With the growing demand for clean energy, the world's installed solar energy capacity has increased substantially in the last few years. But the power output of photovoltaic (PV) panels varies greatly under different weather conditions. To improve PV power stations' prediction accuracy, this paper designs a forecasting system that uses artificial neural networks (ANNs) optimized by Levenberg-Marquardt (LM) algorithm based on weather classification for 1-day ahead hourly forecasting. In this process, first, the weather is divided into three types: A - sunny, B - cloudy and C - rainy by using support vector machine (SVM) method. Then, the correlation between meteorological factors and PV power output is analyzed through Pearson correlation coefficient (PCC) method in order to select the forecasting model's input. Finally, the corresponding LM-ANN forecasting sub-models are established under each weather type. After applying the trained three sub-models to a small PV power plant in Macao, the results prove that the proposed forecasting system achieves better prediction effect than traditional backpropagation ANN model.