Standard Load Profiles (SLP)

• Standard Load Profiles (SLPs) are predefined, normalized electricity demand curves that capture the average consumption behavior of customer groups over specific time intervals.
• They are constructed through methods such as multiplicative decomposition and probabilistic modeling to incorporate seasonality, day-types, and stochastic variations.
• Recent research emphasizes updating legacy SLPs with modern smart meter data and advanced data-driven techniques to accurately reflect evolving consumption patterns.

Standard Load Profiles (SLPs) are predefined, typical electricity demand curves that represent the average consumption behavior of a customer group and are commonly used when interval metering is unavailable or when planners require normalized reference demand shapes. In practice, an SLP is usually defined over a year, day, or week at hourly or 15‑minute resolution, then scaled by annual energy consumption; in the German formulation summarized in recent work, a normalized profile $p(t)$ yields customer load through $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$ (Gabrielski et al., 31 Jul 2025). Within the literature, the term retains this regulatory and operational meaning, but some recent work also extends it toward probabilistic and synthetic profile libraries that preserve the statistical characteristics of real demand rather than a single deterministic shape (Hu et al., 2022).

1. Definition, scope, and operational role

SLPs are used to reconstruct time-resolved load for billing and settlement of non-metered or simply metered customers, support grid planning and operation, provide baseline time series for forecasting, and perform scenario analyses (Gabrielski et al., 31 Jul 2025). Their practical value lies in compressing a large population’s consumption behavior into a small set of standardized curves that can be combined with annual energy, customer class, season, and day type.

In German commercial practice, the standardized VDEW profile system is defined at 15‑minute resolution and includes one household profile, seven commercial building profiles, and three agricultural profiles; each profile set is subdivided into nine daily curves formed by the Cartesian product of three day types—weekday, Saturday, Sunday/holiday—and three seasons—winter, summer, transition (Steens et al., 2020). In Great Britain, typical load profiles are described in closely related terms, but later work notes that such profiles are generally seasonal, not weather sensitive, and lack the diversity needed to aggregate a cohort of premises into a plausible LV substation load (Brash et al., 13 Oct 2025).

The literature also distinguishes between deterministic SLPs and richer probabilistic generalizations. A deterministic SLP specifies one canonical normalized curve for a class. A probabilistic SLP library, by contrast, describes a distribution over plausible curves for a class, season, or feeder. This distinction has become increasingly important as renewables, inverter-dominated grids, rooftop PV, EV charging, and local balancing constraints make short-term variability and coincident peaks operationally material (Anvari et al., 2020).

2. Canonical construction frameworks

The classic German residential SLP construction summarized in recent work follows a multiplicative decomposition. Each measured annual residential time series is first scaled to $1000$ kWh/year, then aggregated to obtain a representative normalized profile. A polynomial “dynamisation factor” $d(t)$ is fitted to daily energy consumption over the year to represent annual seasonality. Daily profiles are then estimated for each combination of season—winter, summer, transition—and weekday type—workday, Saturday, Sunday—after dividing each day by $d(t)$ to avoid double-counting annual seasonality. Recombination yields a continuous annual SLP of the form

$$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$

In the original BDEW convention, national holidays are treated as Sundays (Gabrielski et al., 31 Jul 2025).

The widely used German H0 residential SLP is a specific instance of this general logic. It is a static, generic 15‑minute profile derived from representative measurements collected mostly in the 1970s–1990s, with only $10\%$ of the source data at 15‑minute resolution and about $332$ households aggregated (Anvari et al., 2020). Its intended use is broad: energy system planning, grid operation and network planning, and tariff design and settlement for customers without interval meters.

Commercial building practice in Germany uses VDEW standardized profiles. The paper on forecast-based building load management describes these as having been developed from measured load data of $1209$ different buildings and defines seasonal splits of summer $15/05$–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$0, transition $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$1–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$2 and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$3–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$4, and winter $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$5–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$6 (Steens et al., 2020). A specific commercial example is the G1 profile for buildings with working hours roughly $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$7–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$8.

The canonical construction therefore rests on three structural assumptions: customer classes can be represented by a small set of typical curves; annual energy scaling is sufficient to transfer those curves to individual customers; and season/day-type segmentation captures the dominant regularities. Much of the recent literature tests, relaxes, or replaces these assumptions.

