Synthetic Load Profile Generation for Residential and Commercial Loads: A Comparative Study of Stochastic Models

Synthetic load profiles are essential for distribution network planning, protection sizing, and demand-side management studies. However, most generation methods are validated on a single load typology, and their transferability remains unexamined. Two broad paradigms dominate the generation literature: data-driven approaches that learn the demand distribution directly from historical records (generative adversarial networks, diffusion models, Markov-chain generators) and bottom-up, physically motivated approaches that reconstruct demand from the superposition of discrete appliance ON/OFF events. Same-data, same-metric comparisons across these two families for structurally distinct load typologies remain absent from the literature, and this gap is the one this paper addresses. This paper presents a systematic comparison of three stochastic models (a first-order autoregressive (AR(1)) profile, a physically constrained ON/OFF event model, and a nonlinear-least-squares (NLS) calibrated variant) applied to two fundamentally different load typologies measured with a Class A power-quality recorder at 10-minute resolution: a single-family residential dwelling (13 days, 1860 samples) and an institutional commercial building (9 days, 1333 samples). Evaluation spans six distributional statistics (mean, standard deviation, and the percentiles p50, p90, p95 and p99), the Kolmogorov–Smirnov (KS) statistic, root mean square error (RMSE), and hourly variance profiles. In this two-site study, load typology, not model sophistication, emerges as the dominant factor shaping fit quality. The simple AR(1) reproduces all percentiles within 7% for the near-Gaussian commercial load (skewness 3.47). By contrast, no model reproduces the centre, the dispersion and the upper tail of the highly skewed residential load (skewness 6.56) simultaneously to within 10%. On the residential tail the AR(1) ensemble is the least biased (p99: −5.8%) but by far the most dispersed across realizations (CV = 15.3%), whereas the event-based models are more stable but biased, so model choice on skewed loads is a bias–stability trade-off rather than an accuracy ranking. The NLS-calibrated model attains near-exact residential p95 reproduction (−0.3%) and commercial upper-tail errors below 2%, at a computational cost roughly three orders of magnitude above the AR(1). The dynamic characterization of demand obtained from these models, including the magnitude and frequency of the detected events, provides elements that may be of interest for the sizing and operation of photovoltaic systems in the context considered. Based on these two cases, a preliminary recommendation matrix mapping models to engineering applications is proposed, motivating typology-specific model selection rather than universal approaches; broader validation on additional sites is identified as future work.