Power Trend Prediction Of Hydroturbine Based On Machine Learning

Hydroturbine power curve is an important reference for hydropower station staff to carry out operation dispatching,load distribution and real-time monitoring.Before the prototype hydroturbine put into operation,the reference data of its power curve and other characteristic parameters generally come from the model hydroturbine test in the manufacturer’s laboratory,and the test place and conditions are quite different from the actual field.Therefore,there are some parts to be corrected between the power curve provided by hydroturbine when leaving the factory and the actual operating condition curve of the prototype hydroturbine.It is necessary to obtain the characteristic curve of the unit under the current environment through the field test of the prototype hydroturbine.Field prototype hydroturbines are usually tested by selecting a limited number of operating points under typical operating conditions with relatively limited test data,while increasing the number of tests mechanically results in high operating costs.In order to effectively expand the amount of test data,a hydroturbine power prediction model is built and the operating characteristic curve of the turbine is corrected by using the obtained data to reduce the number of redundant tests and improve the efficiency of prototype turbine tests.In order to effectively establish the hydroturbine power prediction model,this paper firstly discusses the hydraulic model which combines computational fluid dynamics with three-dimensional modeling.It dynamically simulates and predicts the characteristics of hydraulic energy through complex calculation of internal flow field.However,the hydraulic model is predicted under ideal conditions and it is difficult to take into account the actual operating external parameters of the unit.Machine learning model has the characteristics of large data,clear structure and wide application range.Its parameters can be optimized by intelligent optimization algorithm to improve the predictive performance of the model.The machine learning method can quickly establish the turbine power prediction model and provide support for power curve correction.Then,on the basis of collecting a large number of operating data onto different types of hydoturbines under different working conditions,this paper studies the improved multilayer neural network power prediction model based on adaptive anti-normalization strategy and the least squares support vector machine power prediction model optimized by improved grey wolf optimization algorithm.The improved multilayer neural network model uses Rlue activation function to improve the calculation speed and prevent the gradient from disappearing.Combined with L2 regularization to prevent data over fitting,Adam gradient optimizer is used to realize the rapid convergence of the network output.While improving the model,at the same time,in the aspect of data processing,the influence of the anti-normalization process on the accuracy of the actual predicted output value of neural network is studied,and an adaptive anti-normalization interval strategy is proposed to realize the practical application of neural network prediction.In the prediction model of least squares support vector machine,in this paper,an improved grey wolf optimization algorithm based on mixed nonlinear convergence factor and static weight ratio is used to optimize its kernel parameters and regular coefficients,and its convergence and calculation speed are verified by combining Ackley,Rastrigin and schwefel’s problem test functions.Finally,the above two machine learning power prediction models are simulated and analyzed through the small sample Kaplan turbine coordination and non-coordination test data,and the large sample Kaplan turbine and Francis turbine operation data.The simulation results show that the built models can realize the high-precision prediction of hydroturbine output power.The problem of lack of real value interval in actual prediction can be solved effectively by adaptive anti-normalization strategy of improved multilayer neural network model,and significantly improve the prediction accuracy.Through the comparative analysis of large and small sample data,the prediction accuracy of adaptive anti-normalization strategy under small sample data is more obvious.The recognition ability of the original algorithm can be improved by the improved grey wolf optimization algorithm to optimize the least squares support vector machine model for the local optimal value and quickly converge to the optimal value.Compared with the prediction results of particle swarm optimization algorithm,grey wolf optimization algorithm and least squares support vector machine in the case of no optimization,the prediction effect from improved grey wolf optimization algorithm is the best.