIT), signaling limited attention to reliability. Decision-linked outcomes appear in 24%, mainly via cost, reserve, or DR-event success proxies. Day-ahead horizons are most common (47%), followed by intra-day (35%). AMI with weather covariates is near-universal (88%), while explicit DR-event tags occur in 26%. Validation still favors simple splits (53%); rolling-origin (26%) and k-fold (21%) are less used—underscoring our checklist to report quantiles/intervals, calibration, and decision relevance.

The evidence map in Figure 5 plots method families (Stat, Opt, ML, Hybrid) on the x-axis against sectors (Res, Com, Ind, Bldg, Sys, Multi) on the y-axis; bubble size shows how many studies appear in each cell. Three patterns emerge: (1) ML clearly dominates, especially for multi-sector and residential datasets; (2) hybrids are the second most common, concentrated in multi-sector and residential, with emerging use in commercial/industrial; (3) statistical and pure optimization approaches are comparatively sparse and largely confined to multi-sector/system contexts. Building-level evidence is thin across all families. Gaps suggest opportunities for hybrid/ML work in industrial and building sectors with richer datasets.

4. Discussion

Our systematic review of DR potential forecasting reveals a field in transition, where sophisticated modeling techniques are rapidly outpacing the frameworks needed to validate and deploy them effectively. This discussion synthesizes the principal trends, critical limitations, and key research gaps identified from the literature.

4.1. Trends and What They Mean

A dominant trend emerging from our analysis is the clear movement from classical statistical baselines toward more complex machine learning and hybrid architectures. This shift is largely driven by the growing availability of high-granularity residential AMI data, which enables models to capture detailed, nonlinear consumer behaviors. This evolution, however, is not without its challenges. Our review reveals that the methodological sophistication of these models has not been matched by a corresponding evolution in evaluation practices. We found that probabilistic scoring remains uncommon, and robust calibration checks are rare across the included studies. Consequently, the literature is filled with inconsistent superiority claims that are difficult to verify due to a lack of standardized benchmarks and decision-relevant metrics. This points to a critical maturity gap: the field has developed powerful new tools, but it still lacks the consistent validation frameworks needed to reliably assess their true operational value.

4.2. Data and Infrastructure Constraints/Limitations

The practical progress of DR potential forecasting is fundamentally constrained by persistent issues of data quality, access, and representativeness. Typical AMI data cadences range from 15 to 60 min, and the absence of high-quality public benchmark datasets severely limits the reproducibility and comparability of studies. Compounding this issue, we found that crucial contextual information, such as specific tariff structures or DR event logs, is often restricted, forcing models to operate with an incomplete picture of the factors driving consumer response. A particularly critical and under-reported issue is the sampling bias introduced by low or uneven smart meter penetration. Datasets are frequently skewed toward “early adopters,” a demographic that is not representative of the general population. This bias not only leads to overly optimistic estimates of DR potential but also severely limits the transferability of trained models to the broader, late-adopting segments essential for scaling DR programs.

4.3. From Accuracy to Value in DR Potential Estimation

A key finding of this review is that the field’s predominant reliance on traditional point accuracy metrics is insufficient for guiding real-world DR operations. Metrics such as MAPE or RMSE, while useful, cannot guide DR deployment alone, which involves asymmetric financial penalties and discrete operational thresholds for event triggering and settlement. A forecast that is accurate on average can still be operationally detrimental if it fails to predict outcomes at these critical decision points. To bridge this gap, the field must evolve its evaluation paradigm. Evaluations should pair point accuracy with at least one probabilistic loss metric and a formal calibration diagnostic. This more holistic approach is essential for linking a model’s predictive performance to its tangible economic and reliability value, particularly in tasks like reserve provisioning and market bidding under uncertainty.

4.4. Research Gaps Identified from the Literature

Synthesizing these findings, our review identifies a primary research gap: uncertainty is seldom quantified in a decision-relevant way. While many studies report low MAE or RMSE, few provide the probabilistic outputs, such as quantiles or prediction intervals, that grid operators and aggregators need to make risk-aware operational choices. The thresholds used for evaluation are rarely linked to actual market or system operations. Furthermore, our analysis underscores several related gaps that hinder progress, including the scarcity of studies that perform paired comparisons against strong baselines on the same data splits, the lack of a compact public benchmark suite for standardized testing, and the weak connection between reported accuracy metrics and quantifiable operational outcomes like cost reduction or DR event success rates. Addressing these gaps is essential for moving DR potential forecasting from an academic exercise to a trusted and integral component of modern grid management.

5. Fundamentals of Demand Response

5.1. Definition and Classification of DR

The demand response allows end users to adjust their electricity use based on signals sent by the utility or the network operators. These signals may include time-varying prices, explicit incentives, or program notifications from utilities and aggregators [47,48,49]. In line with the U.S. Department of Energy and the European Directive 2019/944, DR is understood as the intentional modification of customer load profiles through manual actions or automated controls. The aim is to support grid stability, lower operating costs, and enable higher shares of renewable generation [3,4]. In practical terms, DR turns formerly passive consumers into active participants who can respond to system needs, sometimes within minutes and sometimes over longer horizons.

DR can be organized by control scheme and by market design. A common split distinguishes active and passive approaches. Active DR relies on real-time responses to grid or market signals, with customers or automated systems operating controllable loads and dispatchable resources. Passive DR relies on price design and tariffs that reshape usage patterns over time rather than direct control in the moment. Deployment can also be viewed through the lens of resource controllability. Dispatchable resources are scheduled by operators and provide targeted curtailment or shifting. Non-dispatchable resources respond automatically to incentives and tariff structures and their effect emerges from the aggregate behavior of many customers [50]. Figure 6 organizes DR into a few broad families. The major types are segregated into reward mechanism, in which incentives are provided to customers for sharing their resources, by energy types, which include various forms of energy, and by region, which is further classified into single area DR and multi-region DR. In this review, our focus will be on DR programs which involve electric power as literature on this domain is ubiquitous.

