Due to the existing large-scale grid-connected photovoltaic (PV) power generation installations, accurate PV power forecasting is critical to the safe and economical operation of electric power systems. In this study, a hybrid short-term forecasting method based on the Variational Mode Decomposition (VMD) technique, the Deep Belief Network (DBN) and the Auto-Regressive Moving Average Model (ARMA) is proposed to deal with the problem of forecasting accuracy. The DBN model combines a forward unsupervised greedy layer-by-layer training algorithm with a reverse Back-Projection (BP) fine-tuning algorithm, making full use of feature extraction advantages of the deep architecture and showing good performance in generalized predictive analysis. To better analyze the time series of historical data, VMD decomposes time series data into an ensemble of components with different frequencies; this improves the shortcomings of decomposition from Empirical Mode Decomposition (EMD) and Ensemble Empirical Mode Decomposition (EEMD) processes. Classification is achieved via the spectrum characteristics of modal components, the high-frequency Intrinsic Mode Functions (IMFs) components are predicted using the DBN, and the low-frequency IMFs components are predicted using the ARMA. Eventually, the forecasting result is generated by reconstructing the predicted component values. To demonstrate the effectiveness of the proposed method, it is tested based on the practical information of PV power generation data from a real case study in Yunnan. The proposed approach is compared, respectively, with the single prediction models and the decomposition-combined prediction models. The evaluation of the forecasting performance is carried out with the normalized absolute average error, normalized root-mean-square error and Hill inequality coefficient; the results are subsequently compared with real-world scenarios. The proposed approach outperforms the single prediction models and the combined forecasting methods, demonstrating its favorable accuracy and reliability.

Compared with the point prediction, the interval prediction of the load could more effectively guarantee the safe operation of the power system. In view of the problem that the correlation between adjacent load data is not fully utilized so that the prediction accuracy is reduced, this paper proposes the conditional copula function interval prediction method, which could make full use of the correlation relationship between adjacent load data so as to obtain the interval prediction result. At the same time, there are the different prediction results of the method under different parameters, and the evaluation results of the two accuracy evaluation indicators containing PICP (prediction interval coverage probability) and the PIAW (prediction interval average width) are inconsistent, the above result that the optimal parameters and prediction results cannot be obtained, therefore, the NSGA-II (Non-dominated Sorting Genetic Algorithm-II) multi-objective optimization algorithm is proposed to seek out the optimal solution set, and by evaluating the solution set, obtain the optimal prediction model parameters and the corresponding prediction results. Finally, the proposed method is applied to the three regions of Shaanxi Province, China to conduct ultra-short-term load prediction, and compare it with the commonly used load interval prediction method such as Gaussian process regression (GPR) algorithm, artificial neural network (ANN), extreme learning machine (ELM) and others, and the results show that the proposed method always has better prediction accuracy when applying it to different regions.