3. Empirical reassessment of legacy assumptions

A central recent result is that German residential SLPs based on data more than $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$9 years old no longer match current consumption patterns. Using smart meter data from the openMeter platform covering more than $1000$0 residential consumers, mainly from North Rhine‑Westphalia, Baden‑Württemberg, Bavaria, Hesse, and Lower Saxony, with measurements from $1000$1 onward and most data from $1000$2–$1000$3, updated residential SLPs show three systematic deviations from the legacy BDEW curves: a less pronounced morning peak, reduced day–night contrast, and a reduced midday peak; the yearly dynamisation factor also exhibits lower seasonal amplitude (Gabrielski et al., 31 Jul 2025). The paper attributes these shifts to behavioral and technological changes, including more flexible working hours, higher female labor participation, reduced use of electric water heaters, more efficient appliances, and the growth of stand-by loads and smart-home devices. Lockdown-period profiles were found to differ substantially from non-lockdown profiles and were therefore excluded from SLP construction.

The same study also tested foundational assumptions rather than discarding them wholesale. Three weekday types remain statistically meaningful: a Random Forest classifier achieved mean accuracy of $1000$4 for distinguishing workday, Saturday, and Sunday, and national holidays mostly resembled Sundays. Three seasons also remain structurally meaningful: time-series k-means clustering on average daily profiles selected three clusters via the silhouette score, mapping clearly to winter, summer, and transition. The proposed refinements are instead local: treat New Year’s Eve as Sunday rather than Saturday, replace abrupt season boundaries by a $1000$5-day linear transition window, and smooth daily SLP curves with a Savitzky–Golay filter. On the normalized $1000$6 kWh/year basis, the reproduced BDEW structure on new data yielded $1000$7 kW per 15‑minute step, the enhanced SLP yielded $1000$8 kW, and the Fourier-based alternative yielded $1000$9 kW (Gabrielski et al., 31 Jul 2025).

At the single-building level, generic standardized profiles can be substantially less accurate than building-specific profile averages. In a German commercial-building study using 5‑minute forecasts, the generic SLP achieved $d(t)$0 kW, $d(t)$1 kW, and $d(t)$2, whereas the personalized standardized load profile (PSLP) achieved $d(t)$3 kW and deep models achieved $d(t)$4 kW for FFNN and $d(t)$5 kW for LSTM (Steens et al., 2020). This does not invalidate SLPs as a class, but it shows that generic profiles are often outperformed by local re-estimation when building behavior is available.

A common misconception is that aggregation alone renders detailed residential dynamics irrelevant. The recent reassessment literature does not support that view. The empirical picture is instead that many legacy structural ideas remain serviceable—seasonality, day-type segmentation, annual scaling—but require recalibration, smoother transitions, and explicit treatment of modern behavioral drift.

4. Data-driven decomposition and stochastic alternatives

A more radical reformulation treats residential demand as the sum of a deterministic trend and stochastic fluctuations rather than as a single smooth profile. Using highly resolved Austrian and German household data from the ADRES and NOVAREF projects, one study constructs an average load profile (ALP) by first averaging four consecutive gap-free weeks and then applying Empirical Mode Decomposition (EMD):

$d(t)$6

Low-frequency intrinsic mode functions form the trend, while high-frequency modes represent fluctuations. For NOVAREF, the decomposition uses $d(t)$7 modes, treats modes $d(t)$8–$d(t)$9 as high-frequency and $d(t)$0–$d(t)$1 as low-frequency, and selects the number of retained low-frequency modes by validation-set MSE. The optimum was $d(t)$2, and $d(t)$3 for all $d(t)$4 (Anvari et al., 2020).

The same work decomposes measured power as

$d(t)$5

The empirical fluctuation term is strongly intermittent, skewed, and heavy-tailed. Its global probability density is fit much better by a $d(t)$6-Maxwell–Boltzmann distribution than by a Gaussian, and a Gaussian severely underestimates large positive deviations. Superstatistical analysis identifies a long time scale $d(t)$7 s over which local fluctuation statistics are approximately stationary and a short relaxation time $d(t)$8–$d(t)$9 s, implying $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$0. Within each window of length $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$1, the fluctuation histogram is well fit by a Maxwell–Boltzmann distribution with window-specific $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$2 and $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$3 (Anvari et al., 2020).