5.2. Demand Response Programs: Price-Based and Incentive-Based

The two principal types of DR programs include price-based and incentive-based approaches. Through price-based control schemes, price-based DR instigates electricity consumers to shift consumption times during fluctuating market prices [51,52,53,54]. Price-related programs such as Time-of-Use (ToU), Real-Time Pricing (RTP), and Critical Peak Pricing (CPP) represent common price-based approaches for DR programs in electricity systems [55,56]. The implemented programs use price-responsive behaviors among consumers to redirect load utilization from peak periods while helping maintain grid balance.

The participants in incentive-based DR programs receive financial payments or utility bill deductions as compensation when they lower their electricity usage during emergencies that result from grid limitations or market price fluctuations [57,58,59,60]. Under Direct Load Control (DLC), enrolled customers allow the utility to remotely curtail or shift specific devices such as electric water heaters and air conditioners. In practice the utility cycles compressors or changes setpoints for short intervals, which eases system stress while keeping comfort within agreed bounds. Alongside DLC, two other DR programs are widely used. Demand bidding invites customers or aggregators to offer curtailable load at stated prices, and emergency DR is called during supply shortfalls or extreme weather when rapid, verifiable reductions are needed [4,50]. Taken together, these mechanisms help preserve grid reliability at relatively low cost because they rely on existing metering and communications infrastructure and can deliver dynamic load adjustments within minutes.

The effectiveness of price-based DR depends strongly on advanced metering and on-site automation. Where smart meters, low-latency data, and programmable controls are widely deployed, customers tend to respond more reliably and with greater persistence across residential, commercial, and industrial settings [57,58]. When responses are manual, impacts are smaller and more variable. As planners integrate DR into resource adequacy and distribution planning, interest is growing in hybrid designs that blend time-varying prices with incentive contracts. Prices can steer routine shifting over many hours, while incentives secure firm, verifiable reductions during system stress [61].

5.3. Key Stakeholders: Utilities, Consumers, Aggregators, Regulators

Several important actors inside the energy sector must work together to achieve effective DR deployment [62,63]. Utilities and grid operators maintain central roles in this system because they launch DR events to stabilize power grids during high demand or supply limitations [62,63,64,65]. Utilities can delay infrastructure expenditures and enhance power grid stability through peak load predictions and signal alerts, allowing for the better integration of wind and solar power assets. Entities watch system performance in real-time to check DR consumer involvement and assess program assessment [3,4].

The total performance of DR programs hinges on how end-users respond to such programs when they participate. These participants belong to residential, commercial, and industrial sectors. They respond to signals combined with incentives by adjusting their electricity usage. Even though residential consumers have limited flexibility options, their large number gives them significant potential to contribute to DR programs. Industrial users can achieve considerable load reductions through proper management, although they need precise coordination to maintain essential business activities. Different communication solutions, together with tailored incentives, are required to engage various customer types in DR initiatives [5,56]. User participation in DR schemes greatly depends on smart meters, home automation systems, and easy-to-use interfaces.

Multiple consumers can work with aggregators that unite their DR resources into stable market-ready resources through intermediary operations. Small-scale participants from the residential sector can reach energy markets through aggregators that unite their flexible loads into market-ready portfolios [62,63]. Aggregators must conduct a DR forecast analysis for aggregated resources and then place market bids while verifying that the scheduled load reductions have been successfully implemented during actual events. Regulatory bodies and policymakers establish the guidelines and economic conditions determining how distribution resources join the market. The regulatory bodies define market regulations while maintaining consumer rights and implement performance-based systems to stimulate involvement from multiple stakeholder entities [50,61].

The functional relationship between these stakeholders presents various obstacles during their exchanges. Conflict exists between stakeholders, such as utilities pursuing reliability needs instead of consumer comfort and aggregator interests driven by profit contributions to DR underutilization. DR’s maximum value becomes reachable across the power system by implementing three fundamental principles: effective collaboration, transparent communication, and well-designed regulatory oversight [49,63].

6. Mathematical Modeling for DR Potential Calculations

6.1. Optimization Models for DR Estimation

Optimization models serve as fundamental tools in estimating the potential of DR by maximizing load flexibility or minimizing operational costs while satisfying system constraints [66,67,68]. A common formulation involves linear programming (LP) or nonlinear programming (NLP) approaches. For instance, a basic DR cost minimization problem can be expressed as follows:

$$\underset{x}{min}\sum _{t=1}^T{C}_t·x_t$$

(1)

subject to

$$L_t^{min}\le x_t\le L_t^{max},\forall t$$

(2)

$$\sum _{t=1}^T{x}_t\ge D_{\mathrm{total}}$$

(3)

$$|x_t-x_{t-1}|\le R_t,\forall t$$

(4)

where $x_t$ is the DR load at time t, $C_t$ is the cost of electricity, $L_t^{min}$ and $L_t^{max}$ are the lower and upper bounds of load reduction, and $R_t$ is the ramp rate limit.

In economic dispatch models, DR is introduced as a flexible resource [69,70,71]. The objective becomes minimizing the total cost of generation and DR as follows:

$$min\left(\sum _{i=1}^N{C}_i\left(P_i\right)+\sum _{j=1}^M{C}_j^{DR}\left(D_j\right)\right)$$

(5)

subject to the following power balance:

$$\sum _{i=1}^N{P}_i+\sum _{j=1}^M{D}_j=D_{\mathrm{load}}$$

(6)

In multi-objective optimization, objectives like cost, emission, and user discomfort are combined [72,73,74,75]:

$$min\left[{\alpha }_1{f}_1\left(x\right)+{\alpha }_2{f}_2\left(x\right)+{\alpha }_3{f}_3\left(x\right)\right]$$

(7)

where $f_1$ is the cost function, $f_2$ represents carbon emissions, and $f_3$ accounts for consumer inconvenience, and ${\alpha }_k$ are weighting factors.

Advanced models integrate DR in unit commitment using Mixed-Integer Programming (MIP) [43,44,76]:

$$min\sum _{t=1}^T\left(\sum _{i=1}^N{C}_i\left(P_{it}\right)+\sum _{j=1}^M{C}_j^{DR}\left(D_{jt}\right)\right)$$

(8)

subject to commitment and spinning reserve constraints.

