Complicated weather conditions lead to intermittent, random and volatility in photovoltaic (PV) systems, which makes PV predictions difficult. A recurrent neural network (RNN) is considered to be an effective tool for time-series data prediction. However, when the weather changes intensely, the long-term sequence of multivariate may cause gradient vanishing (exploding) during the training of RNN, leading the prediction results to local optimum. Long short-term memory (LSTM) network is the deep structure of RNN. Due to its special hidden layer unit structure, it can preserve the trend information contained in the longterm sequence, which is allowed to solve the problems of RNN and improve performance. An LSTM-based approach is applied for short-term predictions in this study based on a timescale that encompasses global horizontal irradiance (GHI) one hour in advance and one day in advance. Inaccurate forecasts usually occur on cloudy days, and the results of ANN and SVR in the literature prove this. To improve prediction accuracy on cloudy days, the clearness-index was introduced as an input data for the LSTM model and to classify the type of weather by k-means during the data processing, where cloudy days are classified as the cloudy and the mixed(partially cloudy). NN models are established to compare the accuracy of different approaches and the cross-regional study is to prove whether the method can be generalizable. From the results of hourly forecast, the R 2 coefficient of LSTM on cloudy days and mixed days is exceeding 0.9, while the R 2 of RNN is only 0.70 and 0.79 in Atlanta and Hawaii. From the results of daily forecast, All R 2 on cloudy days is about 0.85. However, the LSTM is still very effective in improving of RNN and more accurate than other models.

Although the output of a photovoltaic power generation system is significantly positively correlated with solar irradiance, the latter variable is intermittent, random, and volatile. Volatility in solar irradiance is particularly marked when weather conditions are complex, and so, this factor has proved to be difficult to predict. A neural network (NN)-based approach is applied for short-term predictions in this study based on a timescale that encompasses the amount of irradiance each hour throughout the next day. Thus, a backpropagation NN (BPNN), a radial basis function NN (RBFNN), and an Elman NN (ENN) were selected for use in this analysis. A predictive model was established to evaluate the accuracy of different approaches, given variable meteorological conditions. To reduce the influence of solar irradiance, samples used for forecasts were subdivided into spring, summer, fall, and winter, and the forecast results of sunny and rainy as well as cloudy days in different seasons were investigated. The results of this study reveal that the predictive accuracies of the BPNN and RBFNN were poor on rainy and cloudy days, while the efficiency of the ENN was high and stable in variable meteorological conditions.