Uncertainty Forecast Of Load,Runoff And Its Application In Reservoir Optimal Operation

The optimal operation of hydropower energy involves multiple factors such as load distribution of power system,hydrologic runoff process of river basin,optimal operation of cascade power station and so on.It is a huge system engineering problem integrating multiple disciplines.However,with the grid-connected operation of large-scale distributed clean energy system,the supply and demand relationship of power system has undergone profound changes,which brings new challenges to the safe and stable operation of power system.How to improve the accuracy of load prediction of power system is the primary problem to balance the power quantity of power system.In addition,human activities and global climate warming water cycle evolution dual function lead to increasingly complex,exacerbated the imbalance in time and space distribution of water resources situation,further increases the nonlinear non-stationary meteorological hydrological runoff forecast difficulty,at the same time basin reservoir optimal operation has been meteorological hydrological runoff forecast is not accurate,multiple constraints,problem solving,how to improve the uncertainty of meteorological hydrological runoff forecast reliability,further mining under uncertainty forecast runoff reservoir scheduling rule is another problem in the engineering practice.This paper focuses on the key scientific and technical problems of uncertain load,runoff prediction and optimal operation of reservoir.In central China power grid load,hydrological runoff and the three gorges reservoir as the case studies,the integrated use of hydropower,hydrology,statistics,artificial intelligence research frontier science and other multi-disciplinary overlapping fusion method,based on multi-source heterogeneous information fusion driven power grid load and hydrological runoff runoff prediction,considering uncertainty and model parameter uncertainty of reservoir implicit stochastic optimization scheduling model were studied.At the same time,some relevant research results have been applied in the decision support system of central China Power Grid Power regulation center,which provides reliable technical decision support for the production practice of relevant departments of water and electricity regulation.The main research contents and innovative achievements of this paper are summarized as follows:(1)Based on the historical information of power grid load and its related factors,a decompose-predictioning-integrated power grid short-term load prediction model based on variational mode decomposition and long and short memory neural network is proposed.Aiming at the extreme values of related factors that affect the load prediction of power system,the nonlinear mapping relationship method is introduced to construct the nonlinear mapping relational database of power grid load and related factors,which enhances the influence of the extreme values of related factors(such as extreme high temperature,extreme low temperature,holidays,etc.)on the model prediction results.At the same time,to solve the problem that it is difficult to caculate the hyper-parameters of machine learning models and the time cost of single fitness evaluation obtained by traditional intelligent algorithms is high,Bayesian Optimization Algorithm based on Gaussian process is introduced to obtain a best result set of hyper-parameters optimization with a lower fitness evaluation time cost.A case study shows that the proposed model can obtain high-precision short-term load forecasting results,which further enhances the accuracy and robustness of the deterministic short-term load forecasting model for power system.(2)For the power system containing multiple uncertainties,and the traditional deterministic point prediction results is difficult to fully reflect the uncertainty of the future load risk,a multi-step probability density prediction model for power system load based on variational mode decomposition and quantile regression forest is proposed.Firstly,the influence of relevant factors on power system load probability density prediction is considered,and a weighted index of temperature and humidity based on multiple meteorological factors was constructed to further explore the impact of meteorological composite index to short-term load probability density prediction.For the problem of model hyper-parameter optimization,bayesian optimization algorithm based on tree structure Parzen estimator TPE is introduced to seek the optimal model hyper-parameter set.The research results are applied to the 24 h multi-step prediction of grid load in Henan Province in central China.The results show that the proposed model can obtain high quality probability density prediction results and verify the feasibility and effectiveness of the model.Finally,a reservoir dispatching decision support system based on multilingual hybrid programming and distributed microservice framework is constructed based on the design of a short-term load prediction function module for power grids to meet actual engineering requirements.(3)To address the problem of complex modeling and low forecast accuracy of traditional physical models for meteorological hydrological runoff prediction,as well as excessive data quality requirements,the research work proposes a hybrid meteorological hydrological runoff time series interval prediction model based on the correlation of power system load time series prediction model driven by multi-source heterogeneous information fusion,which is based on two-level decomposition and Huber pinball loss function guided by deep gate recurrent unit neural network.The seasonal decomposition and variational mode decomposition are introduced to decompose the hydrological runoff time series in order to analyze the submodal components more thoroughly.At the same time,the traditional runoff forecasting model only focuses on the point forecast results and is difficult to reflect the uncertainty of the forecast,Huber norm enhanced pinball loss function is introduced to replace the loss function of deep gated recurrent unit model,to get the prediction results under different quantile,then the meteorological hydrological runoff time series forecasting model uncertainty point prediction results continuation for uncertainty probability interval prediction results.The proposed model has been applied to reservoir runoff forecast of Qingjiang Shuibuya Hydropower station and Leishui Dongjiang hydropower Station.The results show that the proposed model can obtain the probability interval forecast results of higher interval coverage and lower average interval forecast width under the given confidence interval,and meet the actual engineering requirements.(4)For implicit stochastic optimization scheduling in the process of runoff and reservoir model parameter uncertainty on the result of scheduling rules extraction,with the hydrological runoff time series forecasting model the uncertainty of the probability interval prediction results and the reservoir deterministic optimization scheduling results,on the basis of research work to introduce the monte carlo Dropout approximate bayesian inference to realize the depth of the bayesian learning model parameter uncertainty measurement,is proposed based on the depth of the monte carlo Dropout gating cycle unit implicit stochastic optimization neural network reservoir scheduling model.The model takes into account both the uncertainty of runoff and the uncertainty of model parameters and solves the problem that the traditional deterministic optimal scheduling is difficult to be applied in practice.The results show that the proposed model can obtain both high-precision and narrow-range deterministic dispatching results,which provides a new solution for the uncertain stochastic optimization dispatching of the Reservoir.