Metaheuristic techniques like Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Non-Dominated Sorting Genetic Algorithm (NSGA-II) are used for solving multi-objective nonlinear problems [45,77]. These techniques manage trade-offs between conflicting goals like cost and comfort. Such formulations are applicable across residential, commercial, and industrial domains, providing flexible, scalable, and efficient frameworks for DR estimation [46,77,78]. Figure 7 groups optimization methods for DR into two families. On the left are classic mathematical programs that include linear and nonlinear formulations. These methods work well when the physics and constraints can be written cleanly and when convexity holds. On the right are metaheuristics such as particle swarm optimization, genetic algorithms, simulated annealing, and teaching–learning-based optimization. These are useful when the problem is nonconvex or highly discrete, for example appliance scheduling with comfort rules or mixed device states, where exact solvers may stall. Metaheuristics often find good solutions with modest modeling effort, though they offer no guarantee of global optimality and can be sensitive to tuning and runtime.

6.2. Statistical and Probabilistic Models for DR Estimation

Statistical and probabilistic models play a critical role in estimating DR potential by quantifying uncertainties in load behavior, participation rates, and equipment performance. One of the foundational tasks is constructing a baseline model to estimate what electricity consumption would have been without a DR event. Let the baseline load at time t be denoted as $B_t$, while the actual load is $L_t$. The DR reduction $R_t$ is then given by the following:

$$R_t=B_t-L_t$$

(9)

To estimate $B_t$, regression models are frequently used. A multiple linear regression model incorporating weather, time, and historical usage is defined as follows:

$$B_t={\beta }_0+{\beta }_1{T}_t+{\beta }_2{H}_t+{\beta }_3{D}_t+{ϵ}_t$$

(10)

where $T_t$ is temperature, $H_t$ is historical average load, $D_t$ is a dummy variable for weekday/weekend, and ${ϵ}_t\sim \mathcal{N}(0,{\sigma }^2)$ is Gaussian noise [25,26].

Probabilistic models take this further by modeling the load or DR response as random variables. A common assumption is that DR availability A follows a probability distribution such as follows:

$$A\sim \mathrm{Beta}(\alpha ,\beta )$$

(11)

or, for baseline errors:

$$B_t\sim \mathcal{N}({\mu }_t,{\sigma }_t^2)$$

(12)

Monte Carlo Simulation (MCS) is widely used to model the variability in DR response [20]. Given N scenarios of input variables $X^{\left(i\right)}$, the simulated DR reduction is as follows:

$${\left\{R_t^{\left(i\right)}\right\}}_{i=1}^N=f(B_t^{\left(i\right)}-L_t^{\left(i\right)})$$

(13)

The expected DR potential is computed as follows:

$$\mathbb{E}\left[R_t\right]={\displaystyle \frac{1}{N}}\sum _{i=1}^N{R}_t^{\left(i\right)}$$

(14)

To model joint uncertainties, techniques like Point Estimate Method (PEM) and 2m + 1 approaches are employed in energy hubs with multiple uncertain inputs [26].

In more advanced models, quantile regression is employed to estimate the conditional quantiles $Q_{R_t}\left(\tau \right)$ of the DR response:

$$Q_{R_t}\left(\tau \right|X_t)=X_t^{\top }\beta \left(\tau \right)$$

(15)

This approach enables the estimation of DR potential with confidence bounds. Combined with hierarchical or spatial modeling, such techniques support multi-level DR aggregation analysis [79].

Probabilistic load forecasting and DR estimation provide robust planning under uncertainty, contributing to grid reliability, economic dispatch, and effective participation in DR markets [25,26,79].

6.3. Game-Theoretic Models

Game-theoretic models are essential for forecasting DR potential because they move beyond purely statistical methods to capture strategic interactions among rational, self-interested agents, such as consumers, aggregators, and utilities [80,81,82,83]. In DR, the available flexibility is not a fixed quantity; it emerges from the decisions of participants responding to incentives and market conditions. Game theory provides a formal framework to model these decisions and predict their collective outcome, which is the realized DR potential. The foundational concept is the Nash Equilibrium, which predicts a stable state where no single player can benefit by unilaterally changing their strategy. In the context of DR forecasting, it is used to estimate the aggregate load reduction when numerous consumers simultaneously decide on their participation level in response to a given price signal [84]. Each consumer aims to minimize their costs by balancing the inconvenience of reducing consumption against the financial reward. The equilibrium of this game forecasts the likely, stable level of participation a utility can expect. Let the players be $\mathcal{N}=\{1,\dots ,N\}$ with strategies $s_i\in S_i$ and utility $u_i(s_i,s_{-i})$. A profile $s^{\ast }$ is a Nash equilibrium if no player can gain by deviating.

$$u_i(s_i^{\ast },s_{-i}^{\ast })\ge u_i(s_i,s_{-i}^{\ast })\forall s_i\in S_i,\forall i\in \mathcal{N}.$$

(16)

In DR, a common formulation chooses load reduction $x_i$ at price p by solving

$$x_i^{\ast }\left(p\right)=arg\underset{x_i\ge 0}{min}\left\{C_i\left(x_i\right)-p{x}_i\right\}.$$

(17)

The aggregate response $X\left(p\right)={\sum }_i{x}_i^{\ast }\left(p\right)$ yields a price–reduction curve that forecasters can use as a structural prior on event days.

Many DR programs, however, feature a clear hierarchy. The Stackelberg game is better suited for these leader-follower scenarios, where a utility (leader) sets the price or incentive, and consumers (followers) react. This model is invaluable for forecasting in two ways: first, it predicts how consumers will respond to a given incentive $\left(x^{\ast }\left(p\right)\right)$, and second, it allows the utility to work backward to determine the optimal incentive needed to achieve a specific load reduction target [85]. This transforms the model from a simple predictor into a strategic planning tool for maximizing DR program effectiveness. A leader posts price p and customers best respond with $x^{\ast }\left(p\right)$. The leader then optimizes its objective $U_L$ subject to those responses.

$$x^{\ast }\left(p\right)=arg\underset{x}{max}u(x;p),p^{\ast }=arg\underset{p}{max}{U}_L\left(p,x^{\ast }\left(p\right)\right).$$

(18)

Simulating Equation (18) produces a joint forecast of prices and load reductions under capacity or emissions limits.

