Accurate forecasting of photovoltaic (PV) power generation is important to enhance the efficiency and reliability of renewable energy in the present power grids. Despite significant progress in predicting PV generation, the nature of PV data remains heterogeneous, non-linear, and dynamic, making modelling and predictive ability challenging in their own right. This dissertation proposes a self-adaptive hybrid approach that offers a novel development avenue within the renewable energy forecast paradigm. It utilizes a Nonparametric Dirichlet Process Gaussian Mixture Model (DPGMM) to identify the distinct predictive capabilities of machine-learning algorithms from a validation perspective, thereby informing their employment in the hybrid method. DPGMM was able to identify homogeneous subgroups from a ~3-year PV dataset of 300 households in Sydney, Australia, allowing the development of distinct predictive models for each subgroup. From a statistical validation perspective, contrast testing was completed using a Friedman and Wilcoxon signed-rank procedure. While there exist differences between the models (p < 0.05), it was clear that Gradient Boosting Regression (GBR) outperformed the support vector regression (SVR) and random forest regression (RFR) models on all clusters. Finally, metrics of the mean absolute error (MAE), mean square error (MSE) and root mean square error (RMSE) were employed to examine and measure improvements in accuracy compared to baseline methods. The hybrid approach yielded statistically significant improvement, which included increases of 12% for MAE, 22% for MSE, 15% for RMSE compared against single-stage convolution implementations.