In view of the urgent need to improve the grid connection rate of new energy power generation and the low prediction accuracy of photovoltaic (PV) power generation, a short-term combined forecasting model, namely Temporal Query Kolmogorov-Arnold Network (TQKAN), is proposed by integrating the Maximal Information Coefficient (MIC), Dynamic Time Warping (DTW) and Improved Tianji’s Horse Racing Optimization (ITHRO). Firstly, MIC is used to screen and determine the key meteorological influencing factors of PV output. Secondly, DTW distance is used to determine the historical days with data distribution characteristics similar to the targeted predicted day, in order to generate the suitable training sample set. Next, the optimization performance of Tianji’s Horse Racing Optimization (THRO) algorithm is improved by combining the Good Point Set initialization method and the Experience Exchange Strategy (EES). Finally, the Multilayer Perceptron (MLP) layer of Temporal Query Network (TQNet) is replaced by Kolmogorov-Arnold Networks (KAN), and ITHRO is used to determine the optimal combination of model hyperparameters, leading to the ITHRO-TQKAN combination prediction model. The technical advancement and practical effectiveness of this combined model are verified by the application results on the Australia and Ningxia datasets. The contribution and role of each module in improving the prediction accuracy are fully reflected by a series of ablation experiments, and a good reference is thereby for the construction of high-precision prediction model of PV output.