Beyond predicting responses to existing programs, game theory helps in forecasting the potential of new market structures. Mechanism Design is used to engineer DR programs that are incentive-compatible, ensuring participants are motivated to reveal their true flexibility. By modeling how users report their private valuations $\left(V_i\left(x_i\right)\right)$, these frameworks can forecast a program’s efficiency and guard against strategic misreporting that could undermine the grid’s reliability [86]. Let $V_i\left(x_i\right)$ denote user i’s valuation and $t_i\left(x_i\right)$ the transfer. The social planner solves:

$$\underset{x_1,\dots ,x_N}{max}\sum _{i=1}^N\left[V_i\left(x_i\right)-t_i\left(x_i\right)\right]$$

(19)

Other specialized models address specific forecasting challenges. Congestion games are used to predict DR potential in scenarios with shared resources, such as local distribution networks or clusters of EV chargers, forecasting how users will shift their load to avoid bottlenecks [87]. Evolutionary games model the long-term dynamics of DR participation, forecasting how strategies adapt over time as users learn from repeated interactions. This is crucial for estimating the growth and maturity of DR potential in a developing market.

7. Forecasting Techniques for DR Potential

7.1. Time Series Analysis for DR Forecasting

Classical time series models remain a practical starting point for estimating near-term electricity demand and the share that is flexible under DR. ARIMA and its exogenous variants capture daily and weekly cycles and can incorporate weather and calendar effects as regressors [27,28,29,30]. Seasonal extensions and decomposition methods, for example SARIMA with STL or Fourier terms, often improve fit when multiple seasonalities are present [28,29]. Household-level studies and facility case reports show that these specifications can deliver dependable short-term forecasts that are useful for screening DR capacity and scheduling events when data are clean and regimes are relatively stable [30]. Short-horizon DR applications have also paired ARIMA with renewable generation forecasting to coordinate load and PV in day-ahead and intra-day settings [31].

State-space formulations build on this foundation by allowing parameters to evolve over time and by treating measurement error explicitly. These models tend to adapt faster during regime shifts and can fuse multiple signals through Kalman filtering, which is useful when DER adoption or unusual operating conditions alter load patterns [28]. In practice, forecasters combine exponential smoothing or local-level components with state-space updates to track abrupt consumption changes caused by extreme weather, outages, or event days [29]. For DR planning, probabilistic outputs matter as much as point predictions because operators need availability bands for flexible demand. Recent work therefore reports predictive intervals or scenario paths that support event sizing, portfolio aggregation, and coordination with volatile renewable supply [27,31].

These statistical approaches have limits that matter for DR potential estimation. Linear specifications can miss nonlinear responses to weather, price, social behavior and they may degrade under structural breaks, holiday shifts, or rapid uptake of rooftop PV and storage [88]. Exogenous drivers also introduce forecast-of-forecast error that can propagate into DR estimates when price or PV is itself predicted rather than observed [31]. Even state-space models require careful design when latent dynamics move faster than the filter can track, and many studies emphasize point accuracy while under-reporting calibration of predictive intervals, which reduces operational value for risk-aware DR dispatch [27]. These gaps motivate hybrids that retain the transparency of statistical structure while adding nonlinearity through modern learners, along with consistent reporting of both point and probabilistic performance so that DR potential forecasts are comparable across programs and time frames [28,30].

7.2. Machine Learning and Deep Learning Approaches

The advancement towards higher performance predictors comes through ML approaches because they deliver better accuracy. Artificial Neural Networks (ANNs) provide successful performance regarding complex nonlinear observation relations between weather data, pricing signs, and user actions [32,89,90,91,92]. The simple structure of ANN is shown in Figure 8 where past values of DR potential are feed to linear layers of ANN. ANN acts as function approximator and based on past values of DR, it tries to predict future DR potential values. Residential studies use ANN models that exhibit high accuracy in DR potential prediction as documented in [56]. These prediction models analyze prior DR reactions while adjusting to altering user conduct, making them highly effective for anticipating short and medium-term DR programs.

Long Short-Term Memory (LSTM) networks demonstrate superior performance when analyzing sequential data through their role as Recurrent Neural Network (RNN) (Figure 9) sub-type because they effectively store long-term dependencies [93,94,95,96]. Research proves that models built on LSTM deliver superior results than standard prediction tools when handling erratic or missing data patterns that appear in residential and commercial electricity networks [34]. The predictive model efficiently models DR participation because it requires simultaneous analysis of consumption history and user reaction patterns linked with contextual variables. Due to the recursive nature of these models, they retain sequential information, making them suitable for predicting load variations across various periods [93,94,95,96].

Residential DR programs benefit tremendously from Artificial Neural Networks because these systems recognize energy consumption nonlinearities, which helps optimize residential DR [56,97]. Multiple studies demonstrate that SVM and ensemble learning techniques yield successful results when modeling various building features for commercial DR applications [33,98,99,100,101]. The forecast robustness is enhanced through ensemble approaches that integrate various ML algorithms. Kilimci et al. created a decision integration approach that merges weighted predictions from base learners, including SVR, ARIMA, and deep neural networks, to deliver enhanced forecasting results for retail electricity usage [102].

Deep reinforcement learning (DRL) technology attracts increasing attention in DR management because it goes beyond conventional forecasting capabilities [35,93,96,103]. According to Zhang et al., agents with DRL capabilities discover the most suitable load-shifting methods through environmental interactions that provide rewards. DRL enables smart grids to achieve adaptive decision-making through real-time scheduling decisions by coping with changing renewable energy supply patterns and consumption demands [36]. According to Wang et al., convolutional and pooling RNNs within deep neural architectures enhance generalization abilities in models measuring residential energy consumption since user behaviors are not entirely predictable [104].

In recent DRL-based forecasting, beyond control-oriented scheduling, several 2022–2025 studies use deep reinforcement learning to forecast DR potential or to learn response distributions that feed probabilistic forecasters. Representative works include online transfer forecasters that adapt across heterogeneous users [11], probabilistic DR with explicit calibration [15], RL-augmented schedulers that couple policy learning with event-level forecasts [16], and multi-agent RL in industrial settings where learned value functions provide day-ahead response predictions [22]. These additions ensure our review reflects the latest DRL-based forecasting literature (see also Table 2).

