Accurate prediction of output power of photovoltaic power plant is one of the main ways to promote grid connected photovoltaic power generation and improve the stability of power grid operation. This paper proposed a day-ahead photovoltaic power forecasting method based on particle swarm optimization and least squares support vector machine (PSO-LSSVM). This method used the global search ability of PSO algorithm to obtain the penalty factor and kernel function width of LSSVM, which effectively solved the problem that LSSVM is difficult to find the optimal parameters quickly and accurately. Then, the historical data of numerical weather prediction (NWP) and photovoltaic power were used to train the PSO-LSSVM model, and the trained PSO-LSSVM model was used to forecast the day-ahead photovoltaic power. The forecasting results of PSO-LSSVM, long short-term memory(LSTM), least squares support vector machine(LSSVM) and PSO-BP models were compared and analyzed, and the comparison results show that the forecasting accuracy of PSO-LSSVM model is higher than those of LSTM, LSSVM and particle swarm optimization and back propagation (PSO-BP) models, which proves the superiority of the PSO-LSSVM forecasting model.