Smart meters generate high-resolution data that can feed forecasting models for a range of distribution network applications. However, the literature remains fragmented across forecasting methods, data structures, and operational use cases. No comprehensive taxonomy links forecasting model characteristics to the distinct requirements of grid flexibility, microgrid management, and congestion control. This limits the ability of researchers and system operators to choose models according to data availability, forecast horizon, and grid function. This article develops a taxonomy of smart-meter forecasting models for distribution-level operation. The taxonomy maps model categories, data inputs, forecast horizons, and application domains into a structured classification framework. A systematic literature review of 27 peer-reviewed journal articles published from 2017 to 2024 was conducted. Models were classified by mathematical structure, data requirements, forecast horizon, and relevance to grid flexibility, microgrid energy management, and congestion control. The taxonomy organises forecasting approaches into classical time series, conventional machine learning, deep learning, hybrid and ensemble, and probabilistic forecasting categories. Six tables support the review method, model classification, smart-meter data mapping, application suitability analysis, full taxonomy, and research gap assessment. The taxonomy provides a structured decision-support tool for selecting and designing forecasting models for specific distribution grid tasks. It highlights gaps in cross-domain transferability, real-time deployment, benchmarking, data integration, and operational validation.
Smart-meter data has become a central resource for distribution grid intelligence because it captures granular consumption behaviour, temporal variability, and customer-level demand patterns that were previously unavailable to system operators. Wang, Chen, Hong, and Kang describe smart-meter analytics as a foundation for applications ranging from forecasting to anomaly detection and customer segmentation [1]. Yildiz, Bilbao, Dore, and Sproul similarly show that residential smart-meter data enables detailed analysis of electricity consumption behaviour, but they also emphasise challenges related to variability, household heterogeneity, and data interpretation [2]. Within this context, forecasting models provide the analytical bridge between raw metering streams and operational decisions in active distribution networks.
The literature includes forecasting models that support three increasingly important distribution-level functions: grid flexibility, microgrid energy management, and congestion control. Short-term residential load forecasting using long short-term memory networks demonstrates how smart-meter sequences can support demand anticipation at customer or feeder level [3]. Distributed and privacy-preserving forecasting studies extend this logic toward multi-site coordination, where forecasts can inform flexibility services without necessarily centralising raw customer data [4, 5]. Congestion-oriented work further shows that probabilistic forecasting and power-flow-informed learning can help anticipate overload or voltage-related constraints before corrective action is required [6].
This article contributes a taxonomy review that classifies smart-meter forecasting models by model family, data requirement, forecast horizon, and application suitability. Existing studies provide important technical advances, including adaptive recurrent learning [7], deep probabilistic forecasting [8], and transformer-based forecasting [9], but these contributions are often evaluated within separate methodological or application silos. A taxonomy is therefore needed to make explicit which model categories are most compatible with different smart-meter data types and distribution grid tasks. The review develops this classification without conducting new forecasting experiments, focusing instead on analytical organisation of the 2017–2024 literature.
Figure 1 illustrates how smart-meter forecasting connects granular metering data, forecasting model families, and distribution-level operational functions across grid flexibility, microgrid management, and congestion control.
Figure 1. Conceptual Scope of Smart-Meter Forecasting for Distribution-Level Grid Intelligence
Existing reviews and methodological studies provide valuable coverage of smart-meter analytics, residential load forecasting, and smart-grid applications, but they do not fully integrate model structure, data type, and operational use case into a single classification system. Wang, Chen, Hong, and Kang review broad smart-meter analytics applications, while Kuster, Rezgui, and Mourshed critically assess electrical load forecasting models across building and energy contexts [1, 10]. Dewangan, Abdelaziz, and Biswal focus specifically on smart-grid load forecasting using smart-meter information, yet the classification is not primarily organised around flexibility, microgrid management, and congestion-control suitability [11]. As a result, model choice can remain method-driven rather than task-driven.
