Time series refer to a series of data points indexed in time order, which can be found in various fields, e.g., transportation, healthcare, and finance. Accurate time series forecasting can enhance optimization planning and decision-making support. Time series have multi-scale characteristics, i.e., different temporal patterns at different scales, which presents a challenge for time series forecasting. In this paper, we propose TPRNN, a Top-down Pyramidal Recurrent Neural Network for time series forecasting. We first construct subsequences of different scales from the input, forming a pyramid structure. Then by executing a multi-scale information interaction module from top to bottom, we model both the temporal dependencies of each scale and the influences of subsequences of different scales, resulting in a complete modeling of multi-scale temporal patterns in time series. Experiments on seven real-world datasets demonstrate that TPRNN has achieved the state-of-the-art performance with an average improvement of 8.13% in MSE compared to the best baseline.