Many papers in the reviewed research demonstrate how preprocessing data and selecting features leads to the best results from ML models. The accuracy for commercial DR scenarios improves directly from applying feature engineering practices that focus on selecting data related to weather and occupancy and appliance usage [98].

Among deep learning methods, transformer architectures are a recent addition to forecasting. They performed exceptionally well in natural language processing (NLP) where sequence-to-sequence problems require an entire output sequence rather than a single label [105]. Early systems relied on an encoder–decoder design until the attention layer was introduced [106]. Attention lets the decoder focus on selected parts of the encoder output. In machine translation this means concentrating on specific words in the source sentence. That idea lifted performance across many NLP tasks and became the foundation of the Transformer. A practical benefit is scalability because the architecture supports very large models such as BERT [107]. The core idea behind transformers is attention and it is not limited to language. By directing computation toward the most informative parts of an input, it has improved prediction in domains as different as autonomous driving and image classification. The energy domain is following a similar path. Attention combined with recurrent models has delivered accurate short-term power forecasts [108]. Recent studies now apply attention and Transformer models to electricity load forecasting with encouraging results [37,38]. These findings suggest that attention mechanisms can capture long-range dependencies and variable seasonal patterns that are hard to model with earlier architectures, although careful tuning and adequate data still appear necessary for consistent gains. Figure 10 depicts an encoder-only transformer tailored to predict demand–response potential. Multivariate time series such as load, weather, and tariffs are embedded, and positional encodings inject timing information like hour of day and season. Stacked encoder blocks with multi-head attention and feedforward layers learn long-range dependencies across hours and features, while layer normalization keeps training stable. The concatenated representation feeds a small MLP that outputs the forecasted load or the flexible capacity available over the target horizon.

7.3. Hybrid and Ensemble Methods

Ensemble and hybrid models provide two complementary workflows for forecasting DR potential, as sketched in the Figure 11. On the left, an ensemble trains several distinct learners on the same preprocessed inputs and then fuses their outputs through stacking, voting, or averaging to produce a single prediction; this setup often tempers individual model bias and variance and adapts well to heterogeneous building and weather signals. On the right, a hybrid pipeline first fits a statistical model to capture trend and seasonality, then hands the residual structure to a machine- or deep-learning model that learns remaining nonlinear patterns; the final forecast is the sum of the statistical component and the learned residual. In practice, ensembles are attractive when diverse base models each capture different aspects of the data, while hybrids are useful when classical time series structure is strong but not sufficient.

Combining hybrid and ensemble forecasting models successfully enhances the accuracy and reliability of predicting DR potential [39,40,41,109]. The methods that combine modeling techniques take advantage of their strengths to address the weaknesses found in individual methods. Widespread hybrid models use ARIMA time series techniques and deep learning models when ANNs step into handle the systematic errors that remain after ARIMA processing. Such layered modeling approaches deliver superior forecasting results within dynamic systems with high levels of user demand and market condition uncertainty [110].

Residual correction through machine learning and optimization-based results constitute a hybrid modeling approach. A combined model used SVR and RNN to forecast both short-term and long-term consumption, which delivered superior framework comprehension and dynamic learning capabilities [34]. Hybrid strategies that integrate deep neural networks with reinforcement learning recently have found interesting applications in modern grid systems for DR forecasting and incentive-based programs (RL) [111]. These methods deliver optimal results for detecting complex consumer behavior patterns alongside unpredictable pricing indicators in the energy infrastructure.

The combination of predictions from various forecasting models including ANN and decision trees combined with regression models produces an accurate output. The main reason for ensemble forecasting appears through different specialized models learning separate data characteristics because their collective predictions minimize individual forecast biases and variances [88]. Practical DR scenarios show that ensemble models deliver better performance adapting to building loads, appliance utilization, and weather conditions than single models. Ensemble approaches enable scalable deployments while remaining robust to incomplete data and model specification errors, thus making them suitable for operational smart grid systems [112].

Hybrid couplings combine an optimizer (economic dispatch, bidding, or comfort–cost scheduling) with an ML forecaster/encoder. Recent works link bidding/reserve decisions to learned forecasters (Opt + ML) [20], embed fuzzy or stacked ensembles within scheduling pipelines [24], and use MARL with explicit production constraints for industrial DR [22]. These methods report accuracy (MAE/RMSE) and decision-value outcomes (e.g., profit/cost or comfort penalties), aligning with our taxonomy and evidence tables.

8. Applications and Case Studies

8.1. Industrial Sector DR Potential

Due to the high consumption and flexible operations, industrial facilities have transformed into essential DR program participants. Aluminum smelting, cement production, and steel manufacturing demonstrate high potential to decrease electricity consumption during DR events [113,114,115,116]. Industrial facilities implement aluminum electrolysis and chlor-alkali production processes to reduce energy usage by 40% following emergency or price-based DR call-ups while generating load reductions of multiple hundreds of megawatts [117]. Process configurations that allow fast control adjustments must maintain product quality performance and system operational efficiency for maximum participation [113,114]. Thermal energy storage systems and flexible process control strategies determine the feasibility of DR participation within the cement industry. Such facilities can move their load or decrease their usage following grid signals without interrupting production operations [118]. The industrial utilization of Distribution System Management faces obstacles mainly from concerns about disruption of production, complicated equipment connections, and potential revenue decrease due to process interruptions [113,114].

The necessary solution for handling these obstacles depends on advanced automation technologies and the implementation of specific production schedules. These systems enable industrial participants to join DR programs without interrupting their fundamental business operations. Third-party aggregators and DR service providers are essential facilitators of market participation because they scan through various facilities to establish coordination processes while enhancing response strategy effectiveness [119]. The combination of strategic system integration with intelligent aggregation allows industrial customers to take part in DR markets in a reliable manner, which delivers both economic advantages and operational flexibility.