To reproduce these features, the paper introduces a superstatistical stochastic process built from $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$4 Ornstein–Uhlenbeck components,

$$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$5

and defines

$$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$6

For $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$7, the fluctuation magnitude is exactly Maxwell–Boltzmann distributed locally, while slow variation of $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$8 and $$X(t) = d(t)\cdot \sum_{i=1}^{3}\sum_{j=1}^{3}\bigl(c_i^{wd}(t)\cdot c_j^s(t)\cdot s_{wd,s}(t)\bigr)+\epsilon(t).$$9 across windows reproduces the global heavy-tailed statistics. In this framework, the deterministic ALP plus a stochastic fluctuation profile yields a data-driven load profile that reproduces both mean behavior and the full consumption histogram, especially at large consumption values. The paper’s diversity-factor analysis also shows that coincident peaks remain significant even after aggregation: for $10\%$0 or $10\%$1 households, the ratio of coincident to non-coincident demand can still reach $10\%$2 at some times (Anvari et al., 2020).

5. Probabilistic, generative, and coherent synthetic profiles

A major contemporary development is the reinterpretation of SLPs as distributions of plausible load shapes rather than one deterministic curve. At transmission level, a conditional GAN was trained on $10\%$3 normalized week-long profiles of hourly-sampled bus load, each a vector in $10\%$4, derived from PMU measurements at $10\%$5 samples/s for $10\%$6 buses over $10\%$7–$10\%$8. Conditioning uses a six-dimensional label vector comprising four seasons and two load types—mainly industrial and mainly residential/commercial—so the model learns $10\%$9 (Pinceti et al., 2021). Validation is explicitly SLP-relevant: Wasserstein distance between real and generated distributions decreases toward zero during training; PSDs almost perfectly match, including the dominant daily peak at approximately $332$0, corresponding to $332$1 hours; and LSTM forecasting errors on synthetic and real data are very similar. In the summer residential case, training on synthetic data yielded mean forecast error $332$2 on synthetic test data and $332$3 on real test data; for fall residential, the corresponding values were $332$4 and $332$5. When mapped onto the $332$6-bus Polish system, AC OPF converged for all $332$7 hours in winter and summer weeks with voltages and generator outputs within limits (Pinceti et al., 2021).

At distribution-transformer level, MultiLoad‑GAN addresses a limitation of single-profile generators by synthesizing a group of $332$8 customer profiles simultaneously over a week at 15‑minute resolution, $332$9 time steps, thereby preserving spatial–temporal correlations among loads served by the same transformer (Hu et al., 2022). The generator operates on a $1209$0 representation in which three channels encode load magnitude and one encodes temperature. The paper reports that MultiLoad‑GAN generates more realistic load profiles than existing approaches, especially in group-level characteristics, and that classifier-based realism sharply differentiates joint generation from independent generation: in one experiment the Percentage of Real for MultiLoad‑GAN was $1209$1, compared with $1209$2 for SingleLoad‑GAN (Hu et al., 2022).

At LV substation level, conditional diffusion models are proposed precisely because typical load profiles are described as seasonal, not weather sensitive, and lacking diversity. In one UK study on $1209$3 monitored LV substations, the strongly conditioned diffusion model LVGenWCS, which conditions on weather, calendar, substation metadata, and daily minimum/mean/maximum active and reactive power, achieved $1209$4, $1209$5, Marginal Score $1209$6, MiVo $1209$7, and Wasserstein Distance $1209$8 on unseen substations. In a UKGDS $1209$9-bus network case study, it also produced the best power-flow realism, with bus-voltage MAE $15/05$0 V and phase-angle MAE $15/05$1 (Brash et al., 13 Oct 2025).

A related line of work at MV level identifies a geometric latent structure in standardized daily profiles. PCA of Dutch MV daily profiles showed that three principal components capture about $15/05$2 of variance and that the resulting latent points lie on a thin spherical shell with an arc structure. A principal curve on that sphere defines a continuous ordering of profiles, while a von Mises–Fisher model on the sphere generates continuous mixtures between clusters (Duque et al., 2024). This suggests a move from discrete SLP classes toward continuously parameterized profile families.

6. Inference from aggregated, incomplete, or anonymized data

One response to sparse or privacy-constrained data is to infer SLPs from aggregated measurements. A functional data model for UK substations treats aggregated load as a sum of class-specific typical curves weighted by known customer counts,

$15/05$3

or, with temperature,

$15/05$4

Applied to $15/05$5 Welsh substations with unrestricted domestic (C1) and Economy 7 (C2) customers, the model recovered a C1 curve with the standard UK evening peak and a C2 curve with a strong post‑midnight peak. Model-based clustering then identified two- and three-cluster SLP families with different class-specific mean curves and covariance structures, and BIC favored three clusters (Franco et al., 2021). This framework is directly aligned with SLP estimation when customer-level interval data are unavailable.

