Research On Integrated Forecasting Method For Wind Power,Solar Power,and System Load

In the new power system with wind power and solar power as the main component,the volatility of wind power and solar power,and the demand-side response increase the uncertainty of the power side and load side,which brings great challenges to the safety and economic operation of the power system.One important direction to solve the above problems is to improve the forecasting accuracy of wind power,solar power,and load to support the cooperative dispatching of power-load.Therefore,a study on wind power-solar power-load forecasting is carried out,according to the thought of "temporal-spatial characteristics mining?data cleaning for power forecasting?numerical weather prediction correction?integrated forecasting of power and load".The main contributions include:1.Establishing the quantitative models of wind process time-shifting,wind power volatility,wind and solar power complementarity under multiple temporal-spatial scales to provide the foundation of the following research.First,the delay time and accelerate speed are defined,and then a wind speed time-shifting evaluation model based on wind process extraction is established,which solves the problem of insufficient consideration of spatial correlation in existing indexes.Second,a series of volatility evaluation indexes are proposed based on the distribution of fluctuation frequency and confidence,which complements the description of wind power detailed volatility.Third,a complementarity evaluation index that considering the volatility of wind and solar power is proposed.Then the planning and operation models of the combined system are established according to the optimal real-time complementarity of wind and solar power to optimize the installed capacity ratio and power values,respectively.Taken the resource data and power data of different scales as examples,the wind speed time-shifting,wind power volatility,wind and solar power complementarity are studied under multiple temporal-spatial scales by using the proposed indexes and models.2.Proposing an orientation cleaning method for abnormal data of the power station to provide a comprehensive and representative training data set for wind and solar power forecasting.For the problems with maldistribution and accumulation of abnormal data,3 identification methods,including bidirectional one-sided quartile method,improved K-means method which based on the distance from cluster center to power curve,and double DBSCAN method are proposed to achieve the accurate identification for abnormal data of the power station.On the basis of the above research,an abnormal power reconstruction model based on LightGBM is established.The actual operation data of 30 wind farms and 8 solar plants in China are taken as examples to verify the effectiveness and superiority of the proposed method.3.Proposing a sequence transfer correction method for NWP wind speed/irradiance to improve the accuracy of the most critical inputs for wind/solar power forecasting.First,for the problems with wide distribution and weak regularity of NWP error at the independent time,a sequence transfer correction method for NWP is proposed.The certainty of mapping relationship between inputs and outputs of the correction model is improved by introducing the transfer relationship in wind speed/irradiance time series.Then,STCM is applied to establish 5 NWP correction models based on LR,SVM,BPNN,RF,and RBFNN.The actual operation data of 2 wind farms and 2 solar plants in China are taken as examples to verify the performance of the proposed method.The results show that RMSE of NWP wind speed and irradiation is reduced by 0.20 m/s to 2.59 m/s,15.64 W/m2 to 128.93 W/m2,respectively,compared with the original NWP.4.Proposing an integrated forecasting method of wind power,solar power,and load based on variable attention mechanism and multi-tasking learning to improve the forecasting accuracy by mining and learn the dynamic coupling relationship among three objects.First,the feature extraction model of wind power,solar power,and load is established based on variable attention mechanism.Then,a multi-task learning model in which the loss weight of different tasks can be automatically optimized is constructed.According to this model,an integrated forecasting method for wind power and solar power based on fully connected neural network is proposed,which can complete the forecasting tasks of wind power and solar power at the same time.Finally,a load forecasting model which combined the future information of wind and solar power,and the history information of load is established based on long short-term memory.The operation data of 8 wind farms,6 solar plants,and the load data of a city near the power stations are taken as examples to verify the superiority of the proposed method.The results show that the average RMSE of each wind farm power,solar plant power,and load is reduced by 4.84%,1.86%,and 3.02%,respectively,compared with the corresponding traditional methods.In this thesis,several quantitative models of wind and solar resource and power characteristics under multiple temporal-spatial scales are established at first.Then an orientation identification method for abnormal operation data of wind farm/solar plant,and a sequence transfer correction method for NWP wind speed/irradiation is proposed,respectively,to provide accurate input information for wind and solar power forecasting,including the high-quality training sample set and prediction results of wind speed/irradiation with high-precision.Based on the above research,an integrated forecasting method for wind power,solar power,and load is proposed to provide accurate and reliable forecasting results for the cooperative dispatching of power-load.