A novel hybrid model for multi-step ahead photovoltaic power prediction based on conditional time series generative adversarial networks

Abstract

The accuracy of photovoltaic power forecasting is crucial to the revenue of new energy generation projects in electricity market trading. However, due to the highly stochastic volatility and intermittent characteristics of photovoltaic power, it is difficult to construct high-performance photovoltaic power forecasting models. In this study, a multi-step short-term hybrid prediction model of photovoltaic power is proposed, which combines an improved sparrow search algorithm, Fuzzy c-means algorithm (FCM), improved complete ensemble empirical mode decomposition with adaptive noise (ICCEMDAN), and conditional time series generative adversarial networks (CTGAN). First, a new data clustering method based on FCM and data envelope theory is proposed to divide the dataset based on similar power patterns, and the parameters are optimized by an improved sparrow search algorithm. Second, the grey relational analysis and feature creation are combined to determine the optimal similar day for the forecasting day. Furthermore, the original photovoltaic power time series is decomposed using ICCEMDAN, and the components are reconstructed by sample entropy to reduce the computational cost of forecasting models. Finally, Wasserstein distance, gradient penalty, and Metropolis-Hastings are used to ensure CTGAN training stability. According to the experimental results, it can be concluded that the data envelope clustering method divides datasets more reasonably than meteorological factors. The optimization for the sparrow search algorithm increases its global and local optimization ability to further enhance the performance of FCM. Using CTGAN to generate realistic data that approximates real data distributions can train predictive models with better performance, which increases their adaptability to PV power fluctuations. The proposed hybrid model is validated based on different seasons, different weather conditions, and datasets from different locations, and the results demonstrate the advantage of the proposed 38-step predictive model in accuracy, application, and generalization capabilities over other models involved in this study.