Time series forecasting for hourly photovoltaic power using conditional generative adversarial network and Bi-LSTM

More and more photovoltaic (PV) power generation is incorporated into the grid. However, the intermittence and fluctuation of solar energy have brought huge challenges to the safe and stable operation of the power grid. PV power forecasting is one of the effective ways to solve the above problems, so it has become an important research topic. However, the existing research based on deep learning models mainly focuses on more complex network structures, optimization algorithms, and data decomposition. These hybrid models have encountered a development bottleneck in extracting the inherent features of PV power and related data, and a new idea and method are needed. This paper proposes a novel TSF-CGANs (time series forecasting based on CGANs, TSF-CGANs) algorithm considering conditional generative adversarial networks (CGANs) combined with convolutional neural networks (CNN) and Bi-directional long short-term memory (Bi-LSTM) for improving the accuracy of hourly PV power prediction. We design the generator in the TSF-CGANs network as a regression prediction model, which can extract the features based on historical data and random noise vector by the complex models, and finally use the Bi-LSTM model to output the predicted value. At the same time, the discriminator judges the authenticity of the generated predicted value and the actual value. In the continuous game between the generator and the discriminator, the parameters of the generator are optimized and more accurate prediction results are obtained. The performance of the proposed method is demonstrated with a real-world dataset. Compared with LSTM, recurrent neural network (RNN), back-propagation neural network (BP), support vector machine (SVM), and Persistence models, the values of five performance evaluation indicators, RMSE, MAE, nRMSE, R2, and R, show that the proposed model has better performance in prediction accuracy. Compared with the traditional BP, the TSF-CGANs model reduced the RMSE by 32%, Compared with the Persistence, the forecast skill (FS) of TSF-CGANs is 0.4863. The results indicate that it is feasible to use the generator to realize time series prediction in the proposed TSF-CGANs network. The core idea of TSF-CGANs method is to improve the prediction accuracy of the generator through the continuous game between the generator and the discriminator, which provides a new idea for the training process of the prediction method based on deep learning.