The integration of solar photovoltaic systems into microgrids introduces substantial variability in power generation, making accurate load-profile modelling important for reliable operation and storage sizing. However, long-term, high-resolution load data are often unavailable for new or remote microgrid installations. This study presents a systematic benchmark of four approaches—R-GAN, VAE, ARIMA, and Statistical Sampling—for generating 24-hour load profiles under a small-data regime. The methods are evaluated using eight complementary metrics covering distributional fidelity, statistical consistency, temporal correlation, peak preservation, and variance preservation. Results averaged over three independent runs show that Statistical Sampling provides the best mean performance on seven of the eight metrics, whereas R-GAN achieves the highest mean lag-1 autocorrelation, with a value of 0.620 ± 0.028 compared with 0.610 for Statistical Sampling. R-GAN generally outperforms VAE across the evaluated metrics, while ARIMA provides strong variance preservation. These results demonstrate that classical statistical approaches can remain highly competitive with deep generative models when training data are limited. The findings further show that pooled distributional metrics may favour empirical-marginal resampling without necessarily indicating that temporal dependencies have been learned. Therefore, the selection of synthetic load-generation methods should be guided by the statistical characteristics required for the intended microgrid planning application.