This paper proposes an improved BP neural network prediction model based on hybrid genetic ant colony algorithm （GA－ACO）. Through Pearson correlation coefficient formula， meteorological features with strong correlation with photovoltaic power output are calculated as the input of the training model to reduce the influence of irrelevant meteorological features on the forecast of photovoltaic power output. The genetic algorithm（GA） is used to generate the pheromone distribution in search of the optimal parameter. The ant colony algorithm（ACO） outputs the optimal weight threshold under the condition of the initial pheromone distribution， and the BP neural network is trained twice to output the predicted value. The analysis results show that： taking sunny days as an example， the relative error of GA－ACO－BP neural network model is reduced by 9.47， 4.83 and 3.27 percentage point respectively， compared with the traditional BP neural network model， ACO－BP neural network model and GA-BP neural network model. Therefore， GA－ACO－BP neural network model has better prediction accuracy when applied to photovoltaic power generation prediction.