Research On Artificial Intelligence-Driven Nonparametric Probabilistic Forecasting Of Renewable Energy Generation

Due to the chaotic characteristics of the atmosphere,renewable energy generation exhibits strong inherent stochasticity and fluctuations,which results in the unavoidable forecasting errors.The new forms of the power system,such as the interaction and complex coupling between source and load,increase the complexity of forecasting,which causes multi-uncertainties in forecasting and poses significant challenges to the secure and economic operation of modern power systems with renewable energies.Power system decision-making and power market require the quantification of forecasting uncertainty to obtain more comprehensive probability distribution information of the renewable energy generation for the future.Probabilistic forecasting can effectively quantify the forecasting uncertainty of renewable energy power generation,which has received widespread attentions from both the academic and the industrial worlds.However,traditional probabilistic forecasting models usually rely on prior parametric distributional assumptions and lack the analysis of high-dimensional statistical characteristics of forecasting errors.Besides,the high-dimensional nonlinear mapping relationships between input and output are difficult to be modeled by common statistical learning models.On the one hand,the abundant and diverse multidimensional data enrich the information space for probabilistic forecasting,while on the other hand,they cause data redundancy,which should be further balanced.In view of the above issues,this thesis focuses on artificial intelligence-driven nonparametric probabilistic forecasting of renewable energy generation.The proposed models give full play to the high-dimensional nonlinear mapping ability of advanced artificial intelligence models and discover the higher-order statistical characteristics of forecasting uncertainty.Features of input information are adaptively extracted to realize the nonparametric conditional probability distribution modeling for multi-uncertainties.Thus,it can provide key reliable information for power system decision makers.The main works and innovations of the thesis are included as:(1)An ensemble machine learning-based higher order cumulants method is proposed for probabilistic forecasting of wind power generation.A conditional statistical moments estimation method is established based on bootstrap extreme learning machine.Based on bootstrap machine learning deterministic forecasting models,higher-order conditional statistical moments for both epistemic and aleatory uncertainties are estimated nonparametrically.With the unique additivity and the complete mathematical theory from cumulant generating function to probability distribution characteristic function,a higher order cumulant-based multiple series expansions ensemble method is proposed for conditional probability modeling.It solves the problem that the two types of statistical moments cannot be added directly,and achieves the complete description for nonparametric conditional probability distribution.Therefore,the overall performance and robustness of the model are improved.The proposed method makes a breakthrough in addressing the difficulty that traditional machine learning probabilistic forecasting models rely on the normal distributional assumption and cannot achieve the analysis of higher-order statistics.It builds the higher-order cumulants-based nonparametric probability distribution modeling method,which provides a new perspective for probabilistic forecasting research.(2)A hybrid dual-frequency recurrent neural network model considering spatio-temporal correlations is developed for probabilistic photovoltaic power forecasting.A timelag similarity spatio-temporal correlation discrimination method is proposed.The original time series of distributed power stations are decomposed into low-frequency trend series and high-frequency fluctuation series.The leading correlated stations are identified based on the cross correlation function,which effectively reduces data redundancy and realizes the dynamic discrimination of correlations.Deep recurrent neural networks and deep autoregressive probabilistic neural networks are constructed respectively to forecast the low-frequency trend series and to model the forecasting uncertainty of the high-frequency fluctuation series.It realizes the direct generation of forecasting probability distribution in a nonparametric way.The proposed method builds the direct mapping relationship from the forecasting input to the output probability distribution,which realizes the dynamic acquirement of time varying spatio-temporal correlations,and improves the comprehensive performance of probabilistic forecasting for distributed stations where the meteorological information fails to cover.(3)A hybrid machine learning method based on meteorological feature mining is proposed for probabilistic forecasting of photovoltaic power.The hidden Markov model is used to extract the meteorological characteristics reflected in the numerical weather prediction time series to realize the effective dimension reduction of multidimensional numerical weather prediction data and the effective mining of meteorological features.For different weather conditions,facilitated by the single hidden layer feedforward network structure of extreme learning machine,a machine learning-based quantile regression optimization problem is constructed.By introducing auxiliary variables,the original complex machine learning optimization problem is equally transformed into an efficient linear programming problem,which assures the high training efficiency of high-dimensional nonlinear mapping model.The proposed method succeeds to adaptively extract meteorological characteristics and realize the distinguished construction of machine learning quantile regression models conditioned on different weather conditions,which possesses high training efficiency and superior comprehensive performance.(4)An ensemble deep learning non-crossing quantile regression model is proposed for nonparametric probabilistic forecasting of wind power generation.An exponential stacking mapping method is proposed innovatively,which guarantees the strict non-crossing features of predictive quantiles.The negative summation of quantile elements produced by the deep learning models are mapped into their exponential form as the output quantiles.As a universal method,it can be extendedly applied into other machine learning-based quantile regression models to generate strictly monotonous quantiles.A two-stage ensemble model with the ensemble weight determined by two-norm quantile loss index is proposed to avoid the complex optimization of model ensemble weights.It achieves the adaptive ensemble of both homogeneous and heterogeneous deep learning models,which boosts the generalization ability.The proposed method makes a breakthrough to solve the inevitable quantile crossing problem in traditional machine learning quantile regression models,and realizes the advantage integration of different types of deep learning models for high-dimensional nonlinear mapping and deep feature mining.It ensures the superior comprehensive performance of probabilistic forecasting and enhances the model generalization ability.