Research On Power Generation Dispatching And System Design Of Cascade Hydropower Station Based On Deep Learning

Due to its clean and pollution-free characteristics,hydropower energy has been given priority by all countries in the world.As a rising star in the development of hydropower in the world,China has put into operation the super-hydraulic projects such as the Three Gorges and the South-North Water Transfer in the 21 st century.However,with the scale and cascade trend of hydropower development,the power generation dispatching of cascade hydropower stations needs to take into account the influence of many factors such as hydrometeorology,water demand,and grid security.Traditional dispatching methods have been difficult to adapt to increasingly complex scheduling constraints and comprehensive targets.There is an urgent need to explore theories,models and techniques in the cross-disciplinary field.Thus this paper takes Qingjiang cascade hydropower station as the research object.The deep learning and cascade hydropower station power generation dispatching are combined based on the current artificial intelligence hotspot technology.Cascade hydropower station generation scheduling rules are extracted by the Community based Particle Swarm and Generalized Regression Network(CPSO-GRNN),then an generation forecasting model is proposed based on improved discrete differential evolution algorithm,empirical mode decomposition and long-short-term memory network(MDDE-EEMD-LSTM).At last,a distributed heterogeneous water resource dispatching system is designed and developed based on the models above.The main work content and innovative achievements of the thesis are as follows:(1)According to the idea of stochastic optimization,the deterministic scheduling result of successive optimization algorithm(POA)is used as the training set,and the generation scheduling model of cascade hydropower station based on generalized regression network(GRNN)is established.At the same time,the optimization problem of neural network hyperparameter is solved.A distributed parallel deep learning hyper-parameter optimization framework is designed.What’s more,an improved particle swarm optimization algorithm based on community center of gravity is proposed to optimize GRNN network parameters.The practical application of the Shuibuya Power Station in Qingjiang cascade shows that the improved PSO algorithm proposed in this paper has stronger global optimization ability and convergence speed;the proposed hyperparameter parallel optimization framework accelerates the parameter optimization and training of the model;The GRNN model based on optimal parameters has strong generalization ability and fast calculation speed.At the same time,it can maintain high precision under the condition of insufficient sample set,which can provide decision support for the extraction of power generation dispatching rules for large and medium-sized cascade hydropower stations.(2)For the neural network method,there is no memory function in the prediction of power generation,and it is unable to deal with the problem of long-term dependence.The key factors affecting the power generation are screened by feature engineering,and the power generation based on long-term and short-term memory neural network(LSTM)is established.The EEMD algorithm is introduced to decompose the power generation time series.An improved discrete differential evolution algorithm based on the hyperparametric optimization framework is proposed to optimize the parameters of the model.The simulation results of the Qingjiang cascade Shuibuya and Geheyan two-stage power station show that compared with the conventional time series prediction methods,the stability of the MDDE-EEMD-LSTM model based on the feature analysis of the power generation time series is presented.The generalization ability is strong.Moreover,for large and medium-sized power stations,the single-station prediction is better than the two-stage power station joint prediction model.(3)Focusing on the inconsistency of domestic water conservancy informatization software development standards and poor scalability,this paper separates the front-end and the back-end into micro-services,and realizes agile development and distributed deployment.In the internal structure,combined with the deep learing technology,big data analysis tools and business system,a "micro front end-micro back end-distributed heterogeneous" water dispatch system development model is proposed.This model is to surport the deep learning model optimization and training and a lot of on-line analysis of source heterogeneous data provides.At the same time,a multi-model coupling communication mechanism based on OpenMI standard and message bus is proposed,which runs through the entire system for business scheduling and data interaction,and integrates ELK logging tools into the water disptach system for the first time.The software development application of the Huazhong grid cross-region peaking decision support system shows that the development model proposed in this paper reduces the threshold of deep learning framework application in water regulation business,enhances the system’s ability to handle complex services,and accelerates the parallelism of system development.Moreover,the scalability of the system is enhanced.