A structured taxonomy is necessary because smart-meter forecasting requirements differ substantially across applications, even when similar algorithms are used. Forecasts for demand response may prioritise customer-level peak timing, while microgrid scheduling may require net-load and local-generation forecasts, and congestion control may require feeder, transformer, or voltage-sensitive predictions [4, 6]. Deep learning and ensemble models may be attractive for complex temporal patterns, but their operational value depends on data resolution, missing-data handling, interpretability, and deployment constraints [7, 12]. This taxonomy therefore provides a decision-oriented classification that links forecasting method selection to the operational function it is expected to support.
Figure 2 shows why a taxonomy is needed by contrasting method-centred forecasting studies with task-aligned model selection based on data type, forecast horizon, uncertainty needs, and distribution-grid application.
Figure 2. Rationale for a Taxonomy: From Fragmented Forecasting Studies to Task-Aligned Model Selection
The review used a targeted literature search covering smart-meter forecasting, short-term load forecasting, microgrid energy management, grid flexibility, and distribution-level congestion control from 2017 to 2024. The search prioritised peer-reviewed journal articles from IEEE Transactions on Smart Grid, Applied Energy, Energy, International Journal of Electrical Power & Energy Systems, Renewable and Sustainable Energy Reviews, Energy and AI, and related Q1 energy informatics journals. Search strings combined terms such as “smart meter,” “forecasting,” “load forecasting,” “microgrid,” “demand response,” “distribution congestion,” “deep learning,” and “taxonomy,” reflecting the scope of prior smart-meter analytics and forecasting reviews [1, 11]. The resulting corpus was screened to retain studies that directly informed forecasting model classification or distribution-grid application mapping.
Articles were included when they used, reviewed, or classified forecasting models relevant to smart-meter data, residential or distribution-level load prediction, privacy-preserving forecasting, flexibility applications, microgrid energy management, or congestion-oriented distribution operation. Studies such as Kong, Dong, Jia, Hill, Xu, and Zhang were included because they demonstrate smart-meter-based residential forecasting with recurrent neural networks [3], while Fekri, Grolinger, and Mir were included because federated recurrent forecasting addresses distributed smart-meter settings [4]. Articles were excluded if they were conference papers, policy documents, standards, non-peer-reviewed sources, or studies outside the 2017–2024 window. Studies that focused on power systems but did not provide forecasting relevance to smart-meter analytics or distribution-level operation were not retained unless they contributed to taxonomy dimensions such as data type, model class, or operational suitability.
The classification framework organised the 27 included studies along four axes: forecasting model category, smart-meter data type, application domain, and forecast horizon. Model categories were derived from repeated methodological patterns in the literature, including classical time series, conventional machine learning, deep learning, hybrid and ensemble approaches, and probabilistic forecasting [8, 13]. Data categories reflected smart-meter and distribution-grid observability, including active power consumption, net load, voltage-related measurements, reactive power, and prosumer generation traces [2, 6]. Table 1 summarises the literature search and screening process leading to the final set of 27 included studies.
