Research On Renewable Scenario Generation And Prediction Based On Deep Generative Models

Energy is the basic driving force that promotes the economic development of all countries in the world.In the face of great changes unseen in a century,the whole world is actively deploying and developing renewable energy.Compared with traditional fossil energy,renewable energy sources such as wind energy and solar energy have obvious intermittent,stochastic and volatile characteristics.It is still a challenging research topic to accurately,flexibly and effectively characterize the uncertainty of renewable power generation.Scenario analysis can describe the dynamic characteristics of stochastic processes and is one of the main methods to deal with the uncertainty of renewable power generation.In previous studies,the probabilistic models are mainly adopted in the scenario analysis.However,the probabilistic assumptions are often difficult to accurately describe the actual generation dynamics of renewable energy sources.Therefore,around the essential problem of the uncertainty of renewable power generation,this thesis uses deep generative models of artificial intelligence to study the renewable scenario analysis in a data-driven manner and without additional statistical assumptions.Considering that the scenario analysis based on the deep generative models is different from previous studies,this thesis proposes to divide the renewable scenario analysis into three aspects: scenario generation,scenario prediction and scenario reduction.Scenario generation studies the uncertainty modeling of renewable power generation using deep generative models to generate scenarios consistent with the distribution of historical data.Scenario prediction studies the uncertainty prediction of renewable power generation using deep generative models to generate scenarios representing the dynamics of future power generation.Scenario reduction is to reduce large-scale scenarios to a reasonable computing level.Different from the traditional "point prediction" predicting the most likely realization in the future,"scenario prediction" conducts multiple stochastic predictions for the future so as to represent possible future realizations.The contributions of this thesis are mainly to propose two scenario generation methods and study two scenario prediction problems,and the following innovative research results are obtained.Aiming at the problems that may occur in the process of model training,such as training instability and overfitting,a wind power scenario generation method based on an improved generative adversarial network is proposed.By using an alternative training strategy to enforce the Lipschitz constraint,the problem of gradient explosion or gradient disappearance can be avoided,and the stability of model training can be improved,so that the model can better capture the real distribution of wind power data.The proposed method does not rely on scenario sampling techniques and can directly generate a large number of wind power scenarios.The generated scenarios can not only capture the uncertainty of wind power,but also accurately describe the spatiotemporal relationship of the actual power generation process.Moreover,when the proposed method is tested on a small training set,it can effectively alleviate the overfitting phenomenon and has good generalization ability.Aiming at the problems of the low quality of samples generated by the variational autoencoder and the lack of diversity in the samples generated by the generative adversarial network,a renewable scenario generation method based on the combination of these two models is proposed.By merging the decoder of the variational autoencoder and the generator of the generative adversarial network,it is possible to jointly code,generate and compare samples from the data set,and has the advantage that the variational autoencoder can encode data into the latent space and the advantage that the generative adversarial network can generate high-quality samples.The generated scenarios can not only keep the pattern diversities and statistical similarities,but also capture the temporal correlation of time series and the spatial correlation of different geographic locations.Moreover,the maximum mean difference is proposed to evaluate the quality of the generated samples,and the training process of the model can be evaluated intuitively.Aiming at the uncertainty prediction problem of a single power generation site,a wind power scenario prediction method based on the WGANGP model is proposed.The loss function of the discriminator of the generative adversarial network is trained with gradient penalty,which can make better use of the capacity of the deep neural network,accelerate the training speed of the convergence,and improve the training efficiency of the model for scenario prediction.Combining point prediction information and solving stochastic constraint optimization through the setting of confidence parameters and time scales,a set of scenarios representing future power dynamics can be generated.The generated scenarios are not only close to the point prediction information,but also can fluctuate in a certain range around the central point prediction,so as to describe the dynamic behaviors of wind power generation in different confidence ranges and time scales in the future.Aiming at the uncertainty prediction problem of multiple power generation sites,a deep generative model is proposed to predict spatiotemporal scenarios for wind and solar power.Firstly,the deep generative model is used to model the historical time series data of renewable energy,and the true distribution of renewable power generation dynamics is learned unsupervised without statistical assumption of data distribution.Then,combined with the point prediction information,a stochastic optimization step for future power dynamics is established,and a large number of future scenarios can be generated directly without scenario sampling techniques.The generated scenarios can not only characterize the intermittent,stochastic and volatile characteristics of renewable power generation,but also truly reflect the actual power generation process of renewable energy in the future.Moreover,by adjusting the network structure and several parameters,the dynamic input data of stochastic power generation can be adapted to different regions.