Developing accurate forecasting models is essential across various domains, yet the performance of machine learning models is often hindered by insufficient, limited, or missing data streams. Existing models frequently struggle to manage diverse data streams and capture complex temporal patterns. To solve these, we develop advanced data augmentation techniques that vary noise parameters to generate diverse set of data streams. We also establish a machine learning-based data forecasting pipeline utilizing Fast Fourier Transform (FFT) to produce synthetic frequency-domain data. We then enhance predictive accuracy by creating optimized hybrid machine learning models. Finally, we develop a Generative dynamic adaptive data augmentation (G- DADA) framework to support unsupervised data augmentation across a range of machine learning models for multiple data streams.