Table 1. Literature Search and Screening Summary
Stage | Description | Count |
Records identified through database searches | Records retrieved using targeted searches across smart-meter forecasting, load forecasting, grid flexibility, microgrid management, and congestion-control terms | 214 |
Additional records identified through citation tracking | Additional peer-reviewed journal articles found through backward and forward citation checking of major review and forecasting papers | 36 |
Total records identified | Combined database and citation-tracking records before deduplication | 250 |
Duplicates removed | Duplicate records removed across databases and citation sources | 48 |
Records screened by title and abstract | Records screened for relevance to smart-meter forecasting, distribution-grid operation, model taxonomy, or application suitability | 202 |
Records excluded after title and abstract screening | Records excluded for weak relevance, non-journal format, non-forecasting focus, or lack of smart-meter/distribution-grid connection | 119 |
Full-text articles assessed | Articles assessed for methodological relevance, date range, peer-reviewed status, and application-domain fit | 83 |
Full-text articles excluded | Articles excluded because they were outside 2017–2024, lacked DOI information, were conference papers, or did not support taxonomy construction | 56 |
Studies included in taxonomy review | Final peer-reviewed journal articles used for taxonomy construction | 27 |
Classical time series models, including ARIMA, SARIMA, exponential smoothing, and related statistical methods, remain important baselines in smart-meter forecasting because they provide interpretable temporal structures and relatively low computational cost. Jagait, Fekri, Grolinger, and Mir explicitly use ARIMA within an online ensemble framework, showing how classical models can remain relevant under changing load patterns when combined with adaptive learning [12]. Kuster, Rezgui, and Mourshed also identify traditional statistical models as essential reference points for evaluating newer forecasting approaches [10]. In this taxonomy, classical models are classified as low-complexity, baseline-oriented methods best suited for stable aggregate loads, limited feature spaces, and settings where interpretability is more important than nonlinear representation capacity.
Conventional machine learning models include random forests, support vector regression, gradient boosting, shallow neural networks, and other supervised learning methods that can incorporate weather variables, calendar features, lagged consumption, and customer-level descriptors. Tavassoli-Hojati, Ghaderi, Iranmanesh, Hilber, and Shayesteh demonstrate how neuro-fuzzy modelling can support smart-grid short-term load forecasting while preserving some interpretive structure [14]. Mehdipour Pirbazari, Farmanbar, Chakravorty, and Rong further show that smart-meter forecasting models must be evaluated for generalisation across households and data conditions rather than only local accuracy [15]. In this taxonomy, conventional machine learning occupies an intermediate category between transparent statistical baselines and high-capacity deep learning models.
Deep learning models are prominent in smart-meter forecasting because they can capture nonlinear temporal dependencies, customer heterogeneity, and complex sequential patterns in high-resolution load data. Kong, Dong, Jia, Hill, Xu, and Zhang use LSTM recurrent neural networks for short-term residential load forecasting, while Shi, Xu, and Li develop a pooling deep recurrent architecture for household load forecasting [3, 16]. Fekri, Patel, Grolinger, and Sharma extend this direction through online adaptive recurrent learning, highlighting the need to update models as consumption behaviour changes [7]. Transformer-based and attention-based approaches further expand this category by improving long-range temporal representation and contextual feature weighting [9, 17].
Hybrid and ensemble models combine complementary methods to improve robustness, adaptivity, or feature representation under non-stationary smart-meter conditions. Zhang, Tian, Ma, Li, Lu, and Niu combine variational mode decomposition with stacking ensemble learning, illustrating how signal decomposition and ensemble aggregation can be used to handle complex load patterns [18]. Mujeeb and Javaid propose predictive analytics schemes for electricity load and price, demonstrating the value of combining optimisation-inspired learning and data-driven estimation in smart-grid contexts [19]. In this taxonomy, hybrid and ensemble models are treated as structurally composite methods whose value lies not only in accuracy but also in their ability to combine temporal decomposition, nonlinear learning, and application-specific decision support.
Table 2 summarises the main categories of load forecasting models, highlighting their typical data requirements, key methodological strengths, and inherent limitations, and providing a comparative overview to support appropriate model selection for different operational and data conditions.
