Research On Spatial-Temporal Uncertainty Prediction Method Of Wind Power

With the increasing integration proportion of wind power,the wind power uncertainty has brought significant challenges to the power system's safe,stable,and economical operation.Wind power uncertainty,which diffuses in both spatial and temporal domains,is affected by multiple uncertain factors such as meteorological conditions,wind energy production and transmission,and forecasting models.How to accurately predict the spatial and temporal uncertainty of wind power is still a vital issue for new energy power systems.In this thesis,the spatial and temporal uncertainty prediction methods of wind power are deeply studied based on deep learning theory from the view of spatial,temporal,and spatio-temporal.The main works include:1.The Improved Deep Mixture Density Network(IDMDN)is proposed to avoid density leakage problems and reduce prediction uncertainty by considering the spatial correlation of multiple wind farms in the area.Firstly,a mixture density network module is constructed by Beta distribution,its numerical stabilization training program is proposed at the same time.Then,the IDMDN model is established by combining the proposed mixture density network module and the deep fully connected network,which is used for the joint uncertainty prediction of multi-wind farm's power.The proposed model was validated by the operating data sets of 7 wind farms.The results show that the proposed IDMDN model avoids the density leakage problem and obtains the wind power prediction probability distribution that conforms to the actual situation.Its prediction performance is better than several regional aggregated power prediction models and single wind farm power prediction models.2.A Multi-Source and Temporal Attention Network(MSTAN)is proposed.MSTAN extracts and integrates the implicit temporal pattern and information hidden in the historical data and multi-source numerical weather prediction data.It improves the multi-step uncertainty prediction accuracy of a single wind farm.First,the multi-source Numerical Weather Prediction(NWP)data is introduced,and the temporal error patterns of multi-source NWP are discovered.Secondly,a multi-source variable attention module is proposed for dynamic feature extraction of multi-source NWP by considering the obtained temporal error patterns.Then,a temporal attention module is proposed for feature extraction of long-term dependency in historical observation sequences and multi-source NWP sequences.Finally,a multi?source and temporal attention network model is established to predict wind power uncertainty in the next 1-48 hours based on the mixture density network module in section 1 and the parameter sharing mechanism.The model is tested by the actual operating data of three wind farms.The testing results show that the multi-source NWP improves the prediction accuracy,and the structure design scheme of MSTAN is reasonable.Its deterministic and probabilistic prediction accuracy surpasses the multiple counterparts under two typical technical routes.3.The adaptive spatio-temporal graph convolutional network model(SA-STGCN)is proposed,which strengthens the spatiotemporal feature extraction capability for irregularly arranged wind farm group data,and improves the performance of multi-wind farms and multi-time step joint uncertainty prediction.First,the multi-source variable attention module proposed in study 2 is expanded from the spatial dimension to form a multi-location-multi-source variable screening module,which is used for feature extraction and data fusion of multi-wind farms and multi-source NWP.Secondly,a self-adaptive graph convolution module with dynamic spatial feature extraction capabilities and a three-dimensional temporal attention module are combined as a spatiotemporal feature extraction module.Then,the mixture density network module proposed in study 1 is expanded in the spatial dimension and shared parameters in the temporal dimension.Finally,the SA-STGCN model is established for the joint power uncertainty prediction of multiple wind farms and the next 1-48 hours.The measured operation data of a wind farm cluster in northern China is used to validate the proposed model.The results show that the SA-STGCN model has better adaptability to the complex spatio-temporal correlation data and gets better results than benchmarks in both deterministic and probabilistic prediction.4.A splitting neural network(SplitNN)considering data protection is proposed.The SplitNN indirectly integrates the spatio-temporal information of multiple regions,improves the very short-term prediction performance in each area,and reduces the training time at the same time.First,several typical design modes in Split Learning are summarized.Then a split network model(SplitNN)is established by combining the CSC(Client to Server to Client)mode and MCS(Multi-Client to Server)mode,which is used to predict the very short-term wind power spatial and temporal uncertainty of multiple wind farms under multiple wind power operators.The SplitNN model uses a server network to merge the information from numerous client networks.The client network outputs the wind power probability density according to the fused information and local information.The SplitNN realizes the purpose of collaborative forecasting between multiple wind power operators without migrating the original data.The data of 65 meteorological observatories in the four eastern states of America is applied to validate the proposed model.The results show that the proposed SplitNN improves the prediction accuracy in each state,and the training time is shortened.This thesis systematically studies the wind power spatial and temporal uncertainty prediction methods based on deep learning theory from the view of spatial,temporal,and spatial-temporal wind power uncertainty.Some of the proposed methods and techniques have been applied to the regional wind power prediction systems for engineering purposes and achieved good prediction results.