To improve the short-term prediction accuracy of photovoltaic power generation models with multiple input features, a photovoltaic power prediction ensemble model LGGWO-TCN-MHSA based on optimizing TCN hyperparameters was proposed. The model integrated the levy gold grey wolf optimization (LGGWO), temporal convolutional network (TCN), and multi-head self-attention mechanism (MHSA). First, the Spearman correlation coefficient method extracted the main features that significantly affect photovoltaic power, which were then fed into the TCN prediction model. Then, the proposed multi-strategy LGGWO was applied to the TCN for hyperparameter optimization, which improved the model's prediction performance. Finally, the predicted values were input into the multi-head self-attention model to further boost prediction accuracy. The experiment was verified using original Australian photovoltaic data. By comparing with six groups of models including convolutional neural networks (CNN) and long short-term memory neural networks (LSTM), the mean absolute error (MAE) and root mean square error (RMSE) of the proposed model on the test data set were reduced by 2.03% ~ 82.0% and 10.5% ~ 80.1%, respectively. The results show that the proposed method has high prediction accuracy and good stability.