Table 2. Overview of Forecasting Model Categories
Category | Key Algorithms | Typical Data Requirements | Strengths | Limitations |
Classical time series | ARIMA, SARIMA, exponential smoothing, autoregressive models | Regular historical consumption series, stable sampling interval, limited missing values | Transparent, computationally efficient, useful as baselines, suitable for aggregate stable demand | Weak nonlinear representation, limited ability to handle heterogeneous households, sensitive to non-stationarity |
Conventional machine learning | Random forest, support vector regression, gradient boosting, neuro-fuzzy models, shallow neural networks | Historical load, calendar variables, weather features, customer descriptors, lagged values | Handles nonlinear covariates, flexible feature integration, often easier to train than deep models | Requires feature engineering, may generalise poorly across feeders or customer groups, limited sequence representation |
Deep learning | LSTM, GRU, CNN, sequence-to-sequence RNN, attention networks, transformers | High-resolution sequential data, larger training sets, weather/calendar covariates, sufficient computing resources | Strong temporal pattern learning, suitable for high-dimensional and non-stationary data, adaptable to complex load profiles | Data-hungry, less interpretable, sensitive to drift, requires careful validation and deployment monitoring |
Hybrid and ensemble models | Decomposition plus learning, stacking ensembles, ARIMA-RNN hybrids, adaptive ensembles | Multiple historical series, decomposed signals, engineered features, sometimes external variables | Improves robustness, combines complementary model strengths, useful for non-stationary load behaviour | Can be complex to reproduce, may increase computational burden, difficult to interpret operationally |
Probabilistic and quantile models | Quantile regression, probabilistic neural networks, conditional density models, scenario forecasting | Historical load distributions, uncertainty drivers, sufficient observations across operating conditions | Represents forecast uncertainty, supports risk-aware scheduling and congestion prevention, useful for flexibility markets | Requires probabilistic calibration, harder to communicate to operators, benchmarking is less standardised |
Active power consumption is the most widely used smart-meter data type in short-term load forecasting because it directly represents customer, building, feeder, or aggregation-level demand. Residential consumption studies show that active power traces contain strong daily, weekly, and behavioural patterns, but these patterns can vary substantially across households and seasons [2]. Deep recurrent forecasting studies rely heavily on active power sequences because they provide the temporal signal needed for LSTM, RNN, and attention-based learning [3, 17]. In this taxonomy, active power consumption is classified as the foundational data type for demand response, peak forecasting, aggregate flexibility estimation, and baseline load prediction.
Voltage and reactive power measurements are less commonly foregrounded than active consumption, but they are critical for distribution-level operational applications such as voltage management, congestion detection, and feeder constraint forecasting. Hernandez-Matheus, Berg, Gadelha, Aragüés-Peñalba, Bullich-Massagué, and Galceran-Arellano develop a congestion forecast framework that connects probabilistic power flow and machine learning, indicating that operational forecasting must account for network-state variables rather than consumption alone [6]. Voltage-related forecasting is particularly relevant when distributed energy resources, unbalanced loading, or local congestion risks affect service quality. In this taxonomy, voltage and reactive power data are treated as grid-state data types that support congestion-control and voltage-profile applications more directly than customer-facing flexibility applications.
Net load and prosumer data combine consumption, distributed generation, storage behaviour, and sometimes electric vehicle charging, making them especially important for microgrid management and flexibility provision. Xu, Dai, Li, Wu, and Chen develop an interpretable transformer-based model for probabilistic residential net load forecasting, showing how prosumer-oriented forecasting requires both temporal learning and uncertainty representation [20]. Federated and distributed forecasting studies also show that prosumer and household-level data may be privacy-sensitive, requiring methods that learn from local data without fully centralising it [4, 5]. Table 3 classifies smart-meter data types according to resolution, model compatibility, and preprocessing requirements.
