Time series forecasting has become a dynamic research field that has garnered significant attention from researchers in recent years. Its practical significance lies in enabling stakeholders to make proactive interventions and facilitate informed decision-making for individuals and society by analyzing and predicting various time series data. In this paper, we propose a time series forecasting model named CNN-LSTM Attention DeepAR which designed to predict either points or probability distributions for time series. In this model, we first employ CNN-LSTM to capture multi-scale information from the time series. Subsequently, a position-based attention mechanism is utilized to capture local dependencies among variables. The parameters are estimated through training to infer the distribution parameters, and final predictions are obtained by sampling from the distribution. We conduct experiments on three different datasets from various domains to demonstrate the effectiveness of the proposed approach in time series forecasting.