Another response is to reconstruct or maintain profile libraries despite missing intervals. BERT‑PIN treats load and temperature profiles as token sequences and performs profile inpainting on 15‑minute feeder-level data aggregated from $15/05$6 smart meters. In demand-response baseline estimation, it outperformed stacked autoencoders, LSTMs, and a prior GAN-based inpainting method. For single-segment peak masking, BERT‑PIN achieved $15/05$7, $15/05$8, $15/05$9, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$00, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$01, and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$02, with improvements of up to $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$03 in energy error over the best benchmark (Hu et al., 2023). In SLP terms, this is a maintenance technology for profile libraries derived from imperfect AMI data.

Privacy-preserving derivation of SLP-like profiles has also become a formal research problem. On $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$04 Low Carbon London households with half-hourly data, microaggregation by MDAV produced anonymized group-average time series that preserve aggregated forecast utility across a wide range of group sizes $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$05. The paper reports that at aggregated forecast level, microaggregated data does not significantly compromise accuracy; for example, N‑BEATS at $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$06 achieved $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$07, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$08, and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$09, and information loss and volatility exhibited a practical plateau around $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$10 (Fernandez et al., 8 Jan 2025). This makes cluster-average profiles viable as privacy-compliant SLP surrogates.

When only monthly integrals and a few type-day profiles are available, high-resolution yearly SLPs can be synthesized by combining a 12-periodic Fourier interpolation $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$11, seasonally morphed type-day patterns $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$12, and a smooth scaling factor

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$13

The resulting synthetic profile preserves monthly integrals while retaining daily and weekly structure (Schnake et al., 2019).

7. Bottom-up, behavioral, and end-use grounded SLPs

A parallel tradition derives SLPs from appliance ownership, end-use physics, and occupancy behavior. In Singapore public housing, a bottom-up stochastic appliance model was used to generate daily profiles for $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$14–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$15, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$16, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$17, and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$18-room flats. Appliance starting probability is modeled as

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$19

and household load is

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$20

The resulting monthly energies were close to Energy Market Authority statistics: $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$21 versus $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$22 kWh for $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$23–$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$24 room flats, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$25 versus $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$26 kWh for $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$27-room, $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$28 versus $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$29 kWh for $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$30-room, and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$31 versus $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$32 kWh for $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$33-room flats (Chuan et al., 2015). These are effectively dwelling-type SLPs for a tropical urban environment.

A Saudi hybrid top-down/bottom-up study combined SCADA-derived residential daily and annual curves with appliance statistics and time-of-use assumptions to estimate seasonal end-use shares. Device energy was expressed as

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$34

The resulting composition showed that in summer air conditioning accounts for about $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$35 of residential energy, while in winter water and ambient heating together account for about $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$36; lighting accounted for about $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$37 in summer and $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$38 in winter (Alahmed et al., 2020). Such decompositions are directly relevant when SLPs are used for tariff design or demand-side management rather than only settlement.

For the UK, a statistical appliance review assembled the parameter library needed for appliance-based SLP construction: ownership rates, probability distributions of appliance power demand, and ZIP or polynomial load models including reactive power. It classifies appliances into lighting, resistive loads, switch-mode power supply loads with and without power-factor correction, directly connected motors, and drive-controlled motors, and gives voltage-dependent active and reactive power models

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$39

The paper is explicit that its data are intended as a resource for the development of load profiles for power system analysis (Tsagarakis et al., 2013).

A more recent behavioral-population framework goes further by integrating appliance physics with human activity models calibrated to ATUS and population composition from CPS. For household $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$40, total load is

$$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$41

and aggregated group load is $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$42. Using fixed physical house parameters and identical weather for California and Texas samples, the study still found significant profile differences by income and locality: in California the highest-income bracket reached up to $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$43 above the mean California load during peak hours, while the lowest bracket was up to $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$44 below; in Texas the $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$45 bracket was up to $$X_{\text{customer}}(t)=\frac{E_{\text{year}}}{1000\,\text{kWh}}\,p(t)$$46 above the mean Texas load (Bromley-Dulfano et al., 2022). This suggests that future SLP systems may be segmented not only by tariff or building type, but by population composition and behavior.

Taken together, these bottom-up and behavioral strands indicate an important shift in the meaning of “standard.” The standard profile is no longer necessarily a static national table. It may instead be a reproducible model output: local, periodically recalibrated, covariate-conditioned, uncertainty-aware, and in some cases explicitly stochastic. That broader conception is consistent with the recent recommendation to treat SLPs not as fixed curves, but as the output of an ongoing data-driven process (Gabrielski et al., 31 Jul 2025).