Accurate forecasting of photovoltaic power can effectively promote safe and efficient generation and utilization of photovoltaic power. Accordingly, an ultra-short-term photovoltaic power prediction method combining singular spectrum decomposition (SSD), double-attention mechanism, and bidirectional gating logic unit (BiGRU) time-series modeling is proposed to address the insufficient forecasting accuracy of existing methods. First, SSD is used to reduce the randomness and volatility of photovoltaic signals. A BiGRU network is then adopted to model the time series of the decomposed signals. Additionally, an attention module is designed to simultaneously learn the importance (weight) of the feature and time series by weighting the features extracted by the BiGRU network. The final forecast of photovoltaic power is obtained via the decision-making layer. The experimental results demonstrate that the SSD and attention mechanism can improve the accuracy of forecasts obtained from the deep time-series model. The proposed method is superior to several other conventional methods and is highly practical for different seasons and weather conditions.