Forecasting non-linear time series, such as disease progression and outbreak metrics, presents significant challenges due to noise, complex relationships, and non-stationary attributes, especially in the context of limited and geographically specific datasets. These challenges restrict the development of accurate predictive models for diverse regions, such as in COVID-19 forecasting. To address these issues, this paper introduces an AI-driven framework that integrates deep transfer learning with Long Short-Term Memory (LSTM) models, including multilayer CNN-LSTM and encoder-decoder sequences, for biweekly forecasting of COVID-19 cases and deaths. The proposed framework employs a two-stage process. First, LSTM architectures, utilizing a sliding window technique, capture temporal dependencies to establish a base model trained on COVID-19 data from a source country. In the second stage, transfer learning refines the model by fine-tuning it with data from a target country, adapting the model to local trends, and improving forecasting accuracy. This approach minimizes the need for extensive target country data by leveraging insights from source models. Experimental results using COVID-19 data from eleven countries demonstrated significant improvements in forecast accuracy. Transfer learning models pre-trained on Spain and Belgium COVID-19 datasets achieved substantial reductions in MAPE: the Spain-based model improved predictions for biweekly cases, while the Belgium model excelled in forecasting biweekly deaths. However, pre-trained models from certain countries, such as India, showed limited effectiveness. This framework provides a sustainable, scalable, efficient, and robust solution for disease forecasting, particularly valuable for public health decision-making in regions with limited data. It is adaptable to other health domains and urban management applications.