At the residential scale, Schwarz et al. [120] report a case study examining CBL error structures and demonstrate that small methodological choices (e.g., weather normalization and event day exclusion) yield non-trivial changes in credited DR, underscoring the need for standardized evaluation protocols. Finally, fast HVAC-oriented flexibility has moved from theory toward deployment. Ran et al. [121] demonstrate a virtual-sensor, self-adjusting control that achieves rapid DR actuation in building HVAC, pointing to control-layer enablers that forecasting should represent (e.g., ramp-rate limits and rebound). Collectively, these cases indicate that (i) multi-building calibration, (ii) asset- and vector-specific dynamics such as refrigeration and heat pumps, and (iii) settlement-grade baselines are pivotal to translating forecasted DR potential into realized and auditable flexibility across sectors.

8.2. Residential and Commercial Sector DR Case Studies

The residential DR program operates through user-controlled household appliances such as smart thermostats, electric vehicle chargers, and water heaters. The devices operate automatically following utility signals to achieve significant peak reduction levels. Researchers in Belgium ran a significant DR test, which proved that groups of electric heaters in homes can stabilize the grid frequency during high-demand times [122]. The residential energy users in the United States exhibit up to 15% lower peak demand when they participate in dynamic tariffs through time-of-use and critical peak pricing programs [123]. DR program performance increases because smart home technologies link DR signals to user preferences, which allows users to maintain comfort levels while lowering energy usage.

The primary sources for DR in commercial facilities like offices, malls, and hotels are HVAC management, lighting changes, and energy management system operations. HVAC systems become the preferred choice for load control because they possess operational flexibility. A study reported that commercial buildings have shown peak load reduction of 10–30% by manipulating chiller optimization and thermally pre-cooling strategies [124]. Implementing DR protocols between hotel HVAC systems enabled spinning reserve compliance through load curtailments, reaching up to 37% during peak usage periods [118]. Commercial DR programs implement incentive structures by offering demand bidding and scheduled load reduction to customers while maintaining the uninterrupted operation of their businesses, thus encouraging reliable participation.

8.3. DR in Renewable Energy Integration

The power system integration of renewable energy resources depends heavily on DR as its fundamental operation procedure [125,126,127,128]. Renewable power generation from solar and wind creates primary stability difficulties for maintaining grid stability. DR provides consumers with an adaptable system to manage their electricity consumption, increasing demand at times of renewable power surplus and decreasing demand when renewable power falls below required levels [125,126]. The participation of residential and commercial consumers in DR programs allows them to move their energy usage from peak times to daylight hours and decrease power consumption when wind generation levels drop. The flexible operations of power consumers through DR programs enhance grid stability by stopping renewable energy from detaining operations, which helps sustain reliability [129]. The integration of DR programs leads to stable voltage levels. One such example is the utilization of a data-driven robust optimization model that integrates DR to increase renewable energy consumption within community-integrated energy systems [130]. The method combined building thermal inertia control with DR signals to minimize uncertainty in renewable power generation. The prediction-based consumption shifts helped decrease renewable energy over-production while strengthening the electrical grid’s stability. Implementing generative adversarial networks (GANs) to create renewable output scenarios strengthened renewable prediction accuracy and flexible load scheduling procedures [131,132,133,134]. Such advanced models demonstrate why DR integration is essential in developing operational flexibility while affecting clean energy goals.

8.4. Lessons from Global Implementations

Implementation strategies and market structures from all over the world have become available through international DR programs. PJM and other United States Independent System Operators have developed Peak Hour Rebates and Emergency Load Response Programs, encouraging consumers to reduce power consumption during peak demand. Networks in California showcase the importance of informing customers about Flex Alerts and critical peak pricing through alert systems, which results in increased participation levels [135]. Through these programs, the peak demand has been successfully reduced, and the risk of blackouts has been eliminated significantly, especially when heat waves emerge and renewable generation loses its stability.

The European power grid now utilizes DR through market integration that lets aggregators manage residential and commercial load infrastructures to offer grid services. Through EPEX and Nord Pool, customers can interact with balancing markets by receiving real-time signals that are dispatched through market-based processes [136]. DR technologies can displace traditional spinning reserves according to the examples set by Nordic countries, which offer faster, cheaper flexibility options. Throughout Australia, the Australian Energy Market Operator (AEMO) executes DR testing to alleviate transmission network limitations while enhancing the integration of solar power in communities with abundant solar harvesting capabilities.

Managed demand–response programs work best when three basics are in place: transparent pricing, timely data, and clear regulatory support. Even with those in hand, participation can vary. Local energy-use habits, trust in automation, and access to digital tools often determine participation rates. Early pilots in several Asian markets show mixed outcomes, which appears to reflect differences in market maturity and customer engagement. Evidence also suggests that well-designed dynamic tariffs that adapt prices to household consumption patterns can deliver meaningful savings and load shifting in some regions [137]. Taken together, these lessons point to the conclusion that grid operators should tailor DR program design to local conditions, including socioeconomic profiles, retail market rules, and the available technology stack. At a system level, DR continues to expand because it helps integrate variable renewables, trims peak demand, and adds operational flexibility. Program success usually depends on a few practical ingredients that reinforce one another: supportive regulation, advanced metering and automation, clear customer communications, and market rules that can adjust as conditions change. Where these elements align, DR tends to scale; where any one of them is missing, results are uneven. Some famous DR programs on national level are displayed in Figure 12. China, USA, and various European countries are expanding these programs at rapid pace as these countries are the early adopters of renewable energy and hence there are multiple projects that are currently underway in these countries.

9. Technological Enablers for DR Implementation

9.1. Smart Grid and Advanced Metering Infrastructure (AMI)

Smart grids transform traditional electric power systems into more dynamic and efficient networks, enabling sophisticated management of electricity consumption. AMI’s role as a transformation tool centers on its ability to enable communication between utility providers and end consumers. Real-time power usage and control functions through smart meters, communication networks, and data management systems constitute AMI infrastructure [138,139]. Smart meters deliver detailed usage information to the utility to implement adjustable price signals during DR events and confirm load reduction results [140,141]. AMI delivers multiple advantages, including automated grid operation, fault identification, outage monitoring, and DERs integration. Real-time decisions that AMI technology supports help virtual power plants perform load-balancing operations and cost-reduction activities. AMI is a crucial element for smart city grid reliability since it enables advanced capabilities such as remote billing and efficient load scheduling, resulting in peak shaving success [142]. AMI establishes an upgraded network system that facilitates AI-based forecasting algorithms and improved DR event prediction through enhanced energy management systems [143].