Table 3. Smart-Meter Data Types and Forecasting Model Compatibility
Data Type | Typical Resolution | Suitable Model Categories | Preprocessing Needs |
Active power consumption | 1-minute to hourly readings, often aggregated to 15-minute or hourly intervals | Classical time series, conventional machine learning, deep learning, hybrid models | Missing-value imputation, calendar alignment, outlier handling, lag construction, seasonal decomposition |
Reactive power | 1-minute to hourly readings when available through advanced meters or grid sensors | Conventional machine learning, hybrid models, probabilistic forecasting | Synchronisation with active power, scaling, noise filtering, treatment of low-load intervals |
Voltage magnitude | Feeder, transformer, or customer-level voltage readings at sub-hourly or hourly intervals | Conventional machine learning, probabilistic models, hybrid congestion-oriented models | Phase alignment, topology mapping, detection of abnormal voltage events, removal of sensor artefacts |
Net load | Consumption minus local generation, usually sub-hourly or hourly | Deep learning, probabilistic forecasting, transformers, hybrid models | PV-consumption separation, weather alignment, negative-load treatment, uncertainty calibration |
Distributed PV generation | Sub-hourly or hourly generation traces, often linked to weather and irradiance | Machine learning, deep learning, hybrid weather-informed models | Weather feature integration, daylight filtering, cloud-event handling, normalisation by system size |
Electric vehicle charging load | Irregular event-based or interval-metered charging profiles | Deep learning, hybrid models, probabilistic forecasting | Session detection, event aggregation, missing session reconstruction, behavioural clustering |
Aggregated feeder load | Feeder-level or transformer-level demand, typically 5-minute to hourly | Classical time series, machine learning, probabilistic forecasting, hybrid models | Customer aggregation, topology validation, peak-event labelling, alignment with operational constraints |
Privacy-preserving distributed data | Locally retained household or site-level smart-meter data | Federated recurrent models, distributed deep learning, adaptive models | Local standardisation, secure aggregation design, drift monitoring, site-level quality checks |
Demand response applications require forecasts that identify when demand is likely to peak, how flexible loads may be shifted, and which customer segments can provide operationally useful reductions. Smart-meter analytics literature shows that granular consumption traces can support customer-level and aggregate demand insight, but the forecasting model must be matched to the temporal granularity of the demand response action [1]. Deep recurrent models are suitable when the objective is to capture nonlinear daily and weekly consumption patterns, as shown in residential short-term load forecasting studies using LSTM and pooled recurrent architectures [3, 16]. In this taxonomy, demand response forecasting is therefore classified as a short-horizon, behaviour-sensitive application requiring active power consumption data, calendar variables, and models capable of representing temporal variability.
Dynamic tariff design depends on forecasts that estimate expected demand under time-varying price conditions, peak periods, and possible customer response. Predictive analytics models for electricity load and price show that tariff-oriented forecasting requires both consumption prediction and price-sensitive contextual information [19]. Smart-meter data can reveal household-level behavioural diversity, but Yildiz, Bilbao, Dore, and Sproul show that this diversity also makes tariff forecasting sensitive to customer segmentation and aggregation level [2]. Within the taxonomy, dynamic tariff forecasting is positioned between load prediction and flexibility estimation because the model must anticipate both baseline demand and the likely timing of consumption adjustment.
Flexibility service provision requires forecasts that quantify available upward or downward flexibility at household, feeder, aggregator, or microgrid level. Federated load forecasting studies show that distributed forecasting can support multi-site or customer-level prediction while reducing the need to centralise sensitive smart-meter data [4, 5]. Adaptive federated forecasting further indicates that flexibility-oriented models must handle changes in local consumption patterns across time and participants [21]. In this taxonomy, flexibility service forecasting is treated as an uncertainty-sensitive application where probabilistic, distributed, and adaptive models are more suitable than static point-forecasting approaches.
Peak load prediction is one of the most direct uses of smart-meter forecasting for grid flexibility because it identifies intervals where load shifting, demand response, or local flexibility activation may be needed. Probabilistic residential load forecasting is especially relevant because peak periods are uncertain and often operationally more important than average demand [8]. Enhanced deep networks for short-term and medium-term load forecasting demonstrate how high-capacity models can be used for multiple planning horizons, although the operational interpretation of peak forecasts remains a challenge [22]. Table 4 maps forecasting model categories to grid flexibility use cases and identifies the main modelling challenges for each application.
