Short-term photovoltaic output forecasting is essential for the reasonable formulation of power system production scheduling plan, which can promote the grid-connection and consumption of photovoltaic power generation. Nevertheless, the photovoltaic output is greatly affected by meteorological characteristics, and the generation process has the characteristics of volatility, intermittence and uncontrollability, therefore, the rapid and accurate short-term photovoltaic output forecasting has become a big challenge. To tackle the challenge, this paper proposes a TCN-Attention combination forecasting model based on similar day matching. We employ time series shape-based clustering algorithm and maximal information coefficient to characterize the similarity of photovoltaic output. As a result, the data redundancy caused by all historical data as input can be certainly avoided, we use the temporal convolutional network with parallel computing to learn the photovoltaic output characteristics, and introduce Attention mechanism to highlight the influence of key meteorological characteristics. Thus, the training speed and prediction accuracy of the model can be certainly improved. The results of the conducted experiment based on real data show that the proposed method has the following advantages of direct information extraction, fast training speed and high prediction accuracy.