9.2. IoT and Demand-Side Management (DSM)

DSM experienced a revolution through the IoT with smart devices that automatically execute contextual commands sent by utilities. Smart thermostats, lighting systems, HVAC controllers, and EV chargers leverage real-time pricing or DR event signals to modify consumption patterns [138,144]. IoT platforms with DSM capabilities constantly achieve optimal energy usage at all residential, commercial, and industrial sites by fusing user needs with weather patterns and predictive modeling techniques [140].

IoT technologies deliver edge intelligence while performing device automation; thus, operators and aggregators need less human intervention, and the DR functions more reliably and quickly. Appliances within home area networks (HANs) and building energy management systems (BEMS) use communication abilities to respond instantaneously to grid requests [141,142]. During peak hours, these systems reduce energy usage effectively while keeping comfort levels high throughout the process [145].

Constructs from Smart IoT that link with cloud and edge computing systems allow the development of scalable DSM solutions. The integrated systems contribute to decentralized power networks and cyber threat resilience by using protected communication methods and privacy mechanisms [138,144]. The combination of IoT technology with AMI and AI systems seems to bring about a new future of proactive demand management supporting the smart grid’s evolution.

9.3. Blockchain for DR Transactions and Cloud-Based Big Data Analytics

Blockchain technology developments now enable secure, transparent, and decentralized transactions within the DR framework. The blockchain infrastructure enables direct energy trading between peers. At the same time, smart contracts confirm all DR activities and ensure protected tracking logs [146]. The automated nature of smart contracts handles DR service execution, which extends from baseline estimation to financial settlement through operations flow improvements and service adherence validation [147]. The Hyperledger Fabric platform uses actual DR contracts, which activate procedures based on authenticated consumption records to eliminate user participation disputes as shown in [147]. The decentralized architecture improves energy system resilience by reducing failure points while allowing distributed nodes to validate data autonomously [148]. The ability of blockchain technology to manage micropayments enables the combination of minimal flexibility from consumers, which is an essential factor in increasing residential Distribution Reliability participation [149].

The feasibility of blockchain-powered DR market solutions became evident through successful pilot demonstrations in European and Asian regions. Distributed ledger systems linked with prediction models enable system operators to conduct secure DR signal distributions and immediately trace energy variations according to research by Hashmi et al. [150]. Such distributed systems implement multi-signature methods together with Proof of Stake (PoS) protocols to decrease energy consumption, which otherwise occurs in Proof of Work (PoW) systems [146]. The blockchain-based platform Gridmonitoring protects smart grid infrastructure employing secure monitoring while adjusting DR incentives through user behavior adjustments as per [151].

Blockchain innovation parallels cloud infrastructure and big data analytics, providing digital support systems for immediate DR decision-making processes. The collection and processing of extensive grid data stemming from smart meters, sensors, and IoT devices depend on cloud platforms used by utilities and aggregators, according to [152]. The processing capabilities of cloud-based DR platforms achieve flexible expansions to execute sophisticated forecasting algorithms while performing load categorization and optimization model activities [153]. The implementation of significant data architecture in DR employs distributed fog systems to preprocess data near users’ locations and central cloud facilities managing strategic load management decisions [154]. The tools Hadoop and Spark provide instant streaming analytics to recognize demand anomalies through which they deploy immediate corrective DR signals per [155].

10. Challenges and Barriers in DR Potential Calculation and Forecasting

10.1. Technical Challenges

The rapid growth of DR research has not overcome key technical restrictions that impede precise predictions of potential demand reduction. The main challenge stems from setting dependable load curtailment benchmarks during DR events. These fundamental reference points need to incorporate consumer behavioral patterns and weather conditions because they substantially influence forecasting accuracy [156]. The existing DR estimation techniques face challenges when accommodating modeling uncertainties caused by the dynamic and nonlinear responses of residential, commercial, and industrial loads [157]. Implementing automated DR systems introduces complexity to operations due to problems with delayed communication and faulty equipment, resulting in delayed or failed load-shedding events [158]. System operators face two significant computational challenges when integrating DR into their operational tools. They need simultaneously synchronized models across various time scales and technological domains while dealing with increased computational complexity between regions as described in [159]. When customers participate in price-based DR programs, they face the challenge of controlling different end-use loads like HVAC and industrial processes because these programs may affect product quality or customer comfort levels [156].

10.2. Regulatory and Policy Challenges

The existing regulatory framework outlines the conditions that encourage and restrict distribution resource participation throughout the power system. Several electricity market rules restrict the complete involvement of aggregated DR resources, particularly for small consumers or third-party aggregators [160]. Implementing flat-rate tariffs across multiple regions produces barriers for consumers in using price signals effectively, which diminishes their economic motivation to join DR programs [161]. Dynamic pricing systems should be implemented to improve market responsiveness. This essential factor determines real-time DR capability and enhances forecasting accuracy. Installing DR facilities into capacity markets remains restricted in different regions while providing dependable operational flexibility. The lack of standardized evaluation measurement methods with transparent criteria generates investment uncertainties that hinder the further development of DR resources [157]. The existing market designs tend to prefer supply-based solutions and undervalue the effectiveness and reliability of demand-side flexibility. The potential of DR is limited by three essential policy measures that include real-time settlements and access to consumer data and DR aggregation protocols [158]. Comprehensive regulatory support limits DR scalability, especially when consumers cannot access technologies like advanced metering and programmable load control [160].

10.3. Consumer Participation and Behavioral Aspects

The outcomes of DR programs heavily depend on how consumers behave. Consumer awareness levels and their engagement with programs show slight improvement over time. Most residential users demonstrate limited knowledge regarding DR programs and experience discomfort from program participation whenever their comfort preferences or daily routines require electricity curtailments. Behavioral inertia coupled with distrust in utility operations leads to poor DR program participation [162,163]. The predictive capability and scalability of DR programs become diminished because users choose to opt-out due to their concerns about losing appliance control or privacy or worries about billing mistakes [164,165]. When forecasting DR events, it becomes challenging to predict consumer participation because their actual responses during scheduled outages remain uncertain.