Table 4. Forecasting Models to Grid Flexibility Applications Mapping
Flexibility Application | Suitable Model Categories | Forecast Horizon | Key Challenges |
Demand response event planning | Deep learning, conventional machine learning, hybrid models | 15 minutes to day-ahead | Customer heterogeneity, behavioural uncertainty, baseline estimation, event fatigue |
Load shifting optimisation | Classical time series, machine learning, deep learning | Intra-day to day-ahead | Identifying shiftable demand, separating flexible and inflexible loads, rebound effects |
Dynamic tariff design | Machine learning, hybrid models, probabilistic forecasting | Day-ahead to week-ahead | Price-response uncertainty, segmentation bias, interaction with weather and occupancy |
Flexibility service aggregation | Federated learning, probabilistic forecasting, adaptive recurrent models | Real-time to day-ahead | Privacy constraints, local model drift, aggregation across heterogeneous customers |
Peak load prediction | Deep learning, probabilistic models, hybrid ensembles | 15 minutes to day-ahead | Extreme-event calibration, rare peaks, high penalty for missed overload intervals |
Flexibility reserve estimation | Probabilistic forecasting, quantile models, scenario-based learning | Hour-ahead to day-ahead | Uncertainty communication, reserve confidence levels, coordination with market rules |
The proposed taxonomy is organised around four core dimensions: model category, smart-meter data type, application domain, and forecast horizon. This structure extends broad smart-meter analytics classifications by making the operational purpose of the forecast explicit rather than treating forecasting as a standalone technical task [1]. It also builds on electrical load forecasting reviews that compare model families but do not fully map them to grid flexibility, microgrid management, and congestion-control applications [10]. The taxonomy therefore functions as a decision framework for matching forecasting design to distribution-level operational needs.
Figure 3 presents the proposed taxonomy architecture, showing how model category, smart-meter data type, forecast horizon, and operational application jointly determine suitable forecasting designs.
Figure 3. Integrated Taxonomy Architecture for Smart-Meter Forecasting Model Selection
The taxonomy reveals that model suitability depends on the interaction between data type, modelling structure, and intended application. Active power consumption is sufficient for many demand response and peak prediction tasks, but microgrid management often requires net load and local generation data, while congestion control requires feeder, transformer, voltage, or reactive power information [20, 23]. Deep learning methods are most useful where high-resolution sequential data and nonlinear behaviour dominate, whereas probabilistic and hybrid models become more relevant where uncertainty affects operational risk [8, 18]. This integrated view prevents the misleading assumption that one high-performing model category is universally suitable across all distribution-grid forecasting tasks.
The taxonomy can be used by first identifying the operational decision, then selecting the required data type, forecast horizon, and model family. For example, demand response planning may begin with active power consumption and customer-level features, while congestion control may require feeder aggregation, voltage measurements, and probabilistic outputs [6, 13]. Reviews of smart-meter load forecasting show that model evaluation should account for data quality, missing values, sampling resolution, and application relevance rather than relying only on generic forecast-error measures [11, 24]. Operational use of the taxonomy therefore requires aligning model selection with decision timing, data observability, uncertainty tolerance, and implementation constraints.
A utility seeking short-term residential peak forecasts for demand response could use deep recurrent or hybrid models trained on active power, calendar, and weather data, following the methodological logic of LSTM and adaptive recurrent forecasting studies [3, 7]. A microgrid operator seeking battery scheduling support would require net-load forecasting and may benefit from transformer-based probabilistic models or decomposition-based ensembles because local generation and consumption jointly shape storage decisions [18, 20]. A distribution operator seeking feeder congestion warnings would prioritise probabilistic or machine-learning models that incorporate network-state information and power-flow constraints [6]. Table 5 presents the complete taxonomy by linking model category, data type, application domain, forecast horizon, and performance indicators.
