Data Model Joint Driven Probabilistic Generation-load Forecasting And Its Applications In Power System With Renewables

Power systems with new energy are faced with the challenge of high uncertainty on both generation and load sides.On the generation side,the power system supply capacity with high proportion of new energy is uncontrollable,due to the inherent randomness and volatility of wind speed and solar radiation.On the load side,the construction of smart grid,the integration of distributed new energy/electric vehicles,the implementation of demand response and other measures have fundamentally increased the uncertainty of power system load.In view of the increasing uncertainty factors in new energy power system such as wind power,photovoltaic power,low-voltage load and electricity price under complex situation,power industry urgently needs to solve the uncertainty quantification and analysis problem,which can cope with the challenges of uncertainty factors to the secure,reliable and economic operation of power system.Point forecasting and probabilistic forecasting can provide the expectation value and probability distribution estimation of the future random variables,providing key information support for power system operation and control under significant uncertainty.Probabilistic optimal power flow(POPF)is a basic method to analyze the steady operation state of power system under uncertainty and evaluate the influence of uncertain factors on power system.In order to enhance the renewable power system’s coping ability on uncertainty factors,this dissertation explores systematacially data model joint driven probabilistic forecasting and POPF approaches.The main contributions of this dissertation are as follows:(1)This dissertation proposes a deterministic and probabilistic forecasting appraoch for low voltage load based on hybrid ensemble deep learning.Deep belief networks with multi-layer learning structure are utilized for base forecasting learner.Besides,six primary ensemble deep learning models are constructed by combining different ensemble methods with deep learning.Then,the above six methods are intergated adaptively and the combination weights are determined according to data mining algorithm.For realizing the uncertainty quantification of low voltage load,the point forecasting error is decomposed into model estimation and data uncertainty,which are quantified respectively.Relying on the nonlinear fitting ability of deep learning and the generalization ability of ensemble learning,this approach overcomes the difficulies of lowvoltage load forecasting with strong fluctuation,enhancing the accuracy and adaptability of deterministic and probabilistic forecasting.(2)This dissertation proposes an adaptive ensemble data driven nonparametric probabilistic load forecasting approach.The proposed approach can directly mine the useful information from the historical data,and estimate the probabilistic distribution of the future forecasting target.Similar pattern measurement method and similar pattern quantity determination method are proposed to mine the similar pattern data set that matches the forecasting target.An adaptive weighted ensemble method is established considering the matching degree difference in multiple similar pattern data sets,which enhances the adaptability and robustness of data driven probabilistic forecasting method.This approach has advantages in both the acuuracy and efficiency,independent of regression model construction and probability distribution assumption.(3)This dissertation proposes a data driven nonparametric probabilistic forecasting approach for wind power.Four primary density estimation methods with both accuracy and diversity are constructed by different weighted distance measures and various conditional estimation methods.Besides,the nearest neighbor and probability distribution difference measurement methods are utilized to adaptively combine the four primary estimates above and obtain the final probabilistic forecasting result.This approach improves the accuracy of the primary density estimation method by seting the weight corresponding to the features and similar samples.Different distance measures and estimation methods ensure the diversity of the basic eatimators in ensemble learning.An adaptive weighted combination method is proposed by measuring the forecasting quality of the primary methods.The proposed approach can realize the accurate uncertainty quantification of complex wind power.(4)This dissertation proposes a distribution system probabilistic optimal power flow approach with high-dimensional related random variables.According to the significant spatio-temporal correlation of distributed generation,a conditional probabilistic forecasting method considering spatio-temporal correlation is established.Then the probabilistic forecasting results are taken as the conditional distribution in traditional Gibbs sampling,realizing rapid and accurate sampling under high-dimensional correlation variables.Finally,the probability distribution of distribution system’s operation states and decision variables under uncertainty can be obtained through deterministic optimal power flow of samples.The proposed approach handles the correlation between multi-dimensional variables with the probabilistic forecasting results,overcomeing the shortcomings of low efficiency and inaccurate correlation modeling.By means of the proposed approach,the uncertainty of distribution system’s states and decision variables can be quantified effectively under high-dimensional related uncertain variables.