The “rebound effect” emerges due to behavioral response variability, which causes users to elevate their consumption afterward to counter initial consumption reduction, thus offsetting any initial energy savings [166]. Behavioral modeling has recently started appearing in DR forecasting frameworks to address the diverse behaviors, consumption patterns, and psychological factors that determine user load-shifting behavior [163]. Systems that precisely segment customers according to pricing response and behavioral and participation willingness can design specific reward programs. Educational drives and simple interface tools, including mobile messaging and present-time usage updates, have raised user confidence and elevated engagement numbers [162]. Non-monetary rewards and gamification methods should be used as environmental motivation tools alongside economic incentives to improve their effectiveness. A reliable DR program needs a multi-faceted method that combines behavioral science research with communication approaches and incentives distributed equally among participants.

11. Future Trends and Research Directions

Recent advances in DR potential forecasting have focused on enhancing model fidelity and addressing fundamental tasks like accurate CBL estimation. A precise baseline is essential for quantifying DR, since it defines the counterfactual consumption against which load reductions are measured [167]. However, baselines cannot be directly observed during DR events, making them challenging to predict. To tackle this, new data-driven methods reframe CBL estimation as a time series imputation and forecasting problem. Building on such foundations, researchers are utilizing cutting-edge AI architectures to improve DR potential modeling. Transformer-based models, with their self-attention ability to capture long-range temporal patterns, have demonstrated superior performance in energy demand forecasting, outperforming recurrent networks with lower prediction error [9]. These architectures can effectively learn complex consumption cycles and anomalies, which is crucial for reliable DR potential forecasts. Another promising direction is diffusion probabilistic models that treat forecasting as a generative task. Instead of sequential point predictions, diffusion models generate entire load curves and capture the full distribution of possible demand outcomes [168]. Early studies report that this approach not only achieves state-of-the-art accuracy but also provides a richer representation of uncertainty in power system forecasts, offering novel insights for managing the volatility of demand-side resources.

Emerging machine learning paradigms are further expanding the DR forecasting toolkit. Graph Neural Networks (GNNs) have been introduced to exploit spatial and network correlations in demand data by modeling how different customers or grid nodes influence each other. A recent study proposed an explainable causal GNN (X-CGNN) that integrates causal inference to provide both global and local explanations for multivariate electricity demand forecasts, while achieving state-of-the-art accuracy at household and feeder levels [169]. Such approaches demonstrate the potential to improve prediction fidelity and interpretability by embedding domain structure into the models. Federated learning (FL) is another promising avenue, aimed at addressing data privacy and scalability in model training [170]. Rather than centralizing sensitive smart meter data, FL techniques train DR prediction models locally on edge devices and aggregate only the learned parameter. Such privacy-preserving frameworks are increasingly relevant as vast IoT datasets become available for DR analysis.

Another frontier for DR potential forecasting is the integration of digital twin technologies. A digital twin creates a high-fidelity virtual model of physical assets or systems (e.g., buildings, campuses, or distribution feeders) that updates in real time via sensor data [171]. This approach enables real-time simulation of DR scenarios and proactive forecasting: operators can test control strategies or predict DR outcomes under various conditions in the virtual twin before deploying them on the real grid. In current applications, digital twins serve as enabling tools for enhanced monitoring, visualization, and decision support in DR programs. Future research is expected to scale up these implementations (e.g., to community or city levels) and increase their fidelity for faster response and coordination. As digital twin frameworks mature and integrate with AI forecasting models, they could significantly improve the agility and situational awareness of DR operations.

Despite these technological gains, practical deployment of advanced DR forecasting models faces significant challenges. One critical issue is model interpretability. Grid operators and other stakeholders demand transparency in how forecasts are produced, especially as algorithms grow more complex. This is spurring a growing emphasis on explainable AI techniques for DR—from post hoc methods that attribute predictions to key features, to intrinsically interpretable architectures that yield insight into decision drivers [169]. Another barrier is the integration of novel forecasting tools into existing grid operation and market platforms. Many utility management systems are legacy environments, so bridging the gap between cutting-edge AI models and real-time grid operations is non-trivial [172]. Ensuring that a new DR potential forecasting tool can interface with energy management systems, DR automation servers, or ISO/RTO market signals requires standard data schemas and rigorous validation. Issues of user trust, regulatory support, and standardized evaluation practices can limit the real-world impact of forecasting innovations [173]. Addressing these interdisciplinary challenges in tandem—combining advanced modeling with transparent, user-centric, and secure deployment—defines a critical direction for future research in DR potential forecasting.

12. Conclusions

This review has assessed the main approaches for estimating and forecasting demand response potential across residential, commercial, and industrial sectors. We examined how statistical, analytical, optimization, and AI-based models fit together in practice, highlighting where results are limited by data quality, baseline construction, user behavior, and market rules. The review points to the conclusion that accurate potential estimates depend on sound baselines and transparent evaluation. No single modeling family is sufficient. Our findings indicate a clear trend towards AI-based and hybrid models, which are particularly adept at capturing the complex, nonlinear consumer dynamics that challenge simpler statistical approaches. The most operationally viable strategies, however, often involve hybrid systems that leverage the predictive power of AI to correct the residuals of more transparent statistical models, thereby combining the strengths of multiple paradigms. While these advanced methods show high predictive accuracy in research settings, their ability to directly contribute to grid reliability is contingent on bridging a significant research-to-practice gap. Our review shows that the transition to widespread, real-time operational use is still underway. Key challenges remain, including the need for model interpretability, robust handling of data privacy, and seamless integration with legacy utility systems. Therefore, while these forecasting tools are powerful, they are best understood as enablers that support operational decision-making rather than standalone solutions. In the future, progress will come from better baseline estimation and validation, hybrid and probabilistic forecasting aligned with control and market use, privacy-aware data pipelines, and program designs that support participation and standardized evaluation. These steps can move DR potential forecasting from scattered case studies to repeatable, comparable, and deployable practice at scale. We therefore recommend a community benchmark suite with shared datasets and standard splits, routine reporting of quantiles/CRPS/coverage with calibration checks, paired comparisons to strong baselines, explicit links to operational/cost outcomes, and multi-site, cross-sector validations of emerging hybrid models.