Table 5. Complete Taxonomy of Smart-Meter Forecasting Models
Model Category | Data Type | Application Domain | Forecast Horizon | Typical Performance Indicators |
Classical time series | Active power consumption | Baseline load forecasting, aggregate peak prediction | Hour-ahead to day-ahead | Mean absolute error, root mean square error, mean absolute percentage error, peak timing error |
Classical time series | Aggregated feeder load | Feeder demand monitoring | Day-ahead | Forecast bias, peak error, seasonal residual structure |
Conventional machine learning | Active power plus weather and calendar variables | Demand response planning, tariff design | 15 minutes to day-ahead | Error by time block, peak-period error, feature importance stability |
Conventional machine learning | Voltage and reactive power | Voltage management, transformer monitoring | Real-time to hour-ahead | Voltage deviation error, constraint violation detection, classification accuracy for risk intervals |
Deep learning | High-resolution household consumption | Residential load forecasting, peak load prediction | 5 minutes to day-ahead | Sequence error, peak capture rate, temporal stability, generalisation across households |
Deep learning | Net load and prosumer data | Microgrid balancing, PV-consumption coordination | 15 minutes to day-ahead | Net-load error, ramp forecasting error, probabilistic interval coverage when available |
Transformer and attention models | Long smart-meter sequences with contextual variables | Net-load forecasting, complex prosumer forecasting | Hour-ahead to day-ahead | Forecast accuracy, attention interpretability, uncertainty calibration, peak-event detection |
Hybrid and ensemble models | Decomposed load series, weather features, calendar variables | Battery scheduling, flexible load management | Hour-ahead to week-ahead | Component-wise error, ensemble robustness, scheduling sensitivity |
Hybrid and ensemble models | Feeder load and asset-level aggregation | Congestion warning and transformer loading | Real-time to day-ahead | Overload warning precision, false-alarm rate, missed-event rate |
Probabilistic and quantile models | Active power consumption and uncertainty drivers | Flexibility reserve estimation, peak-risk forecasting | Hour-ahead to day-ahead | Prediction interval coverage, pinball loss, calibration error, sharpness |
Probabilistic and quantile models | Net load, PV generation, storage state proxies | Microgrid scheduling and local energy trading | 15 minutes to day-ahead | Scenario reliability, quantile loss, risk-aware scheduling usefulness |
Federated and distributed models | Locally retained household or site-level smart-meter data | Flexibility aggregation, privacy-preserving microgrid coordination | Hour-ahead to day-ahead | Local model error, cross-site generalisation, communication cost, privacy-preserving utility |
Adaptive online models | Streaming active power and contextual features | Real-time flexibility and drift-sensitive forecasting | Real-time to hour-ahead | Drift response time, online error, model update stability |
Edge-oriented models | Smart-meter and edge sensor measurements | Pre-emptive congestion management, local control | Real-time to hour-ahead | Latency, local forecast error, event detection speed, computational efficiency |
A major gap is the limited evidence that smart-meter forecasting models generalise across households, feeders, microgrids, and congestion-control contexts. Generalisation analysis of short-term load forecasting shows that models trained under one data condition may not transfer reliably to another without recalibration or adaptation [15]. Federated and adaptive forecasting studies address part of this issue, but they also reveal that local data heterogeneity and drift remain persistent barriers to reusable model deployment [4, 21]. Future research should therefore test model transferability across customer groups, network assets, climate zones, and application domains rather than reporting performance in isolated datasets.
Real-time and edge implementation remain underdeveloped relative to the number of studies proposing high-capacity forecasting architectures. Edge sensing data-imaging approaches suggest one path toward local processing, but the literature still provides limited evidence on latency, communication cost, model updating, and operator integration under real grid conditions [25]. Online adaptive recurrent forecasting shows that models can be updated as new smart-meter data arrives, yet practical deployment requires governance over drift detection, retraining frequency, and failure handling [7]. This gap is especially important for congestion control and microgrid resilience, where delayed or unstable forecasts may reduce operational usefulness.
Benchmarking and reproducibility are persistent weaknesses because smart-meter forecasting studies often differ in datasets, aggregation levels, train-test splits, error metrics, and forecast horizons. Reviews of load forecasting techniques in smart grids emphasise that inconsistent evaluation practices make it difficult to compare models fairly across studies [24, 26]. Probabilistic forecasting adds further complexity because calibration, sharpness, and risk-oriented metrics must be considered alongside conventional point-forecast errors [8, 13]. Table 6 maps the main research gaps to future research directions and expected impacts for smart-meter forecasting taxonomy development.
Table 6. Research Gaps and Future Directions
Gap | Description | Potential Research Direction | Expected Impact | Priority |
Cross-domain generalisation | Models are often tested on narrow datasets and may not transfer across households, feeders, or microgrids | Multi-site benchmarking across customer groups, climates, aggregation levels, and grid assets | More reliable model selection for utilities and aggregators | High |
Real-time deployment | Many models are evaluated offline without latency, streaming, or operational-control constraints | Edge-ready forecasting pipelines with latency reporting and online model monitoring | Improved suitability for congestion control and fast flexibility activation | High |
Privacy-preserving forecasting | Customer-level smart-meter data is sensitive, limiting centralised model training | Federated, split-learning, and privacy-preserving forecasting architectures | Wider data access without exposing raw customer profiles | High |
Probabilistic calibration | Uncertainty forecasts are not always calibrated for operational risk decisions | Standard calibration tests, prediction interval evaluation, and risk-aware metrics | Better flexibility reserve estimation and congestion-risk management | High |
Data-quality handling | Missing values, irregular sampling, and meter faults are handled inconsistently | Shared preprocessing protocols and transparent data-quality reporting | More reproducible and robust forecasting workflows | Medium |
Application-specific benchmarking | Generic forecast errors do not always reflect demand response, microgrid, or congestion-control value | Task-specific metrics such as peak capture, overload warning accuracy, and scheduling sensitivity | Better alignment between model performance and operational usefulness | High |
Interpretability | Deep and hybrid models can be difficult for operators to understand | Interpretable forecasting models, feature-attribution methods, and operator-facing explanations | Increased trust and improved operational adoption | Medium |
Integration with grid models | Many smart-meter forecasts are not connected to topology, power flow, or asset constraints | Coupled forecasting and distribution-network simulation frameworks | Stronger support for voltage management and congestion prevention | High |
Lifecycle management | Few studies address model maintenance, drift, retraining, and degradation after deployment | Forecasting model lifecycle protocols for monitoring, retraining, and governance | More sustainable use of forecasting systems in operational settings | Medium |
This taxonomy review classified smart-meter forecasting models according to model category, data type, application domain, and forecast horizon. The resulting framework shows that forecasting model selection should be driven by the operational decision being supported, not by algorithmic novelty alone. Classical, machine learning, deep learning, hybrid, probabilistic, federated, adaptive, and edge-oriented models each occupy distinct positions in the distribution-grid forecasting landscape. The taxonomy provides a structured way to connect smart-meter data with grid flexibility, microgrid energy management, and congestion-control applications.
The review also highlights important gaps in cross-domain generalisation, real-time deployment, privacy-preserving learning, benchmarking, interpretability, and integration with grid models. Future research should move beyond isolated accuracy reporting toward application-specific evaluation and operational validation. Smart-meter forecasting will become more useful for distribution networks when model development is aligned with data quality, uncertainty, forecast horizon, and decision context. A structured taxonomy can support this transition by guiding both researchers and practitioners toward more transparent and operationally relevant forecasting design.
None
None
None
None
Open Access The author(s) retain copyright. This article is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. It may be shared and adapted for non-commercial purposes with appropriate attribution, an indication of changes, and distribution of adaptations under the same license. Third-party material may be subject to separate terms identified in its credit line. View the license at https://creativecommons.org/licenses/by-nc-sa/4.0/.
Institute of